The Unexplained Part of Public Debt

Transcription

1 Iner-American Developmen Bank Banco Ineramericano de Desarrollo (BID) Research Deparmen Deparameno de Invesigación Working Paper #554 The Unexplained Par of Public Deb by Camila F.S. Campos* Dany Jaimovich** Ugo Panizza** *Yale Universiy **Iner-American Developmen Bank, Washingon, D.C. March 2006

2 Caaloging-in-Publicaion daa provided by he Iner-American Developmen Bank Felipe Herrera Library Campos, Camila F.S. Panizza. The unexplained par of public deb / by Camila F.S. Campos, Dany Jaimovich, Ugo p. cm. (Research Deparmen working paper series ; 554) Includes bibliographical references. 1. Debs, Public. 2. Budge deficis. 3. Financial saemens. I. Jaimovich, Dany. II. Panizza, Ugo. III. Iner-American Developmen Bank. Research Dep. IV. Tile. V. Series C dc Iner-American Developmen Bank 1300 New York Avenue, N.W. Washingon, DC The views and inerpreaions in his documen are hose of he auhors and should no be aribued o he Iner-American Developmen Bank, or o any individual acing on is behalf. This paper may be freely reproduced provided credi is given o he Research Deparmen, Iner- American Developmen Bank. The Research Deparmen (RES) produces a quarerly newsleer, IDEA (Ideas for Developmen in he Americas), as well as working papers and books on diverse economic issues. To obain a complee lis of RES publicaions, and read or download hem please visi our web sie a: hp:// 2

3 Absrac 1 This paper shows ha budge deficis accoun for a relaively small fracion of deb growh and ha sock-flow reconciliaion, which is ofen considered a residual eniy, is one of he key deerminans of deb dynamics. Afer having explained he imporance of he sock-flow reconciliaion, he paper shows ha his residual eniy can be parly explained by coningen liabiliies and balanceshee effecs. Keywords: Public Deb, Defici, Balance-Shee Effecs JEL Codes: H63, F34, C82 1 The views expressed in his paper are he auhors and do no necessarily reflec hose of he Iner-American Developmen Bank. The usual caveas apply. Camila Campos: camila.campos@yale.edu, Dany Jaimovich: danyj@conracual.iadb.org, Ugo Panizza: ugop@iadb.org. 3

4 1. Inroducion How do counries ge ino deb? The answer o his quesion may seem rivial. Counries accumulae deb whenever hey run a budge defici (i.e., whenever public expendiure is higher han revenues). In fac, he sandard Economics 101 deb accumulaion equaion saes ha he change in he sock of deb is equal o he budge defici: DEBT DEBT 1 = DEFICIT (1) and ha he sock of deb is equal o he sum of pas budge deficis: DEBT = DEFICIT Whoever has worked wih acual deb and defici daa knows ha Equaion (1) rarely holds and ha deb accumulaion can be beer described as: i= 0 i. DEBT + DEBT 1 = DEFICIT SF (2) where SF is wha is usually called sock-flow reconciliaion. Clearly, Equaion (1) is a good approximaion of deb accumulaion only if one assumes ha of his paper is o describe some of SF is no very large. The purpose SF s main characerisics. The paper shows ha, conrary o wha is usually assumed, he budge defici accouns for a small fracion of he wihin-counry variance of he change in deb over GDP and ha sock-flow reconciliaion plays an imporan role in explaining deb dynamics. The paper also shows ha, on average, SF ends o be posiive and ha here are large cross-counry differences in he magniude of his residual eniy. This suggess ha he magniude of sock-flow reconciliaion is no likely o be purely due o random measuremen error. In paricular, he paper shows ha he problem is especially serious in developing counries and, among his group of counries, he difference beween deb and defici is paricularly large in Lain America and Sub-Saharan Africa. The paper also runs a se of regressions aimed a explaining he main deerminans of he magniude of he sock-flow reconciliaion and finds ha balance-shee effecs due o real depreciaions and coningen liabiliies ha arise a ime of banking crises are srongly correlaed wih he difference beween defici and change in deb. However, he paper also shows ha he regressions can only explain 20 percen of he wihin-counry variance of he sock-flow 4

5 reconciliaion and ha here is sill much ha we do no undersand abou one of he main deerminans of deb accumulaion. While we are no he firs o show ha sock-flow reconciliaion is an imporan par of deb dynamic (see, among ohers IMF, 2003; Marner and Tromben, 2004; European Commission, 2005; Budina and Fiess, 2005), we are no aware of any oher paper ha sysemaically describes he main characerisics of his residual, bu exremely imporan, deerminan of deb accumulaion. The res of he paper is organized as follows. Secion 2 describes our main sources of daa and presens some basic facs on public deb and defici. Secion 3 focuses on a deailed descripion of he sock-flow reconciliaion. Secion 4 runs a se of regressions aimed a explaining he main deerminans of he sock flow reconciliaion. Secion 5 concludes. 2. Daa The purpose of his secion is o describe our daa on fiscal defici and public deb. In his conex, i is worh menioning ha obaining reliable and comparable daa on he sock public deb is a raher difficul exercise. In fac, he IMF Inernaional Financial Saisics (IFS) and IMF Governmen Finance Saisics (GFS), which are he mos common sources of cross-counry daa on governmen saisics, repor daa for a raher limied se of counries. This is even he case for indusrial counries; hese sources do no repor recen daa on public deb for Japan and Ialy, for example. Furhermore, mos cross-counry daases do no make an effor o make he daa comparable across counries (for a discussion of hese issues, see IMF, 2003). 2 Alhough here are now some papers ha aemp o build comparable cross-counry daases on public deb (Cowan e al., 2005; Jeanne and Guscina, 2006; IMF, 2003; Budina and Fiess, 2005), some of hese daa ses are no publicly available and all of hem have a limied counry and ime coverage. As a consequence, we do no rely on hese new daa and only use publicly available sources (hence, he caveas menioned above should be kep in mind). In paricular, we sar wih IFS and GFS and supplemen hem wih daa colleced from naional sources (mosly from he websies or publicaions of he various Minisries of Finance), he UN Economic Commission for Lain America and Caribbean (ECLAC, see Marner and Tromben, 2004), and he Organizaion for Economic Cooperaion and Developmen (OECD). 2 The mos imporan problems include he reamen of sub-naional governmens and he use of gross versus ne deb (for a mehodological noe, see Cowan e al., 2005). 5

