BS Thesis in Economics. Trade Openness and Inequality

BS Thesis in Economics Trade Openness and Inequality An Empirical Analysis Ólafur Kjaran Árnason Advisor: Professor Gylfi Zoëga Faculty of Economics October 2017

Trade Openness and Inequality An Empirical Analysis Ólafur Kjaran Árnason Thesis towards a BS degree in Economics Advisor: Professor Gylfi Zoëga Faculty of Economics School of Social Sciences, University of Iceland October 2017

Trade Openness and Inequality: an Empirical Analysis. This thesis is equivalent to 12 ECTS credits towards a BS degree from the Faculty of Economics at the University of Iceland. 2017 Ólafur Kjaran Árnason This thesis cannot be reproduced without the author s consent. Printing: Samskipti Reykjavík, 2017

Foreword This thesis is equivalent to 12 ECTS credits towards a BS degree from the Faculty of Economics at the University of Iceland. I would like to thank my advisor, Professor Gylfi Zoëga, for his patience and helpful suggestions, and my ever supportive parents, Árni Sigurjónsson and Ásta Bjarnadóttir. They are the finest people. My fellow economist, Maja Jóhannsdóttir my love, jeg glæder mig til at spise dansk frokost sammen i næste uge. Finally, I want to acknowledge the blessing of having had such encouraging conversations with Árni Þór Lárusson, Gunnar Jörgen Viggósson, and my grandmother, Gerður G. Óskarsdóttir, at various stages on the long path of thorns towards writing this thesis. 4

Abstract This paper investigates the empirical relationship between openness to trade and withincountry income inequality. The estimation is carried out with group and time fixed effects, using an unbalanced panel of 112 countries over the period of 1988-2008. A key result is that the impact of increased openness on inequality is negatively related to education levels. For most countries the estimated effect is positive but in countries where the share of population with secondary education is high the effect is reversed and increased openness is expected to lower inequality. Thus, more widespread education may provide protection against increased inequality due to globalization, in direct conflict with the Stolper-Samuelson theorem. This supports the view that rising withincountry inequality in many rich countries in recent decades is driven by other forces than globalization. The results are highly significant, robust to various changes in model specification, and not sensitive to the omission of subsets of the sample. In general, the estimated impact of openness is quite small in terms of changes in the Gini coefficient, but in some cases substantial when examined specifically for individual quintiles of the income distribution. 5

Contents Foreword... 4 Abstract... 5 Contents... 6 Figures... 7 Tables... 7 1 Introduction... 8 2 Literature review... 14 3 Data and methodology... 20 3.1 Dependent variable... 21 3.2 Independent variables... 22 3.3 Empirical strategy... 24 4 Results... 26 4.1 Regression analysis... 26 4.2 Robustness... 28 5 Discussion... 32 References... 38 Appendix 1... 42 Appendix 2... 43 6

Figures Figure 1. World trade, 1870 to 2010.... 10 Figure 2. Global inequality, 1988 to 2008.... 11 Figure 3. Trade and inequality in several countries, 1988 to 2008.... 13 Figure 4. The estimated effect of increased openness conditional on education.... 34 Tables Table 1. Global inequality, 1988 to 2008.... 11 Table 2. Comparison of cross-country studies of openness and income inequality.... 19 Table 3. Number of observations by year and region.... 21 Table 4. Fixed effects regression results: Gini.... 27 Table 5. Fixed effects regression results: Quintile shares.... 28 Table 6. Robustness tests.... 31 Table 7. Interpretation: Gini.... 33 Table 8. Interpretation: Income of bottom quintile.... 34 Table A1. Summary statistics....42 Table A2. Number of observations by country and region....43 7

1 Introduction Most economists agree that increased international trade is an effective means of raising the living standards of people across the globe. 1 The fundamental insight of David Ricardo s trade theory, which he presented 200 years ago, is that because of comparative advantage this should in fact apply equally to countries rich and poor, irrespective of factor endowments (Ricardo, 1817/2001). Many analysts have addressed the issue empirically and there is a vast literature focused on identifying the link between openness and economic growth. Although the empirical evidence has generally been supportive of a positive impact of openness on growth (Hallaert, 2006), there remain significant methodological obstacles, and many economists are skeptical of the results (Rodriguez & Rodrik, 2000). However, while most economists believe in the potential benefits of trade, they also tend to agree that there are winners and losers from trade. According to the Stolper- Samuelson theorem, which follows directly from the Heckscher-Ohlin model of trade, the winners will be those who command factors in which their country is abundant relative to its trading partners and the losers will be those who command factors in which their country is relatively scarce (Stolper & Samuelson, 1941). 2 The implications for income inequality will therefore depend on relative factor endowments of countries and the distribution of factors within countries. A simple version of the model assumes two factors of production, low-skilled and high-skilled labor, where poor countries are relatively abundant in low-skilled labor and rich countries in high-skilled labor, and thus increased trade is expected to lower inequality in poor countries while raising inequality in rich countries. The relationship between openness and inequality has not been studied as extensively as the effects of trade on growth and so far the empirical evidence is inconclusive. The cross-country literature on openness to trade and inequality is reviewed in Chapter 2. This paper will leave aside the question of growth and focus on distributional consequences. Regardless of the relation between globalization and growth, its 1 Openness (to trade), globalization, and (international) trade will be used interchangably in the paper. 2 Factors of production, or factors for short, are any inputs used in the production of goods and services (e.g. low-skilled and high-skilled labor, capital, natural resources, etc.). 8

