War, political turnover, polarization and sovereign default

(1)

Faculty of Economics and Business

Amsterdam School of Economics

Requirements thesis MSc in Econometrics.

1. The thesis should have the nature of a scientic paper. Consequently the thesis is divided

up into a number of sections and contains references. An outline can be something like (this

is an example for an empirical thesis, for a theoretical thesis have a look at a relevant paper

from the literature):

(a) Front page (requirements see below)

(b) Statement of originality (compulsary, separate page)

(c) Introduction

(d) Theoretical background

(e) Model

(f) Data

(g) Empirical Analysis

(h) Conclusions

(i) References (compulsary)

If preferred you can change the number and order of the sections (but the order you

use should be logical) and the heading of the sections. You have a free choice how to

list your references but be consistent. References in the text should contain the names

of the authors and the year of publication. E.g. Heckman and McFadden (2013). In

the case of three or more authors: list all names and year of publication in case of the

rst reference and use the rst name and et al and year of publication for the other

references. Provide page numbers.

2. As a guideline, the thesis usually contains 25-40 pages using a normal page format. All that

actually matters is that your supervisor agrees with your thesis.

3. The front page should contain:

(a) The logo of the UvA, a reference to the Amsterdam School of Economics and the Faculty

as in the heading of this document. This combination is provided on Blackboard (in

MSc Econometrics Theses & Presentations).

(b) The title of the thesis

(c) Your name and student number

(d) Date of submission nal version

(e) MSc in Econometrics

(f) Your track of the MSc in Econometrics

War, Political Turnover, Polarization and

Sovereign Default

Sarah Piassi Machado Lima

(11087862)

MSc in Econometrics Track: Free Track

Date of final version: August 15th, 2016 Supervisor: R.M.W.J. Beetsma

Second reader: J.C.M. van Ophem

Abstract

The widespread increase in public indebtedness that took place after the global financial crisis of 2007/2008 has increased the fear of sovereign default across the globe. This work analyzes the drivers of the likelihood of sovereign debt default. Aiming at studying the effects of war, political turnover and political polarization on sovereign debt default as well as on the amount of debt in default, a unique database is organized in order to, using a sound theoretical background, test whether (1) the probability of default is higher under political turnover and/or the occurrence of war, (2) the amount of debt in default increases under political instability and/or the event of a war.

(2)

Statement of Originality

This document is written by Student Sarah Piassi Machado Lima who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

1 Introduction

The fear of sovereign debt default and its consequences has considerably increased after the global financial crisis that took place around the years of 2007 and 2008 and the widespread increase in public indebtedness that took place thereafter. A well-known example is the Greek government debt crisis, which was constantly on the media around the year of 2012. Greece was not able to meet its debt-servicing obligations, even though it has not defaulted in a legal sense. Furthermore, several emerging market economies have recently experienced episodes of sovereign default, such as Russia in 1998, Ecuador in 1999, Ukraine in 2000 and Argentina in 2001. In these countries, higher sovereign interest rate spread levels and volatility are associated with higher political risk, which suggests that political variables might help to explain sovereign debt default incentives.

This work analyzes the drivers of the likelihood of sovereign debt default and of the amount of debt in default. Aiming at generating positive predictions on the likelihood of sovereign default and its magnitude, a simple theoretical political-economy model is developed. The model describes an small open economy constituted by two agents, which are represented by two political parties with distinct preferences for the two public goods existing in this economy. For each period, the party in power gets to decide the allocation of the goods. Political instability is included in the model by considering the fact that the incumbent party remains in office for the next period with some given probability. The divergence around the allocation of the public goods constitute a measure of political polarization. The endowment for each period depends on whether or not a war is taking place—an event that also happens with a certain given probability. This small economy is able to borrow on the international financial markets and to issue zero-coupon bonds to finance it’s expenditures.

The theoretical model problem is solved and the resulting predictions are tested using the data available on macroeconomic variables, wars and political features for an unbalanced panel constituted of 200 countries over the period of 1980 to 2014. The empirical framework used to test the theoretical forecasts is divided in two parts: a fixed-effects Logit model to explain the probability of default and a Tobit model for the amount of debt in default. More specifically, amongst other theoretical predictions, this paper tests whether the amount of debt in default increases under political instability and/or the event of a war and whether the probability of default is higher under political turnover and/or the occurrence of war.

The remainder of this work is organized as follows: the next section presents a quick literature review on the relevant developments on sovereign debt default theoretical and empirical works. The theoretical framework is described shortly thereafter followed by the description of the database. The section that follows describes the empirical framework here adopted. The results are then reported and discussed. The last section concludes this paper.

(5)

2 Literature review

Politics matter in a government decision to default. Rather than the ability to pay, the willingness to pay, by its turn influenced by politics, plays an important role in debt crises. The existing empirical literature presents evidence of a connection between political variables and sovereign default risk1. Citron and Nickelsburg (1987) [22] build a model of country risk for foreign borrowing which incorporates a number of changes of government over a five-year period, i.e., a political instability variable, and find that political instability is a statistically significant determinant of the probability of default. This finding cast doubts on the wisdom of policies that require economic retrenchment or restructuring as a precondition for rescheduling the external debt payment: restructuring might be related to government changes, which would lead to higher political instability and, hence, to a higher probability of default.

Brewer and Rivoli (1990) [7] test the effects of political instability on a country’s credit-worthiness, that is, on the judgement performed by lenders on the possibility that a borrower may default on his debt obligations. The authors find that political regime instability, measured by changes in the head of government and changes in the governing group, is a statistically significant variable in explaining the probability of sovereign default. Balkan (1992) [5] creates two political risk variables, namely the level of democracy and political instability, and includes them in a Probit model of external debt rescheduling2. The coefficient estimates and the forecasts support the importance of including political events in the assessment of a country’s exposure faced by international lenders. The author finds an inverse relationship between the level of democracy and rescheduling probabilities and a direct relationship between political instability and rescheduling probabilities.

Van Rijckeghem and Weder (2004) [27] also show that politics matter in explaining defaults on external and domestic debt obligations. They analyze a number of political and macroeco-nomic variables using a nonparametric technique to predict safety from default. In particular, a shorter mandate of the executive corresponds to a higher probability of default. Moser (2007) [20] examines twelve Latin American countries from 1992 to 2007 and finds that political in-stability measured by cabinet reshuffles involving key policymakers increases sovereign bond spreads. Hence, as initially stated, the empirical studies suggest that political uncertainty plays a significant role in determining default incentives.

This paper extends the aforementioned findings on sovereign debt default. In order to de-velop a suitable theoretical framework, an important reference is the work of Cuadra and Sapriza (2008) [11]. In their work, Cuadra and Sapriza develop a dynamic stochastic model to explore theoretically and quantitatively some of the channels through which a country’s political process might affect sovereign debt default incentives. Unlike the original model they build upon, which did not allow for countries to default on their debt—they incorporate the possibility of default.

1

Defined as a country’s ability-to-pay and willingness-to-pay its debt.

2

Debt rescheduling relates to a formal delay of debt-service payments and/or the application of extended maturities to the deferred amount.

(6)

The authors model their political system following Alesina and Tabellini (1989) [3], Persson and Svensson (1989) [24] and Ozler and Tabellini (1991) [23]. The sovereign default decision is de-termined using the willingness-to-pay approach by Eaton and Gersovitz (1981) [12]. Carmignani (2003) [9] also follows a similar approach. A large share of the existing literature on sovereign debt default focus mostly on developing countries issues and does not incorporate the effects of war on their models. However, history has shown that wars pose a challenge to the ability of a country to pay its debts and, thus, a war variable is included in both theoretical and empirical frameworks described in the following sections.