6 Using hese various sources, we assemble an unbalanced panel covering 117 counries and consising of approximaely 1,900 observaions. Table A1 in he Appendix liss all he counries included in our daase, he ime coverage for each counry, and summary saisics for deb and defici raios. Our sample includes 24 high-income counries, 59 middle-income counries and 34 low-income counries. The regions wih he larges number of counries are Sub-Saharan Africa (27 counries) and Lain America (25 counries). Souh Asia and Eas Asia are he regions wih he smalles number of counries (five and eigh counries, respecively). While long ime series are available for some counries (e.g., Bahamas, Burundi, Cosa Rica, Iceland, Norway and he US have more han 30 years of daa), for ohers here are very few observaions (Albania, Algeria, Gabon, Sudan, Togo, and Yemen are among he counries wih less han five years of daa). Table 1 shows ha he sample mean of he defici o GDP raio is 4.04 percen and ha average defici ends o decrease wih he level of income. The region wih he highes average defici is Souh Asia (6.5 percen), followed by he Middle Eas (5.6 percen), and Sub-Saharan Africa (4.2 percen). Lain American counries end o have fairly low levels of average defici (jus below he cross-counry average) bu he region is far from being homogeneous and is characerized by he larges variance in he sample. Table 2 repors summary saisics for he deb-o-gdp raio and shows ha he crosscounry average is close o 56 percen. Souh Asia and Sub-Saharan Africa are he regions wih he highes levels of deb (67 and 60 percen, respecively) and Eas Asia and Easern Europe and Cenral Asia are he regions wih he lowes level of deb (35 and 37 percen, respecively). Lain America has a level of deb ha is jus below he sample average and is no much higher han ha of he indusrial counries included in our sample. Again, we find ha Lain America is one of he mos heerogeneous regions in our sample (in his case, second only o Sub-Saharan Africa). As one may expec, we find ha mos of he variance in deb-o-gdp is due o differences across counries (his is he beween sandard deviaion). However, here is also subsanial variance wihin counries. In fac, he wihin sandard deviaion (no repored in he able) is ofen close o 50 percen of he beween sandard deviaion. 6

7 Table 3 focuses on he change in deb divided by GDP ( d, ). 3 If Equaion (1) were o hold, he change in deb should be equal o he budge defici. By comparing Table 2 wih Table 3, we find ha he value of d i, is almos five percenage poins higher han average defici over GDP, indicaing ha more han 50 percen of he average change in deb is no explained by defici. 4 The Table also shows ha while he difference beween i d i, and he defici is fairly small in indusrial counries (abou 0.3 percenage poins), his difference is exremely large in Lain America and Sub-Saharan Africa, where he average defici is abou one-hird he average change in deb. We can now describe he characerisics of he sock-flow reconciliaion by defining he following measure of he difference beween change in deb and defici for counry i a ime. ( DEBT DEBT ) i, i, 1 DEFICITi, δ = 100 (3) Y i, Clearly, δ is jus he sock-flow reconciliaion of Equaion (1) expressed in erms of SF i, GDP ( δ = ). Table 4 describes δ and shows ha he change in deb is nearly five Y i, percenage poins higher han he defici (wih he highes values in Lain America and Sub- Saharan Africa). However, he Table also shows ha here are several counries wih exremely large values of δ (in some cases well above 200 percen). In Lain America, for insance, he difference beween he change in deb and defici has a range of 350 percenage poins (from 73 D 3 D 1 I is imporan o noe ha we do no use he change in he deb-over-gdp raio (i.e., θ, = i 100 ) Y Y 1 D D 1 bu he change in deb divided by GDP a ime (i.e., d, = ( ) i ). As nominal GDP Y Y + g growh (g) ends o be posiive, d i, is usually larger han θ. We use his measure, raher han he sandard θ because we wan o isolae changes in deb from changes in he level of GDP. 4 Using a differen mehodology and a shorer sample, IMF (2003) also finds similar bu less drasic resuls. In paricular, i finds ha more han 25 percen of he increase in he deb-o-gdp raio of a sample of emerging marke counries over he period is due o off balance-shee facors. In a sample of 21 marke-access counries, Budina and Fiess (2005) find ha deb over GDP increased by 22.8 percenage poins from 1994 o 2002, while real GDP grew by 9.3 percen, yielding a change in deb of approximaely 37 percen. The defici (primary plus ineres rae bill) explained abou one-hird of his change while oher facors (including he real exchange rae) explained he remaining wo-hirds. 7