connection with inequality is important enough on its own. If, for example, it is true that globalization can potentially increase the well-being of everyone and it is also true that globalization increases inequality in a systematic way, then it is absolutely necessary to understand how its fruits can be better shared in order to maintain public support for increased globalization. This perspective motivated the International Monetary Fund (IMF), the World Bank (WB), and the World Trade Organization (WTO), all ardent supporters of globalization, to issue a joint report last April on ways to share the gains from trade more widely, stating that trade is leaving too many individuals and communities behind, notably also in advanced economies (IMF, WB, & WTO, 2017, p. 4). The period between 1988 and 2008 was an era of unprecedented globalization and is in many respects ideal for studying the relationship between openness and inequality. Never in history has globalization accelerated so rapidly and so universally, which is why it has been referred to as the era of hyperglobalization (Subramanian & Kessler, 2013). Figure 1 shows this remarkable rise of world trade as a share of world GDP. It is interesting to note that not until the 1970s did globalization reach the same heights as immediately after World War I but thereafter the ratio of trade to world GDP remains fairly stable until shortly before 1990. The late 1980s was a time when many developing countries were integrating into the world economy and the fall of Communist regimes in Eastern Europe and the Soviet Union was about to transform the global political landscape. Together these developments led to hundreds of millions of people entering world markets who previously had been living in autarkic or semi-autarkic countries. The year of 2008 is an appropriate endpoint for the analysis, as the global financial crisis put a halt to the hyperglobalization of the two preceding decades. Apart from the period being very suitable for studying globalization, comparable highquality data on income distribution become much more readily available around 1990, which allows us to study the development of inequality with greater accuracy than before that time (Milanovic, 2016a). For data on inequality, this paper relies on the database constructed by Lakner and Milanovic (2016), which is described in Chapter 3. But what happened to income inequality during the era of hyperglobalization? It is well documented that for the last couple of decades income inequality has been on the rise in many Western countries (see for example OECD, 2015). That is not the whole 9

Figure 1. World trade, 1870 to 2010. Note: World trade as a share of world GDP. The shaded part shows the period that has been referred to as the era of hyperglobalization. Source: Klasing and Milionis (2014) for 1870-1949 and the Penn World Tables (Version 8.1) by Feenstra, Inklaar and Timmer (2015) for 1950-2010. story, however. Economists studying global inequality, measured as if the world was one country, have consistently found that it has been fairly stable in recent decades and that at some point between the late 1980s and the early 2000s global inequality in fact started decreasing (Bourguignon, 2015; Lakner and Milanovic, 2016). 3 As a result, the era of hyperglobalization coincides with the first period since the dawn of the Industrial Revolution that global inequality is on the decline (Bourguignon and Morrison, 2002; Milanovic, 2016a). The concept of global inequality can be thought of as including two components, inequality between countries and inequality within countries, where in today s world the between-component accounts for a larger share of global inequality than the within- component. 4 3 Bourguignon (2015) uses household survey data normalized to GDP per capita, because of the noted discrepancy between household survey means and GDP per capita based on national accounts, and he finds that global inequality started decreasing in the 1990s. Lakner and Milanovic (2016) make no such correction and find that global inequality started to decline in the 2000s. 4 The Theil coefficient is an alternative to the Gini coefficient which has the advantage of being explicitly decomposable into between- and within-components. However, the Gini is preferred in this study due to its widespread use. 10

Table 1. Global inequality, 1988 to 2008. 1988 1993 1998 2003 2008 Global inequality 72.0 71.9 71.8 71.9 69.4 Between inequality (unweighted) 46.3 56.1 55.1 58.2 55.5 Between inequality (population-weighted) 65.9 64.9 64.8 64.5 61.3 Within inequality (unweighted) 35.1 39.6 39.5 39.0 37.6 Within inequality (population-weighted) 32.8 36.6 37.7 38.4 39.2 OECD countries 31.3 34.2 35.0 35.7 36.2 Non-OECD countries 33.2 37.2 38.3 39.0 40.0 Number of countries 72 112 118 130 117 Share of world population 81.0 92.0 91.6 93.2 89.9 Note: Inequality of net personal (PPP-adjusted) incomes in terms of the Gini coefficient. Global inequality consists of inequality between and within countries. The values for between and within inequality are averages across countries. Source: Author s calculations based on inequality data from Lakner and Milanovic (2016) and population data from the World Bank (2017). Figure 2. Global inequality, 1988 to 2008. Note: Inequality of net personal (PPP-adjusted) incomes expressed in terms of the Gini coefficient. Between and within inequality refer to population-weighted averages across countries. Source: Author s calculations based on data from Lakner and Milanovic (2016). 11

