AUTHOR ACCEPTED MANUSCRIPT FINAL PUBLICATION INFORMATION Comparative Advantage, International Trade, and Fertility The definitive version of the text was subsequently published in Journal of Development Economics, 119, 2015-10-23 Published by Elsevier and found at http://dx.doi.org/10.1016/j.jdeveco.2015.10.006 THE FINAL PUBLISHED VERSION OF THIS ARTICLE IS AVAILABLE ON THE PUBLISHER’S PLATFORM This Author Accepted Manuscript is copyrighted by the World Bank and published by Elsevier. It is posted here by agreement between them. Changes resulting from the publishing process—such as editing, corrections, structural formatting, and other quality control mechanisms—may not be reflected in this version of the text. You may download, copy, and distribute this Author Accepted Manuscript for noncommercial purposes. Your license is limited by the following restrictions: (1) You may use this Author Accepted Manuscript for noncommercial purposes only under a CC BY-NC-ND 3.0 IGO license http://creativecommons.org/licenses/by-nc-nd/3.0/igo. (2) The integrity of the work and identification of the author, copyright owner, and publisher must be preserved in any copy. (3) You must attribute this Author Accepted Manuscript in the following format: This is an Author Accepted Manuscript of an Article by Do, Quy-Toan; Levchenko, Andrei A.; Raddatz, Claudio Comparative Advantage, International Trade, and Fertility © World Bank, published in the Journal of Development Economics119 2015-10-23 CC BY-NC-ND 3.0 IGO http://creativecommons.org/licenses/by-nc-nd/3.0/igo http://dx.doi.org/10.1016/j.jdeveco.2015.10.006 © 2016 The World Bank Comparative Advantage, International Trade, and Fertility⇤ Quy-Toan Do Andrei A. Levchenko Claudio Raddatz The World Bank University of Michigan Central Bank of Chile NBER and CEPR May 11, 2015 Abstract We analyze theoretically and empirically the impact of comparative advantage in international trade on fertility. We build a model in which industries di↵er in the ex- tent to which they use female relative to male labor, and countries are characterized by Ricardian comparative advantage in either female-labor or male-labor intensive goods. The main prediction of the model is that countries with comparative advantage in female-labor intensive goods are characterized by lower fertility. This is because female wages, and therefore the opportunity cost of children are higher in those countries. We demonstrate empirically that countries with comparative advantage in industries em- ploying primarily women exhibit lower fertility. We use a geography-based instrument for trade patterns to isolate the causal e↵ect of comparative advantage on fertility. Keywords: Fertility, trade integration, comparative advantage. JEL Codes: F16, J13, O11. ⇤ We are grateful to Raj Arunachalam, Martha Bailey, Francesco Caselli, Francisco Ferreira, Elisa Gam- beroni, Gene Grossman, David Lam, Carolina Sanchez-Paramo, and seminar participants at various insti- ¸ gatay Bircan, Aaron Flaaen, Dimitrije Ruzic, and Nitya Pandalai Nayar tutions for helpful suggestions. Ca˘ provided outstanding research assistance. The project has been funded in part by the World Bank’s Research Support Budget. The views expressed in the paper are those of the authors and need not represent either the views of the World Bank, its Executive Directors or the countries they represent, or those of the Central Bank of Chile or the members of its board. This document is an output from a project funded by the UK Department for International Development (DFID) and the Institute for the Study of Labor (IZA) for the benefit of developing countries. The views expressed are not necessarily those of DFID or IZA. Email: qdo@worldbank.org, alev@umich.edu, craddatz@bcentral.cl. 1 Introduction Attempts to understand population growth and the determinants of fertility date as far back as Thomas Malthus. Postulating that fertility decisions are influenced by women’s oppor- tunity cost of time (Becker, 1960), choice over fertility has been incorporated into growth models in order to understand the joint behavior of population and economic development throughout history (see e.g. Barro and Becker 1989; Becker et al. 1990; Kremer 1993; Galor and Weil 1996, 2000; Greenwood and Seshadri 2002; Doepke 2004; Doepke et al. 2007; Jones and Tertilt 2008). The large majority of existing analyses examine individual countries in a closed-economy setting. However, in an era of ever-increasing integration of world markets, the role of globalization in determining fertility can no longer be ignored. This paper studies both theoretically and empirically the impact of comparative advan- tage in international trade on fertility outcomes. Our conceptual framework is based on three assumptions. First, goods di↵er in the intensity of female labor: some industries employ pri- marily women, others primarily men. This assumption is standard in theories of gender and the labor market (Galor and Weil, 1996; Black and Juhn, 2000; Qian, 2008; Black and Spitz- Oener, 2010; Rendall, 2010; Pitt et al., 2012; Alesina et al., 2013), and as we show below finds ample support in the data. In the rest of the paper, we refer to goods that employ primarily (fe)male labor as the (fe)male-intensive goods. Second, women bear a disproportionate bur- den of raising children. That is, a child reduces a woman’s labor market supply more than a man’s. This assumption is also well-accepted (Becker, 1981, 1985; Galor and Weil, 2000), and is consistent with a great deal of empirical evidence (see, e.g., Angrist and Evans, 1998; Guryan et al., 2008). Finally, di↵erences in technologies and resource endowments imply that some countries have a comparative advantage in the female-intensive goods, and others in the male-intensive goods. Our paper is the first to both provide empirical evidence that countries indeed di↵er in the gender composition of their comparative advantage, and to explore the impact of comparative advantage in international trade on fertility in a broad sample of countries. The main theoretical result is that countries with comparative advantage in female- intensive goods exhibit lower fertility. The result thus combines Becker’s hypothesis that fertility is a↵ected by women’s opportunity cost of time with the insight that this opportu- nity cost is higher in countries with a comparative advantage in female-intensive industries. We then provide empirical evidence for the main prediction of the model using industry- level export data for 61 manufacturing sectors in 145 developed and developing countries over 5 decades. We use sector-level data on the share of female workers in total employment to classify sectors as female- and male- intensive. The variation across sectors in the share 1 of female workers is substantial: it ranges from 8-9 percent in industries such as heavy machinery to 60-70 percent in some types of textiles and apparel. We then combine this industry-level information with data on countries’ export shares to construct, for each country and time period, a measure of its female labor needs of exports that captures the degree to which a country’s comparative advantage is in female-intensive sectors. We use this measure to test the main prediction of the model: fertility is lower in countries with a comparative advantage in female-intensive sectors. The key aspect of the empirical strategy is how it deals with the reverse causality problem. After all, it could be that countries where fertility is lower for other reasons export more in female-intensive sectors. To address this issue, we follow Do and Levchenko (2007) and con- struct an instrument for each country’s trade pattern based on geography and a gravity-like specification. Exogenous geographical characteristics such as bilateral distance or common border have long been known to a↵ect bilateral trade flows. The influential insight of Frankel and Romer (1999) is that those exogenous characteristics and the strong explanatory power of the gravity relationship can be used to build an instrument for the overall trade open- ness at the country level. Do and Levchenko (2007)’s point of departure is that the gravity coe cients on the same exogenous geographical characteristics such as distance also vary across industries – a feature of the data long known in the international trade literature. This variation in industries’ sensitivity to the common geographical variables allows us to construct an instrument for trade patterns rather than the overall trade volumes. Section 3.1 describes the construction of the instrument and justifies the identification strategy at length. As an alternative approach, we supplement the cross-sectional 2SLS evidence with panel estimates that include country and time fixed e↵ects. Both cross-sectional and panel results support the main empirical prediction of the model: countries with a higher female-labor intensity of exports exhibit lower fertility. The e↵ect is robust to the inclusion of a large number of other covariates of fertility, and is economically significant. Moving from the 25th to the 75th percentile in the distribution of the female- labor needs of exports lowers fertility by as much as 20 percent, or about 0.36 standard deviations of fertility across countries. Our paper contributes to two lines of research in fertility. The first is the empirical testing of Becker’s hypothesis that fertility is a↵ected by women’s opportunity cost of time. The key hurdle in this literature is to identify plausibly exogenous variation in this opportunity cost. While the negative correlation between women’s wages and fertility is very well-documented (Jones et al., 2010), it cannot be interpreted causally, since wages are only observed for women who work.1 Some authors have used educational attainment as an instrument for 1 While some studies have argued – implicitly or explicitly – that levels of female labor force participation 2 female wages after estimating a Mincer equation (Schultz, 1986) or directly as a proxy for productivity (Jones and Tertilt, 2008). However, as emphasized by Jones et al. (2010), education and occupational choices are potentially endogenous to fertility: women with a preference for large families might decide to invest less in education or choose occupations with lower market returns. Alternatively, to avoid using endogenous individual characteris- tics, some studies use median and/or mean female wages to proxy for women’s opportunity cost of time (Fleisher and Rhodes, 1979; Heckman and Walker, 1990; Merrigan and St.- Pierre, 1998; Blau and van der Klaauw, 2007). Still, when the wage statistics are computed from the selected sample of working women, they may not be representative of women’s opportunity cost of time when it comes to fertility decisions.2 Our approach avoids these limitations. By constructing country-level measures of female labor needs of exports, and instrumenting these using exogenous (and arguably excludable) geographical variables, we build a proxy for women’s opportunity cost of time that is exogenous to individual fertility, education, or labor force participation.3 Our paper thus provides novel empirical evidence on Becker’s influential hypothesis. The second is the (still sparse) literature on fertility in the context of international in- tegration. Schultz (1985) shows that the large changes in world agricultural prices and the gender division of labor in agriculture a↵ected fertility in 19th-century Sweden, while Alesina et al. (2011) test the hypothesis that historical prevalence of plough use a↵ected fertility. Galor and Mountford (2009) study the impact of initial comparative advantage on the dynamics of fertility and human capital investments. Saur´ e and Zoabi (2011, 2014) examine how trade a↵ects female labor share, wage gap, and fertility in a factor proportions framework featuring complementarity between capital and female labor. Rees and Riezman (2012) argue that when foreign direct investment improves work opportunities for women, fertility will fall. Our framework is the first to combine the Ricardian motive for trade with di↵erences in female-labor intensity across sectors. Our paper also relates to the small but growing literature on the impact of globalization on gender outcomes more broadly (Black and Brainerd, 2004; Oostendorp, 2009; Aguayo- Tellez et al., 2010; Marchand et al., 2013; Juhn et al., 2014). Closest to our paper is Ross are “high enough” in the U.S. so that censoring is not a significant issue (Cho, 1968; Fleisher and Rhodes, 1979), this assumption would be more problematic to make in the context of low and middle-income countries, that typically exhibit low levels of female labor force participation and for which data on female wages are scarce and imprecise in part due to the large size of the informal sector (World Bank, 2012). 2 Heckman and Walker (1990) argue that “[i]t is plausible that in Sweden the wage process is exogenous to the fertility process. Sweden uses centralized bargaining agreements to set wages and salaries” (p.1422). Since this institutional feature is specific to Sweden, this approach is di cult to extend to other contexts. 3 Our methodology is thus similar in spirit to Alesina et al. (2013), who also use a geography-based variable (soil crop suitability in this case) as an instrument for the adoption of a female-labor-intensive technology: the plough. 