Exchange Rate Policy and LDC Foreign Borrowing

Transcription

1 Exchange Rate Policy and LDC Foreign Borrowing Samir Jahjah, Bin Wei, Vivian Zhanwei Yue y July 2010 Abstract This paper empirically analyzes how the exchange rate policy a ects the issuing and pricing of international bonds issued by less developed countries (LDCs). We measure an exchange rate policy by the de facto exchange rate regime and the real exchange rate overvaluation. We nd that countries with a less exible exchange rate regime are less likely to issue bonds and pay higher spreads. Furthermore, we nd that the real exchange rate overvaluation signi cantly increases the bond spread and the bond issuance probability. Moreover, such e ects of the real exchange rate overvaluation tend to be magni ed for countries with a xed exchange rate regime. Keywords: Sovereign Credit Spread, Exchange Rate Regime, Overvaluation, Debt Crisis JEL Classi cations: E58, F31, F33, F34 We are very grateful to two anonymous referees and Pok-sang Lam (the editor) for o ering many insightful comments and suggestions that have improved the paper immensely. We would also like to thank Frank Diebold, Mark Gertler, and Martin Uribe, and the participants at the IMF Institute Seminar for their comments. We thank Carmen Reinhart for providing us with the data on crises. This paper was previously titled Exchange Rate Policy and Sovereign Bond Spreads in Developing Countries. The authors are responsible for all errors and omissions. The views expressed in this paper are those of the authors and do not necessarily represent those of the IMF or IMF policy. y Jahjah is at the International Monetary Fund, th Street, N.W., Washington, D.C sjahjah@imf.org. Wei is at the Department of Economics and Finance, Baruch College, CUNY, 55 Lexington Avenue, New York, NY bin.wei@baruch.cuny.edu. Yue is at the Department of Economics, New York University, 19 West 4th Street, New York, NY zy3@nyu.edu. 1

2 1 Introduction The recent turmoil in the Euro zone has disturbed the economies from Greece to Italy to emerging European countries and raises the wide-spread concerns over the sovereign default and Euro depreciation. Turning our attention to developing countries and historical data, we can nd that the relation between the exchange rate arrangement and debt management has for long been an important policy issue for developing countries. However, the active policy debate on exchange rate policy and country risk has yet to be studied formally in the academic literature. The goal of our paper is to empirically examine how the exchange rate policy a ects the issuing and pricing of foreign debt for less developed countries (LDCs). Developing countries typically have a large amount of debt denominated in foreign currency. Due to the risk of default, 1 developing countries pay a sizable default risk premium on their foreign debt. When the foreign debt is denominated in foreign currency, a weaker local currency can exacerbate debt-service di culties through the balance-sheet e ect and a ect the country spread. Hence, the exchange rate management plays an important role for developing countries foreign debt nancing. At the same time, the choice of exchange rate regime remains an elusive part of macroeconomic policy. In this paper we analyze the impact of the exchange rate policy on foreign borrowing using the primary bond market data on 42 developing countries. Our main methodology is to estimate a Heckman s sample selection model (see Heckman (1979)). In the empirical analysis, we draw on the ndings in the literature to obtain a reasonable set of control variables and include the measures of exchange rate policy as the explanatory variables of bond issuing probability and bond spread. We examine the e ects of exchange rate policy on the issuing and the pricing of international bonds by developing countries. The rst measure of a country s exchange rate policy is its exchange rate regime. It remains as an open question that how the choice of an exchange rate regime impacts a country s foreign debt borrowing. Firstly, there are virtually no comprehensive empirical 1 Reinhart and Rogo (2008) document 71 default episodes for developing countries since 1975 to They also provide a panoramic analysis of the history of nancial crises dating from England s fourteenthcentury default to the current United States sub-prime nancial crisis. 2

3 studies on this question. 2 Secondly, whether a country issues a bond and how the bond is subsequently priced are presumably a ected by the country s overall economic performance. However, the economic literature does not provide unambiguous implications as to which exchange rate arrangement promotes a country s economic performance. The impact of exchange rate regime on the economic performance is probably one of the most controversial topics in macroeconomic policy. 3 Supporters of a exible exchange rate system argue that countries with hard-pegged currencies are more vulnerable to real shocks, which may adversely a ect growth and macro stability. More exible arrangements can better accommodate shocks and thus reduce the uncertainty in the economy. 4 Based on this argument, a xed exchange rate regime results in higher default risk in the context of foreign borrowing. Moreover, by eliminating the monetary policy as a viable policy instrument, hard pegs may force the government to increase its external liabilities, resulting in higher default risk. Gertler, Gilchrist, and Natalucci (2007) show that xed exchange rates exacerbate nancial crises by tieing the hands of the monetary authorities in a nancial accelerator framework. 5 On the other hand, supporters of a xed exchange rate regime argue that this type of exchange rate arrangement provides policy credibility. For example, pegging the exchange rate may help to impose scal discipline on the government. 6 The disciplining e ect of a peg may lead to a reduction in the country risk. Arellano and Heathcote (2010) speci cally show that countries with dollarization face a more favorable borrowing environment because without the monetary policy instrument, these countries value their access to the foreign 2 Obstfeld and Taylor (2003) study the impact of gold standard on country borrowing spreads on the London bond market from the 1870s to the 1930s. Arellano and Heathcote (2010) include a cross-country regression of sovereign credit ratings on the exchange rate volatility from , while they focus on the e ect of dollarization on sovereign debt in their theoretical analysis. 3 See Engel (2009) for a surreny of current research on exchange rate policy. 4 Edwards and Sturzenegger (2005) and Broda (2004) provide some empirical evidence that the terms of trade shocks have a larger e ect on economic performance in countries with more rigid exchange rate regimes, than in countries with a exible exchange rate regime. 5 Gertler, Gilchrist, and Natalucci (2007) focus on the Korean experience during the nancial crisis and quantitatively examine how defending an exchange rate peg may reinforce the nancial crisis. Cespedes, Chang and Velasco (2004) also discuss the role of exchange rate regimes on excerbating nancial crisis in a qualitative analysis. 6 Giavazzi and Pagano (1988) show that a government may choose a particular exchange rate arrangement to buy itself a reputation. 3