While inequality within countries has been rising in both rich and poor countries, and in fact even more so in poor countries, inequality between countries has been on the decline, with the net effect being a slight decrease in global inequality. 5 Figure 2 shows the development of global inequality during the period of 1988 to 2008 and the corresponding numbers are provided in Table 1. The fall in global inequality has certainly accelerated since 2008 due to the much higher economic growth in emerging economies relative to advanced economies. Figure 3 shows the trends in trade and inequality for several countries over the same period. Considering these general trends in trade and inequality, one could hypothesize that the increase in within-country inequality and the decrease in between-country inequality are both connected to the rise in globalization. It is possible, for example, that globalization allows poor countries to catch up with rich countries at a faster rate, thereby lowering between-country inequality, and at the same time it benefits rich people disproportionately at the country level, thereby raising within-country inequality across the board. Obviously, this hypothesis is difficult to evaluate, but motivated by the trends discussed above, this paper sets out to study to what extent the rise in globalization on the one hand and within-country income inequality on the other hand are related. The rest of the paper is organized as follows. Chapter 2 provides a review of previous cross-country studies on openness to trade and inequality and explains how this paper contributes to the literature. Chapter 3 describes the data and methodology used in the empirical analysis and the results are presented in Chapter 4 along with a set of robustness tests. Chapter 5 discusses the findings their implications. 5 It should be noted, however, that this small change in the Gini coefficient of global inequality between 1988 and 2008 conceals what Branko Milanovic describes as the greatest reshuffle of individual incomes since the Industrial Revolution (Milanovic, 2016b). This reshuffle of incomes is perhaps best demonstrated by the so-called elephant curve (Milanovic, 2016a; Milanovic, 2016b). 12

Figure 3. Trade and inequality in several countries, 1988 to 2008. Note: The path of trade and inequality across five benchmark years (1988, 1993, 1998, 2003, and 2008) in a few large advanced economies (Australia, Japan, Germany, South Korea, and the USA) and a few large emerging economies (Brazil, China, India, Indonesia, and South Africa). Data are missing for Australia in 2008 and South Africa in 1988. Source: Trade data from the World Bank (2017) and inequality data from Lakner and Milanovic (2016). 13

2 Literature review The empirical literature on trade openness and within-country income inequality consists of two different approaches. Cross-country studies make use of cross-country variation in openness and inequality, often over a period of time, while case-specific studies usually involve a detailed analysis of trade liberalization episodes in individual countries over a shorter time period. Each approach has its strengths and weaknesses. In their survey of the literature on trade liberalization and income inequality in developing countries, Goldberg and Pavcnik (2007) abstain from relying on crosscountry regressions, primarily because of data constraints, stating that [i]nconsistencies in the measurement of inequality across countries, changes in the household survey response rates over time as incomes rise, and frequent changes in the design of household surveys within the same country make inference based on cross-country evidence ( ) potentially less reliable compared to inference that relies on within-country evidence over shorter periods of time (p. 40). However, if imperfect data and comparability are the main shortcomings of the crosscountry approach, the most obvious drawback of case-specific studies is that they are, indeed, case-specific. As a result, the findings of individual case-specific studies on openness and inequality cannot be easily generalized, and if the case in question is in fact a very special or unique one, even an outlier case, then the study s findings may simply be misleading. Although taking a number of case-specific studies together provides a better chance of reaching some valid, general conclusions, it is still a problematic approach. Nothing ensures that it will be possible to identify which case-specific factors actually explain different results across different case-specific studies. It may also be that most case-specific studies are conducted for countries with particular characteristics, causing a bias in any general conclusions drawn from a survey of such studies. 6 To some extent, the weaknesses discussed above reflect a straightforward trade-off between applicability and accuracy, with cross-country studies gaining greater 6 For example, case-specific studies on developing countries in the 1990s were essentially limited to Latin America, which is indeed reflected in the survey by Goldberg and Pavcnik (2007). What if Latin America is for some reason a special case? Similarily, most case-specific studies on developed countries focus on the USA, which could just as well be a special case among developed countries. If so, then the validity of general conclusions drawn from a survey of such studies does not improve, even as their number increases. 14

