Determinants and Output Growth Effects of Debt Distress

Transcription

1 Determinants and Output Growth Effects of Debt Distress Michael Binder, Sebastian Kripfganz and Tihomir Stucka

2 Determinants and Output Growth Effects of Debt Distress Michael Binder Sebastian Kripfganz Tihomir Stučka December 2014 Preliminary Abstract While the potential perils of debt distress are well understood, considerable uncertainty remains as to the empirical magnitude of the costs of debt distress, particularly with respect to output growth losses induced by debt distress. In this paper, we propose a data-driven yet structured empirical model to characterize the cross-country output growth implications of a country being in debt distress. Our dynamic panel model with debt regime switching and endogenous sample selection in particular allows us to both consider the direct effect of indebtedness on output growth as well as its indirect effect through changes in the distress probability. Our empirical evidence suggests that the debt distress vs. output growth nexus is nonlinear: The direct effect of increases in the debt burden on output growth is either relatively small or insignificant even in episodes of debt distress; an overall significant effect arises through the indirect effect of increases in the debt burden increasing the probability of debt distress. Also importantly, the probability of debt distress is in any case state dependent, varying strongly with a country s institutional and policy track record. Keywords: Debt distress; Output growth; Dynamic panel model; Regime switching; Sample selection; Bias correction; State dependence JEL Classification: C23; H63; O40 We thank Orcun Kaya, participants of the GSEFM Summer Institute and the 19th International Panel Data Conference, as well as seminar participants at Xiamen University, ETH Zürich and Goethe University Frankfurt for helpful comments. Goethe University Frankfurt, mbinder@wiwi.uni-frankfurt.de Goethe University Frankfurt, kripfganz@wiwi.uni-frankfurt.de The World Bank, tstucka@worldbank.org. The views expressed herein are those of the authors and should not be attributed to the World Bank, its Executive Board, or its management. 1

3 1 Introduction The recent acceleration in U.S. federal debt and sovereign debt crises in the Eurozone renewed interest in the relationship between public debt levels, sovereign default, and domestic output growth. This interest goes back to identifying levels of debt that are likely to either start impeding output growth, result in default, or both. The source of sovereign debt default can be idiosyncratic or systemic in nature(kaminsky and Vega-García, 2014), and in conjunction with a wide range of country-specific transmission mechanisms will determine the cost with respect to lost domestic output growth. Country-specific circumstances include, among others, the size of the fiscal deficit, contingent liabilities, overvalued exchange rates, and trade regimes limiting exports. When combined with external shocks such as, for instance, high interest rates, these adverse circumstances can be greatly exacerbated. Public debt default can carry a high cost characterized by many facets of damaged sovereign reputation. Emerging markets hit by a debt crisis experience a drop in the sovereign rating and, thus, a rise in interest spreads (Dell Ariccia et al, 2004). In the 1980s, the first significant post-world War II wave of sovereign debt crises occurred in the emerging markets of Latin America driven by high interest rates and a world recession. Gelos, Sahay, and Sandleris (2011) estimate the loss in market access to have lasted for an average of four years, unable to borrow on normal market terms. During the recent EU crisis, Greece needed four years to regain access to international financial markets after defaulting in 2010, while Ireland and Portugal tapped capital markets after two years. Default or near-default on public debt is likely to extend beyond the sovereign. Debt difficulties of the sovereign often spill over to state owned enterprises and private companies deepening the adverse impact on domestic output growth. Moreover, political reputation can also be damaged beyond repair. The latter occurs with governments forced to undertake politically costly structural reforms, say under IMF programs, often resulting in a change in government (Borensztein and Panizza, 2009). International trade between the debtor and creditor country can also be affected by default dampening positive trade spillovers (Rose, 2005), while Borensztein and Panizza (2010) demonstrate that sovereign defaults hurt disproportionally more export-oriented industries, therefore affecting adversely the source of important positive externalities in an economy. Finally, if spillovers from difficulties with external public debt to the domestic financial system are pronounced, then output growth losses are exacerbated by domestic credit rationing, among others, as demonstrated by the recent crisis in European countries. The aforementioned transmission channels are likely to cause a drop in domestic capital formation, stifling further future 2

4 output growth. After the sovereign defaults in Latin America, the focus of debt problems shifted to Africa. The debt crisis in low income countries arose from a combination of policy actions in debtor countries, macroeconomic shocks in the world economy, and an acceleration of bank lending during More precisely, tightened monetary policy in industrial countries aimed at combating inflationary pressures led to a world-wide recession. The latter caused a drop in export prices and terms of trade in developing countries, which led to the reversal of the interest rate-export growth relationship in the 1980s. The resulting rise in debt burden indicators dried up private lending (Cline, 1990). Official lenders responded to this debt overhang with debt relief efforts in the late 1990s and in When analyzing debt crises, it becomes critical to differentiate the composition of public debt in middle and low income countries. Debt portfolios in middle income countries are associated with predominantly held private debt and limited recourse to official lending. Debt portfolios in low income countries, in contrast, rely predominantly on official debt, which exhibits a much higher grant element compared to commercial debt as reflected in a below-market interest rate, long grace period and long maturities, all best captured through debt levels expressed in present value terms. However, after receiving debt relief from creditors in the late 1990s and in 2005 in exchange for good policies, some low income countries have accessed international capital markets, benefitting from the search for yield in an environment of quantitative easing. Others drew on new commercial credit lines, available largely from China. Furthermore, capacity constraints in low income countries imply differences in managing the debt portfolio when compared to emerging markets. Under such circumstances, debt incidents can be the result of maturity concentrations, for instance, and thus a feature of portfolio composition rather than public debt level. Intuitively, liquidity-driven crises do not occur during normal times, but are rather generated by shocks, at which point the sovereign s borrowing constraint becomes binding (Mendoza, 2002). Such shocks, if they affect international liquidity reminiscent of the recent financial crisis, can affect non-defaulters by preventing their access to capital markets. Such defaults, driven by systemic shocks, in principle have different consequences compared to idiosyncratic shocks. The latter corresponds better to solvency crises, which are typically self-fulfilling in nature as debt stocks reach levels that would be difficult to service even during normal times. This debt overhang problem lies at the core of most debt crises (Kaminsky and Vega-García, 2014). The global financial crisis and the ensuing great contraction featuring critical public debt levels for advanced and developing economies alike, pose new questions for debt sus- 1 See the contributions in Sachs (1989). 3

