Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
PUBLIC DEBT AND ECONOMIC GROWTH: IS THERE A CAUSAL EFFECT?
UGO PANIZZA UNCTAD
ANDREA F. PRESBITERO
Università Politecnica delle Marche MoFiR
CeMaFiR
MoFMoF iR working paper n° iR working paper n° 66 55
April 2012
Public Debt and Economic Growth:Is There a Causal Effect?
Ugo Panizza Andrea F. Presbitero∗
April 2, 2012
Abstract
This paper uses an instrumental variable approach to study whether public debt has a causaleffect on economic growth in a sample of OECD countries. The results are consistent with theexisting literature that has found a negative correlation between debt and growth. However,the link between debt and growth disappears once we instrument debt with a variable thatcaptures valuation effects brought about by the interaction between foreign currency debtand exchange rate volatility. We conduct a battery of robustness tests and show that ourresults are not affected by weak instrument problems and are robust to relaxing our exclusionrestriction.
JEL Codes: F33, F34, F35, O11Keywords: Government Debt, Growth, OECD countries.
∗Ugo Panizza, UNCTAD and The Graduate Institute, Geneva. E-mail: [email protected]. Andrea F.Presbitero, Department of Economics and Social Sciences – Università Politecnica delle Marche (Italy), Moneyand Finance Research group (MoFiR) and Centre for Macroeconomic and Finance Research (CeMaFiR). E-mail:[email protected]; personal webpage: https://sites.google.com/site/presbitero/. We would like to thank,without implications, Alessandro Missale for sharing his data on the composition of government debt, MadhuMohanty for providing us with details on the estimation procedures in his joint paper with Stephen Cecchettiand Fabrizio Zampolli, Aart Kraay for sharing his code for the calculations of Bayesian bounds in IV estimatesand Jack Lucchetti for translating that code in GRETL. We would also like to thank Mr. Denis Pètre of BISand the staff of the debt management offices of several OECD countries for helping us collecting data on debtcomposition. Finally, we would like to thank Mackie Bahrami, Carlo Cottarelli, Alessandro Missale, Andy Powelland Jaejoon Woo for comments on an earlier draft of the paper. The views expressed in this paper are theauthors’ only and need not reflect, and should not be represented as, the views of the United Nations.
1
1 Introduction
Do high levels of public debt reduce economic growth? This is an important policy question. Apositive answer would imply that, even if effective in the short-run, expansionary fiscal policiesthat increase the level of debt may reduce long-run growth, and thus partly (or fully) negate thepositive effects of the fiscal stimulus.
Most policymakers do seem to think that debt reduces long-run economic growth.1 This viewis in line with the results of a growing empirical literature which shows that there is a negativecorrelation between public debt and economic growth in advanced and emerging economies, andfinds that this correlation becomes particularly strong when public debt approaches 100 percentof GDP (Reinhart and Rogoff, 2010a,b; Kumar and Woo, 2010; Checherita and Rother, 2010;Cecchetti, Mohanty and Zampolli, 2011).2
Correlation, however, does not imply causation. The link between public debt and economicgrowth could be driven by the fact that it is low economic growth that leads to high levels ofdebt. Alternatively, the observed correlation between debt and growth could be due to a thirdfactor that has a joint effect on these two variables. Establishing the presence of a causal linkgoing from debt to growth requires finding an instrumental variable that has a direct effecton debt but no direct (or indirect, except for the one going through debt) effect on economicgrowth.
In this paper, we propose a new instrument for public debt and show that instrumentalvariable regressions do not provide evidence that public debt causes growth in a sample ofOECD countries. Our instrumental variables strategy relies on the fact that, in the presence offoreign currency debt, changes in a country’s exchange rate have a direct and mechanical effecton the debt-to-GDP ratio. We thus collect detailed data on the currency composition of publicdebt and match them with data on bilateral exchange rates to build a variable that capturesvaluation effects brought about by exchange rate movements.
The first requisite for a good instrument is relevance: the instrument needs to be correlatedwith the endogenous variable. A battery of weak instrument tests confirm that our instrumentis relevant. Nevertheless, we are aware that, by itself, our valuation effect variable does notsatisfy the second requirement for a good instrument (instrument exogeneity). Valuation effectsare mechanically correlated with the exchange rate which, in turn, may affect economic growth(Rodrik, 2008). Our instrument is also correlated with the share of foreign currency debt, and thelatter may have an effect on economic growth, perhaps through increased volatility (Eichengreen,Hausmann and Panizza, 2005).
However, we see no reason why our valuation effect variable, built using bilateral data, shouldhave a direct effect on economic growth once we control for debt composition and the behaviorof the effective exchange rate.
While we are convinced of the soundness of our exclusion restriction, we cannot provide aformal test of its validity because our model is exactly identified, and exclusion restrictions canonly be tested in overidentified models. We will thus use a Bayesian approach developed byKraay (2012) to show that small violations of our exclusion restriction do not affect our results.
1For instance, the Director of the IMF Fiscal Affairs Department argued that: “in addition to problems forgrowth arising from a debt crisis, one should also be worried about problems for growth arising from high, evenif stable debt.” (Cottarelli, 2011). A related question that we do not address in this paper is whether fiscalexpansions always lead to an increase in debt. For a discussion covering different points of view see: Cafisoand Cellini (2012), DeLong and Summers (2012), Gros (2011), Krugman (2011), Sutherland, Hoeller and Merola(2012), and UNCTAD (2011).
2There is also an older literature which focuses on developing countries and looks at the relationship betweenexternal debt and economic growth. This literature also finds that the debt is negatively correlated with economicgrowth and that this correlation becomes particularly strong when debt reaches a certain threshold (Pattillo,Poirson and Ricci, 2011; Cordella, Ricci and Ruiz-Arranz, 2010).
2
The existing literature found that the negative correlation between debt and growth becomesparticularly strong when debt approaches 100 percent of GDP (Reinhart and Rogoff, 2010a,b;Cecchetti, Mohanty and Zampolli, 2011; Checherita and Rother, 2010). We search for sucha threshold in our data, but do not find strong evidence of a structural break. Since ourfindings do not rule out the presence of a threshold in the longer and larger datasets studiedby other authors, we also experiment with instrumental variable regressions in high and low-debt subsamples. In low-debt countries, we still find that the negative correlation between debtand growth disappears once we instrument debt with our valuation effect variable. However,our instrument does not perform well in the high-debt subsample. We thus use an alternativeidentification strategy based on Fisher’s (1966) covariance restrictions, and find no evidence ofa negative correlation between debt and growth in high-debt countries.
In order to compare our results with those of the existing literature, we build on the influentialpaper by Cecchetti et al. (2011). Our work is also closely related to the historical work ofReinhart and Rogoff (2009, 2010a,b). Our scope, however, is much narrower than that of thesepapers. In particular, we concentrate on government debt and, unlike Cecchetti et al. andReinhart and Rogoff, we do not explore the complex interactions between private and publicdebt.
2 Addressing Endogeneity
In this section we discuss the endogeneity problem that affects the literature on debt and growth,assess the likely direction of the bias, describe how the existing literature dealt with endogeneity,and then illustrate our instrument.
The easiest way to describe the endogeneity problem and assess the likely direction of thebias is to use a simple bivariate model in which growth (G) is a function of debt (D):
G = a+ bD + u, (1)
and debt is a function of growth:D = m+ kG+ v. (2)
The OLS estimator of b is then given by:
b =bσ2v + kσ2uσ2v + k2σ2u
,
and the bias of the OLS estimates is:
E(b)− b =k(1− bk)
σ2v/σ2u + k2
. (3)
Equation (3) shows that OLS estimations are unbiased if k = 0 (i.e., if debt is not endogenous)or if, for some coincidence, bk = 1. If k is negative (as it is likely to be) and bk < 1 (this isalways the case if b is non-negative), OLS estimates are negatively biased.
Of course, we are not the first to recognize that debt is likely to be endogenous. Theexisting literature tries to address endogeneity by using lagged values of the debt-to-GDP ratio(Cecchetti, Mohanty and Zampolli, 2011), GMM estimations with internal instruments (Kumarand Woo, 2010; Presbitero, 2012), and by instrumenting the debt-to-GDP ratio with the averagedebt of partner countries (Checherita and Rother, 2010).
While these are useful first steps, we think that they are inadequate to fully address en-dogeneity. The use of lagged variables is problematic because debt and growth tend to be
3
persistent.3 The high persistence of debt ratios also limits their validity as internal instrumentsin the standard GMM estimators developed by Arellano and Bond (1991), Arellano and Bover(1995), and Blundell and Bond (1998). Moreover, these difference and system GMM estimatorsare poorly suited for macroeconomic datasets with a relatively small number of cross-sectionalunits (Bond, 2002). Instrumenting public debt with that of neighboring countries is problematicin the presence of global shocks and financial and real spillovers. If a shock to country i’s GDPgrowth has an effect on country j’s GDP growth, the evolution of public debt in country i cannotbe used as an instrument for public debt in country j.4
We propose to address the endogeneity issue by instrumenting the debt-to-GDP ratio withthe valuation effect brought about by the interaction between foreign currency debt and move-ments in the exchange rate.
Formally, let us consider a country that has issued public debt denominated in N currencies(this set of N currencies may include the domestic currency). Let us further define Dij as thestock of debt issued by country i and denominated in currency j, and eij as the log of theexchange rate between currency i and currency j. We then propose to instrument the stock ofpublic debt of country i in period t+ 1 with the following variable:
V Ei,t =
j=N∑j=1
Dij,t(eij,t+1 − eij,t)
j=N∑j=1
Dij,t
. (4)
To be a good instrument, V E needs to be relevant (i.e., it needs to be correlated with publicdebt) and exogenous (i.e., it should not belong in the growth regression). We will start bydiscussing relevance and then move to exogeneity.
Relevance may appear to be guaranteed by the fact that there is a mechanical relationshipbetween V E and the debt-to-GDP ratio. However, things are complicated by the fact thatV E is always equal to zero (and therefore useless as instrument) for countries that do not haveforeign currency debt (because, by definition, eij,t+1 − eij,t = 0) or have very stable exchangerates.5 This would not be a concern if we were to focus on emerging market countries whichtend to have large foreign currency debt and high exchange rate volatility (Hausmann, Panizzaand Rigobon, 2006). Relevance may instead be a problem in our sample of OECD countries.
Table 1 reports country-by-country summary statistics of V E. It shows that the variance ofV E is close to zero in France, Germany, Japan, the Netherlands, and the United States. This isnot surprising as these five countries do not have public debt denominated in foreign currency.In the remaining 12 countries, the variance of V E ranges between 0.15 percent and 1.89 percent. The overall variance of V E is approximately 1 percent. The cross-country variance is 0.26percent and the within-country variance is 0.94 percent.
Whether this variance is sufficient to make of V E a relevant instrument for public debt is anempirical question that can be assessed using appropriate statistical tests. While we postponethe discussion of these tests to the next section, we now show that there is a positive and
3Moreover, policymakers who anticipate a growth slowdown may put in place expansionary fiscal policiesaimed at attenuating the slowdown. The increase in debt would thus precede the slowdown but still be causedby the recession.
4Panizza and Jaimovich (2007) show that trading partners’ output growth is a good instrument for GDPgrowth. As a consequence, trading partners’ debt is unlikely to be a good instrument for public debt.
5V E is not necessarily equal to zero for countries that have a fixed exchange rate. This would only be thecase if all the foreign currency debt is denominated in the currency to which these countries fix their exchangerate.
4
statistically significant partial correlation between the lagged value of V E and the stock ofpublic debt.
Our point estimates suggest that a one percentage point increase in lagged V E is associatedwith a 1.5 percentage points increase in the debt-to-GDP ratio (Figure 1). The results are almostidentical if we estimate the model without Japan (Figure 2) and become somewhat stronger ifwe estimate the model by dropping the five countries with the smallest shares of foreign currencydebt (Figure 3). In all cases, the coefficient is positive and statistically significant at the onepercent confidence level.
We now turn to exogeneity. Since our model is exactly identified we cannot test the validity ofour exclusion restriction. Therefore, we need to defend our identification strategy on theoreticalgrounds. We believe that there are two possible reasons why V E may have an effect on economicgrowth. The first has to do with the fact that V E is likely to be correlated with the presenceof foreign currency debt.6 Foreign currency debt, in turn, may reduce a country’s ability toimplement countercyclical macroeconomic policies and thus increase volatility and reduce growth(Eichengreen, Hausmann and Panizza, 2005; Hausmann and Panizza, 2011).
The second reason relates to the fact that V E is a debt-weighted (instead of trade-weighted)effective exchange rate.7 Even though the weights are different, V E could thus be correlatedwith the trade-weighted real effective exchange rate which, in turn, has been shown to have aneffect on economic growth. In particular, Rodrik (2008) shows that the real exchange rate is animportant driver of economic growth and that undervalued exchange rates are associated withhigher economic growth.8
If these are the only two channels through which V E may affect growth, controlling forthe share of foreign currency debt and the level of the real exchange rate will eliminate anydirect or indirect (except for the one that goes through public debt) correlation between V Eand economic growth. In other words, our exclusion restriction is valid as long as we control fordebt composition and the level of the real exchange rate.
While we cannot think of any other channel through which V E can affect economic growth,we are sure that creative skeptical readers will find ways to challenge our exclusion restriction.Therefore, we use Kraay’s (2012) Bayesian approach to build confidence intervals for instrumen-tal variable regressions with weak exclusion restrictions.
Our objective is to follow as closely as possible Cecchetti et al. (2011) who study a sample of18 OECD countries for the period 1980-2005. However, the construction of V E requires detailedinformation on the currency composition of public debt and these data are not generally publiclyavailable. In order to build our instrument, we searched old statistical bulletins, and contactedthe central banks and debt management offices of the OECD countries covered in the sample ofCecchetti et al. (2011). After months of detective work, we were able to obtain data for 17 ofthese 18 countries (the missing country is Greece).9 In the case of Portugal, our sample starts
6However, the direction of the correlation is not clear. In our data, the correlation between these two variablesis positive (0.22) and statistically significant at the 1 percent confidence level.