8 o 281). The indusrial counries have he smalles range, bu even in his case he range is close o 30 percenage poins. These exreme values are due eiher o excepional evens or measuremen error. In he second column of Table 5, he average value of δ is compued by dropping he op and boom 2 percen of he disribuion. Afer dropping hese ouliers, we find ha δ has an average value of 3 percen and ha he average values of δ for Lain America and he Middle Eas drop from 7 percen o 4 and 2 percen, respecively. I is also ineresing o see which counries end o have large values of δ. Table 5 summarizes all he episodes for which δ 10 (a full lis of episodes is repored in Tables A2 > and A3 in he appendix). There are 238 counry-years (corresponding o 13 percen of observaions) for which δ 10, and 50 counry-years (3 percen of observaions) for > whichδ < 10. The indusrial counries, Eas Asia, and Souh Asia are he regions wih he lowes number of episodes (and very few episodes where δ < 10 ). Sub Saharan Africa, he Middle Eas and Norh Africa, and Lain America are he regions wih he larges number of episodes. While his paper focuses on change in deb, we obain he same resuls if we use he sandard decomposiion of he change in deb over GDP (θ). 5 Figure 1 shows ha in mos regions he sock flow adjusmen is he main deerminan of deb growh and inflaion is he main deerminan of deb reducion 3. Deb and Defici The previous secion showed ha simple comparisons of average values of defici over GDP and change in deb indicae ha Equaion (1) is far from being a good approximaion of he main deerminans of deb accumulaion and ha wha is usually considered a residual eniy (he 5 The sandard decomposiion akes he following form: DEBT Y DEBT Y PD DEBT ( gr + ) 1 1 = + i π 1 Y Y 1 ( 1+ g) Y 1 DEBT 1 SF + (1 + g) Y where he firs erm on he RHS of he equaion is he conribuion of he primary defici, he second erm is he ineres bill, he hird erm is he conribuion of nominal growh (which can be spli ino real growh and inflaion) and he las erm is he sock-flow adjusmen. 8

9 sock-flow reconciliaion) is a key deerminan of deb accumulaion. In his secion, we use differen sraegies o provide more evidence in his direcion. 3.1 Regressions Analysis One way o assess he imporance of our large panel o esimae he following fixed effecs regression: SF is o divide deb and defici by curren GDP and use = α + + (4) d, i i β * def, i ε, i where α i is a counry fixed effec (he counry fixed effecs conrol for he fac ha he daa come from differen sources, counries have differen levels of deb, and hey use differen mehodologies for compuing deb and defici) and def, is defici over GDP. If Equaion (1) holds, we expec a high R 2 (he regression s R 2 should be 1 if Equaion 1 holds exacly), α i =0, and β =1. Hence, he regression s coefficiens and R 2 i can be used o asses he relaive (un)imporance of he defici in explaining changes in deb. Table 6 repors he resuls of he esimaion of Equaion (4) for differen sub-samples of counries. Column 1 describes he basic paern. Firs of all, we find ha β is greaer han 1 (bu no significanly differen from 1) indicaing ha a 1 percen increase in he defici o GDP raio ends o ranslae ino a 1.3 percen increase of he deb o GDP raio. More ineresingly, he regression s R 2 shows ha, in our sample of counries, deficis explain less han 8 percen of he wihin counry variance of ha SF explains more han 90 percen of he variance. 6 d, and As he low R2 could be due o he presence of ouliers, in Column 2 we drop 47 ouliers (defined as observaions ha have residuals wih an absolue value greaer han 2.5 sandard deviaions). Afer dropping hese ouliers, β drops o 1.18, bu we sill find ha our model can i only explain 23 percen of he variance of d,. Figure 2 plos he fi of he regression repored in i Column 2 and illusraes ha he low R2 is no due o a few episodes wih a paricularly low fi, bu ha mos counries have observaions ha are far away from he regression s line. Column 3 6 We also ran separae regressions for he 58 counries for which here are a leas 15 years of daa. We found ha β had average and median values of approximaely 1 and ranged beween 1.8 (Zaire) and 5.9 (Rwanda). The regressions R2 had an average value of 0.32, a median value of 0.25, and ranged beween (Egyp) and 0.87 (Ialy). There are only four counries (all indusrial) ha have an R2 above 0.8, 16 counries (11 of hem indusrial) for which he R2 is higher han 0.5, and 18 counries for which he R2 is less han

10 of Table 4 addresses he oulier issues by running he same regression as in Column 1 using a median quanile regression wih boosrapped sandard errors (STATA s BSQREG) and shows ha in his case, he coefficien of he defici variable drops o 0.87 and he R2 goes o The remaining columns run separae regressions for differen regions of he world. Column 4 focuses on 29 counries locaed in Sub-Saharan Africa and finds ha he defici explains only 3 percen of he variance of d,. Columns 5 and 6 show ha in Lain America and i he Caribbean (25 counries) and Souh Asia (5 counries), he defici explains beween 5 and 6 percen of he variance of d,. Columns 7 and 8 focus on Eas Asia (8 counries) and he Middle i Eas and Norh Africa (11 counries) and show ha he defici explains beween 14 and 20 percen of he wihin counry variance of d,. The developing region wih he bes fi is Eas i Europe and Cenral Asia (Column 9, 15 counries). In his case, he defici explains 23 percen of he variance of d,. Only in he sub-group of indusrial counries (Column 10, 24 counries) does i he defici explain more han one-quarer of he wihin counry variaion of d, bu even in his i case, he regression can only explain half of he variance of he dependen variable. 3.2 Theoreical R2 As an alernaive way o describe he paern documened above, we build a measure aimed a deermining which counries have he larges deviaion from he heoreical ideniy Clearly, such a measure canno be he counry average of d = def. δ described in Table 5 because negaive and posiive values of δ would compensae each oher. One possibiliy would be o adop a sraegy similar o he one of he previous secion and run counry-by-counry regressions of Δ DEBT over DEFICIT and use he fi of hese regressions (heir R2) as a measure of how much a counry deviaes from o differeniae counries ha have a good fi in which d = def. One problem wih his sraegy is ha i would no help d = def holds, from counries ha have a good fi bu where he relaionship beween deb and defici can be beer described wih an equaion of he ype: d = α + β * def + ε wih α 0 and β 1. An index ha addresses hese problems and relaes o a regression s R2 can be defined as: 10