4 capital market more and are thus less likely to default. Moreover, a xed exchange rate system is believed by its supporters to foster a more stable environment and faster economic growth. As argued in the literature, hard pegs can lead to lower interest rates and eliminate exchange rate volatility, which stimulates investment and international trade, resulting in faster growth. 7 These growth-enhancing e ects suggest that a xed exchange rate regime may be advantageous to a country s foreign borrowing. As the preceding discussion suggests, determining how a country s exchange rate regime a ects its default probability and its foreign debt borrowing is ultimately an empirical issue that can only be elucidated by analyzing the historical evidence. Our rst main nding is that the choice of exchange rate regime has a signi cant impact on LDC foreign borrowing. Speci cally, the less exible a country s exchange rate regime is, it is less likely to issue foreign bonds and pays higher spreads. The decrease in the bond issuance probability and the increase in the bond credit spread are both statistically and economically signi cant. The marginal e ect from changing a free oating exchange rate regime to an intermediate one on the bond spread is to reduce the bond issuance probability by about 3% and increase the spread by 43 basis points, and further changing from the intermediate one to a xed one decreases the probability by 1.6% and increases the spread by an additional amount of 89 basis points. Our results therefore unambiguously point to the adverse e ect of a xed exchange rate regime on a country s foreign debt nancing, which is consistent with the conclusions from Gertler et al. (2007). Another measure of exchange rate policy is the real exchange rate overvaluation. Real exchange rate as a key relative price is important for the policy analysis because of its implications for international trade and capital ows. 8 In our analysis, we use real exchange rate overvaluation as a second measure of exchange rate policy, which is de ned as the di erence between the actual real exchange rate and its long-run equilibrium level. A country s debt policy may respond to its real exchange rate, especially when the currency 7 See Dornbusch (2001), Rose (2000), and Rose and van Wincoop (2001). Please see Levy-Yeyati and Stuzenegger (2003) for an extensive review. 8 For example, the average level of real exchange rate matters for export-led growth for developing countries, and real exchange rate is a key indicator of incipient currency crises. See Eichengreen (2008). 4

5 is misaligned, for the following reasons. First, an overvalued currency reduces a country s trade competitiveness and weakens the macroeconomic fundamentals. 9 As a result, the default risk may increase, so are the borrowing costs (See Eaton and Gersovitz (1981)). Second, exchange rate overvaluation has been found to be a main cause of currency crises. A vast literature nds that the real exchange rate is overvalued during the period prior to devaluations or crises. 10 When a country borrows in foreign currency, its debt liability becomes more costly to serve following the devaluation and hence the default risk rises. 11 Lastly, the choice of exchange rate regime and real exchange rate overvaluation may have a joint impact on sovereign debt market. 12 An in exible exchange rate regime compounds the adverse e ect of a real overvaluation because the cost of correcting the exchange rate misalignment is higher for a country with a xed exchange rate. The overvaluation has a larger and more persistent impact on the economy for a hard pegger. Therefore, a country with an in exible exchange rate regime is more likely to default on its debt when its currency is overvalued. Our second main nding is that real exchange rate overvaluation signi cantly increases foreign bond issuing probability and generally raises bond spreads for developing countries. The magnitude of this e ect di ers across exchange rate regimes. In our empirical analysis, we use three measures of real exchange rate overvaluation to examine its impact on foreign borrowing. We nd that for all three measures the interaction between a xed exchange rate regime and real exchange rate overvaluation has the biggest e ect on the supply and pricing of international bonds. Quantitatively we nd that a one-standard-deviation increase of real exchange rate overvaluation, measured by the percentage deviation of the real e ective exchange rate from its ten year average, increases the spread by 86 basis points for a country 9 Aghion et al. (2009) nd that countries su ering from real overvaluation experience slower productivity growth. Eichengreen (2008) contains a survey of the literature that document how a competitive real exchange rate fosters growth and real overvaluation slows growth for developing countries. Engel (2010) nds that currency misalignments are ine cient and lower world welfare. 10 See Dornbusch et al. (1995), Edwards (1989), Eichengreen et al. (1995, 1996), Kaminsky et al. (1998), and Goldfajn and Valdes (1999). 11 Schneider and Tornell (2004) nd that balance-of-payments crises are preceded by lending booms and real appreciation in a model with self-ful lling crises and balance sheet e ects. 12 Jahjah and Montiel (2003) nd that a hard peg increases default likelihood, especially in cases of large exchange rate overvaluation. 5

