Does Capital Account Liberalization Affect Income Inequality?

By adopting an identification strategy of difference-in-difference estimation combined with propensity score matching between liberalized and closed countries, this paper provides robust evidence that opening the capital account is associated with an increase in income inequality in developing countries. Specifically, capital account liberalization, in the long run, is associated with a reduction in the income share of the poorest half by 2.66-3.79 percentage points and an increase in that of the richest 10% by 5.19-8.76 percentage points. Moreover, directions and categories of capital account liberalization matter. The relationship is more pronounced when liberalizing inward and equity capital flows.

I.

Introduction
For policymakers worldwide, one of the top concerns is the current debate on how and to what extent a country, especially a developing country, should liberalize its capital account. Therefore, a clear understanding of the related impacts is essential. Specically, capital account liberalization is the external aspect of nancial liberalization and indicates policies that are designed to reduce the constraints of cross-border capital ows into or from foreign economies. Compared with the large body of studies investigating the consequences of capital account liberalization for economic growth and nancial instability, works on its distributional consequences are much less common. In recent decades, a simultaneous increase in income inequality and capital account liberalization has emerged as a signicant phenomenon. A rst glance at the trends of capital account openness and income inequality suggests a positive correlation. Figure 1 shows that income inequality developed in tandem with capital account openness during the period 1970-2015. 1 However, the theoretical hypothesis on the relationship is ambiguous, and much empirical work needs to be done.
This study investigates the relationship between capital account liberalization and income inequality. Specically, it assesses whether and how domestic income inequality changes with the liberalization of cross-border capital ows. Hence, this study focuses on inequality within rather than between countries even though global inequality (i.e., worldwide income distribution) is important. 2 The key ndings are threefold. First, capital account liberalization is associated with an increase in income inequality in developing countries; however, the relationship is insignicant for developed economies. Moreover, the association is stronger over the long term than over the short term: opening the capital account is associated with a short-term rise of 0.07-0.30 standard deviations in the overall Gini coecient and as large as 0.32-0.62 standard deviations over the following ten years. Second, the increase in income inequality is attributable to the considerable increase in the income share of the rich groups and the decrease in that of the poor groups after capital account liberalization. The magnitude of increase in the income share of rich groups is higher than the decrease seen among the poor groups: there is a decrease in the income share of the poorest half by 2.66-3.79 percentage points and an increase in that of the richest 10% by 5.19-8.76 percentage points over the long term. Third, in terms 1 The correlation coecient between the Gini coecient and the Chinn-Ito capital account openness measure is 0.86, and it is statistically signicant with a p-value of less than 0.001. 2 Our choice of focus is based on two reasons. The rst pertains to the diculty of constructing a global inequality index from representative worldwide income data and the second is that we are interested to see whether there are dierent ndings of capital account liberalization with respect to heterogeneity among countries. of the dierent dimensions of capital account liberalization, we nd that both directions and categories are signicant. The strong association with increased income inequality arises mainly from inward capital account liberalization rather than from outward liberalization; moreover, the relationship is the most pronounced in the liberalization of the international equity market while liberalizing foreign direct investment (FDI) shows a much smaller and statistically insignicant relationship with income inequality. All of these ndings do not depend on the selection of specic indicators of capital account liberalization or income inequality. We demonstrate the robustness of our ndings by using Gini coecients and income shares from other databases and other capital account liberalization indicators, and the results do not change qualitatively and are quantitatively even stronger in some specications.
This study makes four substantive contributions to the literature that links external nancial liberalization to income inequality. First, we provide evidence of an association between opening the capital account and the income shares of dierent income groups. Previous studies have largely used the nationwide Gini index as the dependent variable, and thus the use of income share data in this study not only provides overall distributional results but also captures which group changes the most.
Second, we construct a new capital account liberalization index based on existing ones and identify exact liberalizing years for each country based on various capital account openness indicators. By regressing on the other capital account openness indicators, we extend the data of Fernández et al. (2016), including granular data for dierent directions and categories of capital account liberalization. Next, we date the exact liberalizing year for each country and construct a dataset that complements a dierence-in-dierence (DID) analysis. Specically, the liberalizing year is identied when there is a substantial change in the average degree of capital account openness in the ten years before and after, especially when the average openness value changes from negative to positive. Third, we employ the DID approach combined with propensity score matching (PSM) to estimate the relationship between opening the capital account and income inequality within a 20-year window. Thus, we mitigate endogeneity concerns of conventional panel xed eects models, as the DID method aims to construct a quasi-experiment by selecting two groups of similar countries and randomly liberalizing the capital account of the treated group while keeping that of the control group closed. Thus, we can cautiously interpret the ndings of this study one step closer to causality. Fourth, we distinguish between the heterogeneous results of various dimensions of capital account liberalization, which can help narrow the discussion on specic opening policies.
The rest of this paper is organized as follows. Section II reviews the relevant literature. Section III describes the data and variables. Section IV describes the two empirical model specications. Section V presents the estimation results. Section VI concludes. II. Literature Review The interaction between nance and income distribution is necessary to understand the economic impacts of nancial policies and manage the social tensions of inequality. Building on the literature, we rst dene three nested key terms: nancial development, nancial liberalization, and capital account liberalization. We then discuss possible transmission mechanisms between capital account liberalization and income inequality and state our contributions by summarizing relevant studies and novelty of this study.