5 tainability analysis. What are the quantitatively important robust factors driving debt intolerance (Reinhart, Rogoff, and Savastano, 2003)? How do various categories of debt, such as external public debt, domestic public debt, and private debt, interact in driving probabilities of sovereign debt distress? Are there threshold levels of public debt that are critical for a country s output growth performance (Reinhart and Rogoff, 2010; Caner, Grennes, and Köhler-Geib, 2011), and, if so, are these threshold levels in turn dependent on a country s institutional development, public policy record, and/or its financial development? In this paper, we first set out to model empirically the dynamics of debt accumulation to quantitatively characterize the determinants of debt distress. This should be useful in helping to enhance early warning systems for debt distress. To this end, we model in the first stage of our approach the determinants of debt distress using a state-dependent dynamic panel Probit model. We use the latter to quantify the impact of factors reflecting debt burden indicators, macroeconomic fundamentals as well as the institutional and policy track record on the probability that a country is facing an episode of debt distress. Following the methodology of Kraay and Nehru (2006) and Cohen and Valadier (2011), we initially define external public debt distress as default and near-default events related to the stock of arrears on public and publicly guaranteed external debt, the level of IMF financing, and debt restructuring. Our analysis also accounts for sample selection issues when modeling the short- and long-run output growth implications of a country in debt distress. In the second stage of our empirical analysis, the first-stage dynamic panel Probit model is used as a selection equation for the analysis of the output growth implications of a country in debt distress. In other words, the first-stage selection equation is complemented by a dynamic panel model with state-dependent parameters that captures short- and long-run output growth under the various debt regimes. The first- and second-stage equations together form a dynamic panel output growth model with debt regime switching and endogenous sample selection. With the switching regression setup, we consider both the direct effect of indebtedness on output growth and its indirect effect through changes in the probability of debt distress. In addition, our model allows for a rich set of transmission channels that may help explain how country-specific characteristics affect probabilities of debt distress as well as the output growth implications of debt distress. Our estimation methodology corrects for sample selection and incidental parameter biases and is based on Fernández-Val and Vella (2011). We estimate our model using a relatively large cross-country data set involving public debt records for 122 countries. As opposed to the previous literature, our dynamic switching regression model al- 4

6 lows in a structural way for nonlinearity of the output growth effects of the debt burden indicator. The previous literature usually incorporates these nonlinear effects by estimating single-equation models with polynomial, spline, or threshold specifications. Using a sample of 44 countries over a span of two centuries, Reinhart and Rogoff (2010) suggest that at public debt levels of 90 percent of GDP and above output growth drops sharply from between an average of 3 to 4 percent to 0 percent, on average. Moreover, for emerging markets, external public debt at 60 percent is associated with adverse output growth effects. Herndon, Ash, and Pollin (2013), however, identified estimation errors, omission of available data, and random weighting of data resulting in biased outcomes of the Reinhart and Rogoff (2010) paper. They further demonstrate that once advanced countries hit the 90 percent of GDP debt level, output growth rates deteriorate only modestly. Other recent investigations of the debt-growth nexus include Baum, Checherita-Westphal, and Rother (2013), who employ a dynamic threshold panel model on data for 12 euro area countries. They find a positive short run impact of public debt on economic growth below a debt-to-gdp threshold of 67 percent, and a negative effect above a second threshold of 95 percent. Similarly, Cecchetti, Mohanty, and Zampolli (2011) find a threshold for government debt of 85 percent of GDP after which the output growth effect turns bad in their sample of 18 OECD countries. At this level, a further 10 percentage point increase in the ratio of public debt to GDP is associated with a basis point reduction in subsequent average annual growth. They discriminate between different debt components by considering also corporate and household debt. Pattillo, Poirson, and Ricci (2011) estimate growth models with quadratic debt terms or spline expressions for 93 developing countries and find a negative marginal effect of external debt on GDP growth already from about 20 percent of GDP onwards. A similar debt overhang threshold is found by Cordella, Ricci, and Ruiz-Arranz (2010) who in addition estimate a debt irrelevance threshold around 70 to 80 percent of GDP for their sample of 79 developing countries. The authors find also an insignificant debt-growth relationship at very low and very high levels of debt. Our estimation results confirm some findings previously reported in the literature, but also add new insights. We confirm previous results that inertia is the single, most important characteristic of debt distress, while the institutional and policy track record also seems to contribute significantly to explaining cross-country differences in debt distress probabilities. Countries featuring a relatively poor institutional and policy track record experience notably higher probabilities of entering into debt distress for given debt levels. In the context of the interaction between debt regimes and output growth, we also find empirical support that higher values of debt are associated with lower output growth and 5