(7)

3 The theoretical framework

3.1 The model

The theoretical model describes a small open economy with two types of agents represented by two political parties during two periods (t). There are two public goods labelled x and y, of which the allocation amongst agents at t = 1, 2 is determined by the government3. There are two political parties, denoted by X and Y , characterized by their preferences for either good x or y, which are measured by a parameter 0 ≤ α ≤ 1. Both parties care for both good types but assign relatively more weight to the consumption of its own preferred type. Party X prefers good x while party Y has a relative preference for good y. The relative preferences are denoted by αx and αy respectively4. Each period one of the parties is in charge of the executive.

The party in power remains in power during the second period with probability π_p—a mea-sure of political stability. In democracies, elections determine the replacement (or not) of the incumbent party. For non-democracies, the governmental turnover can happen, for instance, as a coup. The party in power knows that being out of office means that the distribution of resources among agents will differ from what it considers to be optimal according to their preferences. Hence, the model captures both political instability, relative to the probability of losing office, and political polarization, the disagreement over the consumption allocated to each group.

The utility for each one of the political parties, during the entire time horizon, is given by (1):

U (α) = αu(x1) + (1 − α)u(y1) + E[αu(x2) + (1 − α)u(y2)] (1)

The utility functions u(.) is well behaved—it is continuous, concave, strictly increasing, twice differentiable—and follows the constant relative risk aversion (CRRA) specification. In the first period, the government has no endowment. However, it can borrow on the international financial markets. In the second period the country must pay its debt5. The second period’s endowment, available for debt repayment and public spending, is labelled e with e ∈ [0, ¯e]. Hence, the incumbent party must take decisions on the allocation of the goods x and y across households and on borrowing and repaying the foreign lenders. More specifically, the country faces three sources of uncertainty in the second period: the election outcome, war and the endowment when the country is not at war.

The endowment e is stochastic and its distribution depends on whether the country is at war or not. War is a discrete event and occurs with probability πw. It is assumed that, if the country

is at war during the second period, it uses all its resources on war spending, completely defaulting on any positive outstanding amount of debt. Also, it cannot spend anything on the consumption of the public goods. Hence, the endowment is e = 0. On the other hand, if the country is not at

3_{An important assumption of the theoretical model is that the government in t − 1 cannot commit to a future}

allocation.

4

Without loss of generalization, it is assumed that αx> αy. 5

(8)

war in period two, the endowment follows the continuous distribution Φn_(˜_{e) = P r(e < ˜}_{e|no war).}

It is assumed that Φn(0) = 0, as the full period utility in (1) is only defined for non-negative consumption of the public goods.

The country issues one-period zero-coupon bonds b6to finance its expending on public goods during the first period. In the second period, the government has to repay its debts. In case of any remaining endowment, it can spend the rest of the endowment in public goods. Hence, the payout to the bondholders in period 2 is given by:

h =    b if e > b e otherwise (2)

The risk neutrality of the investors implies that:

Q(b, πw, r) = 1 − πw 1 + r Z b 0 edΦn(e) + b Z ∞ b dΦn(e) = 1 − πw 1 + r Z b 0 edΦn(e) + b 1 − Φn(b) (3)

where Q(b, πw, r) is the amount risk-neutral investors willing to pay for the payout in (2) given a real

interest rate r. For future use, note that:

∂Q(b, πw, r) ∂b = 1 − πw 1 + r [1 − Φ n_{(b)] ≥ 0} ₍₄₎ ∂Q(b, πw, r) ∂πw = − 1 1 + r Z b 0 edΦn(e) + b 1 − Φn(b) < 0 (5) ∂Q(b, πw, r) ∂r = − 1 1 + rQ(b, πw, r) < 0 (6) ∂2_{Q(b, π} w, r) ∂b2 = − 1 − πw 1 + r φ n_{(b) < 0} ₍₇₎ ∂2_{Q(b, π} w, r) ∂b∂πw = − 1 1 + r[1 − Φ n_{(b)] = −}∂Q(b, πw, r) ∂b (1 − πw) < 0 (8)

where φn_{(.) is the density function associated with Φ}n_(.). _Also, ∂Q(b,πw,r)

∂b > 0 if 0 ≤ b < ¯e and ∂Q(b,πw,r)

∂b = 0 if b = ¯e. The probability of default is given by:

Φ(b) ≡ P r(e ≤ b) = πw+ (1 − πw)Φn(b) (9)

The expected amount of debt in default is given by:

πwb + (1 − πw)

Z b

0

(b − e)dΦn(e) (10) 3.2 The solution

The model is solved backwards, solving first the problem of the political party in office at the second period. The incumbent party chooses x2and y2to maximize:

max x2,y2 α2u(x2) + (1 − α2)u(y2) (11) s.t. x2+ y2=    e − b , if e > b 0 , otherwise (12) 6

(9)

where α2∈ {αx, αy} denotes the preference of the incumbent party in period two. If e ≤ b, the endowment

is completely used to pay the debt and hence the optimal values are x∗₂= y∗₂= 0. If e > b, the solution is given by the first order conditions that implicitly define the optimal values x∗₂ and y₂∗ as functions of α2, b and e.

α2u0(x∗2) = (1 − α2)u0(y2) (13)

y₂∗= e − b − x∗₂ (14) Denoting x∗2= x2(α2, b, e) and y∗2= y2(α2, b, e), note that x2(α2, b, b) = y2(α2, b, b) = 0. The implicit

differentiation gives −1 < dx2 db = − (1 − α2)u00(y2∗) α2u00(x∗2) + (1 − α2)u00(y2∗) < 0 (15) −1 <dy2 db = − (1 − α2)u00(x∗2) α2u00(x∗2) + (1 − α2)u00(y2∗) < 0 (16)

Denoting α1∈ {αx, αy} the preference of the party in office in period one, assuming that the problem

is analogous for both parties and hence with no loss of generality, suppose that party X is the incumbent party in the first period. The party maximizes its expected utility taking into account the uncertainty over the political process, the possibility of war and the uncertainty about second period’s endowment. Since, from the budget constraint, y1 = Q(b, πw, r) − x1, the party chooses x1 and b that solve the

following problem: max x1,b F (x1, b, πp, πw, r) (17) where F (x1, b, πp, πw, r) ≡ αxu(x1)+(1−αx)u(Q(b, πw, r)−x1)+(1−πw)(1−πp)Ωn(αx, αx, b)+πpΩn(αx, αy, b) and Ωn_(α 1, α2, b) =R ¯ e

b[α1u(x2(α2, b, e)) + (1 − α1)u(y2(α2, b, e))]dΦ n_(e)

The first order conditions are

Fx= αxu 0 (x1) + (1 − αx)u 0 (Q(b, πw, r) − x1) = 0 (18) Fb= (1 − αx)u0(Q(b, πw, r) − x1)Qb(b, πw, r) + (1 − πw)(1 − πp)Ωnb(αx, αx, b) + πpΩnb(αx, αy, b) = 0 (19)

were F_x and F_b are the derivatives of F (x₁, b, πp, πw, r) with respect to x1 and b; and Qb and

Ωn_b are the derivatives of Q and Ωn with respect to b respectively. Using dx2

db and dy2 db specified above: Ωn_b(α1, α2, b) = Z ¯e b α1u0(x∗2) dx2 db + (1 − α1)u(y ∗ 2) dy2 db dΦn(e) (20) = − Z e¯ b u0_(x∗ 2)u0(y∗2)[α1λ(y∗2) + (1 − α1)λ(x∗2)] u0_(x∗ 2)λ(x∗2) + u0(y2∗)λ(y∗2) dΦn(e) (21) = − Z e¯ b v(α1, α2, b, e)dΦn(e) (22)

where λ(.) = −u_u0_(.)00(.)2 and λ0(.) =

−u000_(.)u0_(.)+2u00_(.)

u000_(.)3 which is negative if u000(.) ≥ 07. The

second-order sufficient conditions are

∆ = FxxFbb− Fxb2 > 0, Fxx < 0 (23)

7

If u000(.) = 0 the utility is quadratic, if u000(.) > 0 there is precautionary savings or “prudence” (which is fulfilled by CRRA utility).