7Since V E includes domestic debt, it is conceptually similar to a trade-weighted exchange rate multiplied by the
degree of trade openness. In fact, Equation (4) can be rewritten as: V Ei,t =∑N−1
j 6=iDij,t∑N
j=1 Dij,t
∑N−1j 6=i
Dij,t(eij,t+1−eij,t)∑N−1j 6=i
Dij,t.
Where∑N−1
j 6=i Dij,t is total foreign currency debt, and∑N−1
j 6=iDij,t∑N
j=1 Dij,tis the share of foreign currency debt over total
debt.8Note that Rodrik works with the real exchange rate and V E is a nominal (debt-weighted) effective exchange
rate. However, most of the volatility of the real exchange rate is driven by the behavior of the nominal exchangerate (Mussa, 1986). Having said that, in our sample the correlation between V E and the trade weighted effectiveexchange rate is positive but not even close to being statistically significant.
9Our main source of information is based on unpublished data that the Bank for International Settlementsuses to build Table 13 B of its International Debt Securities Statistics. However, these data only go back to 1993.
5
in 1993 because we were not able to find information on the currency composition of public debtin the 1980s and early 1990s. In the next section, we show that our slightly smaller sample canreproduce all the main results of the benchmark paper.
3 Debt and Economic Growth: Causation versus Correlation
To explore whether public debt has a causal effect on economic growth, we build on an influentialpaper of Cecchetti et al. (2011). We stay as close as possible to their work by using the samedata, sample (with the exceptions mentioned in the previous section), and empirical approach.Formally, we use five-year overlapping growth spells to estimate the following regression:
GROWTHi,t+1, t+6 = αyi,t + β
(Debt
GDP
)i,t
+ γ′Xi,t + µi + τt + εi,t, (5)
where y is the log of real GDP per capita, µi and τt are country and year fixed effects, Xi,t
is a matrix of control variables, and GROWTH is average growth of real GDP per capita inpercentage terms (i.e., GROWTHi,t+1,t+6 = (yi,t+6 − yi,t+1) ∗ 100/5).
We are aware that using overlapping periods is not standard in the growth literature (notableexceptions include Bekaert, Harvey and Lundblad (2001, 2005)). However, in our small sampleof 17 countries, using non-overlapping growth spells would have a large cost in terms of theefficiency of our estimations. Therefore, we use overlapping growth periods and do our best tocorrect for the fact that this choice imposes a moving average structure on the error term. Wealso experiment with non-overlapping growth spells and find results that are consistent (albeitmuch weaker) with those of our benchmark specification.
Even though our specification includes lagged GDP, we think that the commonly used dif-ference and system GMM estimators are not well suited for a sample, like ours, with a relativelysmall number of cross-sectional units (Bond, 2002). In fact, we did experiment with differenceand system GMM estimators and found that the specification tests always reject the validity ofour model.10
3.1 Baseline estimations
We start with a baseline model in which GDP growth is regressed on lagged values of the publicdebt-to-GDP ratio, the log of initial GDP per capita, national gross savings (as a share of GDP),population growth, average number of years of secondary education, trade openness, inflation,age dependency ratio (the ratio between people younger than 15 or older than 64 and people inthe 15-64 age range), a banking crisis dummy, and the ratio of liquid liabilities to GDP.11
For previous periods, we consulted national statistical sources (we used the same sources listed in the Appendixof Missale (1999)) and, in some cases, obtained confidential data from national debt agencies and central banks.We deal with missing data by interpolating individual currency shares and then using data from Missale (1999)to measure the overall foreign currency share. Post-1992 data are of of higher quality with respect to the pre-1993data which are from heterogeneous sources and were sometimes interpolated. The fact that measurement erroris larger in the pre-1993 data weakens our instrument but should not lead to any obvious bias. As expected,the within-country correlation between our instrument and debt is stronger in the after 1993 sample than in thewhole sample.
10In particular, we estimated Equation (5) with the difference and system GMM estimators and experimentedwith different sets of instruments and allowed for different lag structures in the instruments. We found thatpublic debt had a negative coefficient (often statistically insignificant, but sometimes significant) and that usingV E as an external instrument had no effect on the magnitude and significance level of the public debt coefficient.However, our models were always rejected by the standard specification tests. In particular, AR2 was alwaysstatistically significant and, in many cases, the Sargan test rejected the validity of the exclusion restrictions.
11These are the same controls used in Column 2 of the top panel of Table 5 and in column 3 of Table A3.1 ofCecchetti et al. (2011).
6
The point estimates of the OLS regression reported in column 1 of Table 2 suggest that a tenpercentage point increase in the debt-to-GDP ratio is associated with an 18 basis point reductionof average growth. This is very close to the baseline estimate of Cecchetti et al. (2011) whofind a 17 basis point effect. The coefficient of public debt appears to be precisely estimated andis statistically significant at the one percent confidence level. Therefore, even though we havea slightly different sample, our baseline results are almost identical to those of the benchmarkpaper.
When we instrument the debt-to-GDP ratio with L.V E, we find a strong positive firststage correlation between the instrument and public debt (column 2 of Table 2; this is the samepartial correlation plotted in Figure 1). More interestingly, the second stage regression of column3 shows that debt is no longer statistically significant, and that its coefficient goes from negativeto positive.
While the lack of statistical significance could be due to the fact that the IV estimator isinefficient, the change in sign is prima facia evidence that the OLS estimates suffer from a largenegative bias. This is exactly what we would expect if there were a negative causal effect goingfrom growth to public debt (k < 0 in the discussion of the previous section).
The bottom panel of Table 2 reports a battery of underidentification and weak instrumenttests. The Kleibergen-Paap LM and Wald statistics reject the null of underidentification. Weare thus confident that our instrument satisfies the rank condition. The Kleibergen-Paap F testyields a value of 9.2. This is close to the Staiger and Stock (1997) rule of thumb value of 10,and it is in between the Stock and Yogo (2005) 5 percent critical values for 10 and 15 percentmaximum bias (these critical values are 16.38 and 8.96, respectively). Therefore, the maximumbias associated with the possible presence of weak instruments can be larger than 10 percentbut it is smaller than 15 percent. Angrist and Pischke (2009) argue that the weak instrumentbias tends to be small in exactly identified models like ours and that the bias disappears whenthe correlation between the instrument and the endogenous variables is greater than 0.1 (theStaiger-Stock rule of thumb requires a correlation of at least 0.3).
Even though the tests reported in the bottom panel of Table 2 and the findings of Angristand Pischke (2009) suggest that our estimations are unlikely to suffer from weak instrumentsbias, we can use Moreira’s (2003) conditional likelihood ratio test (CLRT) to build confidenceintervals that are valid in the presence of weak instruments. When we compare the IV confidenceinterval with the CLRT confidence interval, we find that the latter is wider (-5.31, 19.39 insteadof -6.8, 7.4) but shifted to the right with respect to the IV confidence interval. This is notsurprising because, if anything, the presence of weak instruments may bias the IV estimatortowards the OLS estimator. This is another reason why we think that our estimations provideconvincing evidence that there is no proof that debt has a causal effect on economic growth.
Until now we have not controlled for currency composition and the effective exchange rate.Therefore, the model of Table 2 is likely to be misspecified (see the discussion in Section 2). InTable 3, we augment the baseline specification with a variable that captures the share of foreigncurrency debt and the (trade-weighted) real effective exchange rate index (higher values of theindex are associated with real appreciations). As we include these two variables to eliminateany possible channel through which L.V E may affect growth, we also lag currency compositionand the real exchange rate index.12
As before, OLS estimations find a negative and statistically significant correlation betweendebt and economic growth (Column 1 of Table 3). The point estimate implies that a 10 per-centage point increase in the debt to GDP ratio is associated with a 15 basis point decreasein growth. Again, the coefficient is precisely estimated and is statistically significant at the
12Given that the control variables are lagged with respect to the beginning of the growth episode, V E, currencycomposition, and the real exchange rate index are actually lagged twice.
7
one percent confidence level. As expected, we also find that foreign currency debt and realappreciations are negatively correlated with GDP growth.
The first stage regressions of column 2 shows a strong positive correlation between L.V Eand the debt-to-GDP ratio. The Kleibergen-Paap LM and Wald statistics are now higher thanin Table 2, and always reject the null of underidentification at the five percent confidence level.The Kleibergen-Paap first stage F test now takes a value of 16. This is well above the Staigerand Stock (1997) rule of thumb value, and approximately equal to the Stock and Yogo (2005)threshold for a maximum 10 percent bias (the exact value of the threshold is 16.38). In thesecond stage of the IV regression, we find that the coefficient of public debt changes sign andbecomes statistically insignificant (column 3 of Table 3). As before, we find that controlling forendogeneity eliminates the negative correlation between public debt and GDP growth. At 2.05,the point estimate looks implausibly large. However, the t statistics is 0.5 and the coefficientis not even close to being statistically significant. Again, we find that the CLRT confidenceinterval is shifted to the right and wider than the IV confidence interval.
While fixed effects estimations allow controlling for unobserved time-invariant heterogeneityand attenuate potential omitted variable bias, they also reduce the signal to noise ratio (the signalis the deviation from the mean) and amplify the consequences of measurement error. Moreover,since fixed effects estimations concentrate on how within-country changes in debt affect within-country changes in growth, they may not capture the possible costs of “structurally” high levelsof debt.13 However, our results are not driven by the presence of fixed effects. To the contrary, inthe non-instrumented regressions, the negative correlation between debt and growth is stronger(and more precisely estimated) when we include country and year fixed effects. Moreover, oursample does not include any country with “structurally" high levels of debt (for instance, atthe beginning of our sample, Italy and Japan’s debts were below 70 percent of their respectiveGDP). As a consequence, the within-country variance of the debt-to-GDP ratio is similar to itsbetween-country variance.
We think that our baseline estimations of Tables 2 and 3 provide convincing evidence that inOECD countries there is no causal link going from public debt to economic growth. We alreadyshowed that the lack of a statistically significant effect of debt in the IV regressions is unlikelyto be caused by the presence of weak instruments. We now conduct an extensive battery ofrobustness checks aimed at uncovering possible problems with our estimation strategy.
3.2 Weak Exclusion Restrictions
The key assumption in our estimation strategy is that L.V E does not belong in the growthequation. Since our model is exactly identified, we cannot provide a statistical test of thevalidity of this restriction. We can, however, use Kraay’s (2012) Bayesian approach to checkwhat happens if we relax our exclusion restriction.
To illustrate Kraay’s approach, let us go back to Equation (1) in which growth (G) is afunction of debt (D). Since debt is endogenous, E(D,u) 6= 0 and the OLS estimator is biased.If we can find a variable Z correlated with D but uncorrelated with u, we can recover thestructural parameter of interest. However, the assumption that E(Z, u) = 0 “is likely to be apoor approximation to the actual prior belief of empirical researchers." (Kraay, 2012, p. 112).
Relaxing the assumption that the distribution of E(Z, u) collapses at zero does not preventus from identifying the structural parameters of interest, but it does reduce the precision of thestructural estimates (Kraay, 2012). Kraay’s Bayesian approach quantifies the consequence of
13Our definition of “structurally high” debt is different from that of “structural” debt in Powell (2012). Theformer refers to debt levels that remain persistently high and therefore do not change much over time. The latterrefers to the level of debt net of valuation effects.
8
prior uncertainty on the validity of the exclusion restriction and maps the degree of uncertaintyon the precision of the estimate of the structural parameter.
Specifically, Kraay uses a prior distribution for the correlation between the reduced formerror and the instrument to approximate prior uncertainty about the validity of the exclusionrestriction. In particular, let φ be a uniformly distributed random variable over the support(-1,1) and 1/η the degree of prior uncertainty about the validity of the exclusion restriction.Kraay’s prior distribution g(φ) is then given by:
g(φ) = (1− φ2)η (6)
Prior uncertainty reaches its maximum level (g(φ) is uniformly distributed over -1, 1) whenη = 0. The higher the value of η, the lower the uncertainty. Setting η = 5 is equivalent toassuming a 90 percent prior probability that g(φ) ∈ (−0.46, 0.46). With η = 100, the 90 percentconfidence interval for g(φ) is (-0.12, 0.12), and with η = 500 there is a 90 percent probabilitythat g(φ) ∈ (−0.05, 0.05) (the 90 percent confidence interval drops to -0.04 and 0.04 for η = 1000;see Table 1 in Kraay (2012)).
Having found a method to characterize prior uncertainty, Kraay shows that it is possibleto use numerical sampling to map this prior uncertainty on the posterior marginal distributionof the structural parameters.14 His simulations show that even moderate prior uncertaintycan have a large effect on the precision of the estimated parameters. Perhaps surprisingly, heshows that the loss of precision linked to prior uncertainty is higher in the presence of stronginstruments and large samples. In other words, the stronger the instrument, the more seriousthe consequence of a misspecified model.
When we use Kraay’s approach to examine the consequence of weakening our priors aboutthe validity of our exclusion restriction, we find that a strong violation of our exclusion restrictionyields very large confidence interval (Table 4). In particular, when η ≤ 10 (implying that thereis a 90 percent probability that g(φ) ∈ (−0.34, 0.34)), the confidence intervals are five timeslarger than those of the IV regressions in which we assume that g(φ) = 0. When we set η = 100and thus limit the 90% confidence interval of g(φ) to (-0.12, 0.12), the confidence interval of thestructural estimate becomes only slightly larger than that of the IV estimates (which assumeE(Z, u) = 0).