6 with a xed exchange rate regime, while the same increase only increases the spread by 33 and 29 basis points if the country is in an intermediate and oating exchange rate regime. The same pattern persists when the other two measures are used. Our main results hold in a variety of robustness tests, including allowing for alternative control variables and correcting for endogeneity. To address the endogeneity problem for the exchange rate regime and real overvaluation, we conduct a multi-stage estimation of the Heckman s selection model and use clearly exogenous variables as instrumental variables for the exchange rate regime and overvaluation. We nd that controlling for the endogeneity issue does not change our results qualitatively. These tests make us con dent that our empirical results indeed capture the impact of exchange rate policy on foreign debt for emerging countries. Linking explicitly the exchange rate policy to bonds issuing and pricing is our main contribution to the literature on sovereign default risk in emerging economies. Edwards (1984), Cline (1995), Easton and Rockerbie (1999), and others investigate the determinants of sovereign debt spreads in sovereign loans. Eichengreen and Mody (1998) and Kamin and Kleist (1999) analyze bond spreads on primary market using data on international bonds issued by developing countries. However, none of these empirical works incorporates the impact of exchange rate policy on sovereign bonds pricing and issuing. Edwards (1984) includes nominal exchange rate devaluation as one determinant of spreads, but the impact of devaluation is not signi cant. There are a few empirical analyses and event studies relating the exchange rate policy to the country risk. Reinhart (2002) examines the linkage between default, currency crises, and sovereign credit rating. She nds that defaults usually follow sharp devaluation or are responses to speculative attacks on exchange rate arrangements. Powell and Sturzenegger (2000) evaluate the relation between the elimination of currency risk through dollarization and country risk. Yet their analysis is limited to countries that adopted the Dollar or Euro. This paper is also related to the recent studies on the impact of exchange rate regime and real exchange rate volatility. Levy-Yeyati and Sturzenegger (2003) study the relationship between exchange rate regime and growth, and nd the less exible exchange rate regimes 6

7 are associated with slower growth. Broda (2004) nd that countries with exible regimes are able to bu er terms-of-trade shocks better than those with xed regimes. Aghion et al. (2009) show some empirical evidence that real exchange rate volatility can a ect the longterm productivity growth rate, and nd that the e ect depends critically on a country s level of nancial development. Our work assesses the impact of exchange rate policy on sovereign default risk which is another important dimension for developing countries. In the remainder of the paper, we describe the dataset and our methodology. The main empirical analysis is carried out in Section 3. In Section 4 we summarize the paper and conclude. 2 Data and Methodology 2.1 The Data Bond data come from Capital Data s Bondware and contain the detailed terms of bonds issued in the primary market by 42 developing countries between January 1990 and December The Bondware dataset contains information on the launch spreads, launch dates of international bonds issued in dollars by developing countries. The launch spread is de ned as the di erence between the yields on a bond issued and the U.S. Treasury bond with comparable maturity. We use the Bondware data at the individual bond level at the monthly frequency. There are totally 2,653 bond issues in the sample. The list of countries and the total number of bond issues in the sample period are reported in Table 1. Insert Table 1 Here We work with the primary bond market data because, to the best of our knowledge, there is no secondary market bond dataset that covers a large sample of countries. 14 In 13 There are initially 66 countries covered in the Capital Data s Bondware data during the sample period. Among them, four countries are dropped because they have no RR regime classi cation and twenty countries are further dropped from the sample due to the unavailability of some explanatory variables. 14 J.P. Morgan s EMBI global and EMBI+ are the secondary market datasets constructed for 23 countries starting in 1994 or later depending on the countries. 7

8 addition, using the primary market data allows us to analyze both the issuing and the pricing decisions for developing countries. We use the de facto exchange rate regime as a key explanatory variable in our empirical analysis. We employ the monthly classi cation of the de facto exchange rate regimes constructed by Reinhart and Rogo (2002) (RR) who classify the exchange rate arrangements based on the o cial exchange rate and parallel market rates. We use the de facto exchange rate regime as opposed to the de jure exchange rate regime because the latter is not a good measure of a country s exchange rate arrangement. 15 In most of the analysis, we aggregate the RR exchange rate classi cation into three groups: xed, intermediate, and free oating regimes. 16 The aggregation of exchange rate regimes is summarized in Table In the empirical analysis, we use the exchange rate regime dummies of FIX ( xed regimes), INT (intermediate regimes), and FLOAT (free oating regimes). FIX (resp., INT or FLOAT) takes the value 1 when the country is operating a xed exchange rate regime (resp., an intermediate or free oating regime) and 0 otherwise. Insert Table 2 Here Next, we compute the real exchange rate overvaluation using three measures. 18 The rst two measures of exchange rate overvaluation are computed using the monthly real e ective exchange rates (REER) from the IMF Information Notice System. The REER is a trade-weighted index of multilateral real rates measured by units of foreign goods per domestic goods. The rst measure of the real exchange rate overvaluation is the percentage 15 A country may in practice deviate from its announced exchange rate regime. Calvo and Reinhart (2002) and Alesina and Wagner (2003) study the reasons why countries do not follow their de jure exchange rate regimes. 16 We also conducted the empirical analysis using the exchange rate regimes grouped into four classes: hard peg, conventional peg, intermediate and free oating or grouped into two classes: xed and oating. The di erent grouping methods do not change the results. The estimation is available upon request. 17 Two adjustments are made to the RR classi cation. A free falling regime is de ned as one with a monthly in ation rate greater than 40%. Because the in ation is one regressor in our empirical analysis, we categorize this group using the secondary classi cation. We discard the observations in the dual-market regime because no secondary classi cation is available. Our empirical analysis is robust to the exclusion of these two groups. 18 As reported in Hinkle and Montiel (1999), there is no universal method to compute the exchange rate misalignment or real exchange rate overvalution. 8