(10)

where ∆ is the determinant of the Hessian8_.

3.3 The theoretical model predictions

The total differentiation of the first-order conditions with respect to x1, d, πw, πp and r yields:

" dx1 db # = −1 ∆ " Fbb −Fxb −F_xb Fxx # " Fxπp Fxπw Fxr Fbπp Fbπw Fbr #     dπp dπw dr     (24)

It is easy to see that _dπdb

p = −

FxxFbπp

∆ , as Fxπp = 0. The second-order conditions ensure that

Fxx

∆ < 0 and hence the sign of db

dπp is equal to the sign of Fbπp. Using the previous results:

Fbπp(.) = (1 − πw) Z ¯e b v(αx, αx, b, e) − v(αx, αy, b, e) dΦn(e) (25) where, as specified above

v(α1, α2, b, e) =

u0(x∗₂)u0(y₂∗)[α1λ(y2∗) + (1 − α1)λ(x∗2)]

u0_(x∗

2)λ(x∗2) + u0(y2∗)λ(y2∗)

(26) Tabellini and Alesina (1990) [3] show in their appendix that if λ(x) is decreasing in x, then v(α1, α2, b, e) is hump-shaped in the α2 dimension, with a maximum at α2 = α1. Hence, for

αx 6= αy, and for all b and e, we have that v(αx, αx, b, e) > v(αx, αy, b, e). This yields Fbπp > 0

and hence we have that:

(i): if λ(x) is decreasing in x, debt increases in political instability.

The effect of the probability of war on borrowing is given by: db dπw = FxbFxπw− FxxFbπw ∆ (27) = −αx(1 − αx)u 00_(x 1)u00(Q(b, πw, r) − x1)Qπ wQb ∆ > 0 (28) where F_xb= −(1 − αx)u00(Q(b, πw, r) − x1)Qb > 0, Fxπw = −(1 − αx)u 00_{(Q(b, π} w, r) − x1)Qπw < 0 and Fbπw= (1 − αx)u 00_{(Q(b, π} w, r) − x1)QbQπ w+ (1 − αx)u0(Q(b, πw, r) − x1)Qbπw− (1 − πp)Ω n b(αx, αx, b) − πpΩnb(αx, αy, b) = (1 − αx)u00(Q(b, πw, r) − x1)QbQπ w− 1 1 − πw (1 − αx)u0(Q(b, πw, r) − x1)Qb+ (1 − πw)(1 − πp)Ωnb(αx, αx, b) − πpΩnb(αx, αy, b) = (1 − αx)u00(Q(b, πw, r) − x1)QbQπ h = FxπwQb, with Qbπw= − Qb 1 − πw Hence: 8

Note that the concavity of the utility function ensures that the last second-order condition holds, since Fxx(.) = αxu00(x1) + (1 − αx)u00(Q − x1) < 0.

(11)

(ii): a higher war chance leads to higher debt.

From the expression above it follows that:

(iii): the probability of default is increasing in the size of the debt.

Also, given that P r(e < b | war) > P r(e < b | no war):

(iv): the probability of default is higher under a war than under no war.

Since dP r(e≤b)_π

p = (1 − πw)φ

n_(b)db dπp:

(v): if λ(x) is decreasing in (x), the probability of default is increasing in the likelihood of political turnover.

Since dP r(e≤b)_π

w = 1 − Φ

n_{(b) + (1 − π}

w)φn(b)_dπdb_w > 0:

(vi): the probability of default is increasing in the chance of war.9

The derivative of the expression above with respect to the probability of political turnover is given byπw+ (1 − πw)Φn(b)

_db

dπp. Hence, we obtain:

(vii): if λ(x) is decreasing in (x), the expected amount of debt in default is increasing in the likelihood of political turnover.

The derivative of expected amount of debt in default with respect to the likelihood of war is b[1 − Φn(b)] + [πw+ (1 − πw)Φn(b)]_dπdb_w +

Rb

0 edΦ

n_{(e) where} db

dπw does not depend on e, because

e is integrated out in the first-order condition. Hence, we have:

(viii): the expected amount of debt in default is increasing in the likelihood of war.

9

Note that, unlike item (iv), this item does not deal with the actual event of war and hence can be described as an ex ante effect.

(12)

4 The Database

The database consists of an unbalanced annual panel of 200 countries over the period of 1980 to 2014. It comprises the information on several indicators for each country. The countries are classified according to their income level. The complete list of countries by income class can be found in Table 110. The income classification follows the World Bank official specification as stated bellow11:

“For the current 2017 fiscal year, low-income economies are defined as those with a Gross National Income (GNI) per capita, calculated using the World Bank Atlas method, of $1,025 or less in 2015; lower middle-income economies are those with a GNI per capita between $1,026 and $4,035; upper middle-income economies are those with a GNI per capita between $4,036 and $12,475; high-income economies are those with a GNI per capita of $12,476 or more.”12

The present section explains the construction of the data set. A more summarized description of the variables as well as summary statistics can be found in the Tables section.

4.1 The CRAG

Sovereign debt is a term used to denote a debt issued by a national government. It characterizes a contractual obligation and, as such, failing to meet these obligations to pay interests or principal on the due date constitutes a default. However, sovereign debt defaults might not be so explicit: a default can effectively occur not because the debt service was interrupted, but due to actions that resulted in economic losses by the creditors. The data on sovereign debt defaults for each country was obtained from the Database of Sovereign Defaults of the Bank of Canada’s Credit Rating Assessment Group (CRAG)13. CRAG considers that a default has occurred when debt service was not paid on the due date, payments were not made within the time frame specified under a guarantee and/or if the creditors incurred in any material economic losses on the sovereign debt they hold.

The CRAG database comprises data on debt owed to official and private creditors for all sovereign defaults the CRAG technicians were able to identify between 1970 and 201514. Despite the fact that the database also specifies the reliability of the data available for each country, in this work, all the data is considered regardless of how reliable it is. The data for some transition economies, such as the former Soviet Union states, is not available for the complete period. The

10

Please refer to the last section of this paper—Tables.

11

Note that some countries have no income classification and hence, were not included in the estimations that divide the sample by income classes.

12

https://datahelpdesk.worldbank.org/knowledgebase/articles/906519

13

The full methodology on the collection and organization of the data can be found on the technical report of 2016 by David Beers and Jamshid Mavalwalla [6]

14

The 2015 figures are still provisional and not available to all countries and hence, in this work, we only deal with the period between the years 1980 and 2014.

(13)

values are expressed in nominal U.S. dollars. The CRAG database is used to construct the default dummy, which takes on a value of one in the occurrence of a sovereign default in a given year for a given country and zero otherwise. This database is also used to generate the debt in default as a share of the GDP variable, debtdgdp.