11 T ( δ i, ) = 1 φ i = T (5) 2 ( d di ) i, = 1 2 Noe ha φ i is always non-negaive and naurally relaes o he R2 of a regression of d i, over def. In fac, if we wrie d = α + β * def + ε and, if insead of esimaing he regression s parameer, we force = 0 α and β = 1, he R2 of he model would be 1-φ i. Hence, if he rue parameers describing he relaionship beween deb and defici were α = 0 and β = 1, φ i would be equal o 0. Thus, higher values of φ i indicae larger deviaions of he rue parameers from = 0 α and β = 1. Figure 3 illusraes he heoreical disribuion ofφ i for differen values of β under he assumpions ha α = 0, α = 10, and α = 10. The figure shows ha when α = 0 he disribuion is asymmerical wih φ i rapidly going owards infinie when β ends o 0, and φ i converging o around 1.5 when β goes o infinie, he figure also shows ha φ is equal o 0 when β =1. When α = 10, he disribuion becomes monoone bu sill going o i infinie when β goes o 0 and converging o approximaely 1.5 when β goes o infinie. When α = 10 he disribuion reaches a minimum when β is around 4 and hen sars increasing and, again, converges a around 1.5. Figure 4 shows he values of φ i for our sample of counries. Few counries have a value of φ i close o 0 and mos counries are concenraed in he range. In paricular, 15 percen of counries have values of φ i ha are below 0.5 (he lowes value, 0.009, is for Finland), 30 percen of counries have values ha range beween 0.5 and 1, 35 percen of counries have values ha range beween 1 and 1.5, and he remaining 20 percen have higher values. Table 7 shows ha he mean and median of he disribuion of φ i is approximaely 1 and ha, as expeced, he indusrial counries have he lowes value of φ i and Lain America and he Middle Eas have he highes values of φ i. 7 7 I may seem surprising ha while he heoreical disribuion is highly skewed, he daa of Table 7 indicae ha he mean is idenical o he median. This is due o he fac ha Table 7 does no include four counries ha have values of φ greaer han 4 (hese counries are Esonia, Seychelles, Luxembourg, and Sudan). If we include hese counries, he median goes o 1.05, bu he average jumps o

12 3.3 Deb Explosions So far, we documened ha here are a large differences beween defici and change in deb. Now we explore wheher he difference beween hese wo variables is posiively correlaed wih deb growh. Figure 5 plos he relaionship beween he growh rae of deb over GDP (defined as θ ( D Y D Y ) 100 = i, i, i, 1 i, 1 ρ = def i, di, ) and he raio beween defici and change in deb (defined as ). 8 I shows ha a relaively low levels of deb growh (below 5 percen per year), he defici explains approximaely 80 percen of he change of deb. However, when deb sars growing a a faser rae, he share of deb explained by defici drops dramaically. In paricular, he figure shows ha when annual deb growh reaches 10 percen of GDP, he defici explains less han 40 percen of deb growh. Table 8 regresses θ over ρ (conrolling for counry fixed effecs) and confirms ha here is a negaive and saisically significan relaionship beween hese wo variables. While he fi of he regression is raher poor, he able shows ha he fi improves if exreme values of θ are no considered (compare, for insance, Column 1 wih Column 3 where episodes in which θ >50 are dropped). The able also shows ha he relaionship beween θ over ρ does no vary much across groups of counries. As a las exercise, we look a deb explosions (defined as episodes in which θ >10); Table 9 summarizes he daa and Table A4 liss all he episodes. The firs panel of Table 9 shows ha in he 172 episodes for which θ >10 (9 percen of he counry-years for which we have daa), he average increase in deb over GDP was close o 28 percenage poins, he average change in deb was around 46 percenage poins (he difference beween hese wo values is nominal GDP growh which, in presence of high inflaion, can be very high), and he average raio beween hese wo variables was 70 percen. The fourh column of he able shows ha in our sample of deb explosions, average defici was close o 10 percen of GDP and he raio beween defici and change in deb was abou 27 percen. This is close o one-hird of he same raio during normal imes (when 10> θ >0 he raio beween defici and change in deb is 75 percen). The able also shows ha he regions wih he highes occurrence of deb explosions are Lain America and Sub-Saharan Africa (41 and 66 episodes, respecively) and ha Eas Europe 8 We smooh he curve wih a bandwidh of

13 and Sub-Saharan Africa are he regions wih he lowes average raio beween defici and change in deb (18 and 13 percen, respecively). Since he average values discussed above may be driven by exreme values of θ, we resric he sample in he second panel of Table 9 o 104 episodes for which θ ranges beween 10 and 20 percen. In his case, we find ha he average increase of he deb-o-gdp raio is approximaely 14 percen, he average change in deb is 24 percen and he average raio beween hese wo variables is 68 percen (basically idenical o he op panel of he able). The fourh column of he able shows ha he average defici is 7 percen and ha he raio beween average defici and change in deb is 29 percen, which again is close o he op panel of he able. As before, we find ha Lain America and Sub-Saharan Africa have he highes occurrence of deb explosions (18 and 36, respecively), bu now we find ha he Middle Eas and he indusrial counries have a number of episodes ha are no much lower han hose of Lain America. In fac, we now find ha Lain America has he second lowes (afer he indusrial counries) relaive share of deb explosions. This confirms ha deb explosions in Lain America end o be very large. In fac, Lain America is he only region in he world where here are more episodes in which deb grows by more han 20 percen of GDP han episodes in which deb grows beween 10 and 20 percen of GDP. 4. Wha Drives he Difference? Afer having documened ha here are large differences beween deficis and change in deb, we now run a se of regressions aimed a exploring he deerminans of hese differences. We sar by esimaing he following model: δ = (6) α i βx i, γπ i, ε i, where α i is a se of counry fixed effecs, explain he difference beween defici and change in deb, and X i, a se of counry-year specific variables ha can π is a measure of inflaion (defined as ln(1+inf)). Alhough we do no have a clear heory of how inflaion should affec δ i,, we include his variable because he various componens of δ are nominal variables measured in differen periods of ime (a sock a ime, a sock a ime -1 and wo flow variables measured beween -1 and ). Hence, whenever he defici is differen from he change in deb, 13