9 deviation of the REER from its ten year average (ROV1). The second measure is the percentage change in the REER over the last ve years (ROV2). 19 The third measure is the deviation from a predicted level of the real exchange rate (ROV3). The predicted level of the real exchange rate is based on the equilibrium concept of Purchasing Power Parity and is adjusted from di erences in the relative price of non tradeables to tradeables attributed to di erences in factor endowments (i.e., the Balassa-Samuelson e ect). 20 The PPP real exchange rate is from the Penn World Table (PWT). Following Dollar (1992) and Aghion et al. (2009), we rst perform a pooled OLS regression to obtain the predicted value as an estimate of the equilibrium value of the real exchange rate, and then take the di erence between the actual real exchange rate and its predicted value from the OLS regression as the third measure of real exchange rate overvaluation. In the pooled OLS regression, income per capita relative to that of the United States and geographical and year dummies are used as proxies for factor endowments. We draw on the ndings in the literature to obtain a reasonable set of control variables that have been found to be important determinants of bond spreads. 21 We use real interest rates on ten-year U.S. Treasury bonds (USRATE) and the spreads on the U.S. high yield corporate bonds (HYD) as proxies for the global economic condition. For the domestic economic indicators, we use the GDP growth rate (GDPGR), the GDP per capita in U.S. dollars (GDPPC), the current account as a ratio of GDP (CA2GDP), and in ation (INF). We also include some liquidity and solvency variables, such as, the ratio of debt to GNP (DT2GNP), the ratio of debt service to exports (DS2EX), and the ratio of short-term debt to total debt (SHORTDT). In addition, we employ the regional dummies for countries in Africa (AFRI) and the Latin America (LAT). Our objective is to use a reasonable set of controls to test whether the exchange rate policy has a signi cant impact on the issuing and pricing of 19 These two measures are also used in Frankel and Saravelos (2010). 20 We also measure the exchange rate overvaluation using the di erence between log of the real exchange rate and its H-P trend. The results are robust, but not reported in the paper. They are available upon request. 21 Our baseline speci cation follows closely those reported in Edwards (1984), Eichengreen and Mody (1995), Dell Ariccia et al. (2002), etc. We also include control variables that are not in these earlier studies but have been extensively discussed as important determinants of international bond spread. 9

10 international bonds for emerging markets. We collect data on the macroeconomic indicators, and country-issuer characteristics from the IMF s International Financial Statistics (IFS), the World Bank s World Development Indicators (WDI), the Penn World Table (PWT), the Global Development Finance (GDF) and the Federal Reserve Board. The detailed description of the variables and their sources is are Table A1 in the Appendix. 2.2 The Econometric Methodology This subsection describes the main econometric model that is based on the Heckman s sample selection model. The credit spread of an international bond issued by a developing country is a measure of its default risk. As in Eaton and Gersovitz (1981), Edwards (1984) and the subsequent studies in the literature, we assume that the logarithm of the spread is a linear function of some explanatory variables, X, that a ect the default risk. Formally, log (SPREAD) = X + u; (1) where u is a random error term. The explanatory variables are bond characteristics, exchange rate regime dummies, real exchange rate overvaluation measures, and control variables that summarize the global economic conditions and country characteristics. Because we only observe the bond spread when a bond is issued, a sample selection problem arises. When no spread is observed for a country in a given year, we may assume that the missing spreads are random occurrences and ignore them, but if the gaps occur according to some unknown but systematic selection method, estimating Equation (1) alone leads to biased and ine cient estimates. For example, a country may be excluded from the credit market if its perceived probability of default exceeds a given level, i.e., it reaches a credit-ceiling. 22 Conversely, a country tends to issue international bonds when the borrowing conditions are favorable and its nancing need is high. To deal with the sample selection problem, we create a binary variable for the bond issuance: BI equals 1 when we 22 See Eaton and Gersovitz (1981), Sachs and Cohen (1982), and Sachs (1983). 10

11 observe a nonzero spread for a country at time t, and zero otherwise. We assume BI = 1 fz+v>0g; (2) where Z is a set of observed variables that explain the issuing decision of a country in a given month and v is a random error term. We can think of Z + v as the di erence between bene t and cost from issuing bonds, and Equation (2) indicates that a bond issue is observed if and only if the bene t exceeds the cost. The spread equation (1) and the issuance equation (2) consist of a standard Heckman s (1979) sample selection model. We can estimate Equation (2) as a probit model to get the probability of issuing a bond. Estimating the probit model requires information on those who did not issue bonds. To address this problem, we record a zero for each month and country where no bond issuance is observed. The model can be identi ed by the exclusion requirement for the Heckman selection model. In our empirical analysis, the vector of explanatory variable Z in the issuance equation (2) includes all the variables in X as well as one exclusion variable that is used for identi cation. For the exclusion variable, we use a January dummy in the bond issuance equation. The logic behind using the January dummy as the exclusion variable is the following. Countries are less likely to issue new bonds in January because of the holiday seasons for the major international nancial centers. On the other hand, the January dummy should not enter the spread equation (1) since whether or not the bonds are issued in January should not change the evaluation of the default risk. We use the maximum likelihood method to estimate Equations (1) and (2) jointly under the assumption that the error terms, u and v, follow a bivariate normal distribution. The maximum likelihood method obtains the e cient estimates under a correctly speci ed model. We also check the results by estimating the model using the Heckman s two-stage method. 23 The two procedures give similar results. 23 The two-stage estimation method of the Heckman s model is implemented as follows. In the rst stage, Equation (2) is estimated as a Probit model to get the probability of a bond issue. Then, the value of Mill s ratio (re ecting the conditional probability of the observation being in the observed sample) is incorporated in an OLS regression of (2) using the observed log (spread) only. 11