4.2 The political variables

In 2012 was published the most recent Database of Political Institutions (DPI2012) of the World Bank [14]. In an effort from the researchers of the Inter-American Development Bank [10], the DPI2012 was extended through 2015, adding data for the years 2013, 2014, and 2015—giving birth to the DPI2015, used in this research. From the DPI2015 database were constructed two important indicators used in the present exercise: (i) a dummy variable of political turnover, pturn, which takes a value of one the year after a political turnover has taken place and a value of zero otherwise, where a political turnover is defined as a change in the executive or in the ruling party15; (ii) a dummy variable for capturing political instability, elec, which assumes the value of one if there was an executive or legislative election in a given year and the value of zero otherwise. These indicators were selected to cover the findings in Alesina, Roubini and Cohen (1997) [2], who find evidence of policy changes in the immediate aftermath of elections.

Emanuel Kohlscheen (2009)[15]—while trying to explain the strong difference in default rates between developing countries that have presidential forms of government and those that are parliamentary—argues that parliamentary systems are less prone to default than presidential systems. To account for Kohlscheen’s (2009) potential finding, the dummy pment, from DPI2015 database is used. This variable takes on values of one if the political system is parliamentary, and zero otherwise16.

Alesina and Drazen (1991) [1] suggest that a more polarized government may lead to a game of attrition and delay stabilization. Hence, another important variable for this study, the Index of Competitiveness of Participation, here called polar, which refers to the extent to which alternative preferences for policy and leadership can be pursued in the political arena—a measure of the divergence of policy preferences in a given country—was obtained from the Polity IV Project database [19]. The index ranges from 1 to 5, with 1 standing for a repressed system, 2 for suppressed environment, 3 for a factional political society, 4 for a transitional society and 5 for a stable and competitive system in which secular groups regularly compete for power. A higher degree of polarization corresponds to a lower value of the Competitiveness of Participation Index. In “Political institutions and debt crises” Rijckeghem and Weder (2009) [28] show that polit-ical factors not only matter but also do so in different ways for democratic and non-democratic regimes. The authors find that, amongst other findings, democracies do a better job at avoiding default on external debt when compared to non-democratic regimes. In order to address this result, the polity variable, from Polity IV Project, is used. This variable, also called the Polity

15

For transition governments, like the former republics of the Soviet Union, pturn takes a value of one the year following its independence.

(14)

Score, measures the degree to which a country is a democracy. It ranges from -10 to +10, where -10 stands for a strongly autocratic system, while +10 represents a strongly democratic system.

4.3 The war dummies

The Correlates of War (COW) Project [25] was used to construct a war dummy, war, that assumes the value of one when a country is at war in a given year and zero otherwise. The The Correlates of War (COW) Project defines wars as events of sustained combat, involving organized armed forces, resulting in a minimum of 1000 battle-related fatalities within a twelve-month period. The database includes inter-state wars, intra-state wars and extra-state wars. Inter-state wars occur between recognized states. For instance, the invasion in Iraq in 2003 is an example of an inter-state war. Intra-state wars occur between entities within a state, such as the civil wars in Congo in the 1990’s17. Extra-state wars are fought between a state and a non-state entity outside the borders of the state. An example is the war between Western countries and the Taliban in Afghanistan from 2001 and onwards.

However, the Correlates of War database only comprises data until the year of 2007. Due the lack of reliable and comprehensive databases, the list of wars on Wikipedia18 was used to extend the war dummy until 2015. The Wikipedia list of wars heavily relies on the Conflict Barometer developed by the Heidelberg Institute for International Conflict Research (HIIK)19_{. The HIIK}

has been publishing the Conflict Barometer, an annual analysis of the global conflict events, since 1992. Yet, like previously stated, the HIIK database is not comprehensive. The Conflict Barometer reports list the conflicts that have begun at each year but does not explicitly specify the end of the conflicts20. The conflicts listed on Wikipedia were checked one by one in order to correctly address the countries involved in each conflict. It is possible to see that the conflicts listed also comprise intra-, inter- and extra-state wars—like the Correlates of War Database. However, the data from 2008 onwards is not as perfect, or reliable, as the COW database.

4.4 The macroeconomic variables

From the literature on sovereign defaults, some macroeconomic variables stand out as possible factors affecting the likelihood of default. Manasse, Roubini and Schimmelpfennig (2003) [18], while developing an early-warning model of sovereign debt crises, suggest that low GDP growth, current account imbalances and high inflation can signal a debt crisis. Manasse and Roubini (2009) [17] investigate the economic and political conditions associated to the occurrence of a sovereign debt crisis. By deriving a set of rules of thumb that help identifying some characteristics of defaulters, the authors select ten variables that can predict and classify a sovereign default.

17

However, these wars are not necessarily fought in the territory of the state.

18

See https://en.wikipedia.org/wiki/List_of_wars_2003-10 and https://en.wikipedia.org/wiki/List_ of_wars_2011-present.

19

See http://www.hiik.de/en/index.html.

20

The Conflict Barometer reports are long and a possible improvement for this work would be to perform the time consuming task of reading all the PDF files in order to find all the end dates for each conflict.

(15)

Amongst the ten, the debt to GDP ratio, the real GDP growth and consumer price index (CPI) inflation are also addressed in this work. Kraay and Nehru (2006) [16] empirically examine the determinants of debt distress in developing countries and find that the debt burden and the variability of real GDP growth play an important role. Yet, one must be aware of the fact that real GDP growth may have ambiguous effect on the likelihood of default. A low GDP growth associated with rising government debt may produce solvency problems, however, low GDP growth might reduce imports, diminishing potential liquidity problems.

Following the aforementioned literature, the macroeconomic variables used in this exercise are (i) the debt to GDP ratio, debtratio; (ii) the reserve coverage ratio—an attempt to capture the ability of a country to pay for its imports by means of its foreign reserves—, rim, defined as the difference between reserves and imports as a ratio of GDP; (iii) the ratio of the current account balance to GDP, cab, an alternative measure to account for liquidity; (iv) the GDP growth rates, growth; (v) and the CPI inflation rates, cpi, an attempt to control for the quality of monetary policy21.

For each country, the annual data on GDP, GDP growth rates, total reserves and CPI inflation rate were obtained from the World Bank database22 The debt variable and the current account balance indicator were obtained from the International Monetary Fund (IMF) World Economic Outlook23. It is important to notice that the IMF data only starts at the year of 1980. The data on imports was obtained from the United Nations Conference on Trade and Development (UNCTAD) database24_.

4.5 Descriptive statistics

The descriptive statistics can be found at the section Tables at the end of this paper. Table 2 comprises the variables description while Table 3 presents the summary statistics for the whole sample—alerting for the fact that, as the panel is unbalanced, not all observations on all variables and years are available for all countries. It is possible to see that, for the period between 1980 and 2014, in about half of the sample there was default. The average amount of debt in default as a share of the GDP is about 11.1%. The extreme value of 1968.5% was observed in Liberia in 1995, possibly a reflex of the civil war. The average debt to GDP ratio is 57.7%. The extreme value of 789.8% was also observed in Liberia, but in 2003. The extreme value of 24411% of CPI was observed in Zimbabwe in 2007—possibly the peak of the hyperinflation period. The CPI of -35.8% was observed in Sierra Leone in 2008.

The average GDP growth is about 3.5%. The GDP growth of 149.9% was observed in Equatorial Guinea in 1997, a consequence of the discovery of large oil reserves in 1996 and its

21

When high, the CPI counts as an indicator of mismanagement of monetary policy.

22

More details on these indicators can be found at, respectively, http://data.worldbank.org/indicator/ NY.GDP.MKTP.CD; http://data.worldbank.org/indicator/NY.GDP.MKTP.KD.ZG; http://data.worldbank.org/ indicator/FI.RES.TOTL.CD;http://data.worldbank.org/indicator/FP.CPI.TOTL.ZG.