7 that this relationship is nonlinear. In contrast to the previous literature, we establish that the direct effect of increases in the debt burden on output growth is relatively small or insignificant even in episodes of debt distress. An overall significant impact arises primarily through the indirect effect of increases in the debt burden increasing the probability of debt distress. Also, these effects are state dependent, varying in strength across the complete spectrum of institutional and policy track records. Even under a poor record, the effects are much smaller than suggested by Reinhart and Rogoff (2010), with implications for the current policy debate. The paper proceeds as follows. In the next section, we define debt distress events. In Section 3, we set out the econometric framework for our empirical analysis. We describe in detail the dynamic switching regression panel data model, the necessary bias corrections, and the calculation of the conditional marginal effects of the debt-to-gdp ratio on output growth. Thereafter, in Section 4, we describe the data that we use in the empirical part of this paper, and we discuss the estimation results. In Section 5, we conclude. 2 Debt Distress Events Building on the methodology of McFadden, Eckaus, Feder, Hajivassiliou, and O Connell (1985) and Manasse, Roubini, and Schimmelpfennig (2003), 2 Kraay and Nehru (2006) define episodes of debt distress as default and near default events. More precisely, debt distress occurs when: (i) countries accumulate a stock of arrears on public and publicly guaranteed (PPG) medium and long term external debt exceeding 5 percent of total medium and long term external debt outstanding, (ii) countries use IMF financing (in the form of commitments rather than disbursements) in excess of 50 percent of their quota, or (iii) receive Paris Club rescheduling or debt reduction without adjusting for HIPC Completion Point. Kraay and Nehru (2006) consider only prolonged distress episodes of at least three years that are, in addition, not preceded by another distress period in the previous three years. This aims at the identification of distinct episodes that are not just results of sporadic fluctuations in the above indicators. As a control group, Kraay and Nehru (2006) define normal times as periods of five consecutive years without any debt distress incidence. This definition treats distress and non-distress periods in an asymmetric way, and many country-year observations cannot be used in the empirical analysis because they 2 McFadden et al. (1985) consider repayment problems in the form of arrears, higher-tranche IMF support, or rescheduling requests. Manasse et al. (2003) define a country to be in a debt crisis if it is classified as being in default by Standard & Poor s or if it receives a large nonconcessional IMF loan defined as access in excess of 100 percent of quota. The same definition is used by Bandiera, Cuaresma, and Vincelette (2011). 6

8 are neither classified as one nor as the other. Cohen and Valadier(2011) refine the Kraay and Nehru(2006) definition. With respect to IMF financing, they consider disbursements rather than commitments in excess of 50 percent of a country s quota. With this definition, the exceptional IMF financial support can be interpreted as a substitute for falling into arrears. Furthermore, Paris Club debt relief for countries that reach the completion point of the Heavily Indebted Poor Countries (HIPC) initiative is not considered as a distress event. To treat distress and normal times symmetrically, they define the latter as non-crisis years preceded by three more years without crisis. We follow the definition of Cohen and Valadier (2011) for single year debt distress events, but do not explicitly try to identify prolonged distress episodes. 3 Our strategy rather is to control within our econometric modeling framework (as described in Section 3) for ongoing distress episodes and allow for different impacts of the distress determinants on the probabilities of entering into and exiting from a crisis. Moreover, the leeway of a government might still be narrowed when the above debt distress criteria are already back to normal levels. Therefore, for our empirical analysis we consider a country to be in debt distress if it faces a distress event according to the above definition in the current or previous year. This definition also captures cases where a distress event occurs just at the end of a calender year and thus distorts the economy primarily in the subsequent year. Figure 1 illustrates the construction of the debt distress index for the case of Kenya. In the mid-1970s, Kenya received balance of payments support in the form of IMF commitments exceeding 50% of its quota. 4 However, actual disbursements still remained below this threshold. They reached the threshold triggering a debt distress event for the first time in 1980, remaining significant for four years. After falling back to zero in 1984, Kenya had to rely on IMF assistance again just one year later. Due to the one-year inertia in the construction of our index, we therefore classify the whole time span from 1980 to 1886 as a period of continuous debt distress. The next four years were characterized by a gradual rise in the total arrears on public and publicly guaranteed external debt that finally exceeded the threshold value of 5% in The arrears rose further until 1993 before Paris Club reschedulings came into effect. This second period of prolongued debt distress lasted until The debt renegotiations did not have a long-lasting effect such that additional Paris Club debt relief became necessary in 2000 and 2004, definining the third debt distress episode that was still about to continue at the end of our sample. 3 McFadden et al. (1985) and Bandiera et al. (2011) also use single year observations in their analysis. 4 Compare Kraay and Nehru (2006). 7