applicability by giving up some accuracy, and vice versa for case-specific studies. Neither approach is impeccable but both are important. This paper follows the crosscountry approach to studying the relationship between openness to trade and income inequality, while also taking seriously its aforementioned limitations, which will be addressed further below. First, however, a brief review of the cross-country literature is in order. 7 An early empirical study focused on the question of openness and inequality was conducted by Bourguignon and Morrisson (1990), who utilized a cross-section of 36 countries in 1970. The study reported a negative effect of openness on inequality but was severely constrained by data availability. A few papers on the topic emerged in the late 1990s, hailing from the vibrant debate on the relationship between openness and economic growth which reached its peak at the time. Among those were Edwards (1997); Savvides (1998); Higgins and Williamson (1999); and Spilimbergo, Londono, and Szekely (1999), all of which made use of a new World Bank database on inequality, constructed by Deininger and Squire (1996). 8 More recent studies that have focused exclusively on the effects of openness on income inequality are Reuveny and Li (2003); Milanovic (2005); Gourdon, Maystre, and de Melo (2008); Dreher and Gaston (2008); Bergh and Nilsson (2010); and Jaumotte, Lall, and Papageorgiou (2013). These studies have benefited greatly from better data on inequality as well as other variables. Better data not only improve the quality of crosscountry studies by means of lower measurement errors and greater comparability; increased data coverage also allows for the use of more sophisticated estimation techniques, yielding more reliable estimates. However, the empirical evidence from cross-country studies on openness and inequality is still far from conclusive. Table 2 provides a comparison of the methods and results of a few important papers belonging to this literature. The results of other papers mentioned above but not presented in the table are summarized very briefly in a footnote. 9 Although the results seem to comprise almost the whole range of possible 7 For case-specific studies, see for example Goldberg and Pavcnik (2007) for developing countries, and Autor, Dorn, and Hanson (2013) for the US. 8 Additionally, a few widely cited papers around 2000 addressed the issue as a side note: Li, Squire, and Zou (1998); Barro (2000); Ravallion (2001); Dollar and Kraay (2002); and Lundberg and Squire (2003). 9 The results of other papers mentioned here are as follows: Edwards (1997), Li et al. (1998), and Dollar and Kraay (2002) find an insignificant effect of openness on inequality; Savvides (1998), Barro (2000), and Ravallion (2001) find a positive effect in poor countries; Lundberg and Squire (2003) find a small positive 15

results increasing or decreasing inequality, conforming to or contradicting the Stolper- Samuelson theorem, mostly affecting rich countries or mostly affecting poor countries, and so on there are in fact signs of some consistency. First, most studies find evidence of a positive impact of openness on inequality; out of the fifteen studies discussed in Table 2 and footnote 9, only two report a negative relationship. Second, by and large there is limited empirical support for the Stolper-Samuelson theorem. 10 This matter is not addressed specifically in all of the studies, but it is noticable that only Gourdon et al. (2008) report estimates that are generally in line with the predictions of the Stolper- Samuelson theorem. 11 The absence of empirical evidence in favor of the Stolper-Samuelson theorem, and in fact the Heckscher-Ohlin model of trade more generally, has long been a puzzle. 12 Most problematic is the prediction that increased openness lowers inequality in poor countries while, according to Goldberg and Pavcnik (2007), there is overwhelming evidence (p. 39) supporting the opposite. As a response, some theoretical papers have proposed variations of the Heckscher-Ohlin model that seek to reconcile the theory with the empirical evidence. For example, Wood (1994) provides a Heckscher-Ohlin model with three factors (unskilled, medium-skilled, and high-skilled labor) in which the effect of increased openness on inequality is ambiguous in countries relatively abundant in medium-skilled labor, such as Latin American countries in the 1980s and 1990s. Therefore, he argues, trade liberalization episodes in these countries did not have the Stolper-Samuelson effects expected for countries abundant in unskilled labor. Davis (1996) comes to a similar conclusion assuming two factors and three types of goods that differ in their capital intensiveness. The estimated positive effect of openness on inequality has often been relatively small. This gave rise to a consensus in the late 1990s among labor and trade economists that increased inequality had more to do with factors such as skill-biased technological effect; and Reuveny and Li (2003) find a negative effect of trade openness but a positive effect of financial openness, similar to Jaumotte et al. (2013). 10 According to Goldberg and Pavcnik (2007), there has been no [empirical] support for the predictions of the model, at least not in its strict version (p. 58). 11 The results of Spilimbergo et al. (1999) are in line with Stolper-Samuelson for skilled-labor but not for capital. Dreher and Gaston (2008), and Bergh and Nilsson (2010) find a positive effect of openness on inequality in general, but especially for rich countries, which conforms to Stolper-Samuelson, but they do not find any evidence of a negative impact in poor countries, which should also follow from the theorem. 12 Feenstra (2004) provides a good discussion of the empirical problems of the Heckscher-Ohlin model and the Stolper-Samuelson theorem. Adrian Wood has long argued that these difficulties are in fact much exaggerated, see for example Wood (1994), Wood (1995), and Wood (2017). 16