23_{For more details, please refer to http://www.imf.org/external/pubs/ft/weo/2016/01/weodata/index.}

aspx.

24

(16)

subsequent exploitation. The other extreme observation of the GDP growth, -64%, was observed in 1991 in Iraq. The international economic sanctions on Iraq that followed Iraq’s seizure of Kuwait in August 1990 have drastically reduced its economic activity. The statistics on cab show that, on average, the countries here analyzed have a Current Account Deficit of 3.3%. The variable rim hows that, on average, imports seems to exceed the amount of reserves. War seems to happen in about 11.4% of the observations. As for the political variables, 32.6% of the countries have a parliamentary system. Also, it is possible to observe that elections have taken place in 24.7% of the observations. However, changes in the executive and/or in the ruling party seem to happen in 18.6% of the observations. The Polity Score, that ranges values from -10 to 10 as a way to measure to which extent a country is a democracy, is 1.9: on average, the countries here considered are closer to a democracy than to an autocracy. Looking at the polar dummy, according to the Index of Competitiveness of Participation, on average, the countries here considered are closer to a “factional” position when it comes to the extent to which alternatives preferences for policy and leadership can be pursued in the political arena.

Tables 4 and 5 present the summary statistics for two—High and Upper Middle; Low and Lower Middle—and four— High; Upper Middle; Lower Middle; and Low—income classes respec-tively. From Table 4 it is possible to see that Low and Lower Middle income countries have a 78.5% rate of default, while High and Upper Middle income countries present a rate of 31.8%. The average amount of debt in default is also lower for high income countries, 3.8%. This value corresponds to 21.4% for lower income countries. The debt ratio is also slightly higher for low income countries: 67.9%, in contrast with 51.7%, observed for higher income countries. The average growth rate for both income classes seem to be almost the same: 3.57% for low income countries and 3.47% for higher income countries. The deficit in the Current Account Balance is higher for lower income countries, which present an average deficit of 5.9%. For higher income countries this deficit is about 1.4%. Low income countries also face two times more wars than the higher income countries: 16.3% in contrast with 8.0%.

Almost half of the higher income economies seem to have a parliamentary electoral system, while only 13.0% of the lower income countries seem to have it. The variables elec and pturn behave practically the same for both income classes. However, it is clear that high income countries are closer to a democratic environment, average Polity Score of 4.2, than lower income countries, average Polity Score -0.6. Unlike high income countries that, on average, could be classified between “factional” and “transitional” by the Index of Competitiveness of Participation, lower income countries stand between a “supressed” system and a “factional” system, hence, more polarized.

In Table 5, the contrast between high income countries and low income countries becomes more evident: higher income countries have around 7.5% of its sample in default, while lower income countries show 92.8% of default. The average amount of debt in default is also a lot higher for low income countries, 37.4%, which can be compare to an average of 1.3% for High income countries. The debt ratio as a percent of the GDP is also higher for low income coun-tries. However, the economic growth rates, as observed before, are quite similar. High income

(17)

countries, unlike all the income classes underneath it, present a Current Account Surplus of 1.1% on average. For the bottom income class, the Current Account Deficit is about 6.7%. The incidence of war increases with the decrease of the average income. High income countries face war in approximately 6.3% of its observations while low income countries face war in 19.5% of its observations. Also, only 6.4% of the low income countries have a parliamentary electoral system, while 65.5% of the high income countries have a parliamentary system. It is also possible to see that, the higher the income level of the countries here studied, the higher the Index of Com-petitiveness of Participation and hence, the less polarized the political arena. Also, the higher the income level, the higher the value of the Polity Score and hence, the more democratic the countries comprised in that category on average.

Finally, due to some outstanding values of some of the macroeconomic series, the logs of the macroeconomic variables were used in this work. More specifically log₂(6 + variable) was used. The value of 6 was added to the variables to deal with possible negative values: one can see at Table 3 that the most negative value reaches -558.2%. The motivation behind the choice of the base of the log can be found in the next section where the empirical specifications are explained.

(18)

5 The empirical framework

Before describing the empirical framework, the hypothesis presented in Section 3 are here sum-marized25. Indications on how to test these predictions are also going to be provided. The hypothesis related to the event of default to be tested are:

Hypothesis 1 : the probability of default is increasing in the size of the debt. This is going to be tested by including the debt to GDP ratio, debtratio, in a panel regression for the event of default. To avoid simultaneity bias, this variable will be lagged in one year.

Hypothesis 2 : holding constant the level of debt, the likelihood of default is higher in the event of a war than in the event of no war, which is going to be tested by including the war dummy in the panel regression.

Hypothesis 3 : an increase in the chance of a political turnover raises the likelihood of default. This can be tested by including the dummies pturn and elec to a panel regression for the event of default.

The hypothesis related to the amount of debt in default are:

Hypothesis 4 : debt in default is increasing in the likelihood of a war, which is going to be tested by including the war dummy as explanatory variable to a model for debt in default.

Hypothesis 5 : debt in default is increasing in the likelihood of a political turnover. This can be tested by including the dummies pturn and elec to a panel regression for the amount of debt in default.

In order to test the various predictions from the theoretical model, this section introduces a panel regression model for the occurrence of default and the amount of debt in default. The exact empirical models used to test the predictions discussed above are specified in the next subsections.

5.1 Explaining default

For explaining the probability of the occurrence of sovereign default a fixed-effects Logit model is specified as follows: log pit 1 − pit = αi+ µt+ βxit (29)

(19)

where p_it stands for the probability of default by country i in period t, being default a dummy taking a value of 1 in the case of default, and zero otherwise26. αi is a set of fixed constants,

one for each country. The vector of predictors is given by x_it, which includes the macroeco-nomic variables—all lagged in one period due to possible simultaneity problems—, the political variables—to avoid simultaneity bias, the political variables pturn, elec and polar are lagged in one period27— and the war dummy.

As αi and xit are correlated, αi are treated as fixed parameters, and hence, this setting

constitutes a fixed-effects model. The random-effects model treats α_i and x_it as independently distributed by specifying αias part of an error term. Random-effects models are quite unrealistic

for observational, that is, non-experimental data, as typically omitted and included regressors show multicollinearity. Fixed-effects models are recognized for having an advantage over random-effects models when analyzing panel data because they control for all individual characteristics, measured or unmeasured. However, a major problem of fixed-effects models is the inability to estimate the effect of variables that do not vary within clusters.

5.1.1 The Between-Within method

A fixed-effects analysis uses only within-individuals variation and, for the present exercise, there is no within-country variation for some countries (specially the higher income ones), which results on the exclusion of very important data so the fixed-effects models can be estimated. In order to overcome this problem, it has been proposed to estimate within effects in random-effects models28. Like other fixed effects methods, this hybrid method, also called “between-within” (BW) method, provides a form to control for all cluster-level covariates, whether observed or not. This method is especially attractive for models like ordinal logistic regression, for instance. One might ask: why not just use random-effects estimates or population average methods? Random-effects estimates do not control for unmeasured predictors. Also, fixed-effects estimation is less prone to omitted-variable bias. A population-average coefficient talks about what would happen to the population as a whole if everyone’s predictor variable were increased by one unit— instead of telling what would happen to a single individual if that person’s predictor variable were increased by one unit. For a linear model there is no difference between these two kinds of coefficients, however, for a logistic regression model subject-specific coefficients are typically larger than population-averaged coefficients.