9 3 Econometric Methodology 3.1 Dynamic Switching Regression Panel Data Model We consider the following dynamic panel data model with fixed effects and sample separation. Cross-sectional units are indexed by i = 1,2,...,N, and time periods by t = 1,2,...,T. The sample separation is determined by the debt distress selection equation: d it = I{d it +ǫ it 0}, (1) where I is an indicator function returning 1 if the condition is satisfied and 0 otherwise, and d it = θd i,t 1 +z itβ +α 1i. (2) This specification is dynamic as in McFadden et al.(1985) and Celasun and Harms(2011). The regressors in z it are allowed to have a different impact on the entry and exit probabilities of a crisis, respectively, because their marginal effects depend on the realization of d i,t 1. 5 We additionally include country-specific fixed effects α 1i. The debt distress effects equation is characterized by two regimes, being in distress (d it = 1) or not (d it = 0): ρ (1) y i,t 1 +x it y it = γ(1) +α (1) 2i +u 1it, d it = 1. (3) ρ (2) y i,t 1 +x it γ(2) +α (2) 2i +u 2it, d it = 0 We assume that (ǫ it,u 1it,u 2it ) are independent and identically trivariate normally distributed with zero mean and variance-covariance matrix 1 ζ 1 ζ 2 Σ = ζ 1 σ1 2 σ 12. ζ 2 σ 12 σ 2 2 Equations(3) and(1) form a switching regression model with known sample separation and y it observed in both regimes. Sample separation is endogenous if the covariance between the disturbance term of the selection equation and the errors of the effects equations are nonzero, ζ 1 0 or ζ The regressors x it and z it are predetermined with respect to the disturbances, and the unobserved unit-specific effects α 1, α (1) 2, and α(2) 2 are fixed without any distributional assumption. The initial observations y i0 and the initial states 5 Manasse et al. (2003) include interaction terms d i,t 1z it as part of d it. However, this is not necessary to obtain state-dependent marginal effects because of the nonlinearity of the Probit model. 6 See Maddala (1986). 8

10 d i0 are observed. We apply the two-stage estimation procedure developed by Heckman (1976, 1979). Let d i = (d i1,d i2,...,d it ), Z i = (z i1,z i2,...,z it ), and d (t) i = (d i0,d i1,...,d it ). The partial likelihood function of unit i for the first stage, the selection equation (1), conditional on Z i and the initial state d i0 is given by: L d i(θ,β,α 1i d i,z i,d i0 ) = T f t=1 with the density function of the Probit model f ( d it d (t 1) i,z it ;θ,β,α 1i ), (4) ( d it d (t 1) i,z it ;θ,β,α 1i ) = Φ d it it (1 Φ it ) 1 d it, (5) where Φ it is the cumulative distribution function of the standard normal distribution evaluated at d it. The maximum likelihood estimates are obtained as (ˆθ, ˆβ,{ˆα 1i } N i=1) = arg max θ,β, {α 1i } N i=1 N i=1 t=1 T [d it lnφ it +(1 d it )ln(1 Φ it )]. (6) The endogenous sample separation would lead to an omitted variable bias in the least squares estimation of the debt distress effects equation (3). The properties of the multivariate normal distribution imply where E[u jit d it,d i,t 1,z it,α 1i ] = ζ j λ it, j = 1,2, λ it = (d it Φ it )φ it Φ it (1 Φ it ) is the inverse Mills ratio, and φ it is the probability density function of a standard normal random variable evaluated at d it. Consequently, we can account for the endogenous regime selection by introducing λ it as an additional control variable in equation (3): 7 ρ (1) y i,t 1 +x it y it = γ(1) +ζ 1 λ it +α (1) 2i +e 1it, d it = 1, (8) ρ (2) y i,t 1 +x it γ(2) +ζ 2 λ it +α (2) 2i +e 2it, d it = 0 such that E[e jit y i,t 1,x it,λ it,d it,α (j) 2i ] = 0, j = 1,2. The lagged distress indicator d i,t 1 7 In our estimations we substract y i,t 1 from both sides of equation (8) to obtain a model in the change of y it. This transformation does not affect any of the arguments in this section. (7) 9

11 and the interaction terms d i,t 1 z it in the debt distress selection equation provide valid exclusion restrictions if their coefficients are significantly different from zero. Let y i = (y i1,y i2,...,y it ), X i = (x i1,x i2,...,x it ), Λ i = (λ i1,λ i2,...,λ it ), and y (t) i = (y i0,y i1,...,y it ). Also stack α 2i = (α (1) 2i,α(2) 2i ), and ψ = (ψ (1),ψ (2) ) with ψ (j) = (ρ (j),γ (j),ζ j,σ j ), j = 1,2. Then, the partial likelihood function of unit i conditional on X i, Λ i, the observed states d i, and the initial observation y i0 is given by: L y i (ψ,α 2i y i,x i,λ i,d i,y i0 ) = = T t=1 T t=1 ( ) g y it y (t 1) i,x it,λ it,d it ;ψ,α 2i ( )] [g 1 y it y (t 1) i,x it,λ it ;ψ (1),α (1) dit 2i ( )] [g 2 y it y (t 1) i,x it,λ it ;ψ (2),α (2) 1 dit 2i, (9) whereg isthedensityfory it, andsubscripts1and2denotedensitiesgiventhetworegimes. Consequently, we can identify the parameters for both regimes separately as long as there are no cross-regime restrictions: (ˆψ(1) ),{ˆα (1) 2i }N i=1 = max ψ (1), {α (1) 2i }N i=1 (ˆψ(2) ),{ˆα (2) 2i }N i=1 = max ψ (2), {α (2) 2i }N i=1 N T i=1 t=1 N i=1 t=1 ( ) d it lng 1 y it y (t 1) i,x it,λ it ;ψ (1),α (1) 2i, (10) T ( ) (1 d it )lng 2 y it y (t 1) i,x it,λ it ;ψ (2),α (2) 2i, (11) with ( lng j y it y (t 1) i,x it,λ it ;ψ (j)) = 1 2 ln(2πσ j) 1 2σj 2 e 2 jit, j = 1,2, (12) and e jit defined in equation (8). Furthermore, σ 12 is not identified in the model because we observe only either of the two regimes at a time Incidental Parameters Problem and Bias Correction When the estimation of the parameters of interest cannot be separated from the fixed effects, the presence of the latter leads to inconsistent estimates if the time dimension T is held fixed while the number of units N increases. This is the familiar incidental parameters problem discussed first by Neyman and Scott(1948). It is particularly relevant 8 See also Maddala (1986), Section 2. 10