First, the term nancial development is a broad concept. It involves the establishment and expansion of nancial institutions, instruments, and markets. Most studies on nancial development focus on domestic nancial markets. Theoretically, Becker and Tomes (1979), Galor and Zeira (1993), and Banerjee and Newman (1993) show that nancial market imperfections impede risk-sharing; thus, easing credit constraints and providing the poor access to nancial markets can improve equalized distribution. By contrast, Greenwood and Jovanovic (1990) suggest a nonlinear relationship between nance and income inequality. In the early stages of nancial development, inequality is likely to increase because richer agents have less information friction on risky investments; however, as the nancial sector matures and becomes more extensive, inequality reduces because more participants have access to the nancial market. Generally speaking, empirical studies have measured nancial development by examining how eciently the nancial system fuels the economy. Indicators of nancial development most commonly used in the literature include the gross domestic product (GDP) share of liquidity liabilities such as M2 (Li et al., 1998;Milanovic, 2005;Hamori and Hashiguchi, 2012), the GDP share of credit to the nonnancial sector (Clarke et al., 2006;Beck et al., 2007;Hamori and Hashiguchi, 2012), and stock market capitalization (Baiardi and Morana, 2018;Asteriou et al., 2014;Das and Mohapatra, 2003 Delis et al. (2013), and Li and Yu (2014) nd that nancial liberalization reduces income inequality, but its composition matters, and dierent categories of nancial liberalization can have dierent impacts. Jaumotte et al. (2013) and Zhang and Naceur (2019), however, compare the role of nancial liberalization with that of trade liberalization and nancial access and nd that nancial liberalization increases income inequality.
Third, capital account liberalization is the external aspect of nancial liberalization, and we use this term interchangeably with nancial globalization. Compared with domestic nancial liberalization such as lifting interest rate controls and credit controls, capital account liberalization specically indicates a reduction in crossborder capital ows and investment constraints into or from foreign economies. As the global nancial market has become more integrated over recent decades, studies on nancial globalization have become more common. The literature on the impact of nancial globalization focuses on economic growth (Bekaert et al., 2005;Prasad et al., 2005;Kose et al., 2009) and nancial stability (Berger et al., 2016;Cubillas and González, 2014), and its distributional consequences have thus been underinvestigated until recently.
Several channels could link capital account liberalization to income inequality, but with ambiguous predictions for the sign of the correlation. First, when international capital ows into high-skill industries, opening the capital account would increase wages for high-skilled workers relative to low-skilled workers, thus raising income inequality. This contrasts with the implications of the Stolper-Samuelson theorem (Stolper and Samuelson, 1941), according to which low-skilled workers' wages would increase in developing countries with trade openness because these countries are relatively abundant in low-skilled workers. The Stolper-Samuelson theorem assumes that neither labor nor capital can ow freely across borders. When the movement of cross-border capital ows is allowed, the implication of a reduction in inequality weakens. Second, capital account liberalization can aect income inequality by changing access to nancial resources and the depth of nancial services for dierent income groups. These channels imply that the composition of capital ows matters. For instance, there is evidence that FDI is more inclined to ow into high-skilled sectors and that this tends to increase inequality (Choi, 2006;Acharyya, 2011;Wu and Hsu, 2012;Jaumotte et al., 2013). Meanwhile, Herzer and Nunnenkamp (2013) nd that FDI reduces income inequality over the long run while the short-run eect can be positive, and some studies show that foreign bank lending is likely to be associated with improved nancial access for the poor, which reduces inequality (Fund, 2007). Therefore, whether capital account liberalization is associated with an increase or a decrease in income inequality remains an empirical question.
In this study, the research question concerns whether and how capital account liberalization, which is distinguished by the direction and category of capital ows, is associated with income inequality. The relevant literature has ignored this issue until recently, and the existing ndings are inconclusive. Fund (2007), Jaumotte et al. (2013 and Asteriou et al. (2014) nd that capital account openness is associated with increased inequality. They argue that the dis-equalizing impact increases the premium on high-skilled labor and possibly returns to capital and that this is more signicant in developed countries. In contrast, Dorn et al. (2018) employ an instrumented variable approach and nd a robust and positive link between globalization and the Gini coecient in the case of transition economies versus advanced economies, but they measure globalization in terms of trade, FDI, and social and political globalization and do not consider external nance globalization. Both Jaumotte and Osorio (2015) and Zhang and Naceur (2019) argue that external nancial liberalization policies are related to higher inequality, and Das and Mohapatra (2003) nd that income inequality increases subsequent equity market liberalization.
Other studies nd that the relationship between capital account liberalization and inequality is conditional. Furceri and Loungani (2018) Kim andLin (2011), Han et al. (2012), Kirschenmann et al. (2016), Mah (2013) and Cabral et al. (2016) use the metric of quintile or decile income shares and the income share of the poorest or richest groups to measure income inequality. However, the datasets used in these studies lack international coverage or are restricted to limited years. Using income share data for more countries and years, this study examines which income groups show the largest increase or decrease in income share in relation to capital account liberalization and how such relationships are dierent between income groups.