3 (2008), who shows empirically that oil-abundant countries have lower female labor force participation (FLFP). Ross (2008)’s explanation for this empirical pattern is that Dutch disease in oil-exporting countries shrinks the tradable sector, and expands the non-tradable sector. If the tradable sector is more female-intensive than the non-tradable sector, oil lowers demand for female labor and therefore FLFP. Our theoretical mechanism relies instead on variation in female-labor intensity within the tradable sector. On the empirical side, the e↵ect we demonstrate is much more general: it is present when excluding natural resource exporters, as well as excluding the Middle East-North Africa region. The rest of the paper is organized as follows. Section 2 presents a simple two-country two-sector model of comparative advantage in trade and endogenous fertility. Section 3 lays out our empirical strategy to test the predictions of the model. Section 4 describes the data, while section 5 presents estimation results. Section 6 concludes. All the proofs are collected in Appendix A. 2 Theoretical Framework 2.1 The Environment Consider an economy comprised of two countries indexed by c 2 {X, Y }, and two sectors c c indexed by i = {F, M }. The representative household in c values consumption CF and CM of the two goods, as well as the number of children N c it has according to the utility function c c c ⌘ c 1 ⌘ u ( CF , CM , N c ) = ( CF ) ( CM ) + v (N c ) , with v (.) is increasing and concave. To guarantee interior solutions, we further assume that limN !0 v 0 (N ) = +1.4 We adopt the simplest form of the gender division of labor, and assume that production in sector F only requires female labor and capital, while sector M only requires male labor and capital. Technology in sector i is therefore given by Yic (Ki , Li ) = ic Ki↵ L1 i ↵ , where Li is the sector’s employment of female labor (in sector F ) and male labor (in sector c2{X,Y } M ), Ki is the amount of capital used by sector i, and {ic }i2{M,F } are total factor productivities 4 The assumption that utility is quasi-linear in income is made for analytical tractability. It shuts down the income e↵ect and allows us to focus solely on the substitution e↵ect. For discussions on conditions for the substitution e↵ect to dominate the income e↵ect under more general assumptions, see Jones et al. (2010) and Mookherjee et al. (2012). 4 in the two sectors and countries. Formally, this is the specific-factors model of production and trade (Jones, 1971; Mussa, 1974), in which female and male labor are specific to sectors F and M respectively, while K can move between the sectors. Thus, we take the arguably simplistic view that men supply “brawn-only” labor, while women supply “brain-only” labor, and men and women are not substitutes for each other in production within each individual sector. Of course, there is still substitution between male and female labor in the economy as a whole, since goods F and M are substitutable in consumption.5 The key to our results is the assumption that countries di↵er in their relative productiv- ities F c /M c . For convenience, we normalize (F c )⌘ (M c )1 ⌘ =1 (1) in both countries. Since the impact of relative country sizes is not the focus of our analysis, and the aggregate gender imbalances in the population tend to be small, we set the country endowments of male and female labor and capital to be L ¯c = L ¯ c = 1 and K ¯ c = 1 for M F c 2 {X, Y }. Capital can move freely between sectors, and the market clearing condition c c for capital is KF + KM = 1. Men supply labor to the goods production sector only, and hence supply it inelastically: Lc M = 1. On the other hand, childrearing requires female labor, and women split their time between goods production and childrearing. N c children require ⇥ ⇤ spending N c units of female labor at home, so that N c 2 0, 1 . Female market labor force participation is then LcF = 1 N c. All goods and factor markets are competitive. International trade is costless, while capital and labor cannot move across countries.6 In country c, capital earns return rc and female c c and male workers are paid wages wF and wM , respectively. Let the price of goods i 2 {M, F } be denoted by pi , and set the price of the goods consumption basket to be numeraire: p⌘ 1 ⌘ F pM = 1. (2) 5 The necessary condition for obtaining our results is that in equilibrium, women’s relative wages are higher in the country with a Ricardian comparative advantage in the female-intensive good. This plausible equilibrium outcome obtains under more general production functions in which both types of labor are used in both sectors (see, for instance, Morrow, 2010). On the other hand, our result is inconsistent with models that feature Factor Price Equalization (FPE). FPE is ruled out in our model by cross-country productivity di↵erences in all sectors, which implies that generically FPE does not hold in our model. 6 The assumption of no international capital mobility is not crucial for our results. In fact, our results can be even more transparent with perfect capital mobility. When capital is internationally mobile, relative female wages in the two countries depend only on the relative Total Factor Productivities in the female sector X Y 1/(1 ↵) (when the solution is interior): wF /wF = F X /F Y . This expression relates relative female wages to absolute advantage in the female-intensive sector. Thus, as long as a country’s Ricardian comparative advantage is the same as its absolute advantage (that is, as long as M X /M Y is such that F X /F Y Q 1 ) F X /F Y M Y /M X Q 1), it will have higher female wages, and the rest of the results follow. 5 It will be convenient to express all the equilibrium outcomes of the economy (prices and KFc quantities) as functions of ✓c ⌘ K c c instead of KF . M c2{X,Y } A competitive equilibrium in this economy is a set of prices {pi , rc , wi c }i2{M,F } , capital c c2{X,Y } c c2{X,Y } allocations {✓ } , and fertility levels {N } , such that (i) consumers maximize utility; (ii) firms maximize profits; (iii) goods and factor markets clear. Fertility in both countries and production/consumption allocations are thus jointly de- termined in equilibrium, making it more di cult to handle than the typical model of inter- national exchange in which factor supplies are fixed. For expositional purposes, we describe the equilibrium in two steps. We first characterize the global production and consumption allocations for a given fertility profile {N c }c2{X,Y } . We then endogenize households’ decisions over fertility. 2.2 Production and Trade Equilibrium We first characterize the production and trade equilibrium under a fixed female labor supply ⇥ ⇤ Lc F = 1 N c , for a given N c 2 0, 1 . Firms’ optimization In each of the two sectors i 2 {M, F }, firms rent capital and hire labor to maximize profits: max pi ic K ↵ L1 ↵ rc K c wi L. K,L The necessary and su cient first-order conditions with respect to Kic yield the following ⇣ c ⌘1 ↵ rc c Li expression for the return to capital: pi = ↵i K c . Equalizing the returns to capital i across sectors and assuming that labor markets clear pins down relative prices of the two pF c ✓c 1 ↵ goods: p M =M F c 1 Nc . Under the choice of numeraire (2), prices are equal to 8

1 indicates that country c has a comparative advantage in the female- . intensive good F . The comparative advantage can be decomposed into a technological or Nc Ricardian component c and an occupational or “factor-proportions” component 11 N c, which can exacerbate or attenuate technological di↵erences. We rewrite the two equations (7) and (8) as a system of two equations with two unknowns {✓c , ✓ c } given exogenous model parameters and “pre-determined” values {N c , N c }: ⌘ (1 ⌘ ) ✓c ↵) ⌘ (1 ⌘ ) ✓ c + ( c )⌘(1 = 0 (9) (1 + ✓c )↵ (1 + ✓ c )↵ ✓ c ⇢c c = 1 (10) ✓ 7 Equation (9) implicitly defines a downward-sloping “goods market-clearing curve” in the space (✓ c , ✓c ) and is just⇣a rearrangement ⌘⌘ of equation (7), keeping in mind that normalization M c F cM c c ⌘ (1 ↵) (1) implies that M c = M c F c = ( ) . Since goods produced by the two countries are perfect substitutes, market clearing implies a negative relationship between the size ✓c of the F -sector in country c and its size ✓ c in country c. On the other hand, the upward- sloping “factor market-clearing curve” in the space (✓ c , ✓c ), defined by (10), implies that F -sectors have to be of comparable size in the two countries (i.e. the larger sector F gets in country c, the larger it needs to be in country c as well), otherwise the return to capital will diverge across the F - and M -sectors in each country. Thus, allocations of capital between two sectors in the two countries {✓c }c2{X,Y } are uniquely determined by the system of two equations (9) and (10). Proposition 1: Production and trade equilibrium Consider the endowment struc- ture K ¯ c , Lc c2{X,Y } = {1, 1, 1 ¯ c, L N c }c2{X,Y } . The unique production and consump- M F c c2{X,Y } tion equilibrium is characterized by the vector of prices {pi , rc , wi }i2{M,F } defined by (3)-(6), c c2{X,Y } and capital allocations {✓ } that solve (9)-(10).⌅ The proof of Proposition 1 establishes existence of an intersection of the two “factor market-clearing” and “goods market-clearing” curves, which is therefore unique since the two curves have opposite slopes. 2.3 Fertility Decisions The analysis above is carried out under an exogenously fixed fertility rate or, equivalently, an exogenously fixed level of female labor force participation. We now turn to endogenizing households’ fertility decisions. To pin down equilibrium fertility N c , we proceed in two steps. c First, for a given N c , wF and N c are jointly determined by labor supply and demand. Thus, we must ensure that labor supply is upward-sloping and the female labor market equilibrium is well defined. Second, fertility in the other country a↵ects the labor market equilibrium by shifting female labor demand and hence fertility in country c. We therefore look for a fixed point in {N c , N c } such that the female labor markets are in equilibrium in both countries simultaneously. Fertility choices and female labor supply Taking N c as given and anticipating the production equilibrium prices and quantities, households make fertility decisions accordingly. 8 Namely, they take prices as given and choose N c to maximize their indirect utility: c V c (N ) = r c + wF (1 c N ) + wM + v (N ) . (11) The first-order condition for the representative household’s fertility decision is necessary and su cient and given by 8 0 . ⌅ Thus, an increase in female labor supply in country c increases c’s comparative advantage in the female-labor intensive good (the factor-proportions e↵ect). This will increase ✓c , the size of the F -sector in country c and exert a downward pressure on female wages. By the same token, country c0 s comparative advantage in the female-labor intensive good is reduced, decreasing ✓ c , the size of the F -sector in that country, which in turn will put additional downward pressure on female wages in country c. The female labor demand curve is therefore downward-sloping. Lemma 2: Fertility in partial equilibrium For a given level of the other country’s fertility level N c , there exists a unique N c satisfying both (12) and (13).⌅ In the proof of Lemma 2, we establish that the female labor supply and demand curves either intersect at the corner, i.e. N c = 1 , or in the interior and the solution is also unique since labor supply and demand curves have opposite slopes. Equilibrium fertility Lemma 2 and the labor demand equation (13) imply that the female labor demand curve in country c shifts down when female labor supply in country c goes up. Thus N c (N c ) , the equilibrium fertility rate in country c when that rate in country c is N c , is decreasing; so is N c (N c ) . The following proposition formally establishes that these two “reaction functions” intersect and therefore defines the complete equilibrium of the economy. Proposition 2: Full characterization of the equilibrium Equations (3) to (6), c i2{M,F } (10), and (12) define a vector of prices {pi , rc , wi }c2{X,Y } , capital allocations {✓c }c2{X,Y } and fertility decisions {N c }c2{X,Y } that form the unique equilibrium of the economy.⌅ Comparative statics and cross-sectional comparisons We now consider (✓c , N c ) and ˜c , N (✓ ˜ c ), two equilibrium capital allocations and fertility decisions of the economy when the Ricardian comparative advantage of country c takes values c and ˜ c , respectively. The objective of this section is to compare fertility and the allocation of capital across sectors in these two parameter configurations. 10 Lemma 3: Comparative statics in general equilibrium An increase in compara- ˜ c .⌅ ˜ c and N c N tive advantage exacerbates fertility di↵erences: if c ˜ c , then N c  N From Lemma 3, the main result of the paper is stated in the theorem below: Theorem 1: Cross-sectional comparison If country c has a Ricardian comparative Fc F c advantage in the female-labor intensive sector ( M c > M c ), it will exhibit lower equilibrium fertility: N c < N c .⌅ Theorem 1 is the main theoretical prediction of the model, and one that will be tested empirically. The intuition for this result is as follows. Female wages will be higher in the country with the comparative advantage in the female-intensive sector because of higher relative productivity further exacerbated by a flow of capital to the sector with comparative advantage. Since a higher female wage increases the opportunity cost of childbearing in terms of goods consumption, equilibrium childbearing drops. The theoretical exposition above makes clear what are the measurement and identifica- tion challenges for the empirical work. First, in order to test for the impact of gender-biased comparative advantage on fertility, we must develop a measure of comparative advantage in (fe)male sectors. Fortunately, the model presents us with a way of doing this: observed trade flows. Countries with a comparative advantage in the female-intensive good will export that good. Our empirical strategy thus starts by building a measure of the female intensity of exports based on observed export specialization. Second, the model shows quite clearly that observed specialization patterns, trade flows, and fertility levels are jointly determined. In particular, countries with higher technological comparative advantage in the female sector can potentially accentuate that comparative advantage with a higher female labor supply and will thus e↵ectively exhibit relative factor proportions that also favor exports in the female-intensive sectors. Thus, in order to provide evidence for the causal impact of compar- ative advantage on fertility, we must find an exogenous source of variation in comparative advantage. We describe all parts of our empirical strategy and results below. 3 Empirical Strategy To test for the impact of comparative advantage on fertility, we must first construct a measure of the degree of female bias in a country’s export pattern. We begin by classifying sectors according to their female intensity. Let an industry’s female-labor intensity F Li be measured as the share of female workers in the total employment in sector i. We take this measure as a technologically determined industry characteristic that does not vary across countries. 