12 in nonlinear models because there usually exists no transformation that removes the fixed effects. Hahn and Newey (2004) show that the resulting bias also does not vanish when N and T grow at the same rate towards infinity, and discuss panel jackknife and analytical bias correction. Fernández-Val (2009) provides a large-t expansion of the bias for index coefficients and marginal effects in panel binary choice models. We apply his procedure to our maximum likelihood estimates of the debt distress selection equation. The incidental parameters problem also arises in the estimation of the second stage, the debt distress effects equation. One source of bias is again the presence of the fixed effects, α (1) 2i and α (2) 2i. In linear models, one possibility to deal with unobserved unitspecific heterogeneity is to transform the model such that the fixed effects are wiped out, for example by first differencing. However, in our switching regimes model first differencing will only remove the fixed effects when they are restricted to be the same in both regimes or when regime shifts do not occur within units over time. A second source of bias is due to the inclusion of the inverse Mills ratio as a control variable for the endogenous sample separation. As can be seen from expression (7) this control variable is a function of the first-stage incidental parameters α 1i and needs to be replaced by an estimate ˆλ it = λ(d it,d i,t 1,z it ;ˆθ, ˆβ, ˆα 1i ). Therefore, the estimation error in the first-stage fixed effects transmits to an error in the estimation of λ it. We apply the analytical bias correction for such two-stage fixed effects panel data estimators developed by Fernández-Val and Vella (2011). 3.3 Marginal Effects of Debt on Output Growth In our empirical analysis, the dependent variable y it of the debt distress effects equation is the natural logarithm of real GDP per capita. We can easily obtain a representation of equation (8) with the growth rate of real GDP per capita as the dependent variable by subtracting y i,t 1 on both sides: (ρ (1) 1)y i,t 1 +x it y it = γ(1) +ζ 1 λ it +α (1) 2i +u 1it, d it = 1. (13) (ρ (2) 1)y i,t 1 +x it γ(2) +ζ 2 λ it +α (2) 2i +u 2it, d it = 0 Our key explanatory variable is the debt-to-gdp ratio that enters both the selection and the effects equation. Without loss of generality, let x 1it = z 1it be the debt-to-gdp ratio. 11

13 We are interested in the marginal effect E[ y it I it ] x 1it = 1 d it =0 [ E[ yit (d it ) I it ] x 1it f(d it )+E[ y it (d it ) I it ] f(d ] it) z 1it = γ (1) 1 Φ it +γ (2) 1 (1 Φ it)+e[ y it (d it = 1) y it (d it = 0) I it ]β 1 φ it, (14) where I it = {y i,t 1,x it,z it,d i,t 1 }, and β 1 φ it is the marginal effect of debt on the distress probability. We divide the total marginal effect into a direct and an indirect marginal effect. The direct effect, γ (1) 1 Φ it+γ (2) 1 (1 Φ it), is the probability weighted marginal effect for the two debt regimes. The indirect effect, E[ y it (d it = 1) y it (d it = 0) I it ]β 1 φ it, is the growth differential between the two regimes multiplied by the marginal effect on the distress probability. Due to the presence of Φ it and φ it the marginal effects are nonlinear in the regressor variables x it and z it. That allows us to analyze the shape of the marginal effect over a grid of values for our variables of major interest. For instance, we can calculate conditional average marginal effects for given values of the debt-to-gdp ratio and the debt distress indicator in the previous period by evaluating the marginal effect at these values and averaging over the actual observations of the remaining variables in the sample. State-dependent long-run effects of an increase in the debt-to-gdp ratio on the GDP per capita level can be obtained as E[y i d i,i i ] x 1i γ (1) 1 = /(1 ρ(1) ), d i = 1, (15) γ (2) 1 /(1 ρ(2) ), d i = 0 where an asterisk denotes long-run equilibrium values. An overall direct long-run effect results from weighting the two state-dependent effects with the respective long-run probabilities. The long-run probability for the distress regime is given by Φ i = Φ i (d i, 1 = 1) 1+Φ i (d i, 1 = 1) Φ i(d (16) i, 1 = 0). The denominator results from ruling out regime switches in the long-run. An indirect long-run effect can be computed as E[y i (d i = 1) y i (d i = 0) I i ] Φ i / z 1i, where Φ i z 1i = β 1 [1 Φ i (d i, 1 = 0)]φ i(d i, 1 = 1)+Φ i(d i, 1 = 1)φ i(d i, 1 = 0) [1+Φ i (d i, 1 = 1) Φ i(d i, 1 = 0)]2. (17) A total long-run effect is again the sum of both. 12