Hence, the BW method is adopted in the present exercise in order to deal with the estimation of the fixed-effects Logit model. It is important to alert to the fact that it was never proved that the BW method does what it proposes to do, i.e., to provide consistent estimates of the parameters. The BW method produces the exact same result as the regular fixed-effects method

26_{It is easy to see that p}

it= P r(yit= 1) = e

κit

1+eκit, κit= αi+ µt+ βxit.

27_{It is very unlikely that the political variables polity and pment are influenced by the dependent variable, and}

hence, their contemporaneous values are included.

28

See for instance Allison (2009) [4]; Neuhaus and Kalbfleisch 1998[21]; and Schunck (2013) [26], amongst others.

(20)

for linear models, but for the logistic regression, the BW estimates are not exactly those produced by conditional likelihood. Goetgeluk and Vansteelandt (2008) [13] have proved that the BW method produces consistent estimates for the linear case even though this method can generate biased estimates for some nonlinear models, including logistic regression. However, what they have also concluded is that the biases were small in most practical settings. Hence, considering the model in (29), the BW method decomposes x into a within-cluster component and a between-cluster component. The following generalized linear mixed model is estimated:

log pit 1 − pit = αi+ µt+ βW(xit− ¯xi) + βBx¯i (30)

where ¯xi is a vector of country-specific means. The estimates of βW depends only on

within-cluster variation and hence, it is assumed that it controls for confounding with α and produces consistent estimates of β in equation (29). In general, this assumption is true under the condition that α_i in (29) is a linear function of ¯xi plus a random error that is independent of xit29. Hence,

one can assume that the performance of the BW method for the logistic regression depends to some extent on how this linearity assumption is satisfied.

Considering the following equation—which is exactly the same as the model in (30)—it is easy to see that xit does not have to be expressed as deviations from the means if the means are

included in the equation:

log pit 1 − pit = αi+ µt+ βWxit+ (βB− βW)¯xi (31)

By adding polynomial functions of the means to (31), one could check or allow for non-linearity: log pit 1 − pit = αi+ µt+ βWxit+ θ1x¯i+ θ2x¯2i (32)

If the additional quadratic term is not statistically significant and if the estimate of β_W does not change much, then bias is probably not an issue. However, another problem that might pose a challenge for the estimation of the fixed-effects Logit is the fact that fixed effects models in short panels are generally not estimable due to the incidental parameters problem: the coefficients are usually biased. The incidental parameter problem is typically seen to arise on panel data models when including agent specific intercepts in a regression model. For T held fixed and N large, we have the incidental parameter problem and hence the maximum likelihood might not be consistent. The number of parameters increases directly with the sample size, violating one of the conditions that underlie the asymptotic theory of maximum likelihood estimation.

In particular, it can occur in non-linear models which, unlike linear regression, do not have the property of being unbiased estimators. For Logit models the estimators are consistent,

29

(21)

meaning that, as the ratio of the number of observations to number of parameters increases, the parameter estimates will converge onto their true values as standard errors become arbitrarily small. The problem with fixed-effects is that the number of parameters grows with the number of observations and hence, the parameter estimates can never converge to their true value as the sample size increases. Thus the parameter estimates are severely unreliable.

5.1.2 Interpreting the coefficients

In order to interpret the results of the Logit model it is very helpful to look at the odds ratio as defined below:

eβ = P r(yit = 1|xit+ 1) P r(yit = 0|xit+ 1)

P r(yit= 1|xit)

P r(yit= 0|xit)

The way to interpret the coefficients based on the odds ratio is as follows: all else equal, with increase of x by 1 unit, the odds of y = 1 versus y = 0 increase by factor eβ.30 If eβ = 1.18, it means the odds of y = 1 increase by 18 percent, everything else constant, if the explanatory variable increases in one unit, i.e., the odds of default are multiplied by 1.18.

For the cases where the independent variables are in logs, once the estimated coefficients are exponentiated (eβ), the odds ratio is associated with a b-fold increase in the predictor variable31, where b is the base of the logarithm used to transform the predictor. In this work, in order to make the interpretation of the results more accessible, for log-transforming the predictors, the base of two is used. Hence, the exponentiated coefficients can be interpreted as the odds ratios associated with a doubling of the predictor. Each two-fold increase in the logged variable have the effect on the odds ratio specified in the odds ratio.

5.2 Explaining the amount of debt in default

The amount of debt in default as a ratio of GDP is explained by a Tobit model defined by (33):

debtdgpd∗_it= αi+ βt+ γ1M ACROi,t−1+ γ2W ARi,t+ 1

X

l=0

γ3lP OLi,t−l+ εi,t (33)

where debtdgpd∗_it is a latent variable for the default dummy such that:

debtdgpdit=    debtdgpd∗_it , if debtdgpd∗_it> 0 0 , otherwise

where αi and βt are country and time specific effects, M ACROi,t−1 is a vector comprising

the macroeconomic variables lagged in one period, W AR_i,t comprises the war dummy for each country and for each year and P OLi,t gathers the political variables—some lagged in one period,

as explained in the previous section. The estimated coefficients should be interpreted as the effect of the regressors on the latent variable debtdgpd∗_it. More specifically, the estimated coefficients

30

Where β stands for x’s coefficient.

31

(22)

can be interpreted as the combination of the change in debtdgpd_it of those above the limit, weighted by the probability of being above the limit; and the change in the probability of being above the limit, weighted by the expected value of debtdgpd_it if above the limit. In order to analyse the results of the Tobit estimates, the mean of the marginal effects of a change in the independent variables on debtdgpd_it should be considered: E(debtdgpd_it|0 < debtdgpd_it). In this case, the mean of the marginal effects of a change in the predictor on debtdgpdit (which has a

lower bound in zero) is about ˆγ, everything else constant. Ceteris paribus, for variables in logs at the base 2, ˆγ is the expected change in debtdgpditwhen the independent variable is multiplied

(23)

6 Results

6.1 Explaining default

The estimation results tables can be found in the section Tables. Each table reports the estimated coefficients and the standard errors for each one of the predictors. The superscript “*” indicates that the variable is significant at the level of 10% while “**” and “***” stand for variables that are significant at 5% and 1% respectively. The first table of this section, Table 6, summarizes the coefficients and standard errors generated by the fixed-effects estimation method. It is possible to see that the number of observations is reduced to about one third of what it should be. As already explained in Section 5, a fixed-effects analysis uses only within-individuals variation and, for the present exercise, there is no within-country variation for some countries, which results on the exclusion of very important data so the fixed-effects models can be estimated.

The debt ratio, the CPI, the growth variable and the polar dummy are significant32. Since the variables debtratio and cpi are in logs base 2, the odds ratios reported in the table are associated with a doubling of these variables. Hence, looking at column (1) and considering everything else constant, a 2-fold increase in debtratio is associated with the odds of default increasing by approximately 75.8% percent. Also for the column (1) results, for the case of the CPI, the odds of default increases by 13.8% under a 2-fold increase of this predictor. From the results in column (7), its is possible to see that the higher the level of the Index of Competitiveness of Participation, that is, the lower the level of polarization in the political arena, the lower the probability of default: the impact of an increase in polar is associated with a decrease of 64% in the odds of a default. However, due to the problems associated with the fixed-effects Logit model—as already discussed in the previous sections—the between-within method is also implemented.