Second, this study distinguishes between various perspectives on capital account liberalization. As stated in Asteriou et al. (2014), the composition of nancial ows is signicant for the net eect of globalization on inequality. Building on the new capital account liberalization measure proposed by Fernández et al. (2016), we examine how the liberalization-inequality correlation diers between the liberalization of equities, bonds, FDI, and other capital. We also distinguish between inward and outward capital account liberalization. This study is, therefore, the rst to investigate the association between dierent categories of capital account liberalization and income inequality, thereby having practical implications for policymakers in designing a roadmap of capital account opening.
Third, this study contributes by mitigating endogeneity concerns in the relationship between capital account liberalization and income inequality using a DID model. In addition to the generalized method of moments (GMM) estimation technique used in the panel data model to mitigate the endogeneity of the capital account openness variable, we construct a DID dataset by identifying the exact year of capital account liberalization when a country has experienced a substantial change from a closed capital account to a more liberalized one and by pairing the treated countries with control countries similar to the treated ones before liberalization.
This methodology allows us to compare the change in income inequality between cases with capital account liberalization and those without liberalization based on a quasi-natural experiment. Similarly, the philosophy of identifying episodes of capital account liberalization and conducting DID analysis is applied in Larrain (2014) and Furceri et al. (2019), and our methods dier from theirs in the following respects.
First, we use regressions to identify liberalization years, which show a signicant change in capital account liberalization from an average negative value of capital account openness ten years before to an average positive value of capital account openness ten years after, instead of using simple criteria such as dierences greater than two standard deviations in annual capital account openness. Second, as we are interested in the long-run relationship, this requires considering capital account liberalization episodes of longer than ten years to compare the change in average income inequality ten years before and after liberalization; the studies above, by contrast, focus on short-to medium-term eects occurring within the ve years following liberalization. Third, while the above works do not match treated groups with appropriate control groups over a twenty-year window, we combine the DID setting with the PSM method and construct a quasi-natural experiment of capital account liberalization. Fourth, while the above works do not distinguish between directions and categories of capital account openness, we document dierent years of inward and outward liberalization and those of the liberalization of the equity market, the bond market, FDI, and other investments.

III. Data
This section describes our datasets and the construction of key and control variables.
The key variables are measures of income inequality and capital account liberaliza- The WIID succeeds the dataset compiled by Deininger and Squire (1996) and is commonly used in empirical studies on income inequality, but includes mixed data (i.e., gross versus net, household versus individual, and income versus expenditure data) and less frequent observations. As noted by Gimet and Lagoarde-Segot (2011), merely extrapolating values or extending the data interval based on Deininger and Squire (1996)  The EHII dataset circumvents these problems by deriving the econometric relationship between the Deininger-Squire Gini coecient and a Theil-index-based measure of the industrial sector pay dispersion 5 by controlling the manufacturing employment-to-total-population ratio and other variables. Thus, EHII data account  Table   A1. While each database has its advantages and disadvantages, to the best of our knowledge, the EHII database is the most comprehensive and comparable source of income Gini coecients; therefore, we employ it in our baseline analysis and use data from other sources in our robustness checks.
The time series of the EHII Gini coecient in Figure 1 shows that income inequality is higher in non-OECD 8 countries than in OECD countries and that their trends dier by period. From 1970 to 1987, the Gini coecients of non-OECD countries were declining, whereas those of OECD countries were rising, albeit from a much lower level. From 1987 to 1995, both groups of countries experienced a deterioration in income inequality; however, OECD countries remained stable from 1995 while non-OECD countries could not stabilize until the early 2000s. From 2007 to 2013, income inequality remained stable in non-OECD countries but increased steadily in OECD countries. After 2013, income inequality declined in both groups.
However, reducing the whole income distribution to a single Gini index value can be too simplied to capture the overall distribution structure (Piketty and Zucman, 2014). In addition to the Gini index, we use income share data from the World Inequality Database (WID), which was rst developed by Piketty and Zucman (2014) and later expanded to include the evolution of the national income structure in the 5 Industrial sector pay dispersion data are from the UTIP-UNIDO (United Nations Industrial Development Organization).
6 The construction process of the EHII Gini coecient is described in detail in Galbraith and Kum (2005) and Gimet and Lagoarde-Segot (2011 the WIID data include 1,371 non-missing observations for 93 countries for the same period, although with more granular data for the income share of the rst to fth quintile groups. We also use more granular but smaller coverage income share data from the WIID in our robustness checks.