14 4 Empirical Analysis 4.1 Data Description Data on arrears, the external public and publicly guaranteed (PPG) debt stock, the external private non-guaranteed (PNG) debt stock, current debt service, and the Country Policy and Institutional Assessment (CPIA) index are from the World Bank s (WB) World Development Indicators (WDI) database. The present value calculations of external PPG debt using currency-specific commercial interest reference rates (CIRRs) taken from the OECD database were provided by the Debtor Reporting System database and follow the methodology outlined in Dikhanov (2005). External PPG debt in present value terms captures the financing terms in low income countries which have access to lending with a large grant element. It can therefore be argued that this represents a superior debt burden measure when compared to nominal PPG debt, at least in the context of low income countries. Data on Paris Club negotiations are from the Paris Club website and complemented with data provided by the IMF. Data on IMF financing commitments and disbursements are from the International Financial Statistics (IFS) database. We incorporate an extended set of the domestic public debt data compiled by Abbas and Christensen (2010). Data on GDP per capita, the investment share, and the population growth rate are taken from the Penn World Table 7.0. The dependent variable, y it, in the distress effects model (3) is the growth rate of real GDP per capita, measured as the difference in the natural logarithm of real GDP per capita from period t 1 to t. The explanatory variables, x it, are the ratio of the present value of external PPG debt to GDP, the CPIA index, the investment-to-gdp ratio, and the population growth rate. We further include a deterministic time trend. In additional specifications, we add external PNG debt and domestic public debt as a ratio of GDP to the list of regressors, as well as interaction terms to account for potential nonlinearities in the direct marginal effects. The debt variables are measured in period t 1 to accomodate for potential endogeneity issues. This issue does not arise for the CPIA index which is constructed as an assessment of the outcomes in the previous year. Thus, together with the one-period lagged natural logarithm of real GDP per capita, y i,t 1, these variables represent the state of the economy at the beginning of a year. The investment share and the population growth rate are standard variables in the empirical growth literature that capture the contemporaneous change in the productive input factors capital and labor. The explanatory variables z it in the distress selection model (1) and (2) contain the whole set of explanatory variables from the output growth model. 9 The measurement 9 Interaction terms are not included in the selection equation because interaction effects between the 13

15 of is the same as in the effects equation. 10 Furthermore, the one-period lagged output growth rate enters as a covariate to allow for dynamic feedbacks from the effects equation. Together with the lagged distress indicator this variable serves as an exclusion restriction. We restrict the analysis to developing and emerging market economies. In the main part of the analysis we consider countries for which we have data available on all variables for at least 37 consecutive periods within the time span from 1970 to With this restriction we obtain a slightly unbalanced panel data set for 72 countries. The sample shrinks to 53 countries when we add domestic public debt as a regressor variable. 11 Without this restriction we have data available for 122 countries. Tables 3 and 4 provide summary statistics for the variables used in the analysis for these samples which confirm that the reduced samples have similar properties compared to the corresponding full sample. We also analyze subsamples for low and middle income countries according to their classification by the World Bank. 12 Table 3 reveals that the low income countries have a higher unconditional distress probability (51 percent) compared to middle income countries (37 percent), and a smaller real output growth averaging almost an entire percentage point annually. Both the higher distress probability and smaller real output growth rates are also associated with much smaller CPIA ratings, on average, in low income compared to middle income countries. Low income countries accumulated a higher amount of external public debt relative to GDP in present value as well as nominal terms. However, the difference in the nominal and present value of external debt is 12 percentage points in low income countries, half the size compared to the difference for middle income countries equivalent to 22 percentage points. This reflects the more benign repayment profile in low income countries resulting from more favorable financing terms obtained from official creditors. Debt distress is related to lower output growth rates and higher debt-to-gdp ratios (Table 5). The mean growth rate is seven times smaller in distress, and amounts to 0.3 percent only. Meanwhile, the median growth rate is three times smaller in distress (0.9 percent), emphasizing the sizeable variance in the data. Depending on the measure taken, external PPG debt-to-gdp ratios are between 30 and 40 percentage points higher in the variables arise already from the nonlinearity of the Probit model. 10 One-periodlagsoftheexplanatoryvariablestocircumventpotentialendogeneityproblemsarealsoused by McFadden et al. (1985), Manasse et al. (2003), Kraay and Nehru (2006), and Bandiera et al. (2011). Cohen and Valadier (2011) use two-year lags. 11 Table 1 lists the countries and available years. Note that the dynamic nature of our model reduces the number of effectively available observations by one period per country. The limitation to countries with many consecutive observations helps to reduce dynamic panel data biases that arise in small samples. Also, the efficiency of the bias corrections is limited in strongly unbalanced panels. 12 A country is included in the low or middle income subsample over the whole time span if it was classified as a respective country in at least one the years. Therefore, the two subsamples slightly overlap. 14