11 We then construct for each country and time period a measure of the “female-labor needs of exports”: XI X F LN Xct = !ict F Li , (14) i=1 X where i indexes sectors, c countries, and t time periods. In this expression, !ict is the share of sector c exports in country c’s total exports to the rest of the world in time period t. Thus, F LN Xct in e↵ect measures the gender composition of exports in country c. This measure is meant to capture the female bias in each country’s comparative advantage. It will be high if a country exports mostly in sectors with a large female share of employment, and vice versa.7 Using this variable, we would like to estimate the following equation in the cross-section of countries: N c = ↵ + F LN Xc + Zc + "c . (15) The left-hand side variable N c is, as in Section 2, the number of births per woman, and Zc is a vector of controls. The main hypothesis is that the e↵ect of comparative advantage in female- intensive sectors F LN Xc on fertility is negative ( < 0). The potential for reverse causality is immediate here: higher fertility will reduce women’s formal labor force participation and therefore could also a↵ect the country’s export pattern. To deal with reverse causality, we implement an instrumentation strategy that follows Do and Levchenko (2007), described in the next subsection. We also exploit the time variation in the variables to estimate a panel specification of the type Ntc = ↵ + F LN Xct + Zct + c + t + "ct , (16) where country and time fixed e↵ects are denoted by c and t respectively. The advantage of the panel specification is that the use of fixed e↵ects allows us to control for a wide range of time-invariant omitted variables that vary at the country level, and identify the coe cient purely from the time variation in comparative advantage and fertility outcomes within a country over time. The baseline controls include PPP-adjusted per capita income, overall trade openness, and, in the case of cross-sectional regressions, regional dummies. (We also check robustness of the results to a number of additional control variables.) The cross-sectional specifications are estimates on long-run averages for the period 1980-2007. The panel specifications are estimated on non-overlapping 5-year and 10-year averages. As per standard practice, we take 7 The form of this index is based on Almeida and Wolfenzon (2005) and Do and Levchenko (2007), who build similar indices to capture the external finance needs of production and exports. 12 multi-year averages in order to sweep out any variation at the business cycle frequency. The panel data span 1962 to 2007 in the best of cases, though not all variables for all countries are available for all time periods. 3.1 The Instrument The construction of the instrument exploits exogenous geographic characteristics of countries together with the empirically observed regularity that trade responds di↵erentially to the standard gravity forces across sectors. The exposition here draws on, and extends, the material in Do and Levchenko (2007). For each industry i, we estimate the Frankel and Romer (1999) gravity specification, which relates observed trade flows to exogenous geographic variables: 1 2 LogXicd = ↵i + ⌘i ldistcd + ⌘i lpopc + ⌘ 3 4 5 i lareac + ⌘i lpopd + ⌘ i laread + (17) ⌘6 7 8 i landlockedcd + ⌘i bordercd + ⌘i bordercd ⇥ldistcd + ⌘9 10 11 i bordercd ⇥ popc + ⌘ i bordercd ⇥areac + ⌘i bordercd ⇥popd + 12 ⌘i bordercd ⇥aread + ⌘ 13 i bordercd ⇥landlockedcd + ✏icd , where LogXicd is the log of exports as a share of GDP in industry i, from country c to country d. The right-hand side consists of the geographical variables. In particular, ldistcd is the log of distance between the two countries, defined as distance between the major cities in the two countries, lpopc is the log of population of country c, lareac log of land area, landlockedcd takes the value of 0, 1, or 2 depending on whether none, one, or both of the trading countries are landlocked, and bordercd is the dummy variable for common border. The right-hand side of the specification is identical to the one used by Frankel and Romer (1999). We use bilateral trade flows from the COMTRADE database, converted to the 3-digit ISIC Revision 3 classification. To estimate the gravity equation, the bilateral trade flows Xicd are averaged over the period 1980-2007. This allows us to smooth out any short-run variation in trade shares across sectors, and reduce the impact of zero observations. Having estimated equation (17) for each industry, we then obtain the predicted logarithm of industry i exports to GDP from country c to each of its trading partners indexed by d, \ icd . In order to construct the predicted overall industry i exports as a share of GDP LogX from country c, we then take the exponential of the predicted bilateral log of trade, and sum 13 over the trading partner countries d = 1, ..., C , exactly as in Frankel and Romer (1999): C X ˆ ic = \ X eLogX icd . (18) d=1 d 6= c That is, predicted total trade as a share of GDP for each industry and country is the sum of the predicted bilateral trade to GDP over all trading partners. The instrument for F N LX is constructed using the predicted export shares in each industry i, rather than actual ones, in a manner identical to equation (14): I X F\ LN X c = X bic ! F Li , i=1 X bic where the predicted share of total exports in industry i in country c, ! , is computed from ˆ the predicted export ratios Xic : ˆ ic X X bic ! = PI . (19) X ˆ ic i=1 Note that even though X ˆ ic is exports in industry i normalized by a country’s GDP, every sector is normalized by the same GDP, and thus they cancel out when we compute the predicted export share. 3.1.1 Discussion We require an instrument for trade patterns, not trade volumes, and thus our strategy will only work if it produces di↵erent predictions for X bic across sectors for the same exporter. All of the geographical characteristics on the right-hand side of (17) do not vary by sector. However, crucially for the identification strategy, if the vector of estimated gravity coe cients ⌘ i di↵ers across sectors, so will the predicted total exports X bic across sectors i within the same country. The strategy of relying on variation in coe cient estimates for the same geographical variables bears an a nity to Feyrer (2009), who uses the di↵erential e↵ect of gravity variables on ocean-shipped vs. air-shipped trade to build a time-varying instrument for overall trade openness, and to Ortega and Peri (2014), who exploit the fact that the same gravity variables a↵ect goods trade and migration flows di↵erently to build separate instruments for overall trade openness and immigrant population. This subsection (i) discusses the intuition for how the instrument works; (ii) reviews the existing sector-level gravity literature to provide reasons to expect the gravity coe cients to 14 vary across sectors; (iii) describes the variation in our own gravity coe cients from estimating (17) by sector. The following simple numerical example illustrates the logic of the strategy. Suppose that there are four countries: the U.S., the E.U., Canada, and Australia, and two sectors, Wearing Apparel and Motor Vehicles. Suppose further that the distance from Australia to either the U.S. or the E.U. is 10,000 miles, but Canada is only 1,000 miles away from both the U.S. and the E.U. (these distances are not too far from the actual values). Suppose that there are only these country pairs, and that trade between them is given in Table A1. Let the gravity model include only bilateral distance. The trade values have been chosen in such a way that a gravity regression estimated on the entire “sample” yields a coe cient on distance equal to -1, a common finding in the gravity literature. The gravity model estimated separately for each of the two sectors yields the distance coe cient is -0.75 in Wearing Apparel and -1.25 in Motor Vehicles (this amount of variation in the distance coe cients is reasonable, as we show below). Using these “estimates” of the distance coe cients, it is straightforward to take the exponent and sum across the trading partners as in (18), and to calculate the predicted shares of total exports to the rest of the world in each of the two sectors, as in (19). Now let the share of female labor in Wearing Apparel be F LAP P = 0.71, and of Motor Vehicles, F LM V = 0.09 (these are the actual values of F Li for these two industries). Then, the predicted female labor need of exports of Canada is F\ N LX CAN = 0.18, which is some 40% lower than the predicted value for Australia of F\ N LX AU S = 0.31. The key intuition from this example is that countries located far away from their trading partners will have relatively lower predicted export shares in goods for which the coe cient on distance is higher, compared to countries located close to their trading partners. This information is combined with cross-industry variation in female employment to generate pre- dicted F\N LX . There are several important points to note about this procedure. First, while this simple example focuses on the variation in distance coe cients along with di↵erences in distances between countries, our actual empirical procedure exploits variation in all 13 regression coe cients in (17), along with the entire battery of exporting and destination country characteristics. Thus, to the extent that coe cients on other regressors also dif- fer across sectors, variation in predicted F\ N LX will come from the full set of geography variables. Second, while this simple four-country illustrative example may appear somewhat circular – actual exports and distance a↵ect the gravity coe cient, which in turn is used to predict trade – in the real implementation we estimate the gravity model with a sample of more than 150 countries, and thus the trade pattern of any individual country is unlikely to a↵ect the estimated gravity coe cients and therefore its predicted trade. Third, it is crucial for this procedure that the gravity coe cients (hopefully all 13 of them) vary appreciably 15 across sectors. Below we discuss the actual estimation results for our gravity regressions, and demonstrate that this is indeed the case. Can we support the notion that the gravity coe cients would be expected to di↵er across sectors? Most of the research on the gravity model focuses on the e↵ects of trade barriers on trade volumes. Thus, existing empirical research is most informative on whether we should expect significant variation in the coe cients on distance and common border variables, which are meant to proxy for bilateral trade barriers. Anderson and van Wincoop (2003, 2004) show that the estimated coe cient on log distance is the product of the elasticity of trade flows with respect to iceberg trade costs (commonly referred to as simply the “trade elasticity”) and the elasticity of iceberg trade costs with respect to distance. Thus, the distance coe cient will di↵er across industries if either or both of those elasticities di↵er across industries. A number of papers estimate trade elasticities by sector (see, among many others, Feen- stra, 1994; Broda and Weinstein, 2006; Imbs and M´ ejean, 2013; Caliendo and Parro, 2015). Imbs and M´ ejean (2013) – the most recent and the most comprehensive study – reports sector-level trade elasticity estimates using both of the principal estimation methods pro- posed in the literature. The conservative range of trade elasticities across sectors reported in that paper is from 2 to 20, consistent with the other studies undertaking similar exercises. There is less direct evidence on whether the elasticity of iceberg trade costs with respect to distance varies across sectors. Trade costs do vary significantly across industries. Hummels (2001) compiles freight cost data, and shows that in 1994 these costs ranged between 1% and 27% across sectors in the U.S..8 Hummels (2001, 2007) further provides evidence that the variation in freight costs is strongly related to the value-to-weight ratio: it is more expensive to ship goods that are heavy. Thus, it is plausible that the elasticity of trade costs with respect to distance is heterogeneous across sectors as well. To summarize, there are strong reasons to expect the coe cients in (17) to vary across sectors. It is indeed typical to find variation in the gravity coe cients across sectors, though studies di↵er in the level of sectoral disaggregation and specifications (see, e.g. Rauch, 1999; Rauch and Trindade, 2002; Hummels, 2001; Evans, 2003; Feenstra et al., 2001; Berthelon and Freund, 2008). For instance, Hummels (2001) finds that the distance coe cients vary from zero to -1.07 in his sample of sectors, while the coe cients on the common border variable range from positive and significant (as high as 1.22) to negative and significant (as low as -1.23). 8 In addition to the simple shipping costs, trade costs di↵er across industries in other ways. For in- stance, trade volumes in di↵erentiated and homogeneous goods sectors react di↵erently to informational barriers (Rauch, 1999; Rauch and Trindade, 2002), and to importing country institutions such as rule of law (Berkowitz et al., 2006; Ranjan and Lee, 2007). 