We also find that relaxing our priors on the exclusion restriction has more serious conse-quences for the model in which we control for foreign currency share and debt composition (thelast three columns of Table 4). This is not surprising. Our instrument is much stronger inthe regressions in which we control for foreign currency share and debt composition, and Kraayshows that potential violations of the exclusion restriction are more problematic for models withstronger instruments.
3.3 Other Robustness Checks
Placebo regressions
An alternative way to check whether our results are driven by the presence of a weak instrumentis to substitute our instrument with a randomly generated variable that matches the momentsof L.V E. If this placebo IV regression were to give results similar to those of our IV regressions,we would know for sure that there is a problem with our instrument (Bound, Jaeger and Baker,1995). Figure 4 plots the results of 1,000 replications of placebo IV regressions and Table 5reports the summary statistics for the model that does not control for foreign currency share
14Since the moments of the posterior distribution of the structural parameters are not defined, Kraay (2012)summarizes his results with the median, 2.5th and 97.5th percentiles of the posterior distribution.
9
and exchange rate (column 1) and for the full model (column 2). We find that the mean andmedian values of the placebo regression are similar to those of the OLS estimates but that theyare imprecisely estimate. This is exactly what one would expect with a placebo instrument.More important, for our purpose, the results of the placebo regression are different from thoseof the IV regressions of Tables 2 and 3. While the results of Table 5 do not prove that ourinstrument is valid, they show that the instrument passes the non-robustness test proposed byBound, Jaeger and Baker (1995).
Dealing with autocorrelation
As we use overlapping growth periods, the errors of our model have a moving average structure.So far, we dealt with serially correlated residuals by using the Huber-White sandwich correction.This estimation strategy is appropriate in the presence of within-country time dependence andheteroscedacity of unknown form. An alternative approach consists of using the Newey and West(1987) estimator which allows modeling the autocorrelation process in the error term. This iswhat we do in Table 6. In particular, we experiment with AR(1), AR(5), and AR(10) errorstructures and find results which are basically identical to those of Table 3.15
An alternative measure of debt
We also experiment with an alternative data source for our indebtedness variable. In particular,Table 9 replaces the debt-to-GDP ratio based on flow of funds national balance sheet statistics,with the debt-to-GDP ratio data described in Panizza (2008).16 While the regressions yieldestimates similar to our baseline of Table 3, we now find that the instrument is much weaker. Inparticular, the first stage Kleibergen-Paap F test takes a value of 4.58. This is much lower thanthe Staiger and Stock (1997) rule of thumb and lower than the Stock and Yogo (2005) thresholdfor a 25 percent maximum bias.
Different samples and outliers
While the last growth spell considered by Cecchetti et al. (2011) is 2001-2006, we have morerecent data and estimate regressions that include GDP growth over the 2002-2007 and 2003-2008periods. The difference in coverage could be important because our sample includes the first twoyears of the great recession and the benchmark paper does not. Table 7 drops the 2002-2007 and2003-2008 growth spells and, again, finds results that are essentially identical to our baselineestimates of Table 3. This is also the case if we drop Japan from the sample (Table 8).17
To explore whether our results are driven by outliers, we start by plotting the partial cor-relation between debt and growth obtained with the OLS regressions of column 1, Table 3. InFigure 5 there seem to be four influential observations: Japan 2002 and 2003 and Finland 1987and 1988. We already discussed the case of Japan, we will soon look at Finland.
15Table 6 only reports results for AR(1) and AR(5). AR(10) results are available upon request.16The flow of funds data yield higher debt ratios. The average difference between the two dataset is 12
percentage points with a standard deviation of 10 percentage points. This difference is particularly large forAustralia (22 percentage point), Canada (33 percentage points), Denmark (17 percentage points), Finland (19percentage points), France (14 percentage points), the Netherlands (15 percentage points), and Sweden (17percentage points). The between-country standard deviation of the difference between these two sources of datais 8 percent and the within-country standard deviation is 4 percent.
17We conduct this experiment because Japan has high debt and near-zero variance in V E. We also lookedat what happens if we drop the other four countries with near zero variance of V E (France, Germany, theNetherlands and the United States) and found that the IV results are similar to those of Table 3. However,when we drop these countries, debt is no longer significantly correlated with growth in the OLS regression (thecorrelation remains negative, with a point estimate of -1.05 and a p-value of 0.26).
10
The scatterplot of the partial correlation between debt and growth obtained with the IVregression of the third column of Table 3 (Figure 6) shows several observations which appear tobe influential. Many of these observations have to do with Finland and a few with Sweden andAustralia. This impression is also confirmed by the leverage plot of Figure 7 which shows thatAustralia 1986 and Sweden 1996 have high leverage and large residuals. We thus estimate ourmodel by dropping Finland and the two suspect observations for Australia and Sweden.
In the OLS estimates of the reduced sample, we find the usual strong negative correlationbetween debt and growth (column 1, Table 10; the partial correlation is plotted in Figure 8).When we estimate the IV model we still find that the coefficient of debt is insignificant andpositive (column 3, Table 10; the partial correlation is plotted in Figure 9). However, we nowfind that the point estimate is close to zero and not implausibly large as in the IV regression ofTable 3.
Non-overlapping growth spells
As a final robustness check we look at what happens when we use non-overlapping growth spells.Given that 5-year growth spells only yield 68 observations and 46 degrees of freedom (we haveten control variables, 17 country fixed effects, and 3 period fixed effects), we also work with3-year growth spells (119 observations and 97 degrees of freedom).
When we work with non-overlapping periods, we need to choose a starting year. Choosingthe first available year, as it is usually done, appears to be arbitrary. Therefore, we estimate ourmodel for all available starting years (1981, 1982, and 1983 when we use 3-year growth spells,and 1981, 1982, 1983, 1984, and 1985 when we use 5-year growth spells).
In the 3-year growth spells regressions, OLS estimations always find a negative and statisti-cally significant correlation between public debt and economic growth (Column 1, top panel ofTable 11).18 When we estimate the model starting in 1981, we obtain point estimates which aremuch larger (in absolute value) than those of the model with 5-year overlapping growth spells(-3.24 versus -1.54). Instead, when the start date is set to 1982 or 1983, results obtained arecloser to those of Table 3 and imply that a 10 percentage point increase in debt is associatedwith a 17-18 basis point decrease in growth. As before, the correlation between debt and growthdisappears once we instrument debt with L.V E (Column 2, top panel of Table 11). As in themodel with overlapping growth spells, we also find that IV regressions are often associated witha switch of the sign. In this case, we find a sign switch in two of our three regressions.
When we use OLS regressions for 5-year non-overlapping growth spells, we always find anegative correlation between debt and growth. The point estimates range between -0.92 and-2.08 (Column 1, bottom panel of Table 11) and are not too far from those that we obtained inTable 3. In this case, however, the coefficients are statistically significant in only 2 of our fiveregressions (in one of these two regressions the coefficient is only marginally significant). Whenwe instrument public debt with L.V E, we no longer find a statistically significant associationbetween debt and growth. In three of our five regressions we find that the coefficients becomepositive (column 2, bottom panel of Table 11). We do not want to make too much of theresults of Table 11. The samples are small and the results of the first stage regressions are notencouraging.19 However, it is interesting that the regressions with non-overlapping growth spellsshow the same pattern as the regressions with overlapping growth spells: negative and generally
18While Table 11 only reports the coefficients for the public debt variable, the results reported there wereobtained using the same specification used in Table 3 (with period fixed effects instead of year fixed effects).
19While L.V E has always the expected positive sign, the coefficient is rarely statistically significant (column 3of Table 11). Moreover, the Kleibergen-Paap F tests of the the first stage regressions are often low (column 4 ofTable 11), hinting at a serious weak instrument problem.
11
statistically significant coefficients in the OLS regressions and insignificant and generally positivecoefficients in the IV regressions.
4 Looking for thresholds
We now check whether there is a threshold above which public debt has a larger effect onGDP growth. In a sample of 20 advanced economies for the period 1946-2009, Reinhart andRogoff (2010a,b) find that when public debt is above 90 percent of GDP, median growth isapproximately 1 percentage point lower than in country-years with lower debt. They also showthat, in high-debt countries, average growth is almost 4 percentage points lower than in country-years in which debt is below 90 percent of GDP. Looking at histograms similar to those usedby Reinhart and Rogoff, Cecchetti et al. (2011) do not find a clear correlation between differentdebt levels and GDP growth. They suggest that this could be due to the fact that their smallersample requires more sophisticated econometric techniques.
They thus define DΨ as a dummy variable that takes a value of one when the debt-to-GDPratio is below Ψ and propose to estimate the following threshold regression for Ψ ∈ (50, 120).
GROWTHi,t+1, t+6 = αyi,t + γ′Xi,t + ϕDΨi,t + β1
(Debti,tGDPi,t
DΨi,t
)+ (7)
+β2
(Debti,tGDPi,t
(1−DΨi,t)
)+ µi + τt + εi,t
Next, they compute Hansen’s (1999) likelihood ratio (LR) statistics to identify the thresholdthat yields the best fit of Equation (7) and build confidence intervals around this threshold.20
In the case of public debt, they find that the threshold that yields the best fit of Equation (7)is 96 percent.
When we conduct the same experiment using our data, we find a pattern similar to that ofCecchetti et al. (2011) but with slightly lower values of the LR statistics (the solid line in Figure10). In particular, we find that the threshold that yields the best fit is 92 percent of GDP witha 90 percent confidence interval of 80-98.
While these results are consistent with the presence of a threshold at around 90 percent ofGDP, there are two issues with interpreting Figure 10 as evidence for the existence of such athreshold. The first has to do with the fact that Hansen’s derivations assume a static modelwith iid errors. It is not clear whether his results apply to our model which is dynamic withheteroscedastic errors.
The second, and probably more important, issue has to do with the fact that Hansen’s LRstatistics cannot be used to test for the existence of a threshold. It can only be used to build aconfidence interval around a threshold, under the assumption that such a threshold does exist(Hansen, 1999, p. 351). Testing for the presence of a threshold requires a more complicatedbootstrapping procedure, which we will not pursue here.21
One thing that we can do is to compare the coefficients of β1 and β2 for different values ofΨ. The solid line in the top panel of Figure 11 plots the point estimates of β1 (the dotted linesplot the 95 percent confidence interval) for all possible thresholds between 50 and 120 percent ofGDP. It shows that the point estimates are fairly stable (they range between -1 and -2) and are
20The statistics is: LR(Ψ) = (S(Ψ) − S(Ψ))/σ2. Where S(Ψ) is the sum of squared errors for the model withthreshold Ψ; Ψ is the threshold that minimizes the sum of squared errors; and σ2 is the variance of the errorterm when Ψ = Ψ.
21One reason for not doing this exercise is that we do not know the properties of Hansen’s (1999) method fordynamic panels with non-iid errors.
12
often statistically significant (they are insignificant in the 82 to 92 percent range). The bottompanel plots the values of the point estimates of β2. Again, the estimates are stable (ranging from-1.5 and -2) and, in this case, always statistically significant.
While we find that β2 is estimated more precisely than β1, Figure 11 does not show a largedifference between the two sets of coefficients. For instance, the point estimates suggest thatdebt has the same effect on GDP growth in country-years when the debt-to-GDP ratio is below50 percent (the first observations in the top panel of Figure 10) as in country-years when thedebt to GDP ratio is above 120 percent of GDP (the last observation in the bottom panel ofFigure 11).
We can compute the difference between β1 and β2 and test whether the two parameters aresignificantly different from each other. The solid line in Figure 12 shows that β1 tends to belarger (in absolute value) than β2 when public debt is below 80 percent of GDP and that thesituation reverses when debt surpasses 80 percent of GDP. The dotted line plots the p-value ofa test on the difference between coefficients. It shows that this difference is rarely statisticallysignificant.22 While this is not a true test for the presence of a threshold, Figure 12 providessuggestive evidence that such a threshold may not exist.
Reinhart and Rogoff (2010a) argue that it is hard to find instruments in the presence ofthresholds. We tend to agree with them, but we try anyway. With one threshold, we have twoendogenous variables and we need two instruments. In principle, if L.V E is a good instrumentfor Debt, L.V E ∗ DΨ could be an instrument for DEBT ∗ DΨ, and L.V E ∗ (1 − DΨ) couldbe an instrument for DEBT ∗ (1 − DΨ). The problem is that these instruments may not bepowerful enough to identify the effect of debt on growth.
When we compute Hansen’s (1999) LR statistics using IV estimates, we find an almostflat line (the dashed line in Figure 10). This could be due to the fact that the statistics ismeaningless in the presence of IV estimates; it could also be a sign that the confidence intervalfor the threshold is very large (and therefore there is no threshold); or that there are problemswith our IV strategy.
Figure 13 plots the values of β1 and β2 obtained with the IV estimates. They are unstable,suggesting that our instrument does not work well in the threshold regression of Equation (7).To probe further, we look in detail at the threshold regression with Ψ = 90. The first column ofTable 12 reports the OLS estimates (they correspond to the coefficients plotted in Figure 1 whenthe threshold is 90). We find that the debt-to-GDP ratio has a negative, but not statisticallysignificant, effect on growth when debt is lower than 90 percent of GDP, and that the debt-to-GDP ratio has a larger (in absolute value) and statistically significant effect on growth whendebt is greater than 90 percent of GDP. The difference between the two coefficients, however, isnot statistically significant (Figure 12).
The second and third columns of Table 12 report the first stage estimates. One of the twoinstruments has some power in column 2, but the instruments have no power whatsoever incolumn 3. The underidentification and weak instrument tests reported in the bottom panel ofthe Table confirm that our model is underidentified (the rank condition is not satisfied) andthat the instruments are weak. For sake of completeness, we report the second stage estimates.However, we believe that these estimates are completely unreliable. We report Table 12 tobe fully transparent and make the point that in the threshold regression of Equation (7) ourinstrument is useless. This is why the point estimates of Figure 13 are so unstable.