14 he value of δ should be posiively correlaed wih nominal GDP growh, which is heavily influenced by inflaion. One reason why he change in deb could be higher han he recorded defici is he valuaion effecs due o currency depreciaions in he presence of foreign currency deb. To explore his possibiliy, we sar by focusing on developing counries (indusrial counries do no have large socks of foreign currency deb) and use daa from he World Bank s Global Developmen Finance (GDF) o creae hree dummy variables ha classify all developing counries ino hree groups of equal size. 9 The hree dummies are defined as follows: (i) LOW akes a value of 1 for all counry-years where he exernal deb-o-gdp raio is below 38 percen; (ii) MEDIUM akes a value of 1 for all counry-years where he exernal deb-o-gdp raio ranges beween 38 and 64 percen; (iii) HIGH akes a value of 1 for all counry-years where he exernal deb-o-gdp raio is above 64 percen. Nex, we inerac he hree dummies wih he change in he real exchange rae (DRER, an increase in DRER corresponds o a real depreciaion). Column 1 of Table 10 repors he resuls of our baseline esimaion. As expeced, we find ha inflaion has a posiive and saisically significan coefficien. Furhermore, we find ha currency depreciaions are posiively and significanly correlaed wih δ, a finding ha provides evidence of he presence of balance-shee effecs. More ineresingly, we find ha he effec of currency depreciaions is paricularly large in counries wih high levels of exernal deb. Consider, for insance, a real depreciaion of 30 percen (no an uncommon even in some of he counries included in our sample). In counries characerized by low or medium levels of exernal deb, such a depreciaion is associaed wih an increase of δ of approximaely hree o four percenage poins, bu in counries wih high levels of deb, a similar depreciaion would insead cause δ o increase by more han 10 percenage poins. A he boom of he able we show ha he difference beween coefficiens is also saisically significan (his is no he case for he difference beween he coefficiens associaed wih low and medium exernal deb). Nex, we include indusrial counries and assume ha his se of counries has no foreign currency denominaed exernal deb. Therefore, he regression coefficiens should be inerpreed 9 Since he GDF daa have informaion for oal exernal deb, we are implicily assuming ha mos exernal deb is public (or generaes coningen liabiliies of he public secor). We checked he validiy of his assumpion by compuing he correlaion beween GDF daa on oal exernal deb and IFS daa on public exernal deb and found ha his correlaion is

15 as follows: DRER measures he effec of real depreciaions in indusrial counries; DRER+DRER*LOW measures he effec of a real depreciaion in developing counries wih low levels of exernal deb; DRER+DRER*MEDIUM measures he effec of a real depreciaion in developing counries wih average levels of exernal deb; and DRER+DRER*HIGH measures he effec of a real depreciaion in developing counries wih high levels of exernal deb. Column 2 shows ha he coefficien of DRER is low and no saisically significan, indicaing ha here are no balance-shee effecs in indusrial counries. As before, we find ha balance-shee effecs are imporan in developing counries and ha he effec of a real depreciaion in all hree groups of developing counries is significanly differen (boh in economic and saisical erms) from he effec of a depreciaion in indusrial counries. Finally, we sill find ha balance-shee effecs end o be paricularly imporan in counries wih high levels of deb. Column 3 explores he role of defaul, w expec defauls o be associaed wih deb reducion and hence negaively correlaed wih δ. To capure he effec of defaul, we use daa from Sandard and Poor s and build a dummy variable ha akes a value of 1 around he las year of a defaul episode (in paricular, i akes a value of 1 in he las year of he episode and in he year before and he year afer he las year of he episode). Nex, we build a defaul dummy ha akes a value of 1 in he las year of a Paris club rescheduling and hen anoher dummy ha akes a value of 1 whenever he GDF repors ha a counry has rescheduled is deb. Finally, we build a dummy called DEFAULT ha akes a value of 1 whenever one of he previously described dummies akes a value of 1. Column 3 shows ha he defaul dummy has he expeced negaive sign bu ha he coefficien is small and no saisically significan (we obain similar resuls if we use he hree dummies separaely). Column 4 uses daa from Caprio and Klingebiel (2003) o explore he role of banking crises. These are imporan evens because hey generae a series of coningen liabiliies and oher off-balance shee aciviies ha can ranslae ino deb explosions. As expeced, we find ha he coefficien of he banking crisis dummy is posiive and saisically significan. The coefficien is also quaniaively imporan, indicaing ha he average banking crisis is associaed wih an increase of hree percenage poins in δ. Column 5 joinly includes all he variables discussed above. We find ha he resuls are qualiaively similar o previous ones, bu ha he coefficien of DRER*MEDIUM is no longer saisically significan (however, DRER+ DRER*MEDIUM remains significan) and ha he 15