Capital Account Liberalization
Capital account liberalization is the key explanatory variable in this study. We employ three dierent de jure capital account openness indicators from Chinn and Ito The KAOP EN index developed by Chinn and Ito (2008) is constructed as the rst standardized principal component of k1, k2, SHAREk3, and k4, where k1 is the dummy variable indicating the absence of multiple exchange rates, k2 is the dummy indicating the absence of restrictions on current account transactions, SHAREk3 is the share of a ve-year window in which capital controls were not in eect (k3 is a dummy indicating the absence of restrictions on payments of capital transactions), and k4 is the dummy indicating the absence of requirements to surrender export proceeds. The advantage of the Chinn-Ito index lies in its comprehensive coverage of countries and time periods (i.e., 182 countries from 1970 to 2015); however, there are two concerns. First, three of the four components (k1, k2, and k4) are nancial current account instead of the capital account. Second, the ve-year moving average of the SHAREk3 may subvert the procedure for accurately dating the capital account liberalization year and thus aect later DID estimates. As Chinn and Ito (2002) argue, however, the incorporation of k1, k2, and k4 is based on merit and can be interpreted as the intensity of capital controls because countries may still restrict the ow of capital by limiting transactions on current account or through other systems such as multiple exchange rates and requirements to surrender export proceeds even when the capital transaction is not controlled, and restrictions on the nancial current account ensure that the private sector does not circumvent capital account restrictions. Thus, we continue to use the aggregated Chinn-Ito index in the main analysis, but we present and discuss the dynamic panel estimates using each of the four subcomponents and the DID estimates using the original k3 (without smoothing over the years) in identifying the liberalizing year in appendix Section A5.
The Quinn-Toyoda index is based on a simple textual analysis of text published in the AREAER, which reports on laws used to govern international nancial transactions. This approach measures both the existence (or absence) of restrictions and the magnitude of those restrictions starting from the lowest level (by contrast, k 1 to k 4 in the Chinn-Ito index are dichotomous). The original Quinn-Toyoda index consists of CAP and CUR, which respectively represent openness to capital ows and proceeds from the international trade of goods and services. We only use the CAP, as we focus on the liberalization of capital transactions, and we already have the Chinn-Ito index to account for possible capital transactions made under the category of nancial current account. 9 The advantage of the Quinn-Toyoda index lies in its simplicity and preciseness in capturing capital account openness, which contributes to the dating of capital account liberalization years. Limitations include its more limited coverage of developing countries and updating diculties, which result in lesser availability for recent years, although the authors have made several eorts and expanded the coverage to 126 countries for 1970-2014. 10 The FKRSU dataset provides more granularity by distinguishing the directions and categories of capital ows, unlike the other two indicators. It contains capital control information for ten types of assets: money markets, bonds and other debt securities, equities, collective investments, nancial credits, derivatives, commercial credit, guarantees, sureties and nancial back-up facilities, real estate transactions, and direct investments. To distinguish the direction of capital ows, Fernández et al. (2016) use the buyer's or seller's tax residence information and whether a transaction represents a purchase, sale, or issuance. In this study, we use the aggregated capital controls of each of the four kinds of assets: equities; bonds; direct investments; and other investments, which is the average of the remaining seven types of assets 11 , the aggregated capital controls for the overall capital outow and inow, and the most aggregated capital control of the entire capital account. As the AREAER provides 9 Specically, the Quinn-Toyoda index in this study refers to the CAP component of the original Quinn-Toyoda index.
10 We thank Prof. Dennis Quinn for providing us with the most updated data, which were not publicly available when we were conducting this study.
11 We decide to use this classication because it is generally used in the national Balance of Payments (BOP) Tables and we do not use all ten categories to reduce the diculty in deriving informational ndings and mitigate possible multicollinearity problems in the regressions. sub-categorical information instead of a single type of capital transactions from 1995, the original FKRSU dataset is only available for 100 countries for 1995-2015.
As we show below, a large proportion of capital account liberalization happened in the 1970s and 1980s; from 1995 onwards, the FKRSU index is most limited in comparing impacts before and after liberalization. Thus, we follow Bekaert et al. We also employ the original FKRSU index (starting from 1995 without imputation) and a de facto indicator to measure capital account openness in the robustness checks. The de facto index gauges the actual scale of cross-border capital ows, which may present a dierent pattern from the de jure index especially when the capital control policy is ineective or its implementation is weak. We use the 12 The adjusted R-square is 0.80 for the OECD samples and 0.91 for the non-OECD samples, and the within R-square, which is net of country xed eects, is 0.46 for the OECD samples and 0.42 for the non-OECD samples.
13 However, the possibility of measurement error should be noted, though we believe that the pseudo-FKRSU works well. This is reected in the larger standard error found in the pseudo-FKRSU estimates than in the Chinn-Ito and Quinn-Toyoda estimates shown in Section V. Further details are discussed in appendix Section A2.
14 We report the correlation in the appendix in Table A3. de facto capital account liberalization measure based on Lane and Milesi-Ferretti (2007). Specically, we adopt the ratio of the sum of total external assets and total external liabilities to GDP and its components of the ratio of total equity ows, total debt ows, and total FDI ows to GDP. The correlation between each de jure capital account openness measure and the de facto indicator is much lower than that between the de jure measures but still signicantly positive with coecients ranging from 0.17 to 0.31. 15 Based on the de jure index of capital account liberalization, we can identify the exact year in which a country substantially liberalized its capital account (i.e., converting it from a closed account to a liberalized one). Admittedly, capital account liberalization is not a one-time event but rather a continuous process. However, in certain years, governments were determined to liberalize their capital accounts and removed many constraints on international capital ows. These years mark a substantial shift in capital account liberalization and can form strong beforeand-after contrasts, which we see as a quasi-experiment that suits a DID analysis (described in detail in the next section).