The first stage regressions of Table 12 indicate that our instrument works better for low levels22The p-value is less than 0.05 in 6 of our 70 observations (when the threshold is 84, 85, 86, 89, and 93). Note
that if our 120 observations were the outcome of independent trials (which they are not), we would have expectedto find 3.5 observations with a p-value smaller than 0.05, even in the absence of any true difference between thetwo coefficients.
13
of debt. Therefore, we try splitting the sample.23 Again, we start with the OLS regression. Inthe low debt sample, we find a negative but not statistically significant correlation between debtand growth (column 1 of Table 13). In the first stage regression of column 2, our instrumenthas the right sign but is not statistically significant. The underidentification tests shows thatour instrument satisfies the rank condition, but the F test indicates that the instrument couldbe weak (recall, however, that in an exactly identified model like ours the presence of a weakinstrument may not be such a big problem, Angrist and Pischke (2009)). The IV regression ofcolumn 3 finds that debt is now positive and insignificant. In columns 4-6 we try to gain degreesof freedom by re-estimating the model without controls. The results are qualitatively similar tothose of columns 1-3.
In Table 14, we repeat the experiment of Table 13 by focusing on the country-years for whichdebt is above 90 percent of GDP. As expected, we find that the instrument is weak and that theresults of the IV estimates are meaningless.
One surprising finding of Table 14 is that the coefficient is not even statistically significantin the OLS regression (column 1). This may be due to the fact that the regression does nothave enough degrees of freedom (with 90 observations, 8 country fixed effects, 22 year fixedeffects, and 12 control variables, the regression has 48 degrees of freedom) or due to the factthat our sample is not poolable (i.e., the other controls have different effects on growth in lowand high-debt countries). Column 4 shows that if we try to save degrees of freedom by droppingour set of controls, the coefficient of public debt goes back to values similar to those of Tables2 and 3 and becomes statistically significant at the 10 percent confidence level.
5 An alternative identification strategy
As our instrument does not perform well in high debt countries, we now experiment with analternative identification strategy. Let us go back to our simple model of Equations (1) and (2),but let us now assume that we have an instrument for GDP growth:
G = a+ bD + γTPGROWTH + u (8)
D = m+ kG+ v (9)
Since TPGROWTH does not belong in the debt equation, we can identify the effect ofgrowth on debt (the structural parameter k). Without additional assumptions, we cannot esti-mate b which is our structural parameter of interest.
However, Fisher (1966) showed that if u and v are true structural shocks (i.e., if E(u, v) = 0)a model like that of Equations (8) and (9) has enough restrictions to identify both k and b. Hecalled this identification method, identification through covariance restrictions. Hausman andTaylor (1983) showed that identification through covariance restrictions has an instrumentalvariable interpretation, and Hausman, Newey and Taylor (1987) proposed an augmented threestages least squares (a3SLS) estimator for simultaneous equations with covariance restrictions.
This method is rarely used because the implementation of the a3SLS estimator is cum-bersome. However, Shapiro (1987) shows that, in an exactly identified two equations systemlike that of Equations (8) and (9), the implementation of the a3SLS estimator consist of fourstraightforward steps analogous to two 2SLS. The first step (the first stage of the first 2SLS)consists of regressing G over TPGROWTH (and any other exogenous regressor that belongs to
23By splitting the sample, we relax the assumption that the effect of the controls is the same in low andhigh-debt countries. If this assumption were true, split-sample regressions would be unbiased but inefficient. Ifthis assumption does not hold, split sample regressions are only unbiased.
14
both equations) and then recovering G. In the second step (the second stage of the first 2SLS),we estimate k by regressing D over G (and other exogenous regressors) and then recover v. Byconstruction, v is correlated with D and, by assumption, uncorrelated with G. We can then usev as an instrument for D. Therefore, in the third step (the first stage of the second 2SLS) weregress D over v and other exogenous regressors and recover D. In the fourth step (the secondstage of the second 2SLS), we estimate b (the structural parameter of interest) by regressing Gover D.24
Therefore, if we were to assume that u and v are true structural shocks and could find aninstrument for growth, we would have an alternative way to identify the effect of debt on growth.Panizza and Jaimovich (2007) show that a real external shock consisting of the weighted averageof GDP growth in a country’s export partners is a good instrument for annual GDP growth.They define the real external shock as:
JPi,t =EXPiGDPi
∑φij,t−1GDPGROWTHj,t, (10)
where GDPGROWTHj,t measures real GDP growth in country j in period t, φij,t is the fractionof exports from country i going to country j, and EXPi/GDPi measures country’s i averageexports expressed as a share of GDP.25
The JP instrument is a refinement on the instrument used by Galì and Perotti (2003) ina paper that studies fiscal cyclicality in the Euro area. Variants of the JP instrument havebeen used by Alesina, Campante and Tabellini (2008), Ilzetzki and Végh (2008), and Végh andVuletin (2012). Panizza and Jaimovich (2007) discuss in detail the properties of this instrument.
While there is evidence that the JP instrument works well at the annual frequency, theinstrument has not yet been used at lower frequencies. Therefore, we start by building a variableTPGROWTH which is equal to JP(t+6,t+1) and then run a set of (first) first stage regressions tocheck whether TPGROWTH is a good instrument for growth over five-year periods. When weestimate the model for the whole sample, the results are not encouraging (column 1, top panelof Table 15).26 TPGROWTH is not statistically significant and fails all the underidentificationand weak instrument tests reported in the first column of Table 14.
When we estimate the first stage for the subsample of country-years with debt below 90percent of GDP, we find that the coefficient of TPGROWTH is statistically significant atthe ten percent confidence level and that the tests reject underidentification (Column 2, toppanel of Table 15). The F test is still low, but much higher than that of Column 1. However,TPGROWTH has the wrong sign. The point estimate suggests that high trading partnergrowth reduces country i’s growth. This cannot be right.
Things improve when we look at high-debt observations (columns 3 and 4 of the top panelof Table 15). In country-years with debt-to-GDP ratios above 0.9 (column 3), we find a positiveand statistically significant impact of trading partner growth, good underidentification tests,and an acceptable (albeit still low) first stage F statistic.27
24As the instrument of the second 2SLS is an estimated parameter, we need to adjust the standard errors ofthe second 2SLS. Shapiro (1987) describes how to do it.
25They use a time invariant measure of exports over GDP because a time variant measure would be affectedby real exchange rate fluctuations, and, therefore, by domestic factors. This is not the case for the fraction ofexports going to a specific country (φij,t), because the variation of the exchange rate that is due to domesticfactors has an equal effect in both the numerator and denominator. We build the JP instrument for the sampleof 17 OECD countries using the bilateral trade data published in the Correlates of War Project’s Trade Data Set(Barbieri, Keshk and Pollins, 2012, 2009), which is based on the IMF Directions of Trade Statistics.
26Note that the estimations of Table 15 include all the controls used in Table 2. The coefficients are notreported to save space.
27In fact, we do not know the critical values of these tests for the a3SLS estimator.
15
One problem with restricting the sample to country-years with debt above 90 percent ofGDP is that we only have 90 observations. We also experiment with an 80 percent threshold(119 observations) and still find that the instrument has a positive and statistically significanteffect on growth (column 4). The underidentification and weak instrument tests are worse thanthose of column 3 but still acceptable according to the Angrist and Pischke (2009) criterion.When we include observations with debt levels below 80 percent of GDP, our instrument nolonger works.
On the basis of the results of the top panel of Table 15, we apply Hausman, Newey andTaylor’s (1987) a3SLS estimator to the subsamples of country-years with debt above 80 and90 percent of GDP. In the first column of the bottom panel of Table 15, we report the OLSestimates for the above 90 percent of GDP subsample. As in Table 14, the coefficient is negativebut not statistically significant.28 The a3SLS estimator, instead, finds a positive and statisticallysignificant coefficient. We do not want to make too much of the fact that the coefficient isstatistically significant. But we want to highlight that the results of this alternative estimationstrategy are fully consistent with what we found in the regressions of Tables 2 and 3. WhileOLS regressions yield a negative correlation between debt and growth, instrumental variablesregressions find the opposite.
In the last two columns of the bottom panel of Table 15, we expand our sample to includecountries with a debt-to-GDP ratio of at least 0.8. The OLS regression finds a negative andstatistically significant correlation between debt and growth (column 3) and, again, the a3SLSestimator finds a positive (this time insignificant) effect.
Finding a good instrument is always difficult and we agree with Reinhart and Rogoff (2010a)that finding instruments in the presence of thresholds is particularly hard. However, we foundthat in our subsample of countries the evidence of a threshold effect is not as strong as previouslythought. Moreover, we have an instrument that has some power in the subsample of low-debtcountries and an alternative identification strategy that works in the high-debt subsample, andwe show that both identification strategies suggest that there is no evidence that debt has acausal effect on economic growth.
6 Conclusions
It ain’t what you don’t know that gets you into trouble. It’s what you know for surethat just ain’t so
(attributed to Mark Twain)
We started this paper with a question. What did we find? Does debt have a causal effect oneconomic growth? Unfortunately, our answer to this question is: “We don’t know".
We are well-aware that “we don’t know" papers are not as satisfying as papers with a clearresult. However, we think that there is some value in assessing the degree of our ignorance (seethe Mark Twain quote at the beginning of this section).
We believe that our findings are important for the current debate on fiscal policy. Theremight be many good (or bad) reasons for fiscal austerity, even during recessions. We do notwant to enter that debate here. However, we do not find any evidence that high public debt
28This regression is similar to that reported in column 1 of Table 14. The only difference is that the regressionof Table 15 does not control for the effective exchange rate and foreign currency share. Again, the regressions ofTable 15 include all the controls used in Table 2, but only report the coefficient of the variable of interest.
16
levels hurt future growth in advanced economies.29 Therefore, given the state of our currentknowledge, we think that the debt-growth link should not be used as an argument in support offiscal consolidation.
We recognize that instruments are never perfect and that their use involves tradeoffs betweenconsistency and efficiency. In describing the strengths and weaknesses of our identificationstrategy, we tried to be as transparent as possible and conducted extensive robustness checks.Among other things, we used a new Bayesian approach to check what happens if we relaxour exclusion restriction and experimented with the a3SLS estimator of Hausman, Newey andTaylor (1987). Nevertheless, our exclusion restrictions are non-testable, and we may never beable to convince skeptical readers of their validity. Yet, we are convinced that our instrumentis valid and strong and we were reassured by the fact that the instrumental variable results areconsistent with theoretical arguments suggesting that OLS estimations are negatively biased.Our confidence was also bolstered by the finding that two completely different identificationstrategies yield similar results.
The fact that we do not find a negative effect of debt on growth does not mean that countriescan sustain any level of debt. There is clearly a level of debt which is unsustainable (for instance,when the interest bill becomes greater than GDP) and a debt-to-GDP ratio at which debtoverhang, with all its distortionary effects, kicks in. What our results seem to indicate, however,is that the advanced economies in our sample are still below the country-specific threshold atwhich debt starts having a negative effect on growth.
We believe that there is a subtle channel, which may not be captured in our empiricalexercise, through which high levels of public debt can have a negative effect on growth.30 Inthe presence of multiple equilibria, a fully solvent government with a high level of debt maydecide to put in place restrictive fiscal policies to reduce the probability that a sudden changein investors’ sentiments would push the country towards the bad equilibrium.31 These policies,in turn, may reduce growth (Perotti, 2012), especially if implemented during a recession.32 Inthis case, it would be true that debt reduces growth, but only because high levels of debt leadto contractionary policies. While such an interpretation would justify long-term policies aimedat reducing debt levels, it also implies that countries should not implement restrictive policiesin the middle of a crisis. These policies are the reason for the negative effect of debt on growth.Of course, policymakers under pressure from market participants might not have an alternative.This is why we need prudent fiscal policies and lenders of last resort that can rule out multipleequilibria.
In fact, we suspect that we do not find a negative effect of debt on growth because most ofthe observations included in our dataset are for years in which all OECD countries could usetheir own central banks as lenders of last resort. With the creation of the euro, however, abouthalf of the countries in our sample lost their lender of last resort and soon started facing marketpressure (De Grauwe, 2011).
Our readings of the current empirical evidence linking public debt and economic growth inadvanced economies can be summarized as follows: (i) many papers show that public debt is
29Things are different in developing countries, where a significant fraction of debt is external and the debtoverhang argument (Krugman, 1988; Sachs, 1989) has more bite. For evidence in this direction see Cordella,Ricci and Ruiz-Arranz (2010) and Pattillo, Poirson and Ricci (2011).
30We would like to thank, without implications, Carlo Cottarelli for suggesting this point.31Alternatively, policy makers may decide to implement these restrictive policies as a reaction to market signals
which do not reflect the true fundamentals (we started from the assumption that the government is solvent).32DeLong and Summers (2012) discuss the conditions under which the recessionary effect would be so large that
the contractionary policy would actually be self-defeating and lead to a higher debt-to-GDP ratio. In its massivedowngrade of European sovereigns in January 2012, Standard and Poor’s explicitly mentioned that restrictivepolicies may have a negative effect on debt sustainability (Standard & Poor’s, 2012).
17
negatively correlated with economic growth; (ii) no paper makes a convincing case for a causallink between debt and growth; (iii) our paper suggests that such a causal link does not exist(more precisely, our paper does not reject the null hypothesis that debt has no impact on growth).We realize that our results are controversial, and that some readers will not be convinced by ouridentifications strategy. However, we also believe that the first two points are uncontroversial,and the case that debt has causal effect on growth still needs to be made.
18
References
Alesina, Alberto, Filipe R. Campante and Guido Tabellini. 2008. “Why is Fiscal Policy OftenProcyclical?” Journal of the European Economic Association 6(5):1006–1036.