16 same is rue for banking crisis. In he las column of he able, we conrol for year fixed effecs (which implicily conrol for global shocks) and show ha heir inclusion does no affec our basic resuls. I is ineresing o noe ha he se of conrols included in he regressions of Table 10 explains abou 20 percen of he variance of δ and ha he counry fixed effecs explain abou 30 percen of he variance of δ (see las row of Table 10). This indicaes ha counry specific facors explain mos of he variance of δ and corroboraes he findings of Table 4, which showed ha here are large cross-counry differences in he average value of δ. There are wo possible explanaions for his finding. The firs has o do wih he fac ha measuremen errors ha lead o an underesimaion of he defici are more imporan in some counries han in ohers, which is probably relaed o he fac ha poorer counries have less sophisicaed accouning and budgeing sysems. The oher has o do wih he fac ha he imporance of coningen liabiliies ha lead o deb explosions vary across counries and ha our se of conrols does no capure all hese coningen liabiliies. 10 Table 11 includes GDP growh in he analysis. The firs column shows ha deb ends o grow more han defici during periods of slow GDP growh. Column 2 subsiues GDP growh wih wo dummies variables ha ake a value of 1 during periods of high growh (GOOD TIMES) and periods of slow growh (BAD TIMES). 11 Also in his case, we find ha deb ends o grow faser han he defici during bad imes and slower han he defici during good imes. Column 3 augmens he regression in Column 1 wih he se of conrols in Table 10. We find ha he sign of GDP growh remains negaive bu he coefficien drops by one-hird and is no longer saisically significan. Column 4 uses he se of conrols in Table 10 and he GOOD TIMES and BAD TIMES dummies. In his case, we sill find ha he wo dummies have he opposie sign and are boh saisically significan. In Table 12 we esimae a se of regressions similar o hose in Table 10 bu now subsiue δ wih d and include def in he se of conrols. This is equivalen o esimaing he model of Table 10 by relaxing he resricion ha he coefficien of def is 1. We find ha he def coefficien is always smaller han 1 bu ha ha his coefficien is never significanly differen 10 Anoher key difference is in he size of he regional governmen, which is ofen no well capured by our daa. 11 GOOD TIMES akes a value of 1 when growh is one sandard deviaion above he counry average, BAD TIMES akes a value of 1 when growh is one sandard deviaion below he counry average. REGULAR TIMES is he excluded dummy. 16

17 from 1. All our oher resuls are unchanged (his was expeced because Table 6 already indicaed ha he defici by iself explains an exremely small share of he wihin-counry variance of he change in deb). One problem wih he regressions of Tables 10, 11 and 12 is ha hey assume a linear relaionship beween he dependen variable and he se of independen variables. Therefore, he esimaed resuls migh be driven by exreme values of δ. To address his issue, we relax he lineariy assumpion and run wo ses of Probi regressions. In he firs se of Probis, he dependen variable is a dummy ha akes a value of 1 for all counry years in he op decile of he disribuion of δ. In he second se of Probis, we repea he experimen using he boom decile of he disribuion of δ. 12 Table 13 repors he resuls for evens in he op decile (in his group of evens, δ ranges beween 12.7 and 282 and has an average value of 44.5). We find ha mos of he resuls are similar o hose in Table 10. In paricular, Column 1 shows ha he relaionship beween real depreciaions and he probabiliy of observing an exreme even of δ increases wih he level of exernal deb. Column 2 shows ha in indusrial counries, real depreciaions have a negaive (bu no saisically significan) correlaion wih he probabiliy of observing an exreme even of δ. This column also shows ha in counries wih high levels of exernal deb, depreciaions are highly correlaed wih he probabiliy of observing an exreme even. One puzzling resul of Table 13 is ha he coefficien of he DEFAULT dummy is large, significan, and posiive (Column 3). This is exacly he opposie of wha we expeced, and may have o do wih he fac ha defauled deb is no immediaely subraced from he sock of public deb. The coefficien of he BANKING CRISIS dummy variable insead has he expeced posiive sign. Besides being saisically significan, he impac of his variable is also economically imporan. In paricular, he poin esimaes indicae ha a banking crisis is associaed wih a 10 percen increase in he probabiliy of observing an exreme even of δ. Table 14 focuses on evens in he boom decile of δ (in his group of evens, δ ranges beween -116 and 3.4 and has an average value of -10.9). As expeced, we find ha depreciaions are negaively correlaed wih hese ypes of evens bu he coefficiens are rarely significan. In general, we find ha our model does a very poor job of explaining hese evens. 12 The resuls do no change if we define he dummies using he δ >10 hreshold. 17

18 5. Conclusions The purpose of his paper was o documen he fac ha wha is ofen considered a residual eniy is indeed one of he key deerminans of deb dynamic. Afer demonsraing he imporance of he sock-flow reconciliaion, his paper shows ha his residual eniy can be parly explained by coningen liabiliies and balance-shee effecs. These resuls sugges ha building a safer deb srucure and implemening policies aimed a avoiding he creaion of coningen liabiliies are key o avoiding deb explosions (for conrasing views on how his can be achieved, see Goldsein and Turner, 2004 and Eichengreen, Hausmann and Panizza, 2003). However, his paper also shows ha a large fracion of he variance of he sock-flow reconciliaion canno be explained by balance-shee effecs and our simple regressions One variable ha is likely o be imporan bu ha we do no conrol for is he effec of cour decisions ha force he governmen o make paymens (o public secor workers, for insance) ha were no budgeed. We would like o hank Vio Tanzi for poining his ou. 18