16 distress regime. When output growth rates are compared across four different debt groups, along the lines of Reinhart and Rogoff (2010), we find a negative growth impact after the first debt-to-gdp theshold set at 30 percent (Table 6). The average and median GDP growth rates decline steadily with higher debt groups, in contrast to Reinhart and Rogoff (2010), with the biggest drop in the median output growth rate occuring beyond 90 percent of GDP in nominal terms, and 60 percent in present value terms. Interestingly, in nominal terms the average GDP growth rate remains broadly constant between the second (30, 60) and third (60, 90) debt group, and falls strongly thereafter. Furthermore, in line with our working hypothesis, the unconditional distress probability increases steadily with higher debt groups no matter how we measure debt. The unconditional probability of distress doubles from 25 to 51 percent after countries reach a nominal debt level of 30 percent of GDP. Above the last threshold of 90 percent of GDP in present value terms only three observations in our sample are not distress events. According to the taxonomy applied in the joint IMF-WB debt sustainability framework, we group the CPIA data into weak performers (CPIA 3.25), medium performers (3.25 < CPIA < 3.75), and strong performers (CPIA 3.75). The correlation between output growth and policy performance is highlighted in Table 7, and is distinctly different for each group. Weak policy performance is associated with much lower output growth rates that are on average below 1 percent per annum. The difference in growth performance seems to be somewhat asymmetric between the different CPIA groups, with the increase in the average growth performance moderating as we move from medium to high performers. For external PPG debt, the same relationship does not hold, in line with our expectations. Better policy performers should be able to carry more debt, which we can see for the median debt levels when we shift from weak to medium performers. A shift from medium to high performers does not, however, translate into an increase in the debt level. Together with the output growth description, this seems to suggest that the difference between weak and medium performers is more critical when it comes to economic performance, compared to the value of moving from medium to high performer. This said, when measured against the unconditional probability of distress, it matters whether a country is a weak, medium, or high performer. The distress probability drops sizably between the last two groups. In the following we report the estimation results based on the present value calculations. Using nominal debt instead does not qualitatively alter the results. Quantitatively, the effects become slightly smaller. 15

17 4.2 Debt Distress Determinants The estimation results are presented in Appendix A. Table 8 shows the average marginal effects on the debt distress probability for the samples with many consecutive observations, and Table 9 for the extended samples that also include countries with shorter time series. External PPG debt relative to GDP in present value terms is a significant predictor of debt distress. An increase in the debt-to-gdp ratio of 10 percentage points adds on average 2 percentage points to the probability of experiencing debt distress in low income countries and 3 percentage points in middle income countries. Domestic public debt is also significant in explaining debt distress. For the sample as a whole, the magnitude of both external and domestic debt is close to 2 percentage points. In contrast to the findings of Celasun and Harms (2011), adding external private debt to the model does not help to explain the observed debt distress events among the low and middle income countries in our sample. Whether in the context of a low or middle income country, better quality of policymaking and of institutions matters. A higher CPIA index reduces significantly the distress probability. More precisely, increasing the CPIA rating by one notch reduces the distress probability by an average of 3 percentage points in low income countries, but has much less significance in the context of middle income countries. The latter possibly goes back to the notion mentioned above that shifts from weak to medium performers matter much more than from medium to high performers, two groups in which middle income countries are most likely to reside. The magnitude of the CPIA effect is robust in the presence of other explanatory variables such as private external or domestic public debt. At the same time, the initial level of GDP per capita does not have a significant effect suggesting that it is the institutional development rather than the aggregate economic development that matters for the distress probability. However, a drop in output growth by one percentage point as a proxy of negative macroeconomic shocks significantly increases the debt distress probability by about 2 percentage points in low income countries and 3 percentage points in middle income countries. The investment-to-gdp ratio can be seen as a measure of productive spending of the available resources. Not surprisingly, a larger investment share significantly reduces the probability of facing a debt distress event as it enhances the repayment capacity. Depending on the sample, the magnitude of this effect is almost as large as that of an increase in external PPG debt with opposite sign, or even exceeds it. One possible argumentation is that additional debt does not raise the distress probability if borrowed funds are invested in productive capital projects. 16

18 State dependence matters for the determination of debt distress. The positive and significant average marginal effect of the lagged distress indicator accounts for the inertia of debt distress, although this effect captures in part the construction of our indicator. 13 In this respect, we are able to distinguish between the entry and exit probabilities of debt distress. On average, the probability of staying in distress is about 50 to 55 percentage points higher than the probability of entering into debt distress. An important explanation is that IMF program assistance and debt rescheduling plans typically last longer than a single year. In addition, actual default or the perception of an increased default risk limits the access to credit markets and further tightens the refinancing problems, in particualar in middle income countries with larger private debt exposure. The distress probability over the entire range of external PPG debt, conditional on being already in distress or not is depicted in Figure 2, and the slope of the curve describes the marginal effect. 14 Countries that are not in distress in period t 1 face a probability of entering into distress of about 9 percent for very low debt levels. In the debt-to-gdp range between 60 and 90 percent the probability of debt distress increases from 31 to 49 percent. At the upper limit, a debt level of 120 percent of GDP, our model predicts an average distress probability of 65 percent. The distress probability of countries which are already in distress does not decline below 63 percent even if their debt-to-gdp ratio is close to zero. The level of debt still matters for these countries as the predicted probability rises further to 76 percent at the first threshold of a debt-to-gdp ratio of 30 percent, and to 89 percent for high-debt countries. The average distress probability is clearly state dependent, varying along the whole spectrum of a country s institutional quality and policy track record(figure 3). Comparing countries with the highest and the lowest policy and institutional quality rating, we observe significantly different entry probabilities into distress over the whole range of the debt ratios under consideration. At 30 percent, countries with good institutions face an average distress probability of only 7 percent as opposed to 31 percent in the case of weak policies and institutions. The gap widens further when the second debt threshold is approached at 60 percent of GDP. At this point, the probabilities are 14 and 49 percent, respectively. If a country is already in distress, good governance improves considerably the odds of returning to normal times. At the lower end of the debt ratio range, the difference between the highest and lowest CPIA category is 35 percentage points (42 versus 77 percent). However, 13 This effect is still strongly positive and significant if we consider only current period distress events for the dependent variable. In the calculation of the marginal effects we take into account that the lagged distress indicator is not a continuous but a binary variable. 14 The bootstrap confidence intervals are obtained by resampling the observations with country clusters. We perform 2000 replications. 17