Angrist, Joshua D. and Jorn-Steffen Pischke. 2009. “A Note on Bias in Just Identified IV withWeak Instruments.” London School of Economics.
Arellano, Manuel and Olympia Bover. 1995. “Another look at the instrumental variable estima-tion of error-components models.” Journal of Econometrics 68(1):29–51.
Arellano, Manuel and Stephen Bond. 1991. “Some Tests of Specification for Panel Data: MonteCarlo Evidence and an Application to Employment Equations.” Review of Economic Studies58(2):277–97.
Barbieri, Katherine, Omar Keshk and Brian Pollins. 2009. “Trading Data: Evaluating ourAssumptions and Coding Rules.” Conflict Management and Peace Science 26(5):471–491.
Barbieri, Katherine, Omar Keshk and Brian Pollins. 2012. “Correlates of War Project TradeData Set Codebook, Version 3.0.” Online: http://correlatesofwar.org.
Bekaert, Geert, Campbell R. Harvey and Christian Lundblad. 2001. “Emerging equity marketsand economic development.” Journal of Development Economics 66(2):465–504.
Bekaert, Geert, Campbell R. Harvey and Christian Lundblad. 2005. “Does financial liberalizationspur growth?” Journal of Financial Economics 77(1):3–55.
Blundell, Richard and Stephen Bond. 1998. “Initial conditions and moment restrictions in dy-namic panel data models.” Journal of Econometrics 87(1):115–143.
Bond, Stephen Roy. 2002. “Dynamic panel data models: a guide to micro data methods andpractice.” Portuguese Economic Journal 1:141–162.
Bound, John, David A Jaeger and Regina M Baker. 1995. “Problems with Instrumental VariablesEstimation When the Correlation between the Instruments and the Endogenous ExplanatoryVariables is Weak.” Journal of the American Statistical Association 90(430):443–450.
Cafiso, Gianluca and Roberto Cellini. 2012. “Evidence on Fiscal Consolidations and the Evolu-tion of Public Debt in Europe.” University of Catania.
Cecchetti, Stephen, Madhusudan Mohanty and Fabrizio Zampolli. 2011. The real effects of debt.BIS Working Papers 352 Bank for International Settlements.
Checherita, Cristina and Philipp Rother. 2010. The impact of high and growing governmentdebt on economic growth: an empirical investigation for the euro area. Working Paper Series1237 European Central Bank.
Cordella, Tito, Luca Antonio Ricci and Marta Ruiz-Arranz. 2010. “Debt Overhang or DebtIrrelevance?” IMF Staff Papers 57(1):1–24.
Cottarelli, Carlo. 2011. “The Risk Octagon: A Comprehensive Framework For AssessingSovereign Risks.” http://www.imf.org/external/np/fad/news/2011/docs/Cottarelli1.pdf.
Cottrell, Allin and Riccardo Lucchetti. 2012. “Gretl User’s Guide.”http://sourceforge.net/projects/gretl/files/manual/gretl-guide.pdf/download.
19
De Grauwe, Paul. 2011. The Governance of a Fragile Eurozone. Working Document 346 CEPS.
DeLong, Bradford J and Lawrence H Summers. 2012. “Fiscal Policy in a Depressed Economy.”The Brookings Institution.
Eichengreen, Barry J., Ricardo Hausmann and Ugo Panizza. 2005. The Pain of Original Sin.In Other People’s Money: Debt Denomination and Financial Instability in Emerging-MarketEconomies, ed. Barry J. Eichengreen and Ricardo Hausmann. University Of Chicago Presschapter 1.
Fisher, Franklin M. 1966. The Identification Problem in Econometrics. McGraw-Hill.
Galì, Jordi and Roberto Perotti. 2003. “Fiscal policy and monetary integration in Europe.”Economic Policy 18(37):533–572.
Gros, Daniel. 2011. “Can Austerity be self-defeating.” VoxEU.org.
Hansen, Bruce E. 1999. “Threshold effects in non-dynamic panels: Estimation, testing, andinference.” Journal of Econometrics 93(2):345–368.
Hausman, Jerry A, Whitney K Newey and William E Taylor. 1987. “Efficient Estimation andIdentification of Simultaneous Equation Models with Covariance Restrictions.” Econometrica55(4):849–74.
Hausman, Jerry A and William E Taylor. 1983. “Identification in Linear Simultaneous EquationsModels with Covariance Restrictions: An Instrumental Variables Interpretation.” Economet-rica 51(5):1527–49.
Hausmann, Ricardo and Ugo Panizza. 2011. “Redemption or Abstinence? Original Sin, CurrencyMismatches and Counter Cyclical Policies in the New Millennium.” Journal of Globalizationand Development 2(1):4.
Hausmann, Ricardo, Ugo Panizza and Roberto Rigobon. 2006. “The long-run volatility puzzleof the real exchange rate.” Journal of International Money and Finance 25(1):93–124.
Ilzetzki, Ethan and Carlos A. Végh. 2008. Procyclical Fiscal Policy in Developing Countries:Truth or Fiction? NBER Working Papers 14191 National Bureau of Economic Research, Inc.
Kraay, Aart. 2012. “Instrumental Variables Regressions with Uncertain Exclusion Restrictions:A Bayesian Approach.” Journal of Applied Econometrics 27(1):108–128.
Krugman, Paul. 1988. “Financing vs. forgiving a debt overhang.” Journal of Development Eco-nomics 29(3):253–268.
Krugman, Paul. 2011. “Self-defeating Austerity.” New York Times, July 7.
Kumar, Manmohan S. and Jaejoon Woo. 2010. Public Debt and Growth. IMF Working Papers10/174 International Monetary Fund.
Missale, Alessandro. 1999. Public Debt Management. Oxford: Oxford University Press.
Moreira, Marcelo J. 2003. “A Conditional Likelihood Ratio Test for Structural Models.” Econo-metrica 71(4):1027–1048.
20
Mussa, Michael. 1986. “Nominal exchange rate regimes and the behavior of real exchange rates:Evidence and implications.” Carnegie-Rochester Conference Series on Public Policy 25(1):117–214.
Newey, Whitney K and Kenneth D West. 1987. “A Simple, Positive Semi-definite, Heteroskedas-ticity and Autocorrelation Consistent Covariance Matrix.” Econometrica 55(3):703–708.
Panizza, Ugo. 2008. Domestic And External Public Debt In Developing Countries. UNCTADDiscussion Papers 188 United Nations Conference on Trade and Development.
Panizza, Ugo and Dany Jaimovich. 2007. Procyclicality or Reverse Causality? RES WorkingPapers 4508 Inter-American Development Bank, Research Department.
Pattillo, Catherine A., Helene Poirson and Luca Antonio Ricci. 2011. “External Debt andGrowth.” Review of Economics and Institutions 2(3).
Perotti, Roberto. 2012. The “Austerity Myth" Gain Without Pain? In Fiscal Policy after theFinancial Crisis. NBER Chapters National Bureau of Economic Research, Inc.
Powell, Andrew. 2012. The World of Forking Paths - Latin America and the Caribbean FacingGlobal Economic Risks. Washington, DC: Inter-American Development Bank.
Presbitero, Andrea F. 2012. “Total Public Debt and Growth in Developing Countries.” EuropeanJournal of Development Research forthcoming.
Reinhart, Carmen M. and Kenneth S. Rogoff. 2009. This Time is Different - Eight Centuries ofFinancial Folly. Princeton, New Jersey: Princeton University Press.
Reinhart, Carmen M. and Kenneth S. Rogoff. 2010a. “Debt and Growth Revisited.” VoxEU.org.
Reinhart, Carmen M. and Kenneth S. Rogoff. 2010b. “Growth in a Time of Debt.” AmericanEconomic Review 100(2):573–78.
Rodrik, Dani. 2008. “The Real Exchange Rate and Economic Growth.” Brookings Papers onEconomic Activity 39(2 (Fall)):365–439.
Sachs, Jeffrey D. 1989. The Debt Overhang of Developing Countries. In Debt, Stabilizationand Development, ed. Guillermo A. Calvo, Ronald Findlay, Pentti Kouri and Jorge Bragade Macedo. Oxford: Basil Blackwell.
Shapiro, Matthew D. 1987. Measuring Market Power in U.S Industry. Technical Report 828Cowles Foundation Discussion Paper.
Staiger, Douglas and James H. Stock. 1997. “Instrumental Variables Regression with WeakInstruments.” Econometrica 65(3):557–586.
Standard & Poor’s. 2012. Standard & Poor’s Takes Various Rating Actions On 16 EurozoneSovereign Governments. Ratingsdirect on the global direct portal Standard & Poor’s.
Stock, James H. and Motohiro Yogo. 2005. Testing for Weak Instruments in Linear IV Regres-sion. In Identification and Inference for Econometric Models: Essays in Honor of ThomasRothenberg, ed. Donald W.K. Andrews and James H. Stock. Cambridge University Presschapter 5, pp. 80–108.
21
Sutherland, Douglas, Peter Hoeller and Rossana Merola. 2012. Fiscal consolidation: Part 1. Howmuch is needed and how to reduce debt to a prudent level? OECD Economics DepartmentWorking Papers 932 OECD.
UNCTAD. 2011. Trade and Development Report 2011. Technical report United Nations.
Végh, Carlos A. and Guillermo Vuletin. 2012. How is Tax Policy Conducted over the BusinessCycle? NBER Working Papers 17753 National Bureau of Economic Research, Inc.
22
A Tables
Table 1: Country-by-Country Summary Statistics for V E
Mean Std. Dev. Min Max N. Obs
AUS 0.69% 1.57% -1.03% 5.37% 28AUT -0.01% 0.60% -1.49% 1.43% 28BEL 0.13% 0.93% -1.93% 2.98% 28CAN 0.02% 0.43% -0.85% 0.70% 28DEU 0.00% 0.01% -0.04% 0.04% 28DNK 0.30% 1.21% -1.98% 3.61% 28ESP 0.15% 0.51% -0.45% 1.89% 28FIN 0.84% 1.25% -2.68% 6.51% 28FRA 0.00% 0.00% 0.00% 0.00% 28GBR 0.04% 0.15% -0.28% 0.52% 28ITA -0.02% 0.23% -0.64% 0.46% 28JPN 0.00% 0.01% -0.02% 0.01% 28NLD 0.00% 0.02% 0.00% 0.01% 28NOR 0.23% 0.75% -0.65% 3.02% 28PRT 0.04% 0.44% -0.68% 1.16% 16SWE 0.43% 1.89% -3.41% 7.50% 28USA 0.00% 0.00% 0.00% 0.00% 28
Total 0.17% 0.97% -3.41% 7.50% 464
Notes: Data for Portugal start in 1993. We report VE in percentage term to enhance readability. In the regressions, VE
is used in levels.
23
Table 2: Baseline regressions
(1) (2) (3)
OLS Instrumental Variable Estimates
First Stage IV
Debt/GDP -1.796*** 0.322(0.588) (3.647)
Log Initial GDP PC -20.61*** -1.851** -16.62**(2.708) (0.734) (6.605)
National Gross Savings 0.022 -0.004 0.032(0.029) (0.009) (0.034)
Population Growth 3.630 -3.218 10.440(31.190) (5.542) (40.450)
Schooling 0.710*** 0.013 0.682***(0.179) (0.031) (0.186)
Trade openness 0.0200* -0.001 0.022(0.011) (0.005) (0.014)
Inflation rate 0.028 -0.0108* 0.048(0.033) (0.006) (0.053)
Dependency ratio -0.158*** -0.008 -0.140***(0.047) (0.015) (0.051)
Banking Crisis -0.025 -0.008 -0.013(0.232) (0.026) (0.218)
Liquid liabilities/GDP 0.476 0.050 0.377(0.358) (0.072) (0.490)
L.VE 1.594***(0.525)
N. Obs. 362 362 362R-squared 0.694 0.307 0.435N. of countries 17 17 17
Underidentification and weak instrument tests
K-P LM Stat. 3.82p-value 0.056K-P Wald Stat. 10.68p-value 0.001K-P F test 9.21
TSLS 95% confidence interval for Debt/GDP[-6.8, 7.4]
CLRT 95% confidence interval for Debt/GDP[-5.31, 19.39]
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
24
Table 3: Baseline regressions, controlling for foreign currency share and effective exchange rate
(1) (2) (3)
OLS Instrumental Variable Estimates
First Stage IV
Debt/GDP -1.536*** 2.051(0.480) (3.698)
Log Initial GDP PC -19.70*** -1.869** -12.91*(2.049) (0.752) (7.371)
National Gross Savings -0.019 -0.002 -0.010(0.024) (0.009) (0.032)
Population Growth 10.570 -3.447 22.860(24.020) (5.112) (40.090)
Schooling 0.650*** 0.015 0.601***(0.118) (0.030) (0.174)
Trade openness -0.001 -0.001 0.000(0.012) (0.005) (0.019)
Inflation rate 0.027 -0.011* 0.061(0.025) (0.005) (0.048)
Dependency ratio -0.203*** -0.006 -0.180***(0.041) (0.015) (0.054)
Banking Crisis 0.087 -0.012 0.125(0.194) (0.028) (0.185)
Liquid liabilities/GDP 0.427 0.052 0.247(0.313) (0.069) (0.493)
L.Foreign Currency debt -3.608*** 0.221 -4.610***(1.089) (0.296) (1.655)
L.Effective Exchange rate -0.033*** 0.001 -0.037***(0.007) (0.001) (0.010)
L.VE 1.458***(0.360)
N. Obs. 362 362 362R-squared 0.747 0.315 0.385N. of countries 17 17 17
Underidentification and weak instrument tests
K-P LM Stat. 4.240p-value 0.039K-P Wald Stat. 19.090p-value 0.000K-P F test 15.960
TSLS 95% confidence interval for Debt/GDP[-5.04, 9.39]
CLRT 95% confidence interval for Debt/GDP[-2.89, 28.37]
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
25
Table 4: The consequences of weak exclusion restrictions
Model that does not control for foreign currency Model that controls for foreign currencyshare and debt composition (Table 2) share and debt composition (Table 3)
η Median 95% confidence interval Median 95% confidence interval
5 0.41 -42.22 48.17 2.06 -43.93 59.8910 0.30 -30.32 35.58 1.96 -30.96 45.65100 0.30 -9.91 15.9 1.96 -8.37 23.42200 0.30 -7.69 13.66 1.97 -6.06 21.41500 0.33 -6.13 12.49 1.99 -4.47 20.86999 0.29 -5.01 10.88 1.96 -3.24 19.21
Standard IV 0.32 -6.83 7.47 2.05 -3.83 7.93
Notes: Results are estimated using GRETL (Cottrell and Lucchetti, 2012).