12 In the empirical analysis, we also quantify the impact of exchange rate regime and real overvaluation on the issuing and pricing of the international bonds by calculating the marginal e ects. The marginal e ects consist of two components. There is a direct e ect on the mean of log (SP READ), but also an indirect e ect because the exchange rate regime or real overvaluation a ects the bond issuing decision and hence in uences log (SP READ) indirectly. First, the marginal e ect on the bond spread of changing a country s exchange rate regime from FLOAT to INT is given by 24 E [log (SP READ) j INT log (SP READ) j F LOAT jbi = 1] (4)!!# Z (0;1) Z (0;0) = INT + u " : v v where F IX is the coe cient of F IX in Equation (1) and (x) (x) = (x) is the inverse Mill s ratio in which and are, respectively, the probability density function (PDF) and the cumulative distribution function (CDF) of a standard normal random variable. Let Z (0;0) be the vector of explanatory variables in the bond issuance equation (2) with (F IX; INT ) = (0; 0) and all the other variables at their mean values. Z (0;1) or Z (1;0) is similarly de ned except that (F IX; INT ) is equal to (0; 1) or (1; 0), respectively. Similarly, if the exchange rate regime changes from INT to FIX, then the marginal e ect is given by E [log (SP READ) j F IX log (SP READ) j INT jbi = 1] (5)!!# Z (1;0) Z (0;1) = F IX INT + u " v v where INT is the coe cient of INT in Equation (1). 24 We derive the marginal e ects in Equations (4)-(6) by following Greene (2002). The key is to derive the conditional expectation of log (SP READ) conditioning on the spread being observed, which is given by E [log (SP READ) jbi = 1 ] = X + u ( Z= v) : (3) 12

13 Lastly, the marginal e ect of ROV evaluated at the sample mean in the observed sample is given [log (SP READ) jbi = = ROV ROV u Z v (6) where ROV and ROV denote the coe cients of real exchange rate overvaluation (ROV) in Equations (1)-(2), (x) ( (x)) 2 x (x), and Z is the vector of explanatory variables in the bond issuance equation (2). The marginal e ect of ROV in a given exchange rate regime is similarly de ned. 3 Empirical Analysis In this section we empirically investigate the e ects of the choice of the exchange rate regime (FIX, INT, or FLOAT) and the real exchange rate overvaluation (ROV1-ROV3) on the issuing and the pricing of international bonds by developing countries. We rst report the empirical results in the baseline speci cation, and then report in the next section the results of various robustness tests including the endogeneity tests. 3.1 Empirical Results We explore the e ects of the exchange rate regimes and real exchange rates on LDC foreign borrowing. Because we do not intend to reexamine results profusely analyzed in the empirical sovereign bond spread literature, we choose a relatively noncontroversial set of control variables. 25 We then add the exchange rate regime dummies, FIX ( xed exchange rates), and INT (intermediates), as well as the measures of real exchange rate overvaluation (ROV1-ROV3), in the empirical analysis. We rst estimate the baseline model in which we include the regime dummies (FIX and INT) together with a set of explanatory variables. The estimation results are presented in Table 3. Ignoring the sample selection issue, we rst run a pooled OLS regression using the 25 See Eichengreen and Mody (1998), Edwards (1984). 13

14 bond spread as the dependent variable. The regression results are reported in the second column of Table 3. We then take into account the sample selection issue and estimate the Heckman s model, as speci ed in Equations 1 and 2, using the full sample including the month-country pairs for which there were no bonds issued. The maximum likelihood estimation results are reported in the last two columns of Table 3. Insert Table 3 Here As can be seen, the control variables behave largely as expected. In addition, most of them have the similar coe cients in both the OLS regression and the Heckman s sample selection model. First, the coe cients on AMOUNT and ISSUES are signi cantly positive in the spread equation and signi cantly negative in the issuance equation. As analyzed in Eichengreen and Mody (1998), these variables with the coe cients working in o setting directions can be interpreted as proxies for the supply of bonds. Countries that issued a large number of bonds in a big amount last year have accumulated an unsatis ed appetite for borrowing and tend to supply additional new issues, resulting in an outward shift in the bond supply. Hence a higher borrowing in the past reduces the price of their bonds and increases the spread. Regarding the global economic condition, a higher U.S. real interest rate (USRATE) suppresses the supply of bonds by developing countries due to the higher nancing costs for them, and it has an insigni cant and negative impact on the risk premium. 26 A higher spread on the high-yield corporate bonds (HYD) signi cantly reduces the issuance probability and tends to increase the bond spread. This result con rms the observation that the market requires a similar risk premium on the high-yield corporate bonds and emerging market country bonds. Regarding the issuing country s macroeconomic variables, a high growth rate of per capita GDP (GDPGR) or a high level of GDP per capita (GDPPC) enhances the market demand for international bonds, which increases the issuance probability and decreases the 26 Eichengreen and Moday (1998), Kamin and Keist (1998), and Uribe and Yue (2006) also nd that the US real interest rates reduces the comtemporaneous country spread. 14