Table 7 gives us the estimated coefficients and the standard errors that have resulted from the BW estimation. The number of observations included is now about 3 times higher that what was observed in Table 6. The between estimators (βB) are not very interesting in themselves,

however, it is possible to see that they are a lot larger in magnitude—as already noted by Allison (2009) [4]—than the corresponding within estimators (β_W). The original random-effects models, which does not split the coefficients in between and within estimators, assumes that the deviation coefficients are identical to the cluster mean coefficients. The result shown in Table 7 indicates that this assumption should be rejected suggesting that the fixed-effects approach is in fact superior to random-effects approach—as already stated in the previous sections. Looking at the within estimators, it is possible to see that the results are very similar to what was found in Table 6: looking at the within estimators, column (1) shows that a 2-fold increase in debtratio is associated with and increase of 73.8% in the odds of default while an increase in cpi is related to an increase of 13.4% in the odds of default. Also, from column (7), the lower the level of polarization, the lower the odds of default, which decreases in 54.3% with each increase in the

32

Note that the growth variable is only weakly significant and only in some regressions. Hence, this variable will be left out of the comments.

(24)

Index of Competitiveness of Participation. As expected, Table 8, which shows the results for the estimation specified in Equation (32), reports within estimator very similar to what was found in Table 7. Also, the additional squared terms are non-significant with very few exceptions, what can lead to the conclusion that biased estimated might not be a problem for the BW method here implemented.

The exercise is performed also for two sub-samples: high income countries, which comprises the countries classified as High Income and Upper Middle Income by the World Bank; and low income countries, formed by countries classified as Low Income and Lower Middle Income by the World Bank. The results for the high income countries are shown in Table 9. The effect of the within estimator of debtratio seems to be a lot stronger for high income classes. The within estimator for cpi is no longer significant but now the within estimator for the war dummy is significant33, being associated with a strong increase in the odds of default for the cases where the dummy equals one. The effect of the within estimator of polar also becomes stronger: a low polarization of the political arena is associated with a decrease of 91.9% in the odds of default. Table 10 shows the estimation results of Equation (32) for high income countries. Once again, the results support the thesis that the BW estimates in Table 9 are consistent. Table 11 displays the estimates for low income countries. The results do not seem to be significant to explain the probability of default for low income countries. Table 12 shows evidence that this result is robust.

As for the theoretical predictions, Hypothesis 1 which states that an increase in sovereign debt makes default more likely, seem to be supported by the results explained above, with the exception of the results for the sub-sample of low-income countries. For high income countries, Hypothesis 2, “holding constant the level of debt, the likelihood of default is higher in the event of a war than in the event of no war”, seem to also hold. However, Hypothesis 3, “an increase in the chance of political turnover raises the likelihood of default”, does not seem to hold for any of the income classes.

6.2 Explaining the amount of debt in default

The estimation results for the Tobit model specified in the previous sections can also be found at section Tables. Table 13 shows the estimates for the exercise comprising the whole sample. It is possible to see that only growth and cpi have significant coefficients, even though this happens only for some equations. The political turnover dummy seems to be significant at 10%. Table 14 summarizes the average marginal effect for the independent variable conditioned on the fact that it is censored at zero34. Looking at column (1), everything else constant, a 2-fold increase in growth is associated with an 85.2% decrease on the amount of debt in default. Looking at column (7), a 2-fold increase on the CPI is related to a decrease of 81.6% in the amount of debt in default.

33

Note that, by looking at Table 4, high income countries, here defined as High and Upper Middle income countries, have more observations on the war dummy (3920) when compared to lower income countries (2835).

34_{As already explained in the previous section: E(debdgdp}

(25)

Table 15 gathers the results for high income countries. It is possible to see that all variables are very significant. Table 16 reports the average marginal effects related to the coefficients in Table 15. From column (1) it is possible to see that a 2-fold increase in debtratio increases the amount of debt in default by more than 100%. A 2-fold increase on the CPI decreases the expected amount of debt in default by 97.5%, everything else kept constant. If doubled, growth decreases the expected amount of debt in default by 88.9%. cab and rim also have a negative effect on the expected amount of debt in default. Highlighting the fact that the war dummy and the political variables are not in logs, looking at war, included in column (2), it is possible to see that the occurrence of war in a given year has a positive impact on the expected amount of debt in default.

The political turnover measures, elec and pturn also have a positive impact on the amount of debt in default: the more unstable the political environment, the higher the expected amount of debt in default. This result adds to the literature on sovereign default by showing that political variables can help explaining not only the probability of default—as seen in most of the literature here mentioned—but also the amount of debt in default. The pment dummy also has a positive impact on the expected amount of debt in default, despite the findings of Kohlscheen(2009)[15], who argued that parliamentary systems are less prone to default than presidential systems— apparently, once a default has taken place, the amount of debt in default seems to be higher under parliamentary systems. The variable polity, which measures to which extent a country can be considered democratic, shows a negative coefficient: the more democratic a country, the lower the expected amount of debt in default. Also, the higher the level of the Index of Competitiveness of Participation, i.e., the lower the level of polarization of the political arena, the lower the expected amount of debt in default. Like the estimates for the full sample, the estimates for the low income countries are practically non-significant—as one can see in Table 17 and Table 18.

Despite the fact that, for high income countries, the theoretical predictions related to the expected amount of debt in default, namely, Hypothesis 4 (“debt in default is increasing in the likelihood of a war”) and Hypothesis 5 (“debt in default is increasing in the likelihood of a political turnover”) were fulfilled, these results suggest that the amount of debt in default for low income countries should be explained by a different approach and possibly by the use of different explanatory variables.

(26)

7 Conclusions

Based on the existing literature on sovereign default, the present work deals with the task of identifying the drivers of the sovereign default and the amount of debt in default. In order to generate theoretical insights for the empirical implementations, a theoretical model is developed by including both political instability and wars to a small open economy setting. An empirical framework is developed taking into account the predictions of the theoretical model. The main goal can be divided in two parts: (1) to explain the occurrence of default and (2) to explain the amount of debt in default once a country has defaulted. Despite the fact that the theoretical predictions come from a very simple model that ignores the complexities of the real world and bases itself on very simplified assumptions, this work was able to confirm some of these predictions by making use of an equally imperfect database.

The first insight from the theoretical model is summarized by Hypothesis 1, which states that an increase in sovereign debt makes default more likely. This assumption is supported by the results presented in the previous section, however, it doesn’t hold for lower income countries. Hypothesis 2, “holding constant the level of debt, the likelihood of default is higher in the event of a war than in the event of no war”, is also supported—but for high income countries only. The last insight relative to the event of default, Hypothesis 3, “an increase in the chance of political turnover raises the likelihood of default”, does not seem to hold for any of the income classes or for the whole sample. These results suggest that the model that explains the event of default should be modified in order to be able to capture the effects on the probability of default for low income countries. Also, new or more political variables could be tested.

As for the theoretical predictions related to the amount of debt in default, both hypothesis, namely, Hypothesis 4, “debt in default is increasing in the likelihood of a war”, and Hypothesis 5, “debt in default is increasing in the likelihood of a political turnover”, were fulfilled for high income countries. However, this result is not valid for the whole sample or for lower income countries, which suggest that the amount of debt in default for low income countries should be explained by a different approach and possibly by the use of different explanatory variables.

To summarize, the Logistic regression model developed to deal with the explanation of the occurrence of default does not seem to be including all the necessary variables in order to gen-erate useful predictions on the probability of default. The Logit results, besides alerting for the complexity of dealing with a fixed-effects Logit setting, alerts to the need of specifying new ex-planatory variables for low income countries and to the need of finding new measures of political instability. The Tobit model designed to explain the amount of debt in default seems to perform well when it comes to high income countries. However it does not seem to be the appropriate approach to deal with lower income countries.