Table 5: Placebo regressions
(1) (2)
mean -1.50 -2.50Standard deviation of the mean 24.60 20.095th percentile -28.63 -25.52median -1.59 -1.8295th percentile 29.17 21.46
Model Column 3, Table 2 Column 3, Table 3
Notes: This table shows the distribution of the estimated coefficients of public debt in 1,000 replications of the regressions
of Tables 2 and 3 (column 3) where L.V E was substituted with a randomly generated variable with the same mean and
variance of L.V E.
26
Table 6: Regressions with Newey-West standard errors
(1) (2) (3) (4) (5) (6)Errors have AR(1) structure Errors have AR(5) structure
OLS First Stage IV OLS First Stage IV
Debt/GDP -1.536*** 2.051 -1.536*** 2.051(0.342) (3.840) (0.396) (3.805)
Log Initial GDP PC -19.700*** -1.869*** -12.910* -19.700*** -1.869*** -12.910*(1.709) (0.363) (7.383) (1.910) (0.489) (7.389)
National Gross Savings -0.019 -0.002 -0.010 -0.019 -0.002 -0.010(0.020) (0.005) (0.025) (0.022) (0.007) (0.028)
Population Growth 10.570 -3.447 22.860 10.570 -3.447 22.860(16.990) (4.037) (29.030) (20.130) (4.912) (34.380)
Schooling 0.650*** 0.015 0.601*** 0.650*** 0.015 0.601***(0.151) (0.019) (0.172) (0.152) (0.024) (0.184)
Trade openness -0.001 -0.001 0.000 -0.001 -0.001 0.000(0.009) (0.002) (0.013) (0.010) (0.003) (0.016)
Inflation rate 0.027 -0.0105** 0.061 0.027 -0.0105** 0.061(0.020) (0.005) (0.049) (0.022) (0.005) (0.048)
Dependency ratio -0.203*** -0.006 -0.180*** -0.203*** -0.006 -0.180***(0.038) (0.008) (0.049) (0.042) (0.010) (0.056)
Banking Crisis 0.087 -0.012 0.125 0.087 -0.012 0.125(0.115) (0.026) (0.156) (0.151) (0.028) (0.191)
Liquid liabilities/GDP 0.427* 0.052 0.427 0.052(0.226) (0.048) (0.262) (0.055)
L.Foreign Currency debt -3.608*** 0.221 -4.610*** -3.608*** 0.221 -4.610***(1.097) (0.204) (1.545) (1.125) (0.239) (1.579)
L.Effective Exchange rate -0.033*** 0.001 -0.037*** -0.033*** 0.001 -0.037***(0.006) (0.001) (0.009) (0.006) (0.001) (0.010)
L.VE 1.458** 1.458**(0.684) (0.612)
N. Obs. 362 362 362 362 362 362N. of countries 17 17 17 17 17 17
Notes: The table reports the regression coefficients and, in brackets, the associated Newey-West standard errors. *
significant at 10%; ** significant at 5%; *** significant at 1%.
27
Table 7: Baseline regressions, controlling for foreign currency share and effective exchange rateand last growth spell starting in 2001
(1) (2) (3)
OLS Instrumental Variable Estimates
First Stage IV
Debt/GDP -1.514** 4.095(0.545) (4.435)
Log Initial GDP PC -21.21*** -1.742** -11.350(1.993) (0.663) (8.862)
National Gross Savings -0.018 0.001 -0.022(0.026) (0.007) (0.036)
Population Growth 16.330 -2.089 28.410(21.450) (3.373) (35.860)
Schooling 0.648*** 0.014 0.574**(0.108) (0.031) (0.232)
Trade openness -0.004 0.002 -0.017(0.011) (0.004) (0.027)
Inflation rate 0.031 -0.00962* 0.078(0.022) (0.005) (0.055)
Dependency ratio -0.225*** -0.010 -0.169**(0.050) (0.013) (0.083)
Banking Crisis 0.038 0.011 -0.038(0.203) (0.023) (0.276)
Liquid liabilities/GDP 0.449 0.095 -0.066(0.331) (0.080) (0.796)
L.Foreign Currency debt -3.833*** 0.108 -4.803***(0.960) (0.306) (1.798)
L.Effective Exchange rate -0.035*** 0.001 -0.042***(0.006) (0.001) (0.012)
L.VE 1.331***(0.371)
N. Obs. 328 328 328R-squared 0.76 0.328 0.24N. of countries 17 17 17
Underidentification and weak instrument tests
K-P LM Stat. 3.970p-value 0.046K-P Wald Stat. 15.060p-value 0.000K-P F test 12.870
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
28
Table 8: Baseline regressions, controlling for foreign currency share and effective exchange rate,Excluding Japan
(1) (2) (3)
OLS Instrumental Variable Estimates
First Stage IV
ebt/GDP -1.365* 2.004(0.717) (4.285)
Log Initial GDP PC -18.06*** -1.393* -13.28**(1.603) (0.684) (6.623)
National Gross Savings -0.014 0.003 -0.021(0.024) (0.008) (0.028)
Population Growth -13.010 1.413 -18.210(18.970) (2.223) (16.480)
Schooling 0.604*** 0.025 0.522**(0.124) (0.037) (0.205)
Trade openness -0.002 0.004 -0.016(0.014) (0.003) (0.025)
Inflation rate 0.039 -0.0120* 0.075(0.025) (0.006) (0.057)
Dependency ratio -0.208*** -0.010 -0.173***(0.040) (0.012) (0.057)
Banking Crisis 0.279 -0.006 0.294*(0.164) (0.026) (0.176)
Liquid liabilities/GDP 0.587 -0.013 0.637(0.338) (0.058) (0.398)
L.Foreign Currency debt -3.215** 0.152 -3.902**(1.101) (0.325) (1.549)
L.Effective Exchange rate -0.0288*** 0.001 -0.0303***(0.009) (0.002) (0.010)
L.VE 1.375***(0.347)
N. Obs. 340 340 340R-squared 0.759 0.307 0.425N. of countries 16 16 16
Underidentification and weak instrument tests
K-P LM Stat. 4.410p-value 0.035K-P Wald Stat. 18.480p-value 0.000K-P F test 15.690
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
29
Table 9: Baseline regressions, controlling for foreign currency share and effective exchange rate,alternative measure of public debt
(1) (2) (3)
OLS Instrumental Variable Estimates
First Stage IV
Debt_UP/GDP -1.911*** 3.729(0.473) (7.458)
Log Initial GDP PC -19.490*** -1.394** -11.540(1.952) (0.630) (10.720)
National Gross Savings -0.027 -0.006 0.009(0.023) (0.007) (0.062)
Population Growth 5.370 -5.486 36.250(22.170) (5.425) (65.790)
Schooling 0.615*** -0.007 0.657***(0.115) (0.020) (0.193)
Trade openness -0.004 -0.002 0.006(0.012) (0.004) (0.027)
Inflation rate 0.033 -0.005 0.058(0.026) (0.004) (0.050)
Dependency ratio -0.213*** -0.010 -0.153(0.038) (0.012) (0.100)
Banking Crisis 0.082 -0.012 0.145(0.185) (0.028) (0.223)
Liquid liabilities/GDP 0.426 0.041 0.201(0.295) (0.065) (0.611)
L.Foreign Currency debt -3.746*** 0.120 -4.604**(1.070) (0.179) (1.820)
L.Effective Exchange rate -0.032*** 0.001 -0.039***(0.007) (0.001) (0.013)
L.VE 0.802**(0.374)
N. Obs. 362 362 362R-squared 0.751 0.263 0.213N. of countries 17 17 17
Underidentification and weak instrument tests
K-P LM Stat. 3.340p-value 0.067K-P Wald Stat. 5.450p-value 0.021K-P F test 4.580
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
30
Table 10: Baseline regressions, without outliers
(1) (2) (3)
OLS Instrumental Variable Estimates
First Stage IV
Debt/GDP -1.831*** 0.033(0.382) (4.276)
Log Initial GDP PC -19.480*** -1.712* -16.320**(2.088) (0.811) (7.071)
National Gross Savings -0.037 -0.003 -0.030(0.024) (0.010) (0.024)
Population Growth 4.342 -2.600 9.537(23.370) (5.063) (31.610)
Schooling 0.597*** -0.012 0.620***(0.130) (0.024) (0.161)
Trade openness -0.004 -0.002 -0.001(0.012) (0.005) (0.014)
Inflation rate 0.030 -0.0112* 0.050(0.025) (0.006) (0.044)
Dependency ratio -0.192*** -0.008 -0.179***(0.045) (0.016) (0.045)
Banking Crisis 0.055 -0.017 0.084(0.190) (0.030) (0.180)
Liquid liabilities/GDP 0.441 0.054 0.345(0.277) (0.068) (0.414)
L.Foreign Currency debt -3.231** 0.402 -4.055(1.218) (0.308) (2.589)
L.Effective Exchange rate -0.029*** 0.002 -0.032***(0.007) (0.001) (0.012)
L.VE 1.222*(0.665)
N. Obs. 338 338 338R-squared 0.755 0.272 0.523N. of countries 16 16 16
Underidentification and weak instrument tests
K-P LM Stat. 2.560p-value 0.100K-P Wald Stat. 3.990p-value 0.040K-P F test 3.380
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
31
Table 11: Non-overlapping growth spells
(1) (2) (3) (4)
3-year growth spells
Starting in Second stage First stage K-Pcoefficient of Debt/GDP coefficient of L.VE F testOLS IV
1981 -3.24*** 7.74 0.28 0.20(1.27) (8.43) (1.88)
1982 -1.81** -2.71 2.55 3.35(0.76) (4.61) (1.39)*
1983 -1.74** 2.66 3.85** 4.48(0.85) (3.31) (1.82)
N. Obs. 119 119 119 119
5-year growth spells
Starting in Second stage First stage K-Pcoefficient of Debt/GDP coefficient of L.VE F testOLS IV
1981 -2.08** -9.16 3.12 0.72(0.90) (7.72) (3.67)
1982 -0.93 2.08 1.27 0.58(0.95) (8.33) (1.66)
1983 -1.30 48.36 0.53 0.43(1.32) (73.91) (0.82)
1984 -1.99 6.94 1.98 0.42(1.26) (14.34) (3.04)
1985 -1.73* -0.54 4.28*** 12.27(0.94) (3.02) (1.22)
N. Obs. 68 68 68 68
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
32
Table 12: Regressions with 90% thresholds
(1) (2) (3) (4)
OLS Instrumental Variable Estimates
First Stage IV
D90*Debt/GDP (1-D90)* Debt/GDP
D90*Debt/GDP -0.846 3.784(0.697) (9.518)
(1-D90)* Debt/GDP -1.508** -0.319(0.637) (18.680)
D90 -0.262 0.822*** -1.042*** -2.805(0.902) (0.020) (0.025) (12.210)
Log Initial GDP PC -19.790*** -0.887*** -0.414*** -15.140(1.937) (0.207) (0.159) (15.980)
Nat. Gross Savings -0.013 -0.00475* 0.001 0.011(0.025) (0.002) (0.003) (0.040)
Population Growth 13.550 -3.208* -0.759 29.690(23.470) (1.865) (2.104) (49.370)
Schooling 0.617*** 0.038** -0.0180** 0.461***(0.111) (0.015) (0.008) (0.172)
Trade openness -0.007 0.003** -0.001 -0.023(0.013) (0.001) (0.001) (0.015)
Inflation rate 0.023 -0.002 -0.004** 0.031(0.024) (0.003) (0.002) (0.094)
Dependency ratio -0.197*** -0.014*** 0.011*** -0.146(0.042) (0.005) (0.004) (0.095)
Banking Crisis 0.076 0.021 -0.034* 0.022(0.184) (0.013) (0.018) (0.447)
Liquid liab./GDP 0.524 -0.051** 0.052* 0.704(0.312) (0.024) (0.028) (0.539)
L.For. Curr.debt -3.061** -0.137 -0.002 -2.580(1.078) (0.138) (0.098) (1.578)
L.Eff. Exch. rate -0.033*** 0.001 0.000 -0.037**(0.007) (0.001) (0.001) (0.016)
(1-D90)*L.VE 2.250*** -1.221(0.760) (1.145)
D90*L.VE 0.912 -0.120(0.652) (0.564)
N. Obs. 362 362 362 362R-squared 0.753 0.890 0.940 0.650N. of countries 17 17 17 17
Underidentification and weak instrument tests
F-test on 1st stage 5.46 0.58K-P LM Stat. 0.28p-value 0.59K-P Wald Stat. 0.28p-value 0.59K-P F test 0.13
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
33
Table 13: Regressions with debt below 90% thresholds
(1) (2) (3) (4) (5) (6)
OLS IV OLS IV
First stage First stage
Debt/GDP -0.743 4.563 -0.041 3.927(0.852) (6.388) (1.195) (5.353)
Log Initial GDP PC -18.440*** -0.865*** -13.700** -15.530*** -1.020*** -11.460**(2.612) (0.259) (5.755) (2.356) (0.196) (5.694)
Nat. Gross Savings -0.025 -0.003 -0.006(0.034) (0.004) (0.032)
Population Growth 1.062 -2.607 14.610(26.350) (1.850) (26.440)
Schooling 0.571*** 0.030* 0.407(0.135) (0.015) (0.251)
Trade openness -0.017 0.007*** -0.056(0.018) (0.002) (0.050)
Inflation rate 0.039 -0.006 0.060(0.036) (0.004) (0.046)
Dependency ratio -0.219*** -0.007 -0.175***(0.053) (0.005) (0.054)
Banking Crisis 0.265 0.021 0.158(0.286) (0.014) (0.185)
Liq. liab/GDP 0.438 -0.030 0.611*(0.348) (0.030) (0.313)
L.For. Curr. debt -3.096** 0.016 -3.395***(1.135) (0.152) (1.165)
L.Eff. Exch. rate -0.040*** 0.001** -0.047***(0.007) (0.001) (0.012)
L.VE 1.254 1.443*(0.815) (0.811)
N. Obs. 272 271 271 272 271 271R-squared 0.75 0.555 0.61 0.588 0.455 0.492N. of countries 17 16 16 17 16 16
Underidentification and Underidentification andweak instrument tests weak instrument tests
K-P LM Stat. 3.15 3.91p-value 0.07 0.04K-P Wald Stat. 2.72 3.48p-value 0.09 0.06K-P F test 2.37 3.16
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
34
Table 14: Regressions with debt above 90% thresholds
(1) (2) (3) (4) (5) (6)
OLS IV OLS IV
First stage First stage
Debt/GDP -0.877 -71.530 -1.525* 3.069(0.960) (348.000) (0.766) (5.244)
Log Init GDP PC -17.82*** -4.023*** -301.500 -20.59*** -3.696*** -3.658(3.219) (0.569) (1400.000) (3.241) (0.590) (19.220)
Nat. Gross Sav. -0.024 0.0205** 1.424(0.037) (0.008) (7.241)
Population Gr. -76.140 -5.219 -458.200(90.540) (9.837) (1913.000)
Schooling 0.096 0.025 1.647(0.392) (0.065) (7.589)
Trade openness -0.007 -0.004 -0.285(0.019) (0.003) (1.398)
Inflation rate 0.002 -0.009 -0.609(0.014) (0.006) (3.029)
Depend. ratio -0.015 0.0293*** 2.083(0.068) (0.011) (10.310)
Banking Crisis 0.192 -0.092 -6.284(0.238) (0.056) (32.590)
Liq. liab/GDP 0.545 0.012 1.461(0.473) (0.092) (7.453)
L.For. Curr. debt 2.976 -0.742 -48.530(3.389) (0.525) (246.700)
L.Eff. Exch. rate -0.008 -0.00847*** -0.606(0.010) (0.002) (2.922)
L.VE 0.110 -0.883(0.694) (0.722)
N. Obs. 90 90 90 90 90 90R-squared 0.831 0.817 -29.776 0.777 0.627 0.51N. of countries 8 8 8 8 8 8
Underidentification and Underidentification andweak instrument tests weak instrument tests
K-P LM Stat. 0.04 1.42p-value 0.84 0.23K-P Wald Stat. 0.04 2.04p-value 0.84 0.15K-P F test 0.02 1.49
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%.