15 spread. These variables are proxies for the demand for bonds from international investors because their coe cients work in reinforcing directions in the issuance and spread equations. The debt to GNP ratio (DT2GNP), which is shown to be another proxy for the demand, works in the opposite way. Speci cally, a higher debt to GNP ratio diminishes the market demand, reducing the probability of a bond issue, driving down the price and increasing the spread. The other two indices of a country s external debt (DS2EX and SHORTDT) do not signi cantly a ect the bond spread, but increase the bond issuance probability signi cantly, re ecting a borrowing country s need for liquidity. A higher in ation on the other hand signi cantly increases the bond spread, but does not a ect the likelihood of bond issuance. 27 Lastly, we nd that countries that have a high ratio of current account to GDP (CA2GDP) tends to supply a high volume of bonds. The prices of their issues are thus driven down and the spreads are driven up. Finally, the regional dummies for Africa or Latin America have positive (negative) coe cients in the spread (issuance) equation. The dummy for the January e ect signi cantly reduces the probability of issuing bonds, serving as a valid exclusion variable. The correlation between the error terms in the issuance and spread equations is equal to and signi cantly negative. The negative correlation implies that some unobserved factors that lead to a higher issuance probability also lower the bond spread. Thus these factors should also be interpreted as unobserved determinants of demand. Let us now focus on the impact of the exchange rate regime on the LDC borrowing. We rst discuss the estimation results of the Heckman s sample selection model in the last two columns of Table 3 regarding the role of exchange rate regime. We can see from the table that choosing a less exible exchange rate regime (INT or FIX) decreases the bond issuance probability and increases the bond spread. That is, it is both more di cult and more costly to borrow for countries in intermediate or x regimes, as if these countries were penalized for not choosing a more exible exchange rate regime. Further, the estimated coe cient on FIX is signi cantly higher (lower) than the coe cient on INT in the spread 27 Reinhart and Rogo (2010) document the high correlation between high in ation and the occurrence of debt crisis using data that covers a period of over 200 years. 15

16 (issuance) equation, implying a monotone relation between the exibility of the exchange rate arrangement and the bond spread. The results indicate that a country s exchange rate regime impacts foreign borrowing by shifting the demand curve of its international bonds. Speci cally, the market is less inclined to demand the bonds of a country that has a less exible exchange rate regime. As a result, it is less likely to observe an issue and the corresponding decline in demand increases the spreads on observed issues. The impact of the exchange rate regime is not only statistically signi cant, but also economically signi cant. To see the latter, we quantify the marginal e ect of making a country s exchange rate regime less exible on the bond spread as shown in Equations (4)-(5). In the data, the average spread among the oaters is 319 basis points. From the OLS regression results as in the second column in Table 3, we can see that changing from a oating exchange rate regime to an intermediate one increases the spread by 319*(exp(0.137)-1)=46.7 basis points, and changing from intermediate to xed increases the spread by an additional amount of 92.5 (=319*(exp( )-1)) basis points. The OLS regression ignores the potential sample selection bias. After we take into account the sample selection issue by using the Heckman s model, the marginal e ect from converting a oating exchange rate regime to an intermediate one is 43 basis points, while the marginal e ect from converting the intermediate exchange rate regime further to a xed one increases the spread by an additional amount of 89 basis points. So the direct use of the OLS regression without accounting for the potential sample selection bias tends to slightly overestimate the impact. Using the estimation results of the issuance equation in the last column of Table 3, we compute the marginal e ect from a change in the exchange rate regime on the bond issuance probability. We nd that a country in an intermediate exchange rate regime would be 1.6% less likely to issue a bond if its exchange rate regime had become a xed one, but would be about 3% more likely to issue a bond if it had become a oater. Overall, we nd that countries with a less exible exchange rate regime issue less debt and pay a signi cantly higher bond spread as a result of less demand for the bonds they issued in the international market. Next, we consider the other measure of exchange rate policy in our paper, that is, real 16

17 exchange rate overvaluation. To investigate its impact on the bond issuance and pricing, we estimate the Heckman s model in which we include measures of real exchange rate overvaluation as well as their interaction with the exchange rate regime. As stated in Section 2.1, we use three measures of real exchange rate overvaluation, for which the estimation results are reported in Tables 4A-4C, respectively. Each table contains three columns. We rst use the real exchange rate overvaluation alone as an explanatory variable in the Heckman selection model and report the result in column (I). Column (II) shows that the impact of the real exchange rate overvaluation and the exchange rate regime when both of them are included. Lastly, to better identify their joint impact, we further include the interaction terms between them (Column III), which are the products of the real exchange rate overvaluation and the three exchange rate dummies. By construction these interaction terms sum up to the measure of the real exchange rate overvaluation. Insert Tables 4A-4C Here We nd that the real exchange rate overvaluation signi cantly increases both the bond spread and the bond issuance probability. This result is signi cant and holds for all three measures of real exchange rate overvaluation (ROV1-ROV3). Firstly, an overvalued currency makes a country s export less competitive. Real exchange rate overvaluation is found to be usually associated with low economic growth and loss of government revenue. 28 Hence, the borrowing country may experience greater di culty in servicing its debt. When the gain from correcting the exchange rate misalignment is high and there is little cost associated with default, default probability increases signi cantly. Secondly, a real exchange rate overvaluation is highly likely to be corrected in the form of a currency devaluation or crisis, which increases a country s default risk due to the currency mismatch of the balance sheet. Powell and Sturzenegger (2000) for example nd a strong link between devaluation and default risk. Lastly, a country experiencing real overvaluation tends to borrow more because overvaluation may signal good times (e.g., due to benign real shocks) and developing 28 Prasad et al (2006), Eichengreen (2008), Aghion et al. (2009) study the impact of real exchange rate overvaluation on the economic growth. 17