16 Table A2 reports the cross-sectoral variation in the gravity coe cients in our estimates. For each coe cient, it reports the mean, standard deviation, min, and max in our sample of sectors. The variation in all of the gravity coe cients across sectors is considerable. The distance coe cient, as expected, is on average around 1, but the range across sectors is from -1.65 to -0.53. The common border coe cient has a mean of 1.4, and a standard deviation of 2.5 across sectors. Our instrumentation strategy relies on this variation in sectoral coe cients. There is another potentially important issue, namely the zero trade observations. In our gravity sample, only about two-thirds of the possible exporter-importer pairs record positive exports, in any sector. At the level of individual industry, on average only a third of possible country-pairs have strictly positive exports, in spite of the coarse level of aggregation.9 We follow the Do and Levchenko (2007) procedure, and deal with zero observations in two ways. First, following the large majority of gravity studies, we take logs of trade values, and thus the baseline gravity estimation procedure ignores zeros. However, instead of predicting in-sample, we use the estimated gravity model to predict out-of-sample. Thus, for those observations that are zero or missing and are not used in the actual estimation, we still predict trade.10 In the second approach, we instead estimate the gravity regression in levels using the Poisson pseudo-maximum likelihood estimator suggested by Santos Silva and Tenreyro (2006). The advantage of this procedure is that it actually includes zero observations in the estimation, and can predict both zero and non- zero trade values in-sample from the same estimated equation. Its disadvantage is that it assumes a particular likelihood function, and is not (yet) the standard way of estimating gravity equations found in the literature. Below we report the results of implementing all three approaches. It turns out that all three deliver very similar results, an indication that the zeros problem is not an important one for this empirical strategy. Finally, we stress that unfortunately this instrumentation strategy is only available in the cross-section. In principle, a time-varying instrument for trade patterns could be constructed in this way and used in a panel specification with country and time fixed e↵ects. This precedure would rely on the sector-level gravity coe cients varying over time (di↵erentially across sectors). Our attempt to implement this strategy revealed that there is simply not enough di↵erential time variation in the gravity coe cients for this strategy to be feasible. 9 These two calculations make the common assumption that missing trade observations represent zeros (see Helpman et al., 2008). 10 More precisely, for a given exporter-importer pair, we predict bilateral exports out-of-sample for all 61 sectors as long as there is any bilateral exports for that country pair in at least one of the 61 sectors. 17 4 Data Sources and Summary Statistics The key indicator required for the analysis is the share of female workers in the total em- ployment in each sector, F Li . This information comes from the UNIDO Industrial Statistics Database (INDSTAT4 2009), which records the total employment and female employment in each manufacturing sector for a large number of countries at the 3-digit ISIC Revision 3 classification (61 distinct sectors), starting in the late 1990s. We compute F Li as the mean share of female workers in total employment in sector i across the countries for which these data are available and relatively complete. This sample includes 11 countries in each of the developed and developing sub-samples: Austria, Cyprus, Ireland, Italy, Japan, Lithua- nia, Korea, Malta, New Zealand, Slovak Republic, United Kingdom; and Azerbaijan, Chile, Egypt, India, Indonesia, Jordan, Malaysia, Morocco, Philippines, Thailand, Turkey. Table 1 reports the values of F Li in our sample of sectors. It is clear that there is wide variation in the share of women in sectoral employment. While the mean is 27 percent, these val- ues range from the high of 71 percent in Wearing Apparel and 62 percent in Knitted and Crocheted Fabrics to the low of 8 or 9 percent in Motor Vehicles, Bodies of Motor Vehicles, Building and Repairing of Ships, and Railway Locomotives.11 One may be concerned that F Li could simply be a proxy for skill intensity (since women supply relatively more “brain” than “brawn” labor input compared to men). However, it turns out that F Li is uncorrelated with skill intensity.12 X The export shares !ict are calculated based on the COMTRADE database, which contains bilateral trade data starting in 1962 in the 4-digit SITC revision 1 and 2 classification. The trade data are aggregated up to the 3-digit ISIC Revision 3 classification using a concordance developed by the authors. 11 A potential concern that these values may be very di↵erent across countries in general, and across developed and developing countries in particular. However, it turns out that the rankings of sectors are remarkably similar across countries. The values of F Li computed on the OECD and non-OECD samples have a correlation of 0.9. The levels are similar as well, with the average F Li in the OECD of 0.29, and in the non-OECD of 0.27 in this sample of countries. Pooling all the countries together, the first principal component explains 77 percent of the cross-sectoral variation across countries, suggesting that rankings are very similar. We also experimented with taking alternative averages: medians instead of means across countries; and dropping outlier values of female shares in individual sectors. The results were very similar. Another concern is that F Li is measured based on data from the last 10-15 years, whereas our estimation sample goes back several decades. We are not aware of similar data for earlier periods. Our measure of F Li can be combined with data for earlier time periods as long as there are no “gender-intensity reversals” over time, that is, the ranking of industries by female intensity is stable. 12 The correlation between F Li and the share of skilled workers in the total wage bill is 0.06, and the correlation between F Li and the share of skilled workers in total industry employment is -0.06. The skill intensity data come from Autor et al. (1998), who compute these measures for the U.S.. Unfortunately, we cannot compute skill intensity measures from the UNIDO data used to compute F Li , as these data do not include employment breakdowns by education level. 18 Table 2 reports some summary statistics for the female labor needs of exports for the OECD and non-OECD country groups. We observe that for the OECD, the measure is relatively stable across decades, with an average of about 0.25. For the non-OECD countries, the female labor needs of exports is higher, between 0.27 and 0.30, and, if anything, rising over time. Notably, the dispersion in F N LX among the non-OECD countries is both much greater than among the OECD, and increasing over time. In the OECD sample, the standard deviation is stable at 0.03-0.04, whereas in the non-OECD sample it rises monotonically from 0.08 to 0.12 between the 1960s and the 2000s. Tables 3 reports the countries with the highest and lowest F LN X values. Typically, countries with the highest values of female content of exports are those that export mostly textiles and wearing apparel, while countries with the lowest F LN X are natural resource exporters. Equally important for our empirical strategy are changes over time. Table 4 reports the countries with the largest positive and negative changes in F LN X between the 1960s and today. We can see that relative to the cross-sectional variation, the time variation is also considerable. For the countries with the largest observed increases in F LN X , the common pattern is that they change their specialization from agriculture-based sectors to wearing apparel. For instance, in the 1960s 80% of exports from Cambodia were in the food products sectors (ISIC 151 through 154). By the 2000s, 85% of Cambodian exports are in ISIC 181, “Wearing apparel.” The other countries in the top 10 largest positive changes in F LN X follow this pattern as well. Since food products sectors are right in the middle of the F Li distribution, and “Wearing apparel” is the most female-intensive sector, this type of specialization change will lead to large increases in F LN X . The largest observed decreases in F LN X are driven by the discovery of natural resources. For instance, Niger was an agricultural exporter in the 1960s, with nearly 80% of exports in ISIC 151, “Meat, fish, fruit, vegetables, oils and fats.” By the 2000s, over 60% of Niger’s exports were in “Refined petroleum products” (ISIC 232) and “Nuclear fuel” (ISIC 233). The natural resource-based sectors are among the least female-intensive, with F Li of 0.11- 0.13, which accounts for why countries with major shifts towards natural resources exhibit reductions in their F LN X . It turns out that these two groups of countries experienced very di↵erent changes in fertility. Among the 10 countries with the largest increases in F N LX , fertility fell on average by 3.5 children per woman, from 6.5 to 3 between the 1960s and the 2000s. By contrast, in the 10 countries with the largest decreases in F N LX , fertility fell by only 1.3 children per woman over the same period, from 6.9 to 5.6. Remarkably, while these two groups had similar fertility levels in the 1960s (6.5 and 6.9), their subsequent paths were very di↵erent. 19 This is of course only an illustrative example, and we explore these patterns formally in the next section. Data on fertility are sourced from the World Bank’s World Development Indicators. The baseline controls – PPP-adjusted per capita income and overall trade openness – come from the Penn World Tables. Table 2 presents the summary statistics for fertility (number of births per women) in each decade and separately for OECD and non-OECD countries. There is considerable variation in fertility across countries: while the median fertility after 1980 is 3.3 births per woman in our sample of countries, the standard deviation is 1.8, and the 10th-90th percent range spans from 1.4 to 6.3. The table highlights the pronounced cross- sectional di↵erences between high- and low-income countries, as well as the secular reductions in fertility over time in both groups of countries. Our final dataset contains country-level variables on up to 145 countries. 5 Empirical Results 5.1 Cross-sectional results Table 5 reports the results of estimating the cross-sectional specification in equation (15). Both left-hand side and the right-hand side variables are in natural logs. All of the specifica- tions control for income per capita and overall openness. Column 1 presents the OLS results. There is a pronounced negative relationship between the female-labor need of exports and fertility, significant at the one percent level. By contrast, the coe cient on overall trade openness is zero to the second decimal point and not significant. As is well known, income per capita is significantly negatively correlated with fertility. These three variables absorb a great deal of variation in fertility across countries: the R2 in this regression is 0.63. Column 2 repeats the OLS exercise but including the regional dummies.13 The R2 increases to 0.86, but the female labor need of exports remains equally significant. Figure 1 displays the partial correlation between fertility and F N LX from Column 2 of Table 5. Column 3 implements the 2SLS procedure. The bottom panel displays the results of the first stage. As expected, the instrument is highly significant with a t-statistic of 9.4, and the F -statistic for the excluded instrument of 43 is comfortably within the range that allows us to conclude that the instrument is strong (Stock and Yogo, 2005). Figure 2 presents the partial correlation plot from the first stage regression between F N LX and the instrument. There is a clear positive association between the two variables that does not appear to be 13 The regional dummies correspond to the o cial World Bank region definitions: East Asia and Pacific, Europe and Central Asia, Latin America and the Caribbean, Middle East and North Africa, North America, South Asia, and Sub-Saharan Africa. 20 driven by a few outliers. As expected, the variation in the instrument is much smaller than the variation in the actual F N LX . The instrument is predicting F N LX while throwing out a great deal of country-specific information, and thus the instrument’s predictions for the country-specific F N LX vary much less across countries than do actual values. In the second stage, the main variable of interest, F N LX , is statistically significant at the one percent level, with a coe cient that is about one-third larger in absolute value than the OLS coe cient. Column 4 repeats the 2SLS exercise adding regional dummies. The second- stage coe cient of interest both increases in absolute value and becomes more statistically significant. The OLS and 2SLS results described above constitute the main cross-sectional finding of the paper. Countries that have a comparative advantage in the female-intensive sectors exhibit lower fertility. The estimates are economically significant. Taking the coe cient in column 4 as our preferred estimate, a 10 percent change in F N LX leads to a 4.7 percent lower fertility rate. In absolute terms, this implies that moving from the 25th to the 75th percentile in the distribution of the female content of exports lowers fertility by as much as 20 percent, or about 0.36 standard deviations of average fertility across countries. Applied to the median of 3.3 births per woman in this sample of countries, the movement from the 25th to the 75th percentile in F LN X implies a reduction of 0.64 births per woman. 5.2 Panel Results The cross-sectional 2SLS results are informative, and allow us to make the clearest case for the causal relationship between comparative advantage and fertility. However, because they do not allow the use of country fixed e↵ects, the cross-sectional results may still su↵er from omitted variables problems. As an alternative empirical strategy, we estimate the panel specification (16) on non-overlapping 5-year and 10-year averages from 1962 to 2007. The gravity-based instrumentation strategy is not feasible in a panel setting with fixed e↵ects. On the other hand, country e↵ects allow us to control for a wide range of unobservable time- invariant country characteristics, and identify the coe cient of interest from the variation in F N LX and fertility within a country over time. The results are presented in Table 6. To control for autocorrelation in the error term, all standard errors are clustered at the country level. Column 1 reports the results for the pooled specification without any fixed e↵ects. The coe cient is remarkably similar to the OLS coe cient from column 1 of Table 5. Column 2 adds country fixed e↵ects. The coe cient on F N LX is nearly unchanged, and significant at the one percent level. Column 3 adds time e↵ects to control for secular global trends, while column 4 adds female educational 21 attainment. The results continue to be highly significant. Columns 5–8 repeat the exercise taking 10-year averages instead.14 The coe cients are very similar in magnitude and equally significant. 