19 References Budina, N. and N. Fiess Public Deb and is Deerminans in Marke Access Counries. Washingon, D.C.: The World Bank. Caprio, G., and D. Klingebiel Episodes of Sysemaic and Borderline Financial Crises. Washingon, DC, Unied Saes: World Bank. Mimeographed documen. hp://econ.worldbank.org/view.php?ype=18&id=23456 Cowan, K., E. Levy-Yeyai, U. Panizza and F. Surzenegger Public Deb in he Americas. (In progress). Eichengreen, B., R. Hausmann and U. Panizza Currency Mismaches, Deb Inolerance and Original Sin: Why Are No he Same and Why I Maers. NBER Working Paper Cambridge, Unied Saes: Naional Bureau of Economic Research. European Commission General Governmen Daa. General Governmen Expendiure, Balances and Gross Deb. Brussels, Belgium: European Commission. Goldsein, M., and P. Turner Conrolling Currency Mismaches in Emerging Markes. Washingon D.C.: Insiue for Inernaional Economics. Inernaional Moneary Fund Public Deb in Emerging Markes: Is i Too High? World Economic Oulook, Chaper 3. Washingon, DC, Unied Saes: Inernaional Moneary Fund. Jeanne, O. and A. Guscina Governmen Deb in Emerging Marke Counries. A New Daase. Mimeo, IMF Levy-Yeyai, E., and F. Surzenegger Mehodological Noe on he Consrucion of he Deb Daabase. Buenos Aires: Universidad Torcuao Di Tella. Marner, R., and V. Tromben Public Deb Indicaors in Lain American Counries: Snowball Effec, Currency Mismach and he Original Sin. In: Public Deb. Perugia, Ialy: Banca d Ialia. Reinhar, C., K. Rogoff and M. Savasano Deb Inolerance. Brookings Papers on Economic Aciviy 1:

20 Table 1. Defici over GDP Counry Group μ σ (%) (%) Overall Beween Min (%) Max (%) N. of counries N. of observaions All Counries By Region EAP ECA IND LAC MNA SAS SSA By Income Groups Low Medium High The income group and regional classificaions are hose used by he World Bank Table 2. Deb over GDP Counry Group μ σ (%) Min Max N. of N. of (%) (%) (%) counries observaions Overall Beween All Counries By Region EAP ECA IND LAC* MNA** SAS SSA By Income Groups Low Medium High The income group and regional classificaions are hose used by he World Bank. * Excludes Guyana ** Excludes Israel 20

21 Table 3. Change in Deb over GDP Counry Group μ σ (%) Min Max N. of N. of (%) (%) (%) counries observaions Overall Beween All Counries By Region EAP ECA IND LAC MNA SAS SSA By Income Groups Low Medium High The income group and regional classificaions are hose used by he World Bank Table 4. Change in Deb Minus Defici (δ) Counry μ (%) σ (%) Min Max N. of N. of Group Wihou (%) (%) counries observaions All Ouliers* Overall Beween All Counries By Region EAP ECA IND LAC MNA SAS SSA By Income Groups Low Medium High The income group and regional classificaions are hose used by he World Bank. *Ouliers are he op and boom 2 percen of he disribuion. 21

22 Table 5. Episodes wih δ 10 > Episodes wih δ>5 Episodes wih δ<-5 Number Share of oal Number Share of oal EAP ECA IND LAC MNA SAS SSA All Counries Table 6. Change in Deb over GDP and Defici (regressions wih counry fixed effecs) (1) (2) (3) (4) (5) Defici (0.226)*** (0.052)*** (0.066)*** (0.430)** (0.354)*** N. Obs Nr. Cy R Sample All No Quanile SSA LAC Counries Ouliers Regression (6) (7) (8) (9) (10) Defici (0.295)** (0.361)*** (0.840)*** (0.346)*** (0.056)*** N. Obs Nr. Cy R Sample SAS EAP MNA ECA IND Robus sandard errors in parenhesis. 22

23 Counry Group μ (%) Table 7. Φ Index σ (%) Median (%) Max (%) Min (%) N. of counries All Counries By Region EAP ECA IND LAC MNA SAS SSA By Income Groups Low Medium High Table 8. Change in Deb and ρ (conrolling for counry fixed effecs) (1) (2) (3) (4) (5) θ (0.002)*** (0.003)*** (0.005)*** (0.013) (0.008) Consan (0.030)*** (0.033)*** (0.036)*** (0.121)*** (0.079)*** Observaions Number of Counries R-squared Sample θ>0 0<θ<100 0<θ<50 EAP, θ>0 ECA, θ>0 (6) (7) (8) (9) (10) θ (0.012) (0.004) (0.006)*** (0.003)** (0.008) Consan (0.049)*** (0.061)*** (0.044)*** (0.068)*** (0.179)*** Observaions Number of Counries R-squared Sample IND, θ>0 LAC, θ>0 SAS, θ>0 SSA, θ>0 MNA, θ>0 23

24 Table 9. Deb Explosions θ d θ/d def def/d N Share All Episodes wih θ>10 ALL % % % EAP % % % ECA % % % IND % % % LAC % % % MNA % % % SAS % % % SSA % % % All Episodes wih 10<θ<20 ALL % % % EAP % % % ECA % % % IND % % % LAC % % % MNA % % % SAS % % % SSA % % % 24

25 Table 10: The Deerminans of δ (1) (2) (3) (4) (5) (6) INFLATION (11.454)** (11.199)** (11.285)** (10.775)** (11.346)** (11.581)** DRER*LOW (6.522)** (6.732)* (6.787)* DRER*MEDIUM (5.059)** (5.171)* DRER*HIGH (10.423)*** (10.469)*** (10.588)*** (10.738)** DRER (1.513) (1.589) (3.715)** DEFAULT (2.015) (1.981) (1.963) BANKING CRISIS (1.918)* (1.908) (1.909) R-squared (wihin) Observaions Nr. of Counries Sample Developing Counries All Counries All Counries All Counries All Counries All Counries Fixed Effecs Counry Counry Counry Counry Counry Cry.-Year DRER*LOW=DRER*MED DRER*HIGH=DRER*MED R-squared wih counry FE Robus sandard errors in parenheses. * Significan a 10 percen; ** significan a 5 percen; *** significan a 1 percen. 25