35
Table 15: Identification through covariance restriction at different debt thresholds
(1) (2) (3) (4)
(First) First Stage
TPGROWTH 0.055 -1.392** 0.782** 0.547*(0.232) (0.654) (0.352) (0.304)
R-squared 0.660 0.675 0.836 0.768N. Obs. 362 271 90 119N. of countries 17 16 8 8
Underidentification and weak instrument tests
K-P LM Stat. 0.06 4.35 7.37 3.73p-value 0.81 0.04 0.01 0.05K-P Wald Stat. 0.06 5.16 7.79 4.50p-value 0.80 0.02 0.01 0.03K-P F test 0.06 4.53 4.94 3.24
Sample ALL Debt<90% Debt>90% Debt>80%
OLS and (Second) Second Stage
OLS a3SLS OLS a3SLS
Debt/GDP -0.747 12.960** -1.320* 28.031(1.076) (6.022) (0.632) (26.514)
Sample Debt>90% Debt>90% Debt>80% Debt>80%
Notes: The table reports the regression coefficients and, in brackets, the associated robust (Hubert-White sandwich cor-
rection) standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%. The model includes the same set
of controls used in the benchmark regressions of Table 2. Coefficients are not reported to save space.
36
B Figures
Figure 1: Partial correlation between L.V E and Public Debt over GDP in first stage regression
FIN1992
SWE1996
BEL1986
FIN2003
FIN1989
DNK1987AUS1982
AUS1997
SWE1986
DNK1995
AUS1988
DNK1990
ITA1983
SWE1995
FIN1987BEL1987AUS1991
FRA1983
FIN1990JPN1983
FIN1988AUS1996AUS1995
AUT2003USA1993
AUS2000
DEU1993
SWE2003
BEL1990
ITA1982
FIN2002
NOR1986
AUS1984
NLD1984
NLD1993
FRA1993
FIN2001
AUS1993
CAN1983SWE1990
JPN1993NOR1995
SWE1987
DEU1983
SWE2002
BEL1985USA1983
NLD2000
ESP1986
AUT1995
CAN1997
FRA1982
NOR1990
CAN1993
CAN2000
JPN1982
DNK1993
NOR1987FIN1996GBR1983
NOR1988
FIN1998
NLD1983
ITA1993SWE1994
NOR1989
DEU2000
FIN1985
AUT2001
GBR1993
CAN1999
JPN1999
NLD1999
DNK1994
DEU1982
DNK2003
CAN1982
NOR1992AUT1998AUS1994DEU1999
DNK2002
GBR1982
NLD1982
SWE1991GBR1997FRA1984
USA1982
NLD2001JPN1984
AUT2002PRT2001PRT2002
DEU2001
ITA1984
AUS2002
ESP2001
PRT2003
ITA1986
DNK1992DEU1997
FIN1984
JPN2000
FRA1985
AUS2003DNK1999
AUT1983
NOR1985CAN1991JPN1997
ITA1985
GBR1985
NOR1994GBR2000NOR1991
FRA2000AUT1989
GBR1986
NLD1998
CAN1989
PRT1996AUS2001
ESP1995
CAN1996ESP1998BEL1993
SWE1988USA1986
AUT1985DNK1998
USA1989
USA1994ESP1997
ESP1982GBR1994NOR1999SWE1992FRA1994
SWE1985
ESP1994
GBR1984
CAN2003USA2000GBR1991
USA1991
FRA1999
GBR1995DNK1988NOR1982ESP1999JPN1998ITA1994
JPN1985
DNK1991GBR1999
GBR1989FRA1995
DNK2001FRA1986NLD1994
DNK1985
SWE1998USA1999
DEU2002
JPN1994JPN1986
NLD2002
NLD1997ESP1985BEL2001DEU1994FRA1997DEU1998ESP1989AUT1984
ESP2000BEL1989
AUT1990
SWE1984
AUT1986ESP1984
SWE1989PRT1995
DEU1991
PRT1998USA1995
JPN2001
ESP1996
ESP2002NLD1991
BEL2000PRT1997
GBR1988
DEU1985DNK1989USA1997AUT1993FRA1998
BEL1983
NLD1986
AUS1992FRA1989
FIN1986
DEU1984
NOR1998ITA1989GBR1990
AUT1999USA1985
JPN1995
ITA1995
BEL1995USA1987
USA1984ITA1991
CAN1995
ESP1993
BEL1991
USA1988
NOR1997
GBR1987
DEU1995
SWE2000
NLD1995ESP1987DNK2000
ITA2003
USA2001
BEL1994ITA2001
NOR1996BEL2002AUT1997
JPN2002
ESP1991BEL1988
AUS1998
ITA1999
JPN1989
ITA1998
CAN1988PRT1999GBR1998
ESP2003
PRT1994FRA2001
AUT1987FRA1991
NLD1989
DEU1986
USA1990
AUT1994AUT1992ITA2000
FRA1987
DNK1996
ITA1997BEL1992
ESP1988CAN2001ESP1992BEL1998
NOR1984
USA2002
ITA1996
CAN1998ITA1990FRA1988
DEU2003
DEU1989
ITA1987
USA1992ITA2002
BEL1984
CAN1984AUT1996CAN1986
ESP1983USA1998
FRA2002
ESP1990
GBR2001JPN1991
CAN1994DEU1992
NLD1992
JPN1987
DNK1997
AUS1983
ITA1992CAN1990NLD1988
GBR1996
BEL1999CAN1987
AUT1982CAN1985BEL1997AUT1988
JPN1988
NOR2000
BEL2003
NLD1987
FIN1999
JPN2003
CAN2002USA2003GBR1992AUS1999ITA1988GBR2002SWE1999
USA1996DEU1988FRA1996NOR1983DEU1996JPN1996DEU1987
PRT2000
AUT1991DNK1983
FRA1992NLD1996
NLD1990
FRA1990
JPN1992
FIN1991
FIN1982
NLD2003
NOR2001
CAN1992
FRA2003NOR2002DEU1990
JPN1990
AUT2000
GBR2003
FIN1997
SWE2001
NLD1985BEL1996
NOR2003
SWE1997
AUS1990FIN1993
AUS1985FIN2000FIN1994
SWE1982
NOR1993
SWE1983
BEL1982
DNK1984
DNK1986
DNK1982
AUS1989AUS1987
AUS1986
FIN1983
FIN1995
SWE199
−.2
0.2
.4.6
e( D
EBT2
GD
P | X
)
−.04 −.02 0 .02 .04 .06e( L.VE | X )
coef = 1.5940399, (robust) se = .65984839, t = 2.42
All Countries
37
Figure 2: Partial correlation between L.V E and Public Debt over GDP in first stage regression,excluding Japan
FIN1992SWE1996
BEL1986
FIN2003
FIN1989
DNK1987
AUS1982
AUS1997
SWE1986
DNK1995
AUS1988
DNK1990
ITA1983
SWE1995
FIN1987
FRA1983
AUS1991BEL1987
USA1993
AUT2003
AUS1996
AUS1995
FIN1988
FIN1990
AUS2000
DEU1993
ITA1982
AUS1984
BEL1990
CAN1983
SWE2003
NOR1986
FIN2002
NLD1993
AUS1993
FRA1993
USA1983
DEU1983
NLD1984
FIN2001
BEL1985NOR1995
NLD2000
CAN1993
SWE2002
SWE1987
SWE1990
CAN1997
ESP1986
DNK1993
FRA1982
NOR1990
AUT1995
NLD1983GBR1983
CAN2000
FIN1996
ITA1993
NOR1987
FIN1998
SWE1994
NOR1988
NOR1989
DEU2000
GBR1993FIN1985
AUT2001
DEU1982
CAN1999
NLD1999
NLD1982
DNK1994
DEU1999
GBR1982
CAN1982
USA1982
DNK2003
AUT1998
DNK2002
NOR1992
FRA1984GBR1997AUS1994
ESP2001
NLD2001
SWE1991AUT1983
ITA1984
DEU2001PRT2001AUT2002
PRT2002
FIN1984
DEU1997
DNK1999
ITA1986
FRA1985
GBR2000
AUS2002
FRA2000
GBR1986
PRT2003
ESP1982
ITA1985
NOR1985
NOR1994
GBR1985
DNK1992
NOR1991
CAN1991
AUT1989USA1986
GBR1984
BEL1993
NLD1998
AUS2003
CAN1996
DNK1998
CAN1989
NOR1999
ESP1998
PRT1996AUT1985
USA1994
FRA1999
AUS2001
ESP1997
FRA1994
USA1989
NOR1982
GBR1994
SWE1992
ESP1995GBR1999
USA2000
SWE1988
GBR1995
SWE1985
ESP1999
ITA1994
ESP1985
USA1991
ESP1984
ESP1994
DNK1988
SWE1998
FRA1986GBR1991
USA1999
FRA1995
DNK2001
DNK1991
FRA1997
DEU2002
BEL2001
CAN2003
GBR1989DEU1998
DNK1985
DEU1994AUT1993
NLD2002
NLD1994
NLD1997
BEL1983
BEL1989
AUT1986
ESP1989
PRT1998
ESP2002
ESP2000
AUT1984
SWE1984
PRT1995
DEU1991BEL2000PRT1997
USA1995SWE1989
NLD1986
USA1997
FRA1998AUT1990
DNK1989
GBR1988
AUS1992DEU1985
NLD1991
FIN1986
DEU1984
AUT1999ESP1996
USA1984
NOR1998
FRA1989
USA1985CAN1995
ITA1989
ITA1995
SWE2000
ESP1993
USA1987BEL1995
DNK2000
ITA1999
GBR1987
NOR1997
ITA2001
PRT1999
ESP1987
USA1988
ITA2003
DEU1995
NLD1995
BEL1991
ESP2003
BEL2002
ITA1991
AUT1997
GBR1990
ITA1998
USA2001
ITA2000
AUS1998
BEL1994
BEL1988
FRA2001
NOR1996
NLD1989
AUT1994GBR1998
ITA1997
ESP1983
DEU1986
CAN1988
PRT1994AUT1987
AUT1992
NOR1984
ESP1991DNK1996
FRA1987
CAN1998
ESP1988
BEL1984
FRA1991
AUS1983
USA1990
BEL1998CAN2001
DEU2003
ITA2002
CAN1984DEU1989USA2002
BEL1992
CAN1994
FRA1988
ITA1996
ITA1987
USA1998
GBR2001
FRA2002
USA1992
DNK1997ITA1990CAN1986AUT1996
ESP1992AUT1982
NLD1988
CAN1990
BEL1999
ESP1990
DEU1992
CAN1985
BEL1997
NLD1992
NOR1983
FIN1999
GBR1996SWE1999
NOR2000
NLD1987
ITA1992
AUS1999
AUT1988CAN1987
BEL2003
DNK1983
GBR2002
CAN2002
PRT2000
USA2003
ITA1988
GBR1992
AUT1991FRA1996
DEU1987
USA1996
DEU1996
DEU1988FIN1982NLD1996
FIN1991
FRA1990
FRA1992
NLD1990
NLD2003
FRA2003
NOR2001
DEU1990
CAN1992
NOR2002AUT2000
FIN1997
NLD1985
GBR2003
SWE2001
BEL1996FIN1993
SWE1997
NOR2003
AUS1990
AUS1985FIN1994
NOR1993
FIN2000
SWE1983SWE1982
BEL1982
DNK1984
DNK1982
DNK1986
AUS1989
AUS1987AUS1986
FIN1983
FIN1995
SWE1993
−.2
−.1
0.1
.2
e(
DE
BT
2G
DP
| X
)
−.04 −.02 0 .02 .04 .06
e( L.VE | X )
coef = 1.4546559, (robust) se = .58151304, t = 2.5
Excluding Japan
Figure 3: Partial correlation between L.V E and Public Debt over GDP in first stage regression,excluding countries with no foreign currency debt
FIN1992
SWE1996
BEL1986
FIN1989
ITA1983
FIN2003AUS1982
DNK1987SWE1986