18 countries typically borrow procyclically. 29 Hence, the country supplies more bonds in the market, which in turn drives down the price and results in a higher bond spread. In sum, a larger real exchange rate overvaluation may increase the bond spread and bond issuance probability through these three channels. Based on columns (I) in Tables 4A-4C, we compute the marginal e ect of real exchange rate overvaluation on the spread as speci ed in Equation (6). We nd that if the real exchange rate becomes more overvalued by one sample standard deviation of the overvaluation measure, the average bond spread increases by 47.5, 27.6, and 20.5 basis points when the real exchange rate overvaluation is measured by ROV1-ROV3, respectively. When both the real exchange rate overvaluation and the exchange rate regimes are used in the regression, from columns (II) of Tables 4A-4C the impacts of the real exchange rate overvaluation and the exchange rate regime remain signi cant. A xed or intermediate exchange rate regime has an independent positive e ect on the bond spread and an independent negative e ect on the bond issuance probability. The coe cients on the regime dummies are slightly lower, but remain to be a monotone function of the exchange rate exibility. Lastly, we investigate the combined e ect of real exchange rate overvaluation and exchange rate regime. From columns (III) of Tables 4A-4C, we nd that among the three interaction terms, ROV F IX has the largest and signi cantly positive coe cients in the issuance and spread equations (except that the coe cient becomes insigni cant in the issuance equation for ROV2). This result suggests that the e ects of the real exchange rate overvaluation tend to be magni ed for countries with a xed exchange rate regime. We can think of two possible explanations for these results. First, when a country has a hard peg or limited exchange rate exibility, the real overvaluation tends to be persistent. 30 As a result, servicing foreign debt can be less costly in domestic currency. Hence, countries with a less exible exchange rate arrangement are more likely to borrow in periods of real 29 Arellano (2008), Aguiar and Gopinath (2006), and Yue (2010) document and show the procyclicality of sovereign borrowing in a Eaton-Gersotivz framework. We thank a referee for suggsting this explanation. 30 Edwards (1988) nds that the autonomous forces that move the real exchange rate back to equilibrium operate very slowly, keeping the country out of equlibrium for a long time. 18

19 overvaluation. The increase in the supply of bonds from countries with a xed exchange rate regime and real overvaluation drives down the bond price and results in a higher bond spread. Second, when a country is in a hard-peg regime, the overvaluation has a larger and more persistent adverse impact on the economy. 31 Debt becomes rapidly unsustainable and the probability of default increases. By contrast, for a oater, owing to the exchange rate exibility, nominal devaluation can greatly help to speed up the real exchange rate realignment. Therefore, real exchange rate overvaluation has the least impact on the bond spread for countries with a free- oating regime. We also assess the economic signi cance of the combined e ect by computing the marginal e ect. For example, when the exchange rate overvaluation is measured using ROV1 (see column (III) of Table 4A), we nd that a one-standard-deviation increase of ROV1 increases the spread by 86 basis points for a country with a xed exchange rate regime, while the same increase of ROV1 only increases the spread by 33 and 29 basis points if the country is in an intermediate and oating exchange rate regime. The same pattern persists when the other two measures (ROV2 and ROV3) are used. In summary, we nd that a real exchange rate overvaluation increases both the bond issuance probability and the bond spread, and such e ect takes place mainly when the country has a xed exchange rate regime. 4 Robustness In this section we summarize the various robustness checks that we run to address some of the concerns that our ndings may give rise to. In particular, we discuss: (a) the robustness of our main ndings by including more macroeconomic control variables and the roles played by these additional variables in a ecting the bond issuance/pricing decisions; (b) the endogeneity problem associated with exchange rate regime and real exchange rate overvaluation. 31 Edwards and Levy-Yeyati (2005) argue that the adjustment in equilibrium real exchange rate upon a real external shock takes longer in countries with a xed exchange rate. 19