change and less with globalization (Goldberg, 2015). 13 However, the subject has received renewed interest in recent years as several studies have reported a larger impact of globalization on inequality, 14 leading IMF et al. (2017) to emphasize trade-related adjustment costs that can bring a human and economic downside that is frequently concentrated, sometimes harsh, and has too often become prolonged (p. 4). This paper contributes to the existing literature by paying special attention to the concerns of Goldberg and Pavcnik (2007) highlighted above. First, and most importantly, this is done by using data of higher quality than most previous studies, prioritizing data comparability. Although this emphasis limits the feasible time-frame of the analysis, and thus reduces the number of observations available, the shorter time-frame itself may also increase the accuracy of the results by reducing measurement inconsistencies within countries over time. Moreover, the analysis includes the entire era of hyperglobalization as it has been defined and also covers most of the world s population during that period (see Chapter 3). The inequality database used in this paper, which is described in Chapter 3, relies completely on data taken directly from household surveys. Unfortunately, other studies have usually had to rely on data of lower quality. In most cases, previous studies have used the Deininger and Squire database (Deininger and Squire, 1996), in which observations are much less comparable across countries and (more importantly for fixed effects estimation) over time within countries. 15 Alternatively, some studies have relied on databases that include datapoints not taken from actual household surveys. For example, the SWIID database (Solt, 2009) is based on a missing data algorithm and the EHII database (Galbraith & Kum, 2005) estimates total income inequality based on the systematic relationship between the Deininger and Squire database and data on industrial pay-inequality. It should be noted that the studies by Milanovic (2005) and Jaumotte et 13 Borjas, Freeman, and Katz (1997), studying the effects of trade and immigration on wage inequality, come to a common conclusion in this respect: [A]cceleration of skill-biased technological change, a slowdown in the growth of the relative supply of college graduates, and institutional changes in the labor market are probably more important than immigration and trade in explaining the widening of the U.S. wage structure since the late 1970s (p. 62-63). A recent review of the literature on the relationship between globalization and wage inequality by Helpman (2016) concludes that trade played an appreciable role in increasing wage inequality, but that its cumulative effect has been modest (p. 1). 14 Autor et al. (2013) has perhaps received most widespread attention. The paper reports a sizable impact of rising Chinese import competition on cumulative earnings, employment, and benefits payments across local US labor markets over the period of 1992 to 2007. 15 As noted by the authors of the database themselves, variation in the definition of the variables used to measure inequality gross income or net income, income or expenditure, data per capita or data per household can seriously affect the magnitude of the indicators of inequality and undermine the international and intertemporal comparability of the data (Deininger and Squire, 1996, p. 566). 17

al. (2013) are exceptions, but the former is limited to a much lower number of observations than the present analysis and the latter includes data on much fewer countries. Second, by using fixed effects estimation, persistent data inconsistencies across countries are absorbed by the country fixed effects, and thus do not affect the estimates. This comes at the cost of leaving significantly less variation to be explained in the estimation, since fixed effects estimation only makes use of variation within countries while ignoring variation between countries, but it also produces much more reliable ceteris paribus estimates. It seems to be generally agreed upon in the literature that, if possible, applying fixed effects estimation is the optimal approach. 16 The empirical strategy is discussed further in Chapter 3. 16 Most of the papers in Table 2 use fixed effects estimation, if not as their main approach then at least as a robustness check. Spilimbergo et al. (1999), who are chiefly interested in the impact of relative factor endowments, do not use fixed effects because many countries have few observations and change in relative endowments is relatively slow (p. 85) and Milanovic (2005) does not do it because his panel is very short (p. 27), or only three observations. 18

Table 2. Comparison of cross-country studies of openness and income inequality. Paper Period Observations / Estimation Inequality measure Countries approach Definition of openness (source) Effects of openness on inequality Jaumotte, Lall, and Papageorgiou (2013) 1981-2003 290 / 51 Fixed effects Trade as a share of GDP and 100 minus average tariff rates Gini (Povcal and LIS) Negative, effects of financial openness and technology positive Bergh and Nilsson (2010) 1970-2005 493 / 78 Fixed effects and system GMM EFI and KOF indices Gini (SWIID) Positive, especially in rich countries Dreher and Gaston (2008) 1970-2000 349 / 100 Fixed effects and system GMM KOF index Estimated Gini, not based on household surveys (EHII) Positive, especially in rich countries Gourdon, Maystre, and de Melo (2008) 1975-2000 210 / 64 Fixed effects Ratio of tariff revenues to imports Gini (DS and WIDER) and deciles (WYD) Positive in rich countries, negative in poor countries; positive in skill- and capital-abundant countries Milanovic (2005) 1988-1998 201 / unknown System GMM Trade as a share of GDP Income share of each decile (WYD) Positive in poor countries, negative in rich countries Higgins and Williamson (1999) 1960s-1990s (decade averages) 449 / 44 Pooled OLS and fixed effects Sachs-Warner index Gini and the ratio of top and bottom quintile incomes (DS) Insignificant Spilimbergo, Londono, and Szekely (1999) 1965-1992 320 / 34 Pooled OLS Own index of openness, adjusted for factor endowments Gini (DS) Positive in skill-abundant countries, negative in capital-abundant countries Note: When possible, the number of observations and countries applies to specifications estimated with fixed effects. The openness indices mentioned in the table are the Economic Freedom Index (EFI) by Gwartney, Lawson, and Norton (2008), the KOF index of globalization by Dreher (2008), and the Sachs-Warner (0-1) index by Sachs and Warner (1995). The following databases are mentioned: DS by Deininger and Squire (1996), SWIID is the Standardized World Income Inequality Database by Solt (2009), EHII is the Estimated Household Income Inequality dataset by Galbraith and Kum (2005), WIDER is from the World Institute for Development Economics Research at the United Nations University, Povcal and the World Income Distribution (WYD) are from the World Bank, and LIS refers to data from the Luxembourg Income Study. Source: Author s compilation. 19