5.3 Robustness We now check the robustness of the cross-sectional result in a number of ways. The first set of checks is on how the instrument construction treats zero trade observations. As detailed in Section 3.1, the baseline instrument estimates the standard log-linear gravity specification that omits zeros in the trade matrix, and predicts trade only for those values in which observed trade is positive. We address the issue of zeros in two ways. The first is to predict trade values for the observations in which actual trade is zero based on the same log-linear regression. The second is to instead estimate a Poisson pseudo-maximum likelihood model on the levels of trade values, as suggested by Santos Silva and Tenreyro (2006). In this exercise, the zero trade observations are included in the estimation sample. The results of using those two alternative instruments are presented in columns 5 and 6 of Table 5. It is clear that very little is changed. The instruments continue to be strong, and the second-stage coe cients of interest are similar in magnitude and significant at the one percent level. We conclude from this exercise that the way zeros are treated in the construction of the instrument does not a↵ect the main results. Another concern is that the instrument is constructed based on variables – such as pop- ulation – that do not satisfy the exclusion restriction. Note that the instrument relies on the di↵erential impact of each gravity variable across sectors, as determined by the sectoral variation in non-country-specific gravity coe cients. To further probe into the importance of the country-specific gravity variables, column 7 of Table 5 implements the instrument with- out the exporter population (the population of each particular trading partner is plausibly exogenous to the exporting country’s fertility). The instrument remains strong, as evidenced by the first stage diagnostics, and the main result is robust. Alternatively, column 8 controls for area and population directly. Area is insignificant as a determinant of fertility, and popu- lation comes in with the right sign, but the size of the coe cient, interpreted as an elasticity, is small. The coe cient of interest remains significant and of similar magnitude. Table 7 performs a number of additional specification checks. All columns report the 2SLS results controlling for openness, income, and regional dummies. First, we may expect the impact of F N LX to get stronger with openness. Column 1 checks this by adding an interaction term between F N LX and overall openness. As expected, the interaction 14 To be more precise, these are decadal averages for the 1960s, 1970s, and up to 2000s. Since our yearly data are for 1962-2007, the 1960s and the 2000s are averages over less than 10 years. 22 coe cient is significant: in more open countries the e↵ect we highlight is more pronounced.15 Next, it might be that what matters is the female labor need of net exports. That is, perhaps a country imports a lot of the female-labor intensive goods, in which case its domestic demand for female labor will be lower. This is unlikely to be a major force on average, as import baskets tend to be more similar across countries than export baskets. Most countries specialize in a few sectors, but import a broad range of products. Indeed, in our data the standard deviation of the “female labor need of imports ” (F N LI ) is 3.6 times smaller than the standard deviation of F N LX . Nonetheless, to check the robustness of the results, we use the female labor need of net exports, F N LX F N LI , as the independent variable. Since it can take negative values, we must use levels rather than logs. As the instrument, we use the level of predicted F N LX , rather than log. Column 2 of Table 7 reports the results, and shows that they are robust to using this alternative regressor of interest. Next, we check whether the results are robust to including additional controls. Column 3 controls for female schooling, to account for the possible relationship between education and fertility. Female schooling is measured as the average number of years of schooling in the female population over 25, and is sourced from Barro and Lee (2000). While higher female schooling is indeed associated with lower fertility, the coe cient on F N LX changes little and continues to be significant at the one percent level. Column 4 controls for the prevalence of child labor, since fertility is expected to be higher when children can contribute income to the household. Child labor is measured as the percentage of population aged 10-14 that is working, and comes from Edmonds and Pavcnik (2006). While the prevalence of child labor is indeed positively associated with fertility, the main coe cient of interest remains robust. Column 5 controls for infant mortality, sourced from the World Bank’s World Development Indicators. Countries with higher infant mortality have higher fertility, but our coe cient of interest remains robust. Next, column 6 controls for income inequality, using the Gini coe cient from the World Bank’s World Development Indicators. Higher inequality is associated with higher fertility, but once again the main result is robust. Finally, column 7 controls for the extent of democ- racy, using the Polity2 index from the Polity IV database. The extent of democracy is not significantly associated with fertility, and F N LX is still significant at the one percent level. Table 8 checks whether the finding is driven by particular countries. Column 1 drops 15 The main e↵ect of F N LX is now positive, but of course the overall e↵ect is a combination of the main e↵ect and the coe cient on openness times openness. The distribution of openness in this sample of countries is such that the point estimate of the combined e↵ect of F N LX , which is equal to 1.68 0.49⇥Log (Openness), is positive for all but the bottom 5% least open countries. The table does not report the first-stage coe cients and diagnostics in order to conserve space since there are now two variables being instrumented. The F - statistics associated with both instrumented variables are in excess of 35. 23 outliers: the top 5 and bottom 5 countries in the distribution of F N LX . Column 2 drops the OECD countries, to make sure that our results are not driven simply by the distinction between high-income countries and everyone else.16 Column 3 drops the Middle East and North Africa region, and column 4 drops Sub-Saharan Africa. It is clear that the results are fully robust to dropping outliers and these important country groups. The coe cients are similar to the baseline and the significance is at one percent throughout. Finally, column 5 drops mining exporters, defined as countries that have more than 60% of their exports in Mining and Quarrying, a sector that includes crude petroleum.17 The results are una↵ected by dropping these countries. Finally, one may be concerned that our sample includes only manufacturing sectors. To the extent that some countries export significant amounts of agricultural and mining raw materials, our manufacturing-based F N LX may not accurately reflect the gender bias of a country’s specialization pattern. To address this coverage issue, we also constructed F Li based on data for a single country – the U.S. – using the Labor Force Statistics database of the U.S. Bureau of Labor Statistics (BLS). The BLS has published “Women in the Labor Force: A Databook” on an annual basis since 2005. It contains information on total employment and the female share of employment in each industry covered by the Census, sourced from the Current Population Survey. The data are available at the 4-digit U.S. Census 2007 classification (262 distinct sectors, including both manufacturing and non-manufacturing). In order to construct the share of female workers in total sectoral employment F Li , we take the mean of this value across the years for which the data on the female share of employment are available (2004-2009). After dropping non-tradables, the sector sample includes 78 manufacturing and 15 non-manufacturing sectors. An earlier working paper version of our paper (Do et al., 2012) replicates all of the empirical analysis using this alternative measure of F Li , and shows that the results are robust. Thus, we do not report them here to conserve space.18 16 OECD countries in the sample are: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, Portugal, Spain, Swe- den, Switzerland, the United Kingdom, and the United States. We thus exclude the newer members of the OECD, such as Korea and Mexico. 17 These countries are Algeria, Angola, Republic of Congo, Gabon, Islamic Republic of Iran, Kuwait, Nigeria, Oman, Saudi Arabia, and Syrian Arab Republic. 18 While the U.S.-based alternative F Li measure has the advantage of extending the set of sectors to agriculture and mining, it has two important drawbacks. First, the data are compiled based on individual- level surveys rather than firm- or plant-level data, and thus relies on workers self-reporting their industry of occupation. If the number of individuals in the survey who report working in a particular sector is small, or if workers make mistakes in reporting their industry of employment, the data will be measured with error. And second, the U.S. is only one, very special country, and thus its values of F Li may not be representative of the average country’s experience. For our UNIDO-based measure, averaging the share of female workers across a couple of dozen countries helps alleviate both of these problems. 24 5.4 Mechanisms and Other Outcome Variables The women’s opportunity-cost-of-time hypothesis has a natural counterpart in another use of time, namely female labor force participation (FLFP). We should expect that an increase in comparative advantage in female-intensive sectors, as it lowers fertility, should also increase FLFP. Appendix B discusses this issue at length and estimates the relationship between comparative advantage in female-intensive sectors and FLFP. It appears that comparative advantage in female-intensive sectors increases FLFP, but only for countries with lower levels of income and female educational attainment and higher fertility. We argue that this type of conditional relationship should be expected, given that there is no simple relationship between fertility and FLFP, either in theory or in the data. The results with respect to FLFP are nonetheless supportive of the main hypothesis in the paper. Similarly, one may expect to see a lower gender wage gap in countries with a comparative advantage in female-intensive industries. Unfortunately, testing this hypothesis is even more challenging than for FLFP. First and foremost, there are no reliable and comprehensive data on the gender wage gaps for a large enough cross-section (much less a panel) of countries. Second, even if data on the gender wage gaps were available, actual wage outcomes are a↵ected by worker heterogeneity in a number of ways that would be challenging to account for in estimation. Across countries, there are large di↵erences in the age, work experience, and education distributions of both the male and female labor force. Within countries, wages are only observed conditional on working, which introduces sample selection problem. Controlling concincingly for these major confounding factors would be infeasible in our cross- country context. 6 Conclusion Fertility is an economic decision, and like all economic decisions has long been considered an appropriate – and important – subject of analysis by economists. As trade integration increased in recent decades, there is growing recognition that the impacts of globalization are being felt well beyond the traditional market outcomes such as average wages, skill premia, and (un)employment. This paper makes the case that international trade, or more precisely comparative advantage, matters for this key non-market outcome: the fertility decision. Our results thus emphasize the heterogeneity of the e↵ects of trade on countries’ industrial structures and gender outcomes. At a more conjectural level, to the extent that comparative advantage impacts fertility, it may also impact women’s human capital investments, occu- pational choice, and bargaining power within the household. From a policy perspective, our 25 results suggest that it will be more di cult for countries with technologically-based compar- ative advantage in male-intensive goods to undertake policy measures to reduce the gender gap, potentially leading to a slower pace of women’s empowerment. In an increasingly inte- grated global market, the road to female empowerment is paradoxically very specific to each country’s productive structure and exposure to international trade. At the same time, since our paper points to comparative advantage as a determinant of women’s opportunities, a potential policy lever to a↵ect the gender gap could be through industrial policy promoting female-intensive sectors. Appendix A Proofs Proof of Proposition 1 The “goods market-clearing curve” and “factor market-clearing curve” have opposite slopes. We therefore need to show that they intersect at least once, since if they do, such intersection is unique. A necessary and su cient condition for the two curves to intersect is that the “goods market-clearing curve” be above the “factor market-clearing curve” for low values of f c and below for larger values of ✓c . • As ✓c gets arbitrarily close to 0, equality (9) implies that the “goods market-clearing” curve is bounded below by 1 1 ⌘ , while (10) indicates that the “factor market-clearing” curve converges to 1 < 1 1 ⌘ , and therefore lies below the “goods market-clearing” curve. • On the other hand, when ✓c grows arbitrarily large, the “goods market-clearing” curve converges to 1 1 ⌘ , while the “factor market-clearing” diverges, and hence lies above the “goods market-clearing” curve. Thus, the “goods market-clearing” curve is above the “factor market-clearing” curve in the neighborhood of 1, while the opposite holds for large values of ✓c . Continuity of the two curves implies existence of an intersection.