Hence, a series of future exercises, aiming at producing significant and robust insights for both high and low income countries, can be listed. First, the inclusion of new political variables— in both the Logit model for the probability of default and the Tobit model for the amount of debt in default—should be considered. Perhaps, different variables should be considered for

(27)

different income classes. Second, new exercises considering only highly reliable data from CRAG could be implemented. As explained in Section 4, the CRAG database has different levels of reliability for each one of its countries. Third, further exercises could be made in order to deal with the outstanding values of the macroeconomic variables by using a different approach other than taking logs. Fourth, exercises segmenting the sample into different categories, such as OECD/non-OECD countries and/or democracies/non-democracies, could be performed. Finally, further research could be made in order to replace the Tobit approach by modelling the amount of debt in default under a Heckman selection model frame where a linear regression for the amount of debt in default is estimated as a second step in a model that first specifies if default has in fact occurred.

(28)

8 References

[1] Alesina, A. and Drazen, A. (1991). Why are Stabilizations Delayed? The American Economic Review, 81(5):1170–1188.

[2] Alesina, A., Roubini, N., and Cohen, G. D. (1997). Political cycles and the macroeconomy. [3] Alesina, A. and Tabellini, G. (1989). External debt, capital flight and political risk. Journal

of International Economics, 27(3-4):199–220.

[4] Allison, P. D. (2009). Fixed Effects Regression Models. Thousand Oaks, CA: Sage Publica-tions.

[5] Balkan, E. M. (1992). Political instability, country risk and probability of default. Applied Economics, 24:999–1008.

[6] Beers, D. and Mavalwalla, J. (2016). Database of Sovereign Defaults , 2016. (101):39. [7] Brewer, T. L. and Rivoli, P. (1990). Politics and Perceived Country Creditworthiness in

International Banking. Journal of Money, Credit and Banking, 22(3):357–369.

[8] Brumback, B. A., Dailey, A. B., Brumback, L. C., Livingston, M. D., and He, Z. (2010). Adjusting for confounding by cluster using generalized linear mixed models. Statistics & probability letters, 80(21):1650–1654.

[9] Carmignani, F. (2003). Political Instability, Uncertainty and Economics. Journal of Economic Surveys, 17(1):1–54.

[10] Cruz, C., Keefer, P., and Scartascini, C. (2016). Database of Political Institutions 2015 : Codebook. Dpi2015, 176(January):165–176.

[11] Cuadra, G. and Sapriza, H. (2008). Sovereign default, interest rates and political uncertainty in emerging markets. Journal of International Economics, 76(1):78–88.

[12] Eaton, J. and Gersovitz, M. (1981). Debt with Potential Repudiation: Theoretical and Empirical Analysis. Review of Economic Studies, 48(2):289–309.

[13] Goetgeluk, S. and Vansteelandt, S. (2008). Conditional Generalized Estimating Equations for the Analysis of Clustered and Longitudinal Data. Biometrics, 64:772–780.

[14] Keefer, P. (2012). Database of Political Institutions: Changes and Variable Definitions. Development Research Group, The World Bank.

[15] Kohlscheen, E. (2009). Sovereign risk: Constitutions rule. Oxford Economic Papers, 62(1):62–85.

[16] Kraay, A. and Nehru, V. (2006). When is external debt sustainable? World Bank Economic Review, 20(3):341–365.

(29)

[17] Manasse, P. and Roubini, N. (2009). "Rules of thumb" for sovereign debt crises. Journal of International Economics, 78(2):192–205.

[18] Manasse, P., Roubini, N., and Schimmelpfennig, A. (2003). Predicting Sovereign Debt Crises. International Monetary Fund, 4(2):03–221.

[19] Marshall, M. G., Gurr, T. R., and Jaggers, K. (2014). POLITYTM IV PROJECT Political Regime Characteristics and Transitions, 1800-2013.

[20] Moser, C. (2007). The Impact of Political Risk on Sovereign Bond Spreads - Evidence from Latin America. Proceedings of the German Development Economics Conference, Göttingen 2007 / Verein für Socialpolitik, Research Committee Development Economics, 24.

[21] Neuhaus, J. M. and Kalbfleisch, J. D. (1998). Between- and Within-Cluster Covariate Effects in the Analysis of Clustered. Biometrics, 54(2):638–645.

[22] Nickelsburg, G. and Citron, J.-T. (1987). Country Risk and Political Instability. Journal of Development Economics, 25:385–392.

[23] Ozler, S. and Tabellini, G. (1991). External Debt and Political Instability. NBER Working Paper, 3772.

[24] Persson, T. and Svensson, L. E. O. (1989). Why a Stubborn Conservative would Run a Deficit: Policy with Time- InconsistentPreferences. 104(2):325–345.

[25] Sarkees, M. R. and Wayman, F. W. (2010). Resort to war : a data guide to inter-state, extra-state, intra-state, and non-state wars, 1816-2007. CQ Press.

[26] Schunck, R. (2013). Within and between estimates in random-effects models: Advantages and drawbacks of correlated random effects and hybrid models. The Stata Journal, 13(1):65– 76.

[27] Van Rijckeghem, C. and Weder, B. (2004). The Politics of Debt Crises. Centre for Economic Policy Research, Discussion Paper(4683).

[28] Van Rijckeghem, C. and Weder, B. (2009). Political institutions and debt crises. Public Choice, 138(3-4):387–408.

(30)

9 Tables

9.1 Data and Descriptive Statistics

(31)

(32)

obs mean sd min max default 6683 .5063594 .499997 0 1 debtdgdp 5964 11.11714 64.12173 0 1968.513 debtratio 3732 57.68621 51.43135 0 789.833 cpi 5347 42.46003 564.2845 -35.83668 24411.03 growth 5960 3.508025 6.644223 -64.0471 149.973 cab 5876 -3.292207 11.90808 -242.188 106.836 rim 5365 -21.66886 26.98357 -558.1757 295.5053 war 7000 .1144286 .3183538 0 1 pment 5827 .3258967 .4687491 0 1 elec 5851 .2473082 .4314842 0 1 pturn 5838 .1863652 .3894344 0 1 polity 5392 1.922478 7.178432 -10 10 polar 5117 3.104749 1.45185 1 5

(33)

obs mean sd min max High and Upper Middle

default 3754 .3183271 .4658892 0 1 debtdgdp 3488 3.840751 15.59102 0 300.4605 debtratio 2336 51.76068 38.07318 0 342.666 cpi 3130 25.88593 208.3423 -23.8221 7481.664 growth 3460 3.471091 6.846593 -64.0471 149.973 cab 3365 -1.426458 12.97701 -242.188 106.836 rim 3136 -21.88113 30.02962 -558.1757 295.5053 war 3920 .0806122 .2722734 0 1 pment 3189 .4916902 .5000093 0 1 elec 3197 .2630591 .4403631 0 1 pturn 3190 .2 .4000627 0 1 polity 2864 4.189944 7.154663 -10 10 polar 2830 3.604947 1.457824 1 5

Low and Lower Middle

default 2728 .784824 .4110197 0 1 debtdgdp 2441 21.43801 97.51565 0 1968.513 debtratio 1373 67.92038 67.42266 0 789.833 cpi 2183 63.07737 842.6285 -35.83668 24411.03 growth 2465 3.573005 6.358484 -51.03086 106.2798 cab 2476 -5.866825 9.805194 -90.834 46.067 rim 2194 -21.68735 22.02654 -357.5009 32.65697 war 2835 .1633157 .3697187 0 1 pment 2534 .1302289 .3366215 0 1 elec 2550 .2278431 .4195231 0 1 pturn 2544 .1678459 .3738029 0 1 polity 2448 -.6478758 6.244915 -10 10 polar 2207 2.494336 1.17346 1 5