3 Data and methodology The database on income inequality used in this paper is the Lakner-Milanovic World Panel Income Distribution database (LM-WPID) by Lakner and Milanovic (2016) which derives mostly from PovcalNet and the World Income Distribution (WYD) database. 17 Here, income refers to total personal income, net of taxes and transfers, as recorded in household surveys. The data are given for five benchmark years (1988, 1993, 1998, 2003, and 2008) and come from 550 household surveys in 159 countries, 18 on average representing 95 percent of world GDP and 90 percent of the world s population. When corresponding data have been collected for control variables used in the main model specification, the result is an unbalanced panel of 112 countries and 440 observations, covering on average 85 percent of the world s population. 19 An extended model is also presented for which these numbers are somewhat lower. Countries with only one observation are not included in the panel as they are of no use in a fixed effects estimation, which is the preferred estimation approach. Table 3 gives information about the number of observations by year and region. Member countries of the Organization for Economic Co-operation and Development (OECD) in 2008 are considered as one separate region and other countries are grouped into four geographical regions (all excluding OECD countries): Africa and the Middle East, Asia, Eastern Europe and Central Asia, and Latin America and the Caribbean. As can be seen from Table 3, roughly 30 percent of the data come from OECD countries. Appendix 1 provides summary statistics for all variables used in the main and extended specifications. 17 The PovcalNet was developed for poverty measurement by the Development Research Group at the World Bank and the WYD was constructed by Milanovic. Together, PovcalNet and WYD account for almost 98 percent of the data in the LM-WPID database. Other sources are the Luxembourg Income Study (LIS), the European Union Survey of Income and Living Conditions (SILC), the British Household Panel Survey (BHPS), and the statistics offices of Finland and Portugal. 18 All surveys were conducted within two years of a benchmark year, and at least three and no more than seven years from the previous and next survey. Approximately 75 percent of the surveys were conducted within one year from the benchmark year. 19 Hong Kong and Singapore are excluded from the panel because the relation between trade and inequality is presumably different for cities and city-states than for countries that do not fall into that category. It should be noted that the significance of the results does not depend on this. 20

Table 3. Number of observations by year and region. 1988 1993 1998 2003 2008 Total (%) Total number of observations 57 90 99 99 95 440 100.0 Africa and the Middle East 11 23 23 24 23 104 23.6 Asia 9 12 14 16 15 66 15.0 Eastern Europe and Central Asia 1 6 13 12 13 45 10.2 Latin America and the Caribbean 15 20 20 18 17 90 20.5 OECD 21 29 29 29 27 135 30.7 Population (millions) 3898 4827 5193 5544 5776 Share of world population 76.3 87.1 87.2 87.2 85.4 Note: The number of observations used to estimate the main specification below. Appendix 2 provides a list of the countries included in the estimation and the number of observations for each country. Source: Population data from the World Bank (2017). 3.1 Dependent variable Apart from relying solely on high-quality data drawn from household surveys, the LM- WPID database has many desirable features for the present analysis. 20 First, it covers a large number of countries from all regions of the world during a period of rapid globalization. Second, the database provides information on the average personal income of each decile of the income distribution, which allows for a more detailed study of inequality than synthetic measures such as the Gini coefficient. Third, for a given country the data report either total income (net of taxes and transfers) or total consumption of individuals, which gives a more complete picture of the development of inequality than for example data on wages. 21 As there are important differences between measuring inequality of income and inequality of consumption, it is required that for a given country the data come from either income surveys or consumption surveys, so it is never the case that a country is represented by income data in one year and consumption data in another. 22 20 For a more detailed description of the database, see Lakner and Milanovic (2016). 21 The data are adjusted for purchasing power parity (PPP). This is crucial for comparison of income and/or consumption across countries, and hence for the calculation of global inequality, but it is not particularly important in the current context. 22 Inequality of consumption is generally lower than inequality of income, and therefore the data on inequality are not perfectly comparable across countries. But the requirement mentioned in the text makes the data comparable across time for individual contries, which is critical for the present analysis, since the preferred estimation approach relies completely on the variation within countries (see Section 3.3). 21

It should be noted that a well-known feature of household surveys is their tendency to underestimate top incomes (Korinek, Mistiaen, & Ravallion, 2006) and the data used here will therefore understate the true level of inequality. Also, since the data express only the average income of each decile of the income distribution, all calculations of inequality will fail to take into account inequality within deciles, which may be considerable in the top decile. This has to be kept in mind when interpreting the results. 3.2 Independent variables When possible, independent variables are calculated as five-year averages, so data corresponding to the benchmark year 1988 are the average for 1984-1988. This is standard in the cross-country literature on openness and inequality, and is done for a couple of reasons. Most importantly, this is done to reflect that the independent variables are not assumed to affect the income distribution instantaneously. Another reason is that taking five-year averages may reduce the risk that the results are excessively affected by temporary fluctuations and measurement errors, and moreover, it may help mitigating the endogeneity problem which will be addressed in greater detail in Chapter 4. The main variable of interest, openness to trade, is proxied by the sum of exports and imports as a share of GDP, based on data from the World Bank (2017). The decision to use a de facto measure of actual flows, instead of a de jure measure of trade policy such as barriers to trade, comes from the fact that the aim of this paper is to study how inequality is affected by globalization per se, regardless of whether it is driven by changes in policy or something else. If the focus was specifically on studying the effects of trade policy on inequality, then it would be problematic to use a measure such as the ratio of trade to GDP, since there are obviously many other factors than policy that affect such a measure. The variable is not expressed as a natural logarithm because an increase in the ratio of trade to GDP from 10 to 20 percent is assumed to have the same effect on inequality as an increase from 100 to 110 percent, not the same as an increase from 100 to 200 percent. This applies to all other variables which are denoted in percentage terms. 22