⌅ Proof of Lemma 1 From equation (9), let’s try to characterize the behavior of ✓c when the patterns of compar- ative advantage ⇢ are changing. Dropping the country reference and substituting for ✓ c , f is implicitly defined for every ⇢ by: ✓ ◆↵  ✓ c ⌘ (1 ↵) ↵ ✓ + 1 [⌘ (1 ⌘ ) ✓] + ( ) (1 + ✓) ⌘ (1 ⌘ ) = 0 ⇢ ⇢ that is denoted x (✓, ⇢) = 0. On the one hand, ✓ ◆↵ 1 @ x (✓, ⇢) ↵✓ ✓ (1 ⌘) ✓ = +1 [⌘ (1 ⌘ ) ✓] + ( c )⌘(1 ↵) (1 + ✓)↵ @⇢ ⇢2 ⇢ ⇢2 26 and since x (✓, ⇢) = 0, we can rewrite  @ x (✓, ⇢) (1 + ✓)↵ ✓ ✓ = ( c )⌘(1 ↵) ↵⌘ + (1 ⌘ ) + (1 ↵) (1 ⌘) @⇢ ⇢ ⇢+✓ ⇢ On the other hand, similar derivation yields ✓ ◆⇢ @ x (✓, ⇢) c ⌘ (1 ↵) ↵ ⇢ 1 ↵ [⌘⇢ ✓(1 ⌘ )] ⌘ (1 ⌘) = ( ) (1 + ✓) + @✓ ⇢ (1 + ✓) (✓ + ⇢) [⌘ (1 ⌘ ) ✓] The implicit function theorem indicates that ✓ (⇢) is well defined and continuously di↵er- entiable around ⇢ such that x(✓(⇢), ⇢) = 0; we can now compute the derivative of ✓ with respect to ⇢ : h i ✓ (1 ⌘ ) ✓ ⌘ ✓ (1 + ✓ ) ↵⌘ + (1 ⌘ ) + (1 ↵ ) (1 ⌘ ) ⇢ ✓ 0 ( ⇢) = ⇢ 1 ⌘⇢ [↵ + (1 ↵) (1 ⌘ ) (1 + ✓)] + ✓ (1 ⌘ ) [↵✓ + (1 ↵) ⌘ (1 + ✓)] The second term of the equation is always positive; by virtue of (9) and (10), the first term (1 ⌘ )✓ ⌘ ⇢ 1 > 0. We thus have ✓0 (⇢) > 0. ⌅ Proof of Lemma 2 Having established that the female labor demand curve is downward sloping for every level of country c’s female labor force participation and that the female labor supply curve is upward sloping, we have shown uniqueness of an intersection. We now need to show existence of an intersection. • As N c goes to zero (i.e. female labor supply goes to 1), the labor supply curve defined by (12) diverges given that lim0 v 0 (.) = +1, by assumption. The labor demand curve is on the other hand bounded above since it is downward sloping; it therefore lies below the labor supply curve. • Let’s now let N c get arbitrarily close to 1 , so that ⇢c converges to zero. Equation (10) implies that ✓c will converge to 0, so that, by virtue of (9), ✓ c will converge to some ⇥ ⇤ ¯ c > 0 such that ⌘ + ( c )⌘(1 ↵) ✓ ✓ ¯ c + 1 ↵ ⌘ (1 ⌘ ) ✓ ¯ c = 0. Thus, the labor c demand curve converges to some positive wage w ¯F . Two cases arise: v0 ( 1 ) c – if 0, such that N c (0) = 1 "c . Next, and given that N c (.) is decreasing, we have N c (N ) 2 [0, 1 "c ] , a compact set. Suppose now that N c is set arbitrarily close to 1 . Then, (10) implies that ✓ c converges to 0, uniformly with respect to N c ; (9) in turn implies that ✓c converges towards some ✓ ¯c < 1 such that ↵ ⇥ ⇤ ⌘ + ( c )⌘(1 ↵) ✓¯c + 1 ⌘ (1 ⌘ ) ✓ ¯c = 0. Equation (5) indicates that female wages in country c remain bounded above, so that lim 1 N c (.) > 0. Thus, the curve N c (.) cuts N c (.) at least once, and “from above,” as shown in Figure A2 below. This establishes the existence of an equilibrium N X , N Y . Uniqueness To show uniqueness, we look at the labor market equilibrium. For an interior solution, we note that {(✓c , N c )}c2{X,Y } are implicitly defined by the intersection of labor supply and demand, i.e. ✓ ◆↵ ✓ ◆1 ⌘ (1 ↵) v 0 (N c ) 1 ✓c = (1 ↵) . (A.1) 1 + ✓c 1 Nc N c can thus be expressed as a function N (.) of ✓c and exogenous parameters only such that N (.) is continuously di↵erentiable and simple algebra yields for an interior solution: 1 dN (✓) 1 N (✓ ) 1 1 ⌘ (1 ↵) 1+✓ ↵ ✓ = v ”[N (✓ )] 1 N (✓ ) 0 (A.2) d✓ ✓ v 0 [N (✓ )] 1 ⌘ (1 ↵) We now turn to the system of equilibrium conditions (9) and (10) that are conditional on la- bor endowments 1 NX, 1 N Y . On the one hand, (9) defines a negative unconditional relationship between ✓c and ✓ c ; on the other hand, we rewrite (10) as ✓c c ✓ c = (A.3) 1 Nc 1 N c that can be written uc (✓c ) = c u c (✓ c ) , where uc (✓) = 1 ✓ N (✓ ) . Inequality (A.2) implies c that u (.) is increasing, so that (A.3) defines a positive unconditional relationship between ✓c and ✓ c . Thus, the two equilibrium conditions for capital define two curves with opposite slope, implying a unique intersection, given that existence was established above. Uniqueness of capital allocation across sectors implies uniqueness of fertility decisions.⌅ 28 Proof of Lemma 3 The ratio of female wages in the two countries and use (10) to obtain the following equality: ✓ ◆↵ v 0 (N c ) 1 + ✓c = ( c )1 ⌘(1 ↵) . (A.4) v 0 (N c ) 1 + ✓ c v 0 (N˜ c) c v 0 (N c ) 1+✓ c 1+✓˜c Equality (A.4) implies that if ˜ c then either v 0 (N c ) v 0 (N˜ c) or 1+✓ c 1+✓˜ c, or both. In other words, a change in comparative advantage triggers either a change in fertility choices in either or both countries (N c  N ˜ c and/or N c ˜ c ), or a reallocation of N capital across sectors in either or both countries (✓ c ˜c ✓ and/or ✓ c  ✓ ˜ c ). However, since c = 1/ c , a stronger comparative advantage in the F -good in country c is associated with a weaker comparative advantage in country c, vice and versa. Therefore, if a change in comparative advantage positively (resp. negatively) a↵ects fertility in country c, it will simultaneously negatively (resp. positively) a↵ect fertility in country c. The same holds for capital allocation. Thus, we can state the following: ⇣ ⌘ ⇣ ⌘ c c c ˜ ˜ =) N  N and N c c N˜ c or ✓ c ˜c ✓ and ✓  ✓c ˜ c (A.5) Finally, to see that both fertility and capital allocation respond to an exogenous change in comparative advantage, we note that the right-hand side of (A.1) is increasing in ✓c , while the left-hand side is decreasing in N c . The following equivalence therefore holds: ✓c ˜c () N c  N ✓ ˜ c. (A.6) That is, a higher inflow of capital in the F -sector is associated with higher female labor force participation and hence lower fertility in equilibrium. Equivalence (A.6) implies that the last term in (A.5) is therefore redundant and we can simply write ⇣ ⌘ c ˜ c =) N c  N ˜ c and N c N ˜ c . (A.7) ⌅ Proof of Theorem 1 To move from comparative statics to cross-sectional comparisons, we set ˜ c = 1. Interior solutions Equilibrium conditions (9) and (10) and labor market clearing equa- ˜ c = N 0, ˜c = N tions (A.1) are thus symmetric in both (N c , N c ) and (✓c , ✓ c ), implying N 0 ˜c ˜ c 1 where N satisfies (A.1) with ✓ = ✓ = 1 ⌘ . Implication (A.7) becomes for ˜ = 1: c c 1 =) N c  N 0  N c . Corner solutions Finally, since the arguments leading to Proposition 4 assume interior solutions for equilibrium fertility in both countries, we now address the cases in which the 29 1 labor market equilibrium is at a corner (i.e. N c = or N c = 1 ). Without loss of generality, suppose that c 1. • If N c = 1 , i.e. the F -sector in country c disappears, then N c < 1 (since N c = 1 implies that ✓c = 0, and (9) does not hold for ✓c = ✓ c = 0), and the proposition trivially holds. Indeed, if c0 s comparative advantage in the F -sector is large enough, then c will end up producing all the F -goods in the economy. 1 • Alternatively, suppose that N c = and N c < 1 . Female wages are given by ✓ c ◆1 ⌘ (1 ↵) ✓ ◆ c c ✓ 1 0 1 wF = (1 ↵) c  v 1 N ✓ c ◆1 ⌘ (1 ↵) ✓ 1 c wF = (1 ↵) c = v0 N c 1 N and since N c < 1 , and v 0 (.) is decreasing, we have v 0 (N c )>v 1 so that wF c > wF c . This implies c < 1, a contradiction. • Finally, N c = N c = 1 cannot be an equilibrium since no production would take place, thus pushing female wages in both countries to infinity. This concludes the proof.⌅ Appendix B Female Labor Force Participation The theoretical model in Section 2 connects comparative advantage to fertility through the opportunity cost of women’s time. This mechanism is related to female labor force participa- tion (FLFP). This section presents a set of empirical results on how comparative advantage a↵ects FLFP. To clarify the connections between these and the baseline results, we preface the empirics with a theoretical discussion of the relationship between fertility and FLFP. B.1 Theoretical Discussion In the simple model of Section 2, fertility is perfectly negatively correlated with FLFP, which, if taken literally, conveys the impression that comparative advantage a↵ects fertility “through” FLFP. However, the notion that fertility is a↵ected by the opportunity cost of women’s time is distinct from women’s labor supply for a series of reasons. First, the elasticity of FLFP with respect to women’s wage is not simply the negative of the elasticity of fertility with respect to the wage. Suppressing the country superscripts, let N , as before, be the number of children, and denote FLFP by LF = 1 N . Denote the @ x wF elasticity of a variable x with respect to the female wage by "x ⌘ wF x . It is immediate that "LF = "N 1 NN . Thus, for a finite "N , the elasticity of FLFP with respect to the wage 30 approaches zero as childrearing time goes to zero, either because of low or low N . This suggests that in countries with already low fertility, or in countries with low (for instance, due to easily accessible childcare facilities, as in many developed countries) the impact of (log) opportunity cost of women’s time on (log) FLFP may not be detectable.19 Second, even in levels the negative linear relationship between fertility and labor supply is an artifact of the assumption that working in the market economy and childrearing are the only uses of women’s time. More generally, suppose that there is another use of women’s time, Q, which can stand for leisure, investments in quality of the children (as opposed to quantity N ), or non-market housework. Suppose further that the indirect utility, instead of (11), is now represented by: V (N, Q) = r + wF (1 N µQ) + wM + v (N ) + z (Q) , (B.1) where µ is number of units of a woman’s time required to produce one unit of Q. On the one hand, this addition leaves unchanged the first-order condition with respect to fertility, (12), embodying the notion that fertility is a↵ected by the opportunity cost of women’s time. On the other hand, there is now another first-order condition that relates women’s op- portunity cost of time to Q: z 0 ( Q) wF = . (B.2) µ Thus, the relationship between FLFP and wF is now LF = 1 (v 0 ) 1 ( wF ) µ(z 0 ) 1 (µwF ), and the elasticity of FLFP with respect to the wage is N µQ "L F = "N "Q . 1 N µQ 1 N µQ It is immediate that FLFP and fertility are no longer inversely related one-for-one. Depending on the curvatures of v (.) and z (.), FLFP could be more or less concave in wF than N , even as (12) continues to hold and the wage-fertility relationship is una↵ected. When "Q is di↵erent from "N , and µQ is high relative to N , "LF can look very di↵erent from negative "N even when women’s labor supply is far away from 1.20 Third, the simple model above assumes that the marginal utility of income is always constant at 1. Departing from that assumption and introducing diminishing marginal utility c of income will make the relationship between FLFP and wF even more complex, and possibly non-monotonic, due to income e↵ects. While in all of the cases above, FLFP and fertility were still negatively correlated, with income e↵ects it is possible to generate a positive relationship 19 To give a stark example, suppose that v (.) is CES: v (N ) = N 1 1/⇣ /(1 1/⇣ ), so that the elasticity of fertility with respect to the wage is simply constant: "N = ⇣ . In this case, we will always be able to detect the e↵ect of (log) wage on (log) fertility at all levels of fertility or income, whereas the impact of (log) wage on (log) FLFP will go to zero as income rises/fertility falls. 20 As an example, when v (.) and z (.) are CES: v (N ) = N 1 1/⇣ /(1 1/⇣ ) and z (Q) = Q1 1/⇠ /(1 1/⇠ ), N µQ "Q and "N are simply constants, and "LF = ⇣ 1 N µQ + ⇠ 1 N µQ , which can obviously be very di↵erent from ⇣ . 31 between FLFP and fertility at high enough levels of income, for instance through satiation in goods consumption. Finally, when it comes to measurement of FLFP, an additional challenge is that the model is written in terms of the intensive margin (i.e. hours), whereas the FLFP data are recorded at the extensive margin (binary participation decision). This implies that, especially for countries with already high FLFP, in which in response to fertility women adjust hours worked rather than labor market participation, our data will not be able to accurately capture the interrelationships between FLFP and fertility.21 To summarize, the insight that fertility is determined by the opportunity cost of women’s time does not have a one-to-one relationship to FLFP. One can easily construct examples in which the wage elasticities with respect to fertility and FLFP are very di↵erent.22 In addition, even the simple baseline model above implies that the elasticity of female labor supply with respect to the opportunity cost of women’s time is not constant, and approaches zero as time spent on childrearing falls. This suggests that the impact of comparative advantage in female-intensive goods on FLFP will be attenuated, and potentially di cult to detect in countries with high income and low fertility. B.2 Empirical Results With those observations in mind, Table A3 explores the relationship between F N LX and FLFP. FLFP data come from the ILO’s KILM database, and are available 1990-2007. All shown specifications include controls for per capita income and openness, and regional dum- mies. Column 1 presents the OLS regression. The coe cient on FLFP is positive but not significant. Column 2 reports the 2SLS results. The coe cient becomes larger, but not significant at conventional levels (p-value of 11%). However, as argued above the elasticity of FLFP with respect to F N LX should not be expected to be constant across a wide range of countries. Thus, in columns 3 and 4 we re-estimate these regressions while letting the impact of F N LX vary by income. The di↵erence is striking. Both the main e↵ect and the interaction with income are highly significant, and the impact of F N LX is clearly less pro- nounced for higher-income countries. Column 5 reports the 2SLS results in which F N LX is interacted with fertility, and column 6 with female educational attainment. In both cases, all of the coe cients of interest are highly significant.23 Of course, the main e↵ect of the F N LX is now not interpretable as the impact of F N LX on FLFP. To better illustrate how the impact of F N LX on FLFP varies through the distri- bution of income, fertility, and educational attainment, we re-estimate the specification with quartile-specific F N LX coe cients, rather than the interaction terms (that is, we discretize 21 Unfortunately, data on hours worked are not available for a large sample of countries. 