We mainly follow Braun and Raddatz (2007) with some supplements and revisions to nd liberalizing years using each of the three capital account openness indicators. Methodologically, we use regressions to identify the year showing a substantial change of the 10-year average in the 20-year window centered around that year. 16 Due to space limitations, we present specic methods and tables reporting the identied year for each country based on each of the three indicators in appendix Section A3.

Control Variables
Following the recent literature on nance and income inequality (Asteriou et al., 2014;Johansson and Wang, 2014;Seven and Coskun, 2016), in all estimations shown below, we control for a set of conventional variables including GDP per capita, the square of GDP per capita, ination, trade openness, education, the age dependency ratio, government consumption, private credit, money supply, and unemployment.
However, to relieve concerns that the main ndings may be based on selective control variables, we report additional results with no controls or with basic macroeconomic controls only (i.e., GDP per capita and its squared term and ination) in Tables A11-A14 in the appendix.
In addition, studies show that institutional quality and corruption are also useful determinants of income inequality (Lin and Fu, 2016;Chong and Gradstein, 2007;Li et al., 2000;Gupta et al., 2002). To measure institutional quality and corruption, we use P olity2 from the Polity IV datasets and the corruption index from the International Country Risk Guide (ICRG) database published by the Political Risk Services (PRS) Group. However, these two additional variables start from a recent year (1984) and are not available for the 1970s and early 1980s, when many instances of capital account liberalization happened. Including them in the regression, thus, results in much fewer observations and lowers the credibility of estimates of the DID analysis, which requires ten years of data before and after liberalization. Therefore, we do not control them in the baseline analysis but use them in the robustness checks.

IV.
Empirical Methodology

Panel Fixed Eects Model
We rst apply the conventional panel xed eects model with the following specication: where i is the country and t is the year. For the dependent variable Inequality i,t , we use both the Gini coecient and the income share of dierent groups, and we control for its lagged term to account for possible persistence. In the panel xed eects model, CapitalAccountLiberalization i,t represents the capital account openness indicators, and we use the Chinn-Ito, Quinn-Toyoda, and pseudo-FKRSU indicators in the baseline regressions and the original FKRSU and de facto openness indicators in the robustness checks. We also employ the ner subcategory indicators of the pseudo-FKRSU dataset to investigate the role of the dierent dimensions of capital account liberalization on income inequality. In X i,t , we include control variables such as GDP per capita and its squared term, ination, private credit, unemployment, money supply, education, government consumption, urbanization, the age dependency ratio, and trade openness. Finally, we control for country xed eects.
As it is a dynamic panel model, OLS estimates using xed eects can be biased.
We thus estimate Equation (1) using the GMM proposed by Arellano and Bond (1991) and Blundell and Bond (1998). We treat the capital account openness indicator as endogenous and the lagged dependent variable as pre-determined, and we use their lagged terms as instrumental variables. In choosing between the dierence GMM and system GMM, we use the latter in the baseline analysis because there could be weak instrument issues for dierence equations when lagged levels are only weakly correlated with the subsequent rst dierences. This is more likely when the panel units are relatively large and the periods considered are short, and thus adding level equations can correct for potential bias using dierence equations only.
However, as pointed out by Roodman (2009) Coecient β 1 bears the highest interest. When the dependent variable is the Gini coecient, a signicantly positive β 1 indicates that capital account liberalization is associated with an increase in income inequality and vice versa. When the dependent variable is the income share, a signicantly positive β 1 indicates that capital account liberalization is associated with an increase in the income share of a certain group and vice versa.

DID Model
Taking advantage of the identied years of capital account liberalization, we can simulate a quasi-randomized experiment and conduct a DID analysis. The standard DID specication is as follows: The coecient of interest is γ 2 on the interaction term of P OST T and T REAT ED i,T . Vector X i,t contains a group of control variables that are the same as those in the panel xed eects model. φ i is the country xed eect, which can be used to control for a range of omitted variables.
We want to establish the long-term relationship between opening the capital account and inequality while reducing the inuence of short-term dynamics on the estimation. Therefore, we use the 10-year average of all variables before and after capital account liberalization. Specically, for each treated country x that liberalized its capital account in year xt and each of its control countries xj 1 , xj 2 , ..., xj n , we take the averages of two periods, [xt − 10, xt) and [xt, xt + 10). Thus, the value of variable P OST T is 0 for the average of period [xt − 10, xt) and 1 for the average of period [xt, xt + 10). We identify the treated countries as those having experienced a capital account liberalization event, and their value of T REAT ED i,T is 1. The key question is to nd the best control groups for each treated country (i.e., countries with T REAT ED i,T = 0). We adopt two approaches following Levchenko et al. (2009). First, we clean and select treated countries. When the countries have two periods of capital account liberalization (i.e., experiencing a reversal after the rst round of liberalization), we treat them as two separate observations if the gap between the two liberalization periods is more than ten years and remove such cases if the reversal happened within ten years. We also require that capital account liberalization periods last longer than ten years. As the identication of a signicant breakpoint near edge years can be unstable, we remove the case when the identied liberalization year lies within the rst two years of the country sample. We also omit cases for which the capital account is always open, leaving countries that have liberalized their capital accounts from a closed state and those retaining a closed capital account throughout the sample period. We then generate the variable LIB, which takes a value of 1 for the former and 0 for the latter. We consider countries that experienced capital account liberalization as the treated countries where variable T REAT ED equals 1 if LIB = 1. Second, we nd the group of control countries for each treated country using two approaches: the broad approach and the PSM approach. For the former, for each treated country x that liberalized its capital account in year xt, we use two criteria to determine control countries xj1, xj2 . . . xjn. First, their capital accounts should be closed during the 20-year window [t-10, t+10), including countries whose capital accounts are always closed (i.e., LIB = 0) and those that experienced capital account liberalization (i.e., LIB = 1) but with a year of liberalization jt later than xt + 10 or earlier than xt − 10. The second criterion is that they should be OECD countries if the treated country is an OECD country or non-OECD countries if the treated country is a non-OECD country.