AUS1997
ITA1982
AUS1988
DNK1995
SWE1995AUS2000
AUS1993CAN1983
AUS1995
DNK1990
FIN1987
ITA1993
CAN1993
AUS1991
NOR1986
DNK1993
AUT2003
FIN1988
FIN1990
BEL1987
FIN2002
FIN2001
AUS1996
CAN2000
AUS1984
GBR1993
CAN1997
NOR1995
GBR1983
ESP1986
BEL1985
SWE2003SWE2002
CAN1999SWE1990
BEL1990
AUT1995
NOR1987
FIN1998
AUT2001
CAN1982
BEL1993
SWE1987
ITA1984
ESP1982
SWE1994
AUT1993DNK1994GBR1997
NOR1988SWE1991
AUT1983
GBR1982
FIN1985
BEL1983NOR1990
DNK1999
FIN1996
AUT1998NOR1989
PRT2001
ESP2001
NOR1999
FIN1984
AUS1994
GBR2000
DNK2002
AUT2002
ESP1984
ITA1985
SWE1984
NOR1994
NOR1992
ITA1986
ESP1993
PRT2002
ESP1999
BEL2000
ESP1983
NOR1982ESP2000GBR1984
NOR1991
GBR1999
AUS2001
GBR1989
ITA1994
AUS2002
CAN1991
ESP1994
GBR1995
GBR1994
DNK2003
DNK1998
GBR1986
AUT1984
SWE1985
GBR1991BEL2001
GBR1985
ESP1995
ESP1997
ESP1998AUS1983
CAN1989AUT1989
DNK1992
NOR1985
PRT1998
SWE1988
NOR1998DNK2001NOR1997
DNK1988
AUT1985
PRT2003
PRT1997
ESP1985
DNK1985
AUT1999
PRT1995
DNK2000
DNK1991SWE1998CAN2003
ITA1995
CAN1996
AUS2003
PRT1996
ESP2002
BEL1989
SWE1989
DNK1989
PRT1999
AUT1986
BEL1984
SWE2000
AUT1997PRT1994
AUT1994
BEL1994
ESP1989
GBR1988
GBR1998
CAN1995
ITA1991
GBR1990
FIN1986
DNK1983
BEL1991
BEL2002
ITA1989CAN2001
ITA1999
BEL1995
SWE1992
CAN1998
ITA2000
ESP1996
AUT1992
GBR1987
AUS1998
AUT1990
CAN1984
ITA1997
NOR1983
ITA2001
CAN1986
ITA1998
ESP1991NOR2000
FIN1999
DNK1997
BEL1998
BEL1997
CAN1988
NOR1996
ESP2003BEL1999
NOR1984
AUS1992
AUT1982CAN1994
AUS1999SWE1999
GBR2001
BEL1992
ESP1987
PRT2000
CAN2002BEL1988ESP1992
FIN1982
DNK1996
AUT1987
AUT1996
ITA2002ITA1996
ITA2003
CAN1985
ITA1987
ITA1992CAN1987
ESP1988
ITA1990
AUT2000
AUT1991
GBR1996
GBR2002
BEL2003
NOR2001AUT1988
FIN1993
FIN1991GBR1992
ESP1990
ITA1988
CAN1990
FIN1997
NOR2002
NOR1993
SWE1983
SWE2001
GBR2003CAN1992
FIN2000
SWE1997
BEL1996
SWE1982
AUS1985
NOR2003
BEL1982
AUS1990
FIN1994
DNK1984
DNK1982
DNK1986
AUS1989
AUS1987AUS1986
FIN1983
SWE1993
FIN1995
−.2
−.1
0.1
.2
e(
DE
BT
2G
DP
| X
)
−.04 −.02 0 .02 .04 .06
e( L.VE | X )
coef = 1.6977435, (robust) se = .62239094, t = 2.73
Excluding Countries without Foreign Currency Debt
38
Figure 4: Placebo regressions
−100
−50
050
100
dGR
/dD
0 200 400 600 800 1000n
Regressions without controls for FC and RER
−100
−50
050
100
dGR
/dD
0 200 400 600 800 1000n
Regressions with full set of controls
Notes: This figure plots the point estimates (with 95 percent confidence intervals) of the regressions described in Table 5.
To enhance readability we only plot observations for which the confidence interval ranges between -100 and +100.
Figure 5: Partial correlations between debt and growth: OLS estimates (column 1, Table 3)
FIN1991
NLD2003
NOR1990SWE2001
SWE2003
BEL1982
NLD2002
SWE2002
BEL2003ITA1984
NLD2001
ITA1983
FIN1986
NLD2000
ITA1986
BEL1983
ITA1985
JPN1984JPN1989
JPN1994
USA2002
JPN1982BEL1984FIN1990
JPN1993
GBR1992
USA2001
JPN1983
DNK2003
ITA1982SWE2000
ITA1987
CAN2003
JPN1987
NOR1989
JPN1995
JPN1988
BEL1985SWE1999FIN1985
JPN1985
FIN1987
BEL1986
JPN1992
JPN1986
DNK2002AUT1984
BEL2002AUT1986
GBR1991ESP1986NLD1997JPN1990DNK2001
ESP1985USA2000
ESP1987AUS2000
ESP1988NOR1995
SWE1998USA2003NOR1996
SWE1997
USA1982
AUT1985
USA1983
GBR2001
JPN1991NLD1999DNK2000NLD1996
CAN2002
GBR1993
GBR1999
AUS2001AUS2002FIN1992
GBR2002ESP1984
ITA1988
AUT1997
FRA1986
DEU1986
USA1998CAN1982
DNK1993
NOR1998CAN1983BEL2001
AUS2003
DEU1987
NLD1995
GBR1994
FIN1988
FIN1984
NOR1991
AUS1999
ESP1982DNK1999
ESP1983FRA1993FRA1987
NLD1998
USA1999
CAN2001
SWE1996NOR1988
AUT1987DNK1991DNK1990BEL2000
BEL1987
PRT1996AUS1998
SWE1990
AUS1997
SWE1995
GBR2000
GBR1996
PRT1997
SWE1993
ESP1992
NOR1997
CAN2000FRA1994
DNK1998
DEU1988
FRA1992FRA1988DEU1985
ESP1989AUT1982
ESP1993
DNK1989
CAN1984
GBR1998
FIN1989JPN1996
ITA1989
AUS1991DNK1994
FRA1991
DEU1994
NLD1994PRT1995ESP1994
NOR1999PRT1998
SWE1991
CAN1992
FRA1989
GBR1997
FRA1985
GBR2003AUS1992
AUT1983
USA1997AUT1996SWE1989USA1984
DEU1989GBR1995
NOR1992FRA1996
ESP1990AUT1995
BEL1998ESP1991DEU1993
DEU1999
USA1996
ITA1990
PRT1999
AUS1990DEU1984
DEU1998
FIN1982
GBR1990
ITA1992DEU1996
ITA1991
FRA1990AUT1998DNK1997
FIN1983
DNK1992CAN1999USA1995
ITA2003
ESP1995BEL1999NOR1994AUS1996
AUS1984NOR1983
SWE1992
DNK1988
DEU1995
DEU1997AUT1993FRA1997DNK1995FRA1995DNK1996
FRA1984
BEL1988DNK1982
PRT1994
CAN1998AUT1994AUT1988PRT2000ITA1993AUS1993
CAN1985
SWE1994
DEU2000
BEL1997
CAN1986
FRA1998DEU2002USA1985
FRA2000
CAN1990
CAN1989NOR1982
NOR1984AUS1994DEU2001AUS1995
FRA1999DEU1983NLD1985CAN1991
FRA2001
NLD1987DEU2003
CAN1997
BEL1996
CAN1987
NOR2000
FIN2003
GBR1989
PRT2003CAN1994
USA1986
FIN1993
AUT2003
DEU1992
SWE1988
FIN1995
AUS1982AUT1989
AUT1999
AUS1983NOR2001
ITA1994
DNK1983
CAN1988
NOR1993ESP1996
AUT1992
FIN2002
AUS1985
USA1994
AUT1990
AUS1989
DEU1990JPN1997FRA2002
PRT2001
JPN1998
BEL1995DEU1982NLD1988AUT2000ESP1997USA1993PRT2002FRA1983NLD1993
DEU1991
CAN1993NLD1986AUT2001
NOR1985
AUT2002
FIN1994
ITA2002
AUT1991
NOR2002
AUS1986
CAN1996
GBR1986
USA1992
FRA2003GBR1988
FIN2001
FIN1996
ITA1999ITA2000NLD1982
ESP2003NLD1989
CAN1995
USA1987ESP1998
SWE1987
NOR1987
NLD1983
DNK1987
ESP1999
USA1991
GBR1985ITA2001
AUS1988
NLD1992
ESP2000AUS1987GBR1984
GBR1987
NLD1990BEL1989ESP2002
USA1990FIN1999FIN2000NLD1991
FRA1982
USA1989USA1988BEL1994ITA1998
GBR1983
JPN1999
SWE1986BEL1993FIN1997
ESP2001
ITA1997GBR1982
NOR2003
DNK1984
ITA1995ITA1996FIN1998
NOR1986BEL1992SWE1985
DNK1985
BEL1991
DNK1986
NLD1984
JPN2000
BEL1990
SWE1982
SWE1984
SWE1983
JPN2001
JPN2002
JPN2003
-2-1
01
2
GD
P G
row
th |
X
-.2 0 .2 .4 .6Public Debt to GDP | X
coeff=-1.536, robust t-stat=3.20
39
Figure 6: Partial correlations between debt and growth: IV estimates (column 3, Table 3)
FIN1992
SWE1996BEL1986
FIN2003
FIN1989
DNK1987
AUS1988
SWE1995AUS1982SWE1986
DNK1990
DNK1995
AUS1997
FIN1987
AUT2003
BEL1987
FIN1988
AUS1991
ITA1983
USA1993
FRA1983
DEU1993FIN1990AUS1996
AUS1984AUS1995
JPN1983
BEL1990NOR1995
NLD1993
FIN1985BEL1985FRA1993AUS2000
NLD1984JPN1993
CAN2000CAN1983ESP1986
SWE2003
FIN2002
SWE1990
ITA1982
SWE1987
DEU1983
CAN1997
AUS1993SWE1994NLD2000
FIN2001
USA1983
SWE2002
DNK1994
FRA1982
NOR1982
FIN1984
CAN1999
AUT2001
FIN1996
PRT2002AUT1983
AUT1995
PRT2003
DNK1993
CAN1993
ITA1993
DEU2000
NOR1990
NOR1992
PRT2001
NLD1983
GBR1993NOR1986
NOR1989
JPN1999
GBR1983JPN1982
AUT2002
NLD1999
NOR1988GBR1997
FIN1998
CAN1982NOR1987
DEU1982
DEU1999
NOR1994
AUS1994
JPN2000
ESP2001
DNK1988
NLD2001
NLD1982FRA2000
AUT1984
JPN1997
USA2000GBR2000
AUT1985
GBR1982AUT1998
USA1982
DEU1997
CAN2003
SWE1998
DEU2001
GBR1991
FRA1984
DNK1992
FIN1986
CAN1996
ITA1986
SWE1991
GBR1994DNK1991FRA1985
DNK1985
SWE1988
CAN1991
JPN1998
NLD1998
ESP1985
ESP1998BEL1993
USA1994
USA1999
FRA1994FRA1999
USA1991
NLD1994JPN1984DNK1989
GBR1985
ESP1982
GBR1999
CAN1989SWE1985
USA1989
ITA1985
AUS2003
ITA1984
GBR1986
ESP1999DEU1994
NOR1991
ESP1995
ITA1994
AUS2002NLD1997
JPN1994
ESP1997
GBR1989FRA1997ESP1984
DEU1998
DNK2002ESP1994DNK1998BEL1983
USA1986
JPN1985
ESP2002
GBR1995USA1997
ESP2000
JPN1986
AUT1986
NOR1996
PRT1998
NLD1991
NLD2002
DEU1991
FRA1986
AUS2001JPN2001
AUT1989
DNK1999
FRA1995PRT1996DEU2002
DNK2003GBR1990NOR1997
GBR1988
BEL2001FRA1998
BEL1989ESP1989
ITA2003
ITA2001
USA1995SWE1989
NOR1985
NLD1986
GBR1984
DEU1985
SWE1984
AUT1982NOR1999
USA2001
ESP1996
PRT1999AUT1999
FRA