Following Bergh and Nilsson (2010), three control variables are included in all model specifications. These are the log of mean income, the dependency ratio and the share of population over 15 years old with secondary education. 23 Mean income is included to correct for changes in the income distribution related to income levels. 24 The data on mean income are taken from the LM-WPID database, like the data on inequality, and represent average net income as recorded in household surveys. The variable is expressed in natural logarithm since proportional changes in mean income, rather than absolute changes, are assumed to have the same effects on inequality. The dependency ratio, defined as the share of population under 15 years old and over 64 years old, enters the model to control for changes in demography. Higgins and Williamson (1999) find evidence for strong effects of demography on inequality where countries with large mature working age cohorts have lower inequality than countries with large young working age cohorts. The theoretical reasoning is simple. Large cohorts tend to receive lower earnings than small cohorts due to supply-effects in the labor market, and life-cycle earnings are highest for mature working age individuals. Therefore, when large cohorts are on top of the age-earnings curve, this contributes to lower inequality. Although the dependency ratio does not account for this directly it is assumed to be associated with higher inequality following similar logic. Data on demography are from the World Bank (2017). The share of population over 15 years old with secondary education is included to account for the skill-level of the workforce. The source of the data is the Barro-Lee Educational Attainment Dataset by Barro and Lee (2013) and the variable is defined as the share of people who have completed secondary education, therefore also counting people with tertiary education. 25 The effect of a higher skill-level on inequality is theoretically ambiguous. As more people receive a skill premium, inequality may go up, 23 To be exact, Bergh and Nilsson (2010) include the share of population over 25 years old with tertiary education, but here the share of population over 15 years old with secondary education is preferred instead. 24 Many empirical studies have included mean income along with the square of mean income, based on the well-known Kuznets hypothesis of inequality (Kuznets, 1955). However, since the Kuznets hypothesis is not the subject of this paper, and adding the square of mean income did not much improve the fit of the models estimated below, it is not included in the present analysis for the sake of simplicity. 25 The dataset provides data for five-year intervals, starting in 1950. Therefore, the data on education used here correspond to the years 1985, 1990, 1995, 2000, and 2005. 23

but a greater supply of skilled workers may also lead to a decrease in the skill premium itself and thereby lower inequality. An extended model specification includes three additional control variables. These are foreign direct investment (FDI), democracy, and the share of industry in GDP. FDI enters the model as the stock of (inward) FDI as a share of GDP and is included to allow for different effects of financial openness in addition to openness to trade. Democracy is expected to lower inequality by leading to greater redistribution via the median voter hypothesis and the share of industry in GDP is added to control for the structure of the economy. The sources for these variables are the United Nations Conference on Trade and Development (UNCTAD, 2016) for FDI, the Center for Systemic Peace (CSP) for democracy, and the World Bank (2017) for the share of industry in GDP. The democracy variable takes values from -10 (strongly autocratic) to 10 (strongly democratic) and since absolute changes, not proportional changes, are assumed to have the same effect on inequality, the variable is not expressed in natural logarithm. 3.3 Empirical strategy All the empirical models estimated in this paper are fixed effects regression models of the form ' where i denotes country and t denotes time, X it is a matrix of control variables, C i and T t represent country and time fixed effects, respectively, u it is an idiosyncratic error term, and the coefficients of interest are b trade and b interaction. To begin with, the model is estimated without an interaction term but in other specifications trade is allowed to interact with the log of mean income or the share of population with secondary education. This is done to investigate whether the impact of trade on inequality depends on these variables in a systematic way. All model specifications are estimated with the Gini coefficient as the dependent variable and the main specification is also studied in greater detail by estimating the model separately for the income share of each quintile of the income distribution. The country fixed effects absorb all factors that are country-specific and do not vary over time. In this context, the inclusion of country fixed effects is arguably very important, since there are many factors that affect the level of inequality and do not vary over time. For example, geography and history can be thought of as purely time-invariant 24