19 accumulating debt is more costly for good governance countries in this case because the distress probability immediately starts rising while it remains virtually constant, though on an already high level, for countries with weak policies. The difference between the two groups remains significant up to a debt ratio of 45 percent of GDP. 4.3 Output Growth Effects of Debt Distress In situations of debt distress output growth is on average two percentage points lower and external PPG debt two to three times larger than in normal times (Table 5). Between external debt and output growth we observe a negative relationship not only unconditionally (Table 6) but also conditional on the debt distress regime and additional covariates. In the baseline case, adding 10 percentage points of external PPG debt relative to GDP reduces annual output growth by 0.1 percentage points in the distress regime and by 0.3 percentage points in normal times, significant at the 10 percent level (Table 10). As long as higher external PPG debt does not lead to a regime change, it may still increase the perceived risk of future debt distress which in turn deteriorates the investment and consumption climate. This effect remains unchanged when we add external private debt and an interaction term of the two debt categories as additional regressors (Table 11). However, while an increase in the external private debt-to-gdp ratio does not affect the probability of debt distress, it enhances output growth. An explanation consistent with both observations is that private debt is associated with direct capital investment, while additional public debt is not uncommonly used for populist spending, in particular in the run-up to elections. The effect of external private debt is more pronounced in the distress regime for low income countries which reflects the mechanism that private activity can step in as a growth driver when the government is not in a position to stimulate the economy. In middle income countries, the output growth effect of private debt is of a similar magnitude in both regimes but significant at the 10 percent level in normal times only (Table 12). The third debt component, domestic public debt, does not help to explain the output growth differences among the two regimes (Table 13). Among the other regressors, the effect of the quality of policy and institutions is significantly positive and slightly larger in the debt distress regime. While strong governance generally help to boost economic growth it seems to be even more important when a country faces a debt crisis. An improvement in the CPIA rating by one notch adds 1.5 percentage points to output growth in the distress regime, and 1 percentage point otherwise. The initial GDP level has a significantly negative effect on output growth which is consistent with the conditional convergence hypothesis in the economic growth literature. 18

20 The coefficient of the investment share is significant and positive, as expected, while the coefficient of the population growth rate is insignificant in most specifications. Only in low income countries, a significantly negative link between the output and population growth rates is observed. The literature on the debt-growth nexus emphasizes that this relationship is nonlinear. Conditional on not being in debt distress, we find evidence that the direct marginal effect of external PPG debt accumulation is significantly negative for initially low debt-to-gdp ratios but vanishes with higher initial debt ratios (second specification in Table 10). When starting at zero debt, the first ten percentage points of accumulated debt are associated with a cost of 1.7 percentage points in annual output growth. 15 However, allowing for such a polynomial specification increases the estimation uncertainty, and the output growth effect of external PPG debt turns insignificant in the debt distress regime. The direct effect of an increase in the debt-to-gdp ratio on output growth conditional on the regime reveals only part of the picture. More external PPG debt also increases the chance of falling into the debt distress regime that comes along with lower output growth. This indirect effect adds up with the direct effect to the total effect of rising debt-to-gdp ratios on output growth. Figure 4 shows these three effects for our baseline specification with external PPG debt as the only debt component. 16 The direct effect is given by the probability-weighted average of the debt coefficient estimates for the two regimes in the output growth regression. At the 95 percent confidence level it is insignificant over the whole range of external PPG debt. Turning to the indirect effect, we can observe a significant expected loss in output growth as a consequence of a higher debt distress risk when additional external PPG debt is accumulated. The different shape of the output growth effects conditional on being in debt distress in the previous period or not is due to the strong state dependence of the debt distress probability. When a country is initially not in debt distress, the indirect effect is significant for debt-to-gdp ratios below the second threshold of 60 percent. If a country already faces a distress episode, additional debt still triggers a significant loss in output growth until the first threshold of 30 percent because it reduces the probability of exiting the distress regime. In both situations, the total effect remains significant beyond the respective thresholds even though both the direct and indirect effect turn individually insignificant. In this upper range, only the combination of the direct and the indirect effect yields an overall significant 15 In further estimations not shown here we also add interaction terms between external public debt and the additional regressors, in particular the CPIA index, but the respective coefficients turn out to be insignificant. 16 As before, the bootstrap confidence intervals are obtained by resampling the observations with country clusters 2000 times. 19