22 Indeed, in the data there is no simple negative relationship between fertility and FLFP. For instance, Ahn and Mira (2002) show that it is not stable even among the OECD countries: FLFP was was negatively correlated with fertility until the 1970s and 1980s, and but since then the correlation changed sign, and fertility is now positively correlated with FLFP. 23 In order to conserve space, Table A3 does not report the first-stage coe cients and diagnostics. With the income, fertility and educational attainment interactions, two variables are being instrumented, which would require reporting multiple coe cients and F -statistics. All of the F -statistics in these specifications are above 25. 32 income, fertility, or female educational attainment into quartiles, and allow the F N LX co- e cient to di↵er by quartile). Figure A3 reports the quartile-specific coe cient estimates, with the bars depicting 95% confidence intervals. The top panel presents the results by quartile of income. There is a statistically significant positive e↵ect of F N LX on FLFP in the bottom quartile of countries, with the coe cient estimate of 0.53. In the second quartile, the coe cient is positive at 0.36, but no longer significant. In the top half of the income distribution, the coe cient estimates are close to zero and not significant. The second panel presents the same result with respect to fertility. As expected, the impact of F N LX on FLFP is most pronounced at high levels of fertility. The top quartile estimate is statistically significant at the 1% level, and the third quartile coe cient is signif- icant at the 10% level. Finally, the bottom panel presents the results with respect to female educational attainment quintiles. The impact of F N LX is strongly positive in the bottom quartile, and close to zero elsewhere. To summarize, the results with respect to FLFP are suggestive that the impact of com- parative advantage on fertility is concomitant with a female labor supply response, but only in some countries. As argued above, this is should be expected, given that the relationship between FLFP and fertility is not straightforward. 33 References Aguayo-Tellez, Ernesto, Jim Airola, and Chinhui Juhn, “Did Trade Liberalization Help Women? The Case of Mexico in the 1990s,” July 2010. NBER Working Paper No. 16195. Ahn, Namkee and Pedro Mira, “A note on the changing relationship between fertility and female employment rates in developed countries,” Journal of Population Economics, 2002, 15 (4), 667–682. Alesina, Alberto, Paola Giuliano, and Nathan Nunn, “Fertility and the Plough,” American Economic Review, Papers and Proceedings, May 2011, 101, 499–503. , , and , “On the Origins of Gender Roles: Women and the Plough,” Quarterly Journal of Economics, May 2013, 128 (2), 469–530. Almeida, Heitor and Daniel Wolfenzon, “The E↵ect of External Finance on the Equilibrium Allocation of Capital,” Journal of Financial Economics, January 2005, 75 (1), 133–164. Anderson, James and Eric van Wincoop, “Gravity with Gravitas: a Solution to the Border Puzzle,” American Economic Review, 2003, 93 (1), 170–192. and , “Trade Costs,” Journal of Economic Literature, 2004, 42 (3), 691–751. Angrist, Joshua D. and William N. Evans, “Children and Their Parents’ Labor Supply: Evidence from Exogenous Variation in Family Size,” American Economic Review, June 1998, 88 (3), 450–77. Autor, David H., Lawrence F. Katz, and Alan B. Krueger, “Computing Inequality: Have Computers Changed The Labor Market?,” Quarterly Journal of Economics, November 1998, 113 (4), 1169–1213. Barro, Robert J. and Gary S. Becker, “Fertility Choice in a Model of Economic Growth,” Econometrica, March 1989, 57 (2), 481–501. and Jong-Wha Lee, “International Data on Educational Attainment: Updates and Impli- cations,” 2000. CID Working Paper No. 42. Becker, Gary S., “An Economic Analysis of Fertility,” in Gary S. Becker, ed., Demographic and Economic Change in Developed Countries., Princeton, NJ: Princeton University Press, 1960. , “A Theory of the Allocation of Time,” Economic Journal, September 1965, 75 (299), 493–517. , A Treatise on the Family, Cambridge, MA: Harvard University Press, 1981. , “Human Capital, E↵ort, and the Sexual Division of Labor,” Journal of Labor Economics, January 1985, 3 (1), S33–S58. 34 , Kevin M. Murphy, and Robert Tamura, “Human Capital, Fertility, and Economic Growth,” Journal of Political Economy, October 1990, 98 (5), S12–S37. Berkowitz, Daniel, Johannes Moenius, and Katharina Pistor, “Trade, Law, and Product Complexity,” The Review of Economics and Statistics, May 2006, 88 (2), 363–373. Berthelon, Matias and Caroline Freund, “On the Conservation of Distance in International Trade,” Journal of International Economics, July 2008, 75 (2), 310–320. Black, Sandra E. and Alexandra Spitz-Oener, “Explaining Women’s Success: Technological Change and the Skill Content of Women’s Work,” Review of Economics and Statistics, April 2010, 92 (1), 187–194. and Chinhui Juhn, “The Rise of Female Professionals: Are Women Responding to Skill Demand?,” American Economic Review, Papers and Proceedings, May 2000, 90 (2), 450– 455. and Elizabeth Brainerd, “Importing equality? The impact of globalization on gender discrimination,” Industrial and Labor Relations Review, July 2004, 57 (4), 540–559. Blau, David M. and Wilbert van der Klaauw, “The Impact of Social and Economic Policy on the Family Structure Experiences of Children in the United States,” December 2007. Broda, Christian and David Weinstein, “Globalization and the Gains from Variety,” Quarterly Journal of Economics, May 2006, 121 (2), 541–85. Caliendo, Lorenzo and Fernando Parro, “Estimates of the Trade and Welfare E↵ects of NAFTA,” Review of Economic Studies, January 2015, 82 (1), 1–44. Cho, Lee-Jay, “Income and di↵erentials in current fertility,” Demography, 1968, 5 (1), 198– 211. Do, Quy-Toan and Andrei A. Levchenko, “Comparative Advantage, Demand for External Finance, and Financial Development,” Journal of Financial Economics, December 2007, 86 (3), 796–834. , , and Claudio Raddatz, “Comparative Advantage, International Trade, and Fertility,” February 2012. RSIE Discussion Paper 624. Doepke, Matthias, “Accounting for Fertility Decline During the Transition to Growth,” Jour- nal of Economic Growth, September 2004, 9 (3), 347–383. , Moshe Hazan, and Yishay Maoz, “The Baby Boom and World War II: A Macroeconomic Analysis,” December 2007. NBER Working Paper No. 13707. Edmonds, Eric and Nina Pavcnik, “International Trade and Child Labor: Cross Country Evidence,” Journal of International Economics, January 2006, 68 (1), 115–140. Evans, Carolyn, “The Economic Significance of National Border E↵ects,” American Economic Review, 2003, 93 (4), 1291–1312. 35 Feenstra, Robert C, “New Product Varieties and the Measurement of International Prices,” American Economic Review, March 1994, 84 (1), 157–77. Feenstra, Robert, James Markusen, and Andrew Rose, “Using the Gravity Equation to Di↵erentiate Among Alternative Theories of Trade,” Canadian Journal of Economics, 2001, 34 (2), 430–447. Feyrer, James, “Trade and Income – Exploiting Time Series in Geography,” Apr 2009. NBER Working Paper No. 14910. Fleisher, Belton M. and George F. Rhodes, “Fertility, Women’s Wage Rates, and Labor Supply,” American Economic Review, March 1979, 69 (1), 14–24. Frankel, Je↵rey A. and David Romer, “Does Trade Cause Growth?,” American Economic Review, June 1999, 89 (3), 379–399. Galor, Oded and Andrew Mountford, “Trading Population for Productivity: Theory and Evidence,” Review of Economic Studies, October 2009, 75 (4), 1143–1179. and David N. Weil, “The Gender Gap, Fertility, and Growth,” American Economic Review, June 1996, 86 (3), 374–387. and , “Population, Technology, and Growth: From Malthusian Stagnation to the De- mographic Transition and beyond,” American Economic Review, September 2000, 90 (4), 806–828. Greenwood, Jeremy and Ananth Seshadri, “The U.S. Demographic Transition,” American Economic Review, Papers and Proceedings, May 2002, 92 (2), 153–159. Guryan, Jonathan, Erik Hurst, and Melissa Kearney, “Parental Education and Parental Time with Children,” Journal of Economic Perspectives, Summer 2008, 22 (3), 23–46. Heckman, James J and James R Walker, “The Relationship between Wages and Income and the Timing and Spacing of Births: Evidence from Swedish Longitudinal Data,” Economet- rica, November 1990, 58 (6), 1411–41. Helpman, Elhanan, Marc Melitz, and Yona Rubinstein, “Estimating Trade Flows: Trading Partners and Trading Volumes,” Quarterly Journal of Economics, May 2008, 123 (2), 441–487. Hummels, David, “Towards a Geography of Trade Costs,” 2001. Mimeo, Purdue University. , “Transportation Costs and International Trade in the Second Era of Globalization,” Journal of Economic Perspectives, Summer 2007, 21 (3), 131–154. ejean, “Elasticity Optimism,” September 2013. Forthcoming, Imbs, Jean and Isabelle M´ American Economic Journal: Macroeconomics. 36 Jones, Larry E., Alice Schoonbroodt, and Michele Tertilt, “Fertility Theories: Can They Explain the Negative Fertility-Income Relationship?,” in “Demography and the Economy” NBER Chapters, National Bureau of Economic Research, Inc, October 2010, pp. 43–100. and Michele Tertilt, “An Economic History of Fertility in the U.S.: 1826-1960,” in Peter Rupert, ed., Frontiers of Family Economics, Vol. 1, Bingley, UK: Emerald Group Publish- ing, 2008, pp. 165–230. Jones, Ronald, “A Three-Factor Model in Theory, Trade, and History,” in Jagdish Bhagwati, ed., Trade, Balance of Payments, and Growth: Papers in International Economics in Honor of Charles P. Kindleberger, Amsterdam: North Holland, 1971. Juhn, Chinhui, Gergely Ujhelyi, and Carolina Villegas-Sanchez, “Men, women, and machines: How trade impacts gender inequality,” Journal of Development Economics, January 2014, 106, 179 – 193. Kremer, Michael, “Population Growth and Technological Change: One Million B.C. to 1990,” Quarterly Journal of Economics, August 1993, 108 (3), 681–716. Marchand, Beyza Ural, Ray Rees, and Raymond G. Riezman, “Globalization, Gender and Development: The E↵ect of Parental Labor Supply on Child Schooling,” Review of Eco- nomics of the Household, 2013, 11 (2), 151–173. Merrigan, Philip and Yvan St.-Pierre, “An econometric and neoclassical analysis of the timing and spacing of births in Canada from 1950 to 1990,” Journal of Population Economics, 1998, 11 (1), 29–51. Mookherjee, Dilip, Silvia Prina, and Debraj Ray, “A Theory of Occupational Choice with Endogenous Fertility,” American Economic Journal: Microeconomics, November 2012, 4 (4), 1–34. Morrow, Peter M., “Ricardian–Heckscher–Ohlin comparative advantage: Theory and evi- dence,” Journal of International Economics, November 2010, 82 (2), 137 – 151. Mussa, Michael, “Tari↵s and the Distribution of Income: The Importance of Factor Speci- ficity, Substitutability, an Intensity in the Short and Long Run,” Journal of Political Economy, 1974, 82, 1191–1203. Oostendorp, Remco, “Globalization and the Gender Wage Gap,” World Bank Economic Review, January 2009, 23 (1), 141–161. Ortega, Francesc and Giovanni Peri, “Openness and income: The roles of trade and migra- tion,” Journal of International Economics, March 2014, 92 (2), 231 – 251. Pitt, Mark M., Mark R. Rosenzweig, and Mohammad Nazmul Hassan, “Human Capital Investment and the Gender Division of Labor in a Brawn-Based Economy,” American Economic Review, December 2012, 102 (7), 3531–60. 37 Qian, Nancy, “Missing Women and the Price of Tea in China: The E↵ect of Sex-Specific Earnings on Sex Imbalance,” Quarterly Journal of Economics, July 2008, 123 (3), 1251– 1285. Ranjan, Priya and Jae Young Lee, “Contract Enforcement And International Trade,” Eco- nomics and Politics, July 2007, 19 (2), 191–218. Rauch, James E., “Networks Versus Markets in International Trade,” Journal of International Economics, June 1999, 48 (1), 7–35. and Vitor Trindade, “Ethnic Chinese Networks in International Trade,” Review of Eco- nomics and Statistics, 2002, 84 (1), 116–130. Rees, Ray and Ray Riezman, “Globalization, Gender, And Growth,” Review of Income and Wealth, March 2012, 58 (1), 107–117. Rendall, Michelle, “Brain Versus Brawn: The Realization of Women’s Comparative Advan- tage,” September 2010. mimeo, University of Zurich. Ross, Michael L., “Oil, Islam, and Women,” American Political Science Review, February 2008, 102 (1), 107–123. Santos Silva, J.M.C. and Silvana Tenreyro, “The Log of Gravity,” Review of Economics and Statistics, 2006, 88 (4), 641–658. e, Philip and Hosny Zoabi, “When Stolper-Samuelson Does Not Apply: International Saur´ Trade and Female Labor,” May 2011. mimeo, Swiss National Bank and Tel Aviv University. and , “International Trade, the Gender Wage Gap and Female Labor Force Participa- tion,” Journal of Development Economics, 2014, 111 (1), 17–33. Schultz, T. Paul, “Changing World Prices, Women’s Wages, and the Fertility Transition: Sweden, 1860-1910,” Journal of Political Economy, December 1985, 93 (6), 1126–54. , “The Value and Allocation of Time in High-Income Countries: Implications for Fertility,” Population and Development Review, 1986, 12, 87–108. Stock, James H. and Motohiro Yogo, “Testing for Weak Instruments in Linear IV Regres- sion,” in Donald W.K. Andrews and James H. Stock, eds., Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg, Cambridge University Press Cambridge, UK 2005, pp. 80–108. Willis, Robert J., “A New Approach to the Economic Theory of Fertility Behavior,” Journal of Political Economy, March-April 1973, 81 (2), S14–S64. World Bank, World Development Report 2012: Gender Equality and Development, World Bank, 2012. 