Under the broad approach, limiting the control countries to those with closed capital accounts in the same period and belonging to either the same OECD or non-OECD group of treated countries can help easily pair the treated country with many control countries. However, the control country can still be dierent from the treated one. The PSM method thus allows us to select the most similar countries from the control groups drawn from the above broad approach.
Specically, we use the following steps to conduct PSM. First, we estimate the propensity score dened as the conditional probability of receiving capital account liberalization treatment for each country i in year t given characteristics Y from a logit model: where OP EN i,t equals 1 if the capital account of country i is open during year t. For countries whose capital accounts have always been closed, OP EN i,t takes a value of 0. For treated countries that have experienced a shift from a closed capital account to a liberalized capital account, the value of OP EN i,t is 1 if t lies in the liberalization period and 0 otherwise. Y represents a group of covariates. We follow Levchenko et al. (2009) and use the logarithm of GDP per capita (LGDP P ER), the standard deviation of GDP per capita growth for the past ve years (V OLAT ILIT Y ), trade openness (T RADE), and the chief executive's number of years in oce (Y RSOF F C). 17 These variables are signicant determinants of capital account liberalization according to the literature. We favor this parsimonious specication because the purpose of this step is not to predict liberalization as precisely as possible but to obtain a distribution of propensity scores that allows us to match the treated and potential control countries.
Again, we estimate the OECD and non-OECD countries separately. Thus, we obtain the propensity scores of capital account liberalization for each country i in year t. To conrm the balancing hypothesis, the statistical test reported in the appendix in Figure A3 shows that all of the covariates are insignicantly dierent between the matched treated and control countries, and the standardized percentage bias across the four covariates is roughly 0 for the matched countries and much larger for the unmatched ones.
Next, we keep the propensity scores of the ve years before capital account liberalization for each treated country and potential control countries identied using the broad method. Then, we construct the control group for each treated country using a proximity measure based on the propensity score. Specically, we compute the proximity between liberalized country i and another potential control country j as the average of the squared dierence between pscore i,t and pscore j,t for the ve-year period before capital account liberalization. 18  18 proximity i,j = 1 5 Σ ti t=ti−4 (pscore j,t − pscore i,t ) 2 , where t i is the liberalization year of treated countries j according to their proximity to country i and use the ve most proximate countries as the control countries for each treated country. 19 To better illustrate the process of nding our control groups using broad matching and PSM, we provide a concrete example in appendix Section A4 and report the full PSM matching results for each country in the supplementary data.
However, it should be noted that we cannot say that the estimates based on PSM-DID dominate those based on broad matching DID. First, the restrictive requirements of the PSM process substantially reduce the number of observations, which is only one-fth of that through broad matching. Second, although PSM does a good job of nding similar groups of treated and control groups, the long list of control variables in the regression is also eective in generating reliable results for the broad matching sample under the condition that other determinants are similar or remain unchanged. From the later estimations, we observe that the control variables are all almost signicant in the broad matching sample while many are insignicant in the PSM sample. Therefore, it is useful to interpret the DID results using both broad matching and PSM samples.

Capital Account Liberalization and the Gini Coecient
We rst discuss short-term dynamics between capital account liberalization and the Gini coecient with estimates of the panel xed eects model. Table 2 reports the results of estimating Equation (1) with the Gini coecient as the dependent variable.
Odd columns report the results from the xed eects model and even columns report those from the system-GMM model. We use all three capital account liberalization indices (the Chinn-Ito, Quinn-Toyoda, and pseudo-FKRSU) as shown in the column titles; this is to show that the main ndings do not depend on the selection of specic indicators. Moreover, we estimate separately for the subsamples of non-OECD and OECD countries, which are shown in the rst and last six columns.
As stated in Section IV, the GMM estimation has been criticized for its weak instrument variable (Roodman, 2009) and sensitive results. We apply a two-step system GMM estimation and conduct Windmeijer correction for the two-step standard errors. To better evaluate and interpret the results, we describe our criteria country i.