while factors such as institutions, culture, and ethnic or linguistic fragmentation can be thought of as being at least close to time-invariant over the period 1988 to 2008. Similarly, the time fixed effects absorb all factors that are time-specific and do not vary among countries, for example factors such as global business cycles. Therefore, allowing for fixed effects can be expected to greatly reduce bias due to omitted variables although it is true, of course, that omitted variables which do vary over time may still cause bias in the estimates. This can only be dealt with by including more relevant independent variables in the model, which is done in the extended specification and also in the robustness tests in Chapter 4. Including fixed effects is equivalent to assuming that the marginal effects of the independent variables are the same across countries but that there is a different intercept for each country. By doing so, the fixed effects estimator uses only variation within countries while ignoring variation between countries, which is why it is also known as the within estimator. As a result, the method does not work well with data for which the variation within countries is little or for variables that change very slowly over time. This may be an issue here since the average number of observations is quite low, for instance it is 3.9 for the main specification. The critical assumption needed for the fixed effects estimator to be consistent is strict exogeneity, meaning that the idiosyncratic error terms u it have zero mean conditional on all explanatory variables and fixed effects at any point in time. The validity of this assumption is evaluated in the robustness tests in Chapter 4. For the estimator to be efficient, the idiosyncratic errors have to be homoscedastic and not serially correlated. But since this is not likely to be the case, all regressions are run with heteroscedasticity and autocorrelation consistent standard errors, assuming that the errors are clustered at the country level. The method used is from Arellano (1987) and is optimal when the time-dimension is relatively short and the cross-section is relatively large. 26 (Stock and Watson, 2015; Wooldridge, 2001) 26 The estimations are carried out in R using the plm-package. The robust standard errors are calculated by the Arellano method and are of type sss, which corresponds to the small-sample correction used by Stata. 25

4 Results 4.1 Regression analysis Table 4 displays the estimation results of four model specifications in which the dependent variable is the Gini coefficient. The first column does not include an interaction term, so trade is assumed to have the same marginal effects on inequality across all countries. In the second and third columns trade is allowed to interact with income levels and the share of population with secondary education, providing a test of the simple version of the Stolper-Samuelson theorem described in Chapter 1. In columns 1 and 2, the variables of interest are not statistically significant at any reasonable level of significance. In column 3, however, both trade and the interaction term become significant at the 1 percent significance level suggesting that the effect of trade on inequality is highly dependent on education levels. Interestingly, the signs of the coefficients are in stark contrast with those predicted by the Stolper-Samuelson theorem. Instead of a negative trade coefficient and a positive interaction coefficient, with the net effect being positive in countries abundant in skilled-labor and negative in countries abundant in unskilled-labor, the results presented here indicate the opposite. That is, trade has less of a positive effect on inequality as the share of its population with secondary education increases until at some point the net effect is fully reversed so that more trade is expected to lower inequality. This turning point occurs when approximately 55 percent of the population have secondary education, such as in France and the UK in 2008. The results carry over to the extended specification in column 4 where additional control variables are included in the model. These are the stock of inward FDI as a share of GDP, democracy, and the share of industry in GDP. The coefficients for trade and the interaction term are similar in size and become even more significant than in column 3, now at the 0.1 percent significance level. The additional control variables increase the fit of the model as measured by R 2 but none of them is statistically significant. The main specification in column 3 will now be studied in greater detail by estimating the model separately for each quintile of the income distribution. 27 27 Estimating the extended specification for each quintile of the income distribution yields very similar results, so focusing on the main specification should be good enough. 26

Table 4. Fixed effects regression results: Gini. (1) (2) (3) (4) Gini Gini Gini Gini Trade 0.0191 0.0982 0.0758** 0.0816*** (1.22) (1.34) (3.25) (3.40) Log of mean income 3.87* 4.57* 3.67* 4.58** (2.58) (2.65) (2.48) (2.82) Dependency ratio 0.223 0.254 0.284' 0.159 (1.33) (1.45) (1.69) (0.88) Share secondary 0.0508 0.0591 0.187** 0.0816*** (1.29) (1.48) (2.91) (2.65) Trade * log of mean income -0.0100 (1.16) Trade * share secondary -0.00139** -0.00164*** (3.18) (3.42) Foreign direct investment -0.00145 (0.09) Democracy -0.0528 (0.39) Share industry -0.0481 (0.55) Number of observations 440 440 440 374 Number of countries 112 112 112 101 R 2 (within) 0.0667 0.0698 0.0934 0.115 Significance levels: *** 0.1 percent, ** 1 percent, * 5 percent, ' 10 percent Note: The dependent variable is the Gini coefficient. Numbers in parentheses are t-statistics, computed from robust standard errors (see footnote 26). The independent variables are trade as a share of GDP, the log of mean income, the dependency ratio, the share of population over 15 years old with secondary education, the stock of (inward) foreign direct invest as a share of GDP, a measure of democracy from -10 to 10, and the share of industry in GDP. The results are presented in Table 5 where the dependent variable is the income of each quintile as a share of mean income. A clear pattern emerges from the table. For quintiles one to three (the bottom 60 percent of the income distribution) the coefficient for trade is negative and the interaction coefficient is positive, while for the top quintile the coefficients have the opposite signs. There is no evidence that trade affects the income share of the fourth quintile in a systematic way. Note that the results are particularly stark for the bottom 20 percent and the top 20 percent of the income distribution. Again, the turning point occurs when the share of population with secondary education is approximately 55 percent. 27