38 Table 1. Share of Female Workers in Total Employment, Highest to Lowest ISIC Code Sector Name F Li 181 Wearing apparel, except fur apparel 0.71 173 Knitted and crocheted fabrics and articles 0.62 192 Footwear 0.49 172 Other textiles 0.47 321 Electronic valves and tubes and other electronic components 0.46 332 Optical instruments and photographic equipment 0.45 191 Leather and leather products 0.43 323 TV and radio receivers, sound or video apparatus 0.43 333 Watches and clocks 0.42 319 Other electrical equipment n.e.c. 0.42 182 Fur and articles of fur 0.41 154 Other food products 0.39 331 Medical appliances and instruments 0.38 369 Manufacturing n.e.c. 0.38 322 TV and radio transmitters; telephony and telegraphy apparatus 0.38 171 Spinning, weaving and finishing of textiles 0.37 242 Other chemical products 0.36 151 Meat, fish, fruit, vegetables, oils and fats 0.36 223 Reproduction of recorded media 0.35 315 Electric lamps and lighting equipment 0.34 300 O ce, accounting and computing machinery 0.34 160 Tobacco products 0.33 221 Publishing 0.33 311 Electric motors, generators and transformers 0.32 313 Insulated wire and cable 0.32 312 Electricity distribution and control apparatus 0.30 222 Printing and service activities related to printing 0.29 293 Domestic appliances n.e.c. 0.28 252 Plastics products 0.27 314 Accumulators, primary cells and primary batteries 0.26 152 Dairy products 0.25 372 Recycling of non-metal waste and scrap 0.25 155 Beverages 0.23 251 Rubber products 0.23 210 Paper and paper products 0.23 243 Man-made fibres 0.22 359 Transport equipment n.e.c. 0.21 39 Table 1 (cont’d). Share of Female Workers in Total Employment, Highest to Lowest ISIC Code Sector Name F Li 343 Parts and accessories for motor vehicles and their engines 0.21 153 Grain mill, starch products, and prepared animal feeds 0.20 361 Furniture 0.20 261 Glass and glass products 0.19 289 Other fabricated metal products 0.19 202 Products of wood, cork, straw and plaiting materials 0.18 371 Recycling of metal waste and scrap 0.17 201 Sawmilling and planing of wood 0.16 291 General purpose machinery 0.16 269 Non-metallic mineral products n.e.c. 0.16 241 Basic chemicals 0.15 353 Aircraft and spacecraft 0.15 292 Special purpose machinery 0.14 231 Coke oven products 0.14 232 Refined petroleum products 0.13 272 Basic precious and non-ferrous metals 0.13 273 Casting of metals 0.12 281 Structural metal products, tanks, reservoirs, steam generators 0.12 233 Nuclear fuel 0.11 271 Basic iron and steel 0.10 341 Motor vehicles 0.09 351 Building and repairing of ships and boats 0.09 352 Railway and tramway locomotives and rolling stock 0.08 342 Bodies for motor vehicles; trailers and semi-trailers 0.08 Mean 0.27 Min 0.08 Max 0.71 Notes: This table reports the share of female workers in total employment by sector, averaged across coun- tries. 40 Table 2. Summary Statistics for Female Labor Need of Exports and Fertility OECD NON-OECD Panel A: Female Labor Need of Exports Mean St. Dev Countries Mean St. Dev Countries 1960s 0.263 0.043 20 0.275 0.077 102 1970s 0.256 0.044 20 0.274 0.082 103 1980s 0.255 0.047 20 0.284 0.100 103 1990s 0.261 0.042 21 0.302 0.109 123 2000s 0.256 0.032 21 0.293 0.122 128 Panel B: Fertility Rates Mean St. Dev Countries Mean St. Dev Countries 1960s 2.80 0.460 20 6.15 1.367 102 1970s 2.13 0.457 20 5.75 1.593 103 1980s 1.74 0.261 20 5.13 1.758 103 1990s 1.63 0.248 21 3.99 1.847 123 2000s 1.64 0.254 21 3.38 1.704 128 Notes: This table reports the summary statistics for F N LX and fertility, by country group and decade. Table 3. F N LX : Top 10 and Bottom 10 Countries, 1980-2007 Highest F N LX Lowest F N LX Lesotho 0.650 Algeria 0.146 Haiti 0.572 Angola 0.144 Bangladesh 0.557 Kazakhstan 0.141 Mauritius 0.528 Venezuela, RB 0.140 Sri Lanka 0.525 Saudi Arabia 0.138 Honduras 0.486 Kuwait 0.138 Cambodia 0.485 Nigeria 0.137 El Salvador 0.471 Gabon 0.137 Nepal 0.465 Iraq 0.135 Dominican Republic 0.461 Libya 0.134 Notes: This table reports the 10 countries with the highest, and 10 countries with the lowest F N LX . 41 Table 4. F N LX : Top 10 and Bottom 10 Changers since 1960s Largest Increase in F N LX Largest Decrease in F N LX Cambodia 0.410 Mozambique -0.097 Honduras 0.311 Rwanda -0.112 Haiti 0.269 Sudan -0.112 Sri Lanka 0.225 Ecuador -0.129 Tunisia 0.211 Congo, Rep. -0.132 Albania 0.210 Chad -0.147 Morocco 0.196 Angola -0.159 El Salvador 0.186 Yemen, Rep. -0.160 Madagascar 0.182 Niger -0.170 Nicaragua 0.169 Timor-Leste -0.281 Notes: This table reports the 10 countries with the largest increases and the largest decreases in F N LX . Change is calculated as the di↵erence between the F N LX in the 2000s and that in the 1960s. 42 Table 5. Cross-Sectional Results, 1980-2007 (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS Dependent Variable: (Log) Fertility Rate (Log) Female Labor -0.29*** -0.20*** -0.37*** -0.47*** -0.57*** -0.56*** -0.28*** -0.38*** Need of Exports (0.080) (0.057) (0.128) (0.085) (0.131) (0.137) (0.095) (0.115) (Log) Openness -0.00 0.01 -0.01 0.01 0.01 0.01 0.01 -0.01 (0.037) (0.032) (0.037) (0.032) (0.034) (0.034) (0.030) (0.037) (Log) GDP per capita -0.39*** -0.26*** -0.40*** -0.27*** -0.28*** -0.28*** -0.26*** -0.27*** (0.020) (0.023) (0.020) (0.023) (0.024) (0.025) (0.022) (0.022) Log (Area) 0.02 (0.016) Log (Population) -0.04*** (0.017) Constant 5.48*** 4.17*** 5.81*** 5.23*** 5.61*** 5.57*** 4.47*** 5.18*** (0.296) (0.314) (0.480) (0.362) (0.514) (0.540) (0.436) (0.766) 43 R2 0.630 0.859 First Stage Dependent Var. (Log) FLNX (Log) Predicted FLNX 3.23*** 3.04*** 3.32*** (0.342) (0.373) (0.548) (Log) Predicted FLNX 2.43*** (out of sample) (0.469) (Log) Predicted FLNX 1.00*** (Poisson) (0.201) (Log) Predicted FLNX 3.03*** (No Population) (0.547) F-test 43.02 34.69 32.21 27.24 24.77 29.78 First Stage R2 0.400 0.534 0.402 0.392 0.461 0.548 Region Dummies no yes no yes yes yes yes yes Observations 145 145 145 145 145 145 145 145 Notes: Robust standard errors in parentheses; * significant at 10%; ** significant at 5%; *** significant at 1%. All variables are averages over the period 1980-2007 and in natural logs. Variable definitions and sources are described in detail in the text. Table 6. Panel Results, 1962-2007 (1) (2) (3) (4) (5) (6) (7) (8) Five-Year Averages Ten-Year Averages Dependent Variable: (Log) Fertility Rate (Log) Female Labor -0.37*** -0.34*** -0.22*** -0.22*** -0.38*** -0.36*** -0.24*** -0.23*** Need of Exports (0.067) (0.077) (0.058) (0.061) (0.069) (0.093) (0.069) (0.072) (Log) Openness -0.02 -0.18*** -0.02 -0.00 -0.02 -0.18*** -0.02 -0.00 (0.028) (0.041) (0.031) (0.034) (0.028) (0.049) (0.036) (0.039) 44 (Log) GDP per capita -0.38*** -0.35*** -0.18*** -0.18*** -0.38*** -0.37*** -0.20*** -0.19*** (0.019) (0.051) (0.043) (0.047) (0.019) (0.059) (0.048) (0.051) (Log) Female -0.00 -0.01 Educational Attainment (0.038) (0.041) Country FE no yes yes yes no yes yes yes Year FE no no yes yes no no yes yes R2 0.576 0.885 0.937 0.936 0.584 0.895 0.943 0.942 Observations 1,247 1,247 1,247 1,102 627 627 627 554 Notes: Standard errors clustered at the country level in parentheses; * significant at 10%; ** significant at 5%; *** significant at 1%. All of the variables are 5-year averages (left panel) or 10-year averages (right panel) over the time periods spanning 1962-2007, and in natural logs. Variable definitions and sources are described in detail in the text. Table 7. Alternative Specifications and Controls: Cross-Sectional 2SLS Results, 1980-2007 (1) (2) (3) (4) (5) (6) (7) Dependent Variable: (Log) Fertility Rate (Log) F N LX 1.69** -0.41*** -0.40*** -0.30*** -0.34*** -0.42*** (0.820) (0.092) (0.096) (0.089) (0.089) (0.093) (Log) F N LX ⇥(Log) -0.49** Openness (0.192) F N LX F N LI -0.02*** (0.004) (Log) Openness 1.66** 0.01 0.03 0.07 -0.00 -0.03 0.01 (0.651) (0.034) (0.041) (0.044) (0.028) (0.042) (0.034) (Log) GDP per capita -0.26*** -0.31*** -0.25*** -0.27*** -0.13*** -0.29*** -0.26*** (0.023) (0.027) (0.032) (0.033) (0.036) (0.031) (0.030) (Log) Female -0.11** Educational Attainment (0.046) Child Labor Indicator 0.01*** (0.002) 45 (log) Infant Mortality 0.20*** (0.047) Gini Coe↵ 0.78*** (0.302) Polity 2 Indicator 0.00 (0.005) Constant -2.27 4.23*** 4.88*** 4.55*** 2.77*** 4.72*** 4.97*** (2.883) (0.295) (0.438) (0.449) (0.702) (0.372) (0.439) First Stage (Log) Predicted F LN X 2.97*** 2.99*** 3.07*** 3.12*** 3.04*** (0.362) (0.457) (0.449) (0.507) (0.427) Predicted F LN X 2.77*** (0.493) F-test 22.52 31.74 29.45 35.39 20.98 35.05 First Stage R2 0.531 0.558 0.513 0.538 0.527 0.548 Region Dummies yes yes yes yes yes yes yes Observations 145 145 125 103 144 102 144 Notes: Robust standard errors in parentheses; * significant at 10%; ** significant at 5%; *** significant at 1%. All variables are averages over the period 1980-2007. Variable definitions and sources are described in detail in the text. Table 8. Subsamples: Cross-Sectional 2SLS Results, 1980-2007 (1) (2) (3) (4) (5) Sample: no no no Sub- no Middle East No mining outliers OECD Saharan Africa & North Africa exporters Dependent Variable: (Log) Fertility Rate (Log) Female Labor -0.48*** -0.47*** -0.59*** -0.42*** -0.47*** Need of Exports (0.121) (0.082) (0.161) (0.087) (0.102) (Log) Openness 0.02 0.04 0.01 0.01 0.01 (0.034) (0.037) (0.053) (0.031) (0.033) (Log) GDP per capita -0.26*** -0.32*** -0.29*** -0.29*** -0.28*** (0.025) (0.026) (0.030) (0.024) (0.024) 46 Constant 5.17*** 5.44*** 5.85*** 5.27*** 5.35*** (0.499) (0.348) (0.713) (0.365) (0.433) First Stage (Log) Predicted FLNX 2.69*** 3.14*** 2.55*** 2.94*** 2.85*** (0.400) (0.407) (0.398) (0.400) (0.406) F-test 32.81 30.62 32.84 35.59 34.24 First Stage R2 0.439 0.547 0.542 0.497 0.474 Region Dummies yes yes yes yes yes Observations 135 125 104 129 135 Notes: Robust standard errors in parentheses; * significant at 10%; ** significant at 5%; *** significant at 1%. All variables are averages over the period 1980-2007. Variable definitions and sources are described in detail in the text. Figure 1. Partial Correlation Between Fertility and F N LX TKM .5 TJK OMN SAU GTM MYS IRL AGO LAO TUR NOR PNG AFG PAK PHL GIN LBY ARE KHM FRA GAB SWE GBR HND KAZ BOL PRY KGZ DNK FIN NER IRQ TCD CIV VEN NLD SEN USA ALB COG CMR CHE HTI NGA ZMB KWT Ln(Fertility) BEN MEX KEN UGA ZWE RWA MLI MWI BFA AZE ECU PER UZB AUT MRT SLE JOR MOZ YEMAUS NZL MNGNIC DOMSLV ARG NPL ETH GMB GRC 0 IDN SWZ PRT POL ESP DEU MKD SDN PAN CRI LBR TGO COL BRA VNM BDI JAM MDG HUN ZAF ARMITA SOM NAM RUS IRN IND GHA CAN BGD CAF DZA CHL SYR TZA SGP ROM ISR TTO SVKCZE GNB SVN EST BLR URY BWA GEO JPN THA EGY HRV LTU BGRERI KOR MAR LVA MDA LKA HKG LBN TUN LSO UKR CHN BIH −.5 MUS CUB −1 −.5 0 .5 1 Ln(FNLXc) Notes: This figure displays the partial correlation between F N LX and fertility, in logs, after controlling for openness, per capita income, and regional dummies (see Column 2 of Table 5). 47 Figure 2. First Stage: Partial Correlation between F N LXc and F\ LN X c 1 LSO MUS TUN MAR HTI ISR .5 LBN HND MDG BWA SWZ DOM SLV CRI TUR GTM GRC PRT JAM Actual Ln(FNLXc) KHM ALB HKG IRLNIC SLE CHE MNG CAF MRT JORBGD MKD LKA MWI GMB URY MLIGINBFA HRV DNK ETH CUBNAM ITA USA BDI SOM ROM PHLUGA HUNARM ERI CHN BIH BLR NLD SVN MDA EST THA FRA TZA KGZ LTU GBR KOR AUT BGR FIN PAK CIV LAO KEN ZAF ZWESEN POL EGY NPL RWA GNB 0 MYS PRYDEU VNM ESP SWE SDN PAN NZL SVK SYR CZE SGP BEN TCDUZBARE COL ARG AUS OMN MOZ LVA GHA CAN BRA MEX IND JPN AFG PNG PER NOR IRN YEM KWT DZA SAU LBY TGO IDN TJK CHL UKR ECUGEO RUS TKM BOL IRQ AZE −.5 COG CMRLBR ZMB GAB KAZ AGO NER TTO NGA VEN −1 −.2 −.1 0 .1 .2 Instrument (Predicted Ln(FNLXc)) Notes: This figure presents the partial correlation plot from the first stage regression between the actual value of F N LXc and the instrument. 48 Table A1. An Illustration of the Instrumentation Strategy Sector Exporter Destination Distance Exports F Li Apparel Canada EU 1000 2500 0.71 Apparel Canada US 1000 4500 0.71 Apparel Australia EU 10000 850 0.71 Apparel Australia US 10000 415 0.71 Motor Vehicles Canada EU 1000 25000 0.09 Motor Vehicles Canada US 1000 15000 0.09 Motor Vehicles Australia EU 10000 1000 0.09 Motor Vehicles Australia US 10000 1150 0.09 49 Table A2. Variation in Gravity Coe cients Across Sectors Coe cient Mean Std. Dev. Min Max Ln(Distancecd ) -1.115 0.238 -1.651 -0.532 Ln(P opc ) -0.083 0.359 -0.986 0.367 Ln(Areac ) -0.138 0.226 -0.507 0.393 Ln(P opd ) 0.723 0.227 0.404 1.424 Ln(Aread ) -0.144 0.120 -0.568 0.050 Landlockedcd -0.538 0.439 -2.590 0.644 Bordercd 1.398 2.520 -6.814 5.957 Bordercd ⇥ Ln(Distancecd ) 0.200 0.236 -0.462 0.674 Bordercd ⇥ Ln(P opc ) 0.239 0.178 -0.236 0.665 Bordercd ⇥ Ln(Areac ) -0.194 0.150 -0.542 0.158 Bordercd ⇥ Ln(P opd ) -0.214 0.193 -0.596 0.364 Bordercd ⇥ Ln(Aread ) 0.019 0.119 -0.360 0.283 Bordercd ⇥ Landlockedcd 0.398 0.281 -0.290 1.180 50 Table A3. FLFP: Cross-Sectional Results, 1980-2007 (1) (2) (3) (4) (5) (6) OLS 2SLS OLS 2SLS 2SLS 2SLS Dependent Variable: (Log) FLFP (Log) F N LX 0.07 0.20 1.63*** 2.53*** -0.94*** 1.34*** (0.078) (0.126) (0.580) (0.913) (0.346) (0.489) (ln) F LN X *(ln) GDP per capita -0.18*** -0.27*** (0.070) (0.103) (ln) F LN X * (ln) Fertility 0.88*** (0.248) (ln) Fertility -2.95*** (0.869) (ln) F LN X * (ln) Fem. Educ. Attainment -0.67** 51 (0.269) (ln) Fem. Educ. Attainment 2.34** (0.927) (Log) Openness 0.03 0.04 0.63*** 0.92*** 0.04 -0.01 (0.029) (0.031) (0.227) (0.342) (0.043) (0.038) (Log) GDP per capita -0.02 -0.02 -0.01 -0.01 0.00 -0.08 (0.053) (0.054) (0.056) (0.060) (0.060) (0.048) Constant -0.80* -2.00*** -5.98*** -9.83*** 1.68 -4.69*** (0.465) (0.671) (1.929) (3.149) (1.292) (1.738) R2 0.577 0.599 Region Dummies yes yes yes yes yes yes Observations 145 145 145 145 145 125 Notes: Robust standard errors in parentheses; * significant at 10%; ** significant at 5%; *** significant at 1%. All variables are averages over the period 1980-2007, except FLFP, which is averaged over 1990-2007. Variable definitions and sources are described in detail in the text. Figure A1. Female Formal Labor Market Equilibrium wF wF Labor& Labor& supply&& supply&& Labor& Labor& demand& demand& 1 - λN 1 - λN Interior&solu3on& Corner&solu3on& 52 Figure A2. Equilibrium Female Labor Force Participation Nc 1 λ N −c (N c ) N c (N −c ) 1 Nc λ 53 Figure A3. Impact of F N LX on FLFP by Quartile 1 Coefficient on FNLX 0 −.5 .5 First Second Third Fourth Income Quartile (a) By Income 1 .5 Coefficient on FNLX 0 −.5 −1 First Second Third Fourth Fertility Quartile (b) By Fertility 2 1.5 Coefficient on FNLX .5 0 −.5 1 First Second Third Fourth Female Educational Attainment Quartile (c) By Educational Attainment Notes: This figure displays the quartile-specific coe cients on F N LX in the 2SLS regressions with log FLFP as the dependent variable, and the controls/regional dummies as in Table A3. Panel (a) displays the coe cients by income quartile, panel (b) by fertility quartile, and panel (c) by female educational attainment quartile. 54