19 Also, we can follow the rst neighbor method by keeping the nearest country only so that each treated country has only one control country. The results obtained when using the one-for-one method have much fewer observations but are robust with the ve-for-one method. To save space we do not include these tables in the paper, but they are available upon request.
for generating the GMM estimates as follows. First, the coecients of the lagged dependent variable should lie between coecients from the pooled OLS and xedeects models. 20 Second, the null hypothesis of second-order autocorrelation should be rejected. Third, the model should pass the Hansen and Sargan over-identication test. The Hansen test is robust but may be weakened by many instruments, so we also conduct the Sargan test, which is not robust but not weakened by many instruments. Fourth, we collapse the instruments to combine instruments through addition into smaller sets and limit the lag depth to avoid having too many instruments. We limit the number of instruments to be less than or close to the number of groups (countries in this study) and take Hansen test statistics away from 1 but larger than 0.20 as a safe sign.
The results dier between developed and developing economies. Liberalizing the capital account tends to be associated with higher income inequality only in developing economies, as the coecients of capital account liberalization are positive and signicant in both xed eects and GMM estimates for the non-OECD subsample, but statistically insignicant for the OECD subsample. In terms of magnitudes, the xed eects estimates of capital account liberalization for the OECD subsample are lower than one-third of those for the non-OECD subsample, and the GMM estimates for the OECD subsample are less than one-tenth of those for the non-OECD subsample. In addition, the results shown in Table A10 in the appendix, which are from an interaction specication of each variable interacted with a dummy indicating non-OECD countries, also reject the equality of capital account coecients between OECD and non-OECD countries (except for xed eect estimates when we use the Quinn-Toyoda index to measure capital account liberalization). This reiterates Eichengreen (2001), who argues that developing countries are more likely to suer the negative eects of capital mobility on income distribution due to weak institutions or regulations. Besides, as shown by Figure 1 and liberalizing years shown in the data le, OECD countries had capital account openness for a longer period and experienced liberalization earlier than non-OECD countries. For our sample period, i.e., post-1970, it is more appropriate to use non-OECD countries to study capital account liberalization. Additionally, we have fewer observations for the developed economies. Thus, we focus on non-OECD countries and only report their results in the following analysis.
Specically, the results given in columns (1)-(6) of Next, we investigate the long-term impact of capital account liberalization on the Gini coecient using estimates from the DID model, which allows us to compare the average Gini coecient ten years before and after nancial liberalization for the paired treated and control countries with similar characteristics. Table 3 presents the DID estimates for the non-OECD countries. The odd and even columns dier in their matching methods: the odd columns report the results using the sample based on broad matching, and the even columns report the results using the sample based on PSM. Columns (1)-(2), (3)-(4), and (5)-(6) present the estimates using different capital account liberalization indices (Chinn-Ito, Quinn-Toyoda, and pseudo-FKRSU, respectively) to identify the treated countries and post-liberalization years.
Again, we nd an association between capital account liberalization and increased inequality for developing economies, as the interaction terms of P OST and T REAT ED are positive and signicant across all specications and matching methods, though the signicance obtained when using the Quinn-Toyoda index is weaker, as it only arises with a p-value of less than 0.1. Specically, the following three main results are found. 21 First, our methods of matching control groups and assigning pseudo-post-treatment years work well because the falsied treatment shows an insignicant impact for the control groups; meanwhile, the real treatment shows a signicant impact for the treated group. Second, the 95% condence intervals are 21 A visualized version of the ndings can be found in Figure A4 in the appendix, which shows the average marginal eects of liberalizing a country's capital account with 95% condence intervals. all above zero for the treated countries, indicating that liberalizing capital account is associated with higher inequality. Moreover, the economic signicance of the impact is considerable: a capital account liberalization event is associated with an increase in the Gini coecient by an average value ranging from 1.77 to 3.37 over 10 years, which is equivalent to 0.32 to 0.62 standard deviations of the Gini coecient observed in the sample.

Capital Account Liberalization and Income Share
In addition to the increase in the Gini coecient, we nd that capital account liberalization is associated with a decrease in the income share of the poor and an increase in the income share of the rich. We replace the dependent variable with income shares for the bottom 50%, middle 40%, and top 10%, and rerun the analysis using the panel and DID models. Similarly, we estimate the specication using all three capital account liberalization indices and both the xed eects model and system GMM model.
As shown in the upper panel of Table 4 The corresponding association between a full liberalization and the income share of the poorest half is a decrease of 0.92 to 6.53 percentage points (equivalent to 0.15 to 1.08 standard deviations), and that with the income share of the richest 10% is an increase of 2.13 to 9.85 percentage points (equivalent to 0.18 to 0.82 standard deviations). Concerning the average income shares of the bottom 50% and top 10% income groups, which are 15.91% and 46.71%, respectively, the impact is considerable. Moreover, the statistical signicance of GMM estimates is stronger for the income share of the top 10% than that for the bottom 50%, as the latter only stands out at the 10% signicance level. These results imply that capital account liberalization is associated with an increase in the income share of the rich and a decrease in that of the poor in developing economies.
Next, we present estimates of the long-term relationship between capital account liberalization and the income share based on the DID model in the lower panel of Table 4. 22 Consistent with the ndings obtained from the panel xed eects model, 22 We also visualize the results in Figure A5 in the appendix by plotting the marginal eect of capital account liberalization on income shares with 95% condence intervals.