20 In the rst robustness check, we add more macroeconomic control variables. We include the debt crisis dummy (DCRISIS), debt rescheduling dummy (DRES), and total reserve to GNI (RES2GNI) as additional regressors. Because of the data availability, there are 40 countries left in the sample when these controls are used. The debt crisis dataset is taken from Reinhart and Rogo (2008). The debt rescheduling dummy, constructed from GDF, is equal to unity if there is a non-zero amount of debt rescheduled for a country and zero otherwise. The results are summarized in columns (I) and (II) in Table 5A. In both the OLS regression and the Heckman s model, the dummy for debt rescheduling (DRES) enters the spread equation signi cantly and positively. It also picks up the e ect from debt crises, making the debt crisis dummy (DCRISIS) insigni cant. Further, although the coe cients of the dummies for both debt rescheduling and debt crises are insigni cant in the issuance equation (see column (II) in Table 5A), they are positive, implying that a country that is in crisis or is experiencing debt rescheduling nds it more di cult to issue new bonds. Moreover, such a country is considered by investors to have higher default probability, and thus the country has to provide a higher spread on its bond if it chooses to issue one. The ratio of total reserve to GNI (RES2GNI) decreases both bond spreads and the likelihood of bond issuance signi cantly. It suggests that a country that has relatively large reserve tends to supply a low volume of bonds and consequently the prices of their issues are driven up and the spreads are driven down. Lastly, from the comparison between Table 5A and Table 3, we still nd that the exchange rate regime a ects the issuing and pricing of international bonds after we control for debt crisis, rescheduling, and the ratio of reserve to GNI. In the second robustness check, we deal with the concerns that some variables, such as the exchange rate regime and real exchange rate overvaluation, may be endogenous. In the previous analysis, we treat all the variables as strictly exogenous for both bond issuance and spread determination. But one concern is that the relation we nd in the data may be caused by the reversed causality. In particular, the choice of exchange rate regime may be a response to a debt crisis or a mechanism to lower borrowing costs. In the subsequent analysis, we deal with this potential endogeneity problem. 20

21 As a rst attempt at the endogeneity issue, we single out observations associated with countries with de facto pegs throughout our sample period (FIXALL) by following Levy- Yeyati and Sturzenegger (2003) and include the dummy, FIXALL, in the OLS regression and the Heckman s model (see columns (I) and (II) in Table 5A). As argued by these authors, since this group of countries corresponds to economies within long-standing currency unions, it seems reasonable to assume that the original regime choice is independent from their growth performance and from the bond issuance/pricing decisions. As can be seen from columns (I) and (II) in Table 5A, the positive (negative) impact of a xed exchange rate regime on the spread (the likelihood of bond issuance) is signi cant for this group of countries relative to the rest of the countries in our sample. This presents initial evidence that the main ndings in our paper are not severely contaminated by the endogeneity problem. We next correct for the endogeneity of the exchange rate regime and real exchange rate overvaluation using a feasible generalized two-stage IV (2SIV) estimator and report the regression results in column (III) of Table 5A and in Table 5B. To correct for the endogeneity of the exchange rate regime, we rst run a multivariate logit model of the exchange rate regimes choice, R, which can take the value of FIX, INT or FLOAT. The multinomial logit model assumes that the probability of one outcome can be expressed as follows: Pr (R = F IX) = Pr (R = INT ) = Pr (R = F LOAT ) = exp (Y 1 ) 1 + exp (Y 1 ) + exp (Y 2 ) exp (Y 2 ) 1 + exp (Y 1 ) + exp (Y 2 ) exp (Y 1 ) + exp (Y 2 ) where Y is the vector of variables used to explain the choice of an exchange rate regime. s are the associated coe cients. The relative probability of choosing FIX (INT) to the FLOAT is exp (Y t 1 ) (exp (Y t 2 )). Similarly, to deal with the potential endogeneity problem associated with real exchange rate overvaluation, we run three OLS regressions on the variables Y to obtain the tted values for ROV1-ROV3, and then we use these tted values as well as those for exchange rate 21

22 regime dummies that are obtained from the above multinomial logit regression to estimate the Heckman s sample selection model. The key issue is to nd suitable instrumental variables for the exchange rate regime and the real overvaluation. For the exchange rate regime, following Levy-Yeyati and Sturzenegger (2003) we use the ratio of the country s GDP over the U.S. GDP (SIZE), the geographical area of the country (AREA), an island dummy (ISLAND), the ratio of reserve to monetary base (RESBASE), and a regional exchange rate indicator (REGEXCH) that is equal to the average exchange rate regime of the country s neighbors de ned as those under the same IMF department. We also control for the potential endogeneity issue for the real overvaluation. As in Prasad et al. (2006) and Eichengreen (2008), we use the share of working-age persons in the population (WORKPOP) and a dummy variable for oil-exporting countries (OILEX) as the instrumental variables for the real exchange rate overvaluation. We use all the exogenous regressors in the baseline model and additional instrumental variables in auxiliary regressions to obtain tted values for the exchange rate regime and overvaluation. The second and third columns in Table 5C report the result of the multinomial logit regression of the exchange rate regime over all the instruments. The coe cients are interpreted as a variation in the relative probability of choosing one regime over a free- oating regime. The last three columns show the estimates of the OLS regressions on the di erent measures of real exchange rate overvaluation. Most variables are highly signi cant and have the expected signs. On the choice of the exchange rate regime, smaller countries tend to be more open and thus are more likely to choose a xed exchange rate regime. A high initial level of reserves helps a country to overcome the fear of oating. Finally, the regional exchange rate indicator may indicate explicit or implicit exchange rate coordination among neighboring countries. 32 Regarding the OLS regressions for the real overvaluation, a higher share of working-age population reduces the likelihood of real overvaluation See Levy-Yeyati and Sturzenegger (2003) for more details on the multinomial logit model for the exchange rate regime. 33 Prasad et al. (2006) argue that a rapidly growing labor force should lead to undervaluation due to the pressure on policy makers to maintain a competitive real exchange rate in order to absorb additional workers into employment. Eichengreen (2008) also documents a similar relation between the share of working age population and real overvaluation. 22