26 Table 11. The Deerminans of δ (1) (2) (3) (4) INFLATION (11.130)** (10.838)** (12.533)** (11.305)** DRER*LOW (7.496)** (6.276)** DRER*MEDIUM (5.526) (5.874) DRER*HIGH (11.147)*** (10.440)*** DRER (1.814) (1.828) DEFAULT (2.062) (1.860) BANKING CRISIS (2.519) (1.979) GDP GROWTH (0.118)*** (0.130) GOOD TIMES DUMMY (0.857)** (0.847)* BAD TIMES DUMMY (1.241)*** (1.200)** Observaions Nr. of Counries R-squared (wihin) Fixed Effecs Counry Counry Counry Counry Sample All Counries All Counries All Counries All Counries 26

27 Table 12. The Deerminans of d (1) (2) (3) (4) (5) (6) DEFICIT/GDP (0.185)*** (0.143)*** (0.148)*** (0.149)*** (0.144)*** (0.153)*** INFLATION (11.486)** (11.213)** (11.342)** (10.824)** (11.343)** (11.559)** DRER*LOW (6.461)** (6.505)* (6.558)* DRER*MEDIUM (5.040)** (5.190)* DRER*HIGH (10.378)*** (10.497)*** (10.615)*** (10.724)** DRER (1.545) (1.613) (3.729)** DEFAULT (2.011) (1.977) (1.966) BANKING CRISIS (1.927)* (1.914) (1.917) R-squared (wihin) Observaions Nr. of Counries Sample Developing All All All All All Counries Counries Counries Counries Counries Counries Fixed Effecs Counry Counry Counry Counry Counry Cry.-Year DRER: LOW=MED DRER: HIGH=MED R-squared wih counry FE Robus sandard errors in parenheses. * Significan a 10 percen; ** significan a 5 percen; *** significan a 1 percen. 27

28 Table 13. Probi Regressions for Episodes in Top δ Decile (1) (2) (3) (4) (5) (6) INFLATION (0.084)*** (0.072)*** (0.060)*** (0.077)*** (0.055)** (0.064)** DRER*LOW (0.169) (0.159) (0.158) (0.179) DRER*MEDIUM (0.115)* (0.122)** (0.120)** (0.128) DRER*HIGH (0.136)*** (0.136)*** (0.129)*** (0.147)** DRER (0.075) (0.080) (0.099) BANK CRISIS (0.029)*** (0.028)*** (0.026)* DEFAULT (0.032)*** (0.032)*** (0.033)*** Observaions Nr. of Counries Sample Developing Counries All Counries All Counries All Counries All Counries All Counries FE NO NO NO NO NO YEAR Sandard errors in parenheses. * Significan a 10 percen; ** significan a 5 percen; *** significan a 1 percen. 28

29 Table 14. Probi Regressions for Episodes in Boom δ Decile (1) (2) (3) (4) (5) (6) INFLATION (0.035) (0.030) (0.029) (0.028) (0.032) (0.032) DRER*LOW (0.184) (0.210) (0.216) (0.211) DRER*MEDIUM (0.168)* (0.201) (0.204) (0.210) DRER*HIGH (0.130) (0.169) (0.165) (0.187) DRER (0.120) (0.125) (0.147) BANK CRISIS (0.026) (0.026) (0.028)** DEFAULT (0.026)** (0.026)* (0.026)** Observaions Nr. of Counries Sample Developing Counries All Counries All Counries All Counries All Counries All Counries FE NO NO NO NO NO YEAR Sandard errors in parenheses. * Significan a 10 percen; ** significan a 5 percen; *** significan a 1 percen. 29

30 Table A1. Counries Included in he Sample Counry Code Region Iniial year Final year Deb/GDP Defici/GDP δ φ FIJI* FJI EAP INDONESIA IDN EAP KOREA KOR EAP MALAYSIA MYS EAP MONGOLIA MNG EAP PAPUA NEW GUINEA PNG EAP SOLOMON ISLANDS* SLB EAP THAILAND THA EAP ALBANIA ALB ECA BELARUS BLR ECA CROATIA HRV ECA CYPRUS CYP ECA CZECH REPUBLIC CZE ECA ESTONIA EST ECA GEORGIA GEO ECA HUNGARY HUN ECA LATVIA LVA ECA LITHUANIA LTU ECA POLAND POL ECA RUSSIA RUS ECA SLOVAK REPUBLIC SVK ECA TAJIKISTAN TJK ECA TURKEY* TUR ECA AUSTRALIA AUS IND AUSTRIA AUT IND BELGIUM BEL IND CANADA CAN IND DENMARK DNK IND FINLAND FIN IND FRANCE FRA IND GERMANY DEU IND GREECE GRC IND ICELAND ISL IND IRELAND IRL IND ITALY ITA IND JAPAN JPN IND LUXEMBOURG* LUX IND MALTA* MLT IND NETHERLANDS NLD IND NEW ZEALAND NZL IND NORWAY NOR IND PORTUGAL PRT IND SPAIN ESP IND SWEDEN SWE IND SWITZERLAND CHE IND UNITED KINGDOM GBR IND UNITED STATES USA IND ARGENTINA ARG LAC BAHAMAS, THE BHS LAC BARBADOS BRB LAC BOLIVIA BOL LAC BRAZIL* BRA LAC CHILE CHL LAC COLOMBIA COL LAC COSTA RICA CRI LAC ECUADOR ECU LAC EL SALVADOR SLV LAC GRENADA GRD LAC GUATEMALA GTM LAC GUYANA GUY LAC