DID estimates also suggest that capital account liberalization is associated with a decrease in the income share of poorer groups and an increase in the income share of richer groups and thus increased income inequality. Dierent from results obtained from the dynamic panel xed eects model, results obtained from the DID model suggest that capital account liberalization is also signicantly associated with a reduction in the income share of the middle group, and its long-run association with income share is only positive for the top 10%. Moreover, the gaps between estimates of the poor and rich groups are more signicant. As valid proof of classifying the treated and control groups, the post-liberalization eect is insignicant for the control groups and the 95% condence intervals only lie within the same above-zero or  Third, the most rich-biased association is found from the liberalization of the international equity market, which is associated with an increase in the income share gap between the richest 10% and the remaining 90% by roughly 8.53 percentage points while estimates of liberalizing the bond market and other investments are 6.95 and 5.71 percentage points, respectively. The insignicant impact of FDI seems to contrast with the ndings of Choi (2006), Acharyya (2011), Wu and Hsu (2012), and Jaumotte et al. (2013), who suggest a signicantly positive relationship between FDI and income inequality; however, this could be reconciled because we use de jure measurements of FDI liberalization while these studies use actual values of FDI ows, which have more portfolio capital characteristics, and greeneld investments have given way to mergers and acquisitions as argued in Mody and Murshid (2005).
In addition, we are interested in the long-run relationship while these other studies focus on the short term, and Herzer and Nunnenkamp (2013) show that the relationship between FDI and income inequality could be positive over the short term and negative over the long run.
These results have three implications. First, the relationship between increased income inequality and inward capital account liberalization reiterates past ndings in the literature that international capital tends to ow into high-skilled labor or sectors, suggesting that policymakers should be cautious of augmented skill-biased inequality from capital inows. Second, liberalized international equity markets are less likely to expand nancial access for the poor, but oer more intensive benets for those who are already rich. Third, direct investments, which tend to be long term and more stable than the rest, are more likely to display benets of nancial integration as predicted by neoclassical economic growth theory, as they do not show signicant correlations with higher inequality.

Robustness Checks
We conduct various robustness checks and report the results in appendix Section A6.
Here, we simply summarize the ndings. First, we use the non-imputed original de jure FKRSU, which starts in 1995, and the de facto indicator that captures actual cross-border capital ows as well as its dis-aggregation into three types of assets (i.e., equity, debt, and FDI) from Lane and Milesi-Ferretti (2007)  Third, we additionally control for the country's institutional quality and corruption. Studies have shown that institutional quality and corruption are associated with income inequality. We use P olity2 from the Polity IV dataset, which captures regime authority characteristics with a higher value indicating a more democratic political regime and higher institutional quality, and the corruption index from the ICRG database. 24 These two additional variables are limited and not controlled in the baseline analysis because they are only available for more recent years from 1984 while many capital account liberalizations happened in the 1970s and 1980s; therefore, we lost substantially valuable observations and are unable to use them in the DID estimation, which requires ten years of data before and after liberalization.
When estimating the dynamic panel xed eects model with additional controls of institutional quality and corruption, we add them in both linear and non-linear ways to account for a possible U-shaped relationship as shown in Li et al. (2000). For institutional quality and corruption, we nd that democratic regimes seem to be associated with higher Gini coecients, and corruption shows a U-shaped relationship with the Gini coecients; meanwhile, they do not have a consistent association with income shares. More importantly, the main conclusions that capital account liberalization is associated with higher income inequality and specically with higher Gini coecients, smaller income shares for the bottom 50%, and a larger income share for the top 10% do not change, though the statistical signicance is weaker in the system GMM estimates.
24 Besides the index from the ICRG, two other datasets are widely used to measure corruption: the corruption perceptions index published by Transparency International and the control of corruption index of the World Governance Indicators. However, as these start from more recent years (1995), we use the ICRG's corruption index, which starts from 1984. VI.

Conclusion
The relationship between capital account liberalization and income inequality has been gaining increasing attention in recent years. This has opened a relatively new area of study in nancial globalization besides its relationship with economic growth and nancial stability. However, the existing ndings are inconclusive. This study thus uses two empirical strategies, a dynamic panel xed eects model and a DID model, to revisit this question. Our ndings suggest that capital account liberalization is associated with greater income inequality in developing economies.
First, we document that changing the capital account from fully closed to fully liberalized in developing countries is associated with a rise of 0.07-0.30 standard deviations of the Gini coecients for the short term and a rise as much as 0.32-0.62 standard deviations of the Gini coecients for the ten years after liberalization.
Second, this increased income inequality involves the shrinking of the income share of the poor versus the expansion of that of the rich. When comparing the ten years before capital account liberalization with the ten years thereafter, the liberalizing event is associated with a decrease in the income share of the poorest 50% by 2.66-3.79 percentage points and an increase of the share of the richest 10% by 5. 19-8.76 percentage points. Third, we nd that the direction and category of capital account liberalization are essentially important. Inward capital account liberalization is more associated with income equality than outward liberalization, and equity market liberalization is associated with a larger increase in the income share of the rich and a larger decrease in that of the poor; meanwhile, we do not nd any signicant association between the liberalization of FDI and income inequality.
While we acknowledge that the mechanism through which capital account liberalization aects income inequality is important, we do not discuss it in this study.
To investigate this channel, we need more detailed micro-level data on household income such as the wages and compensation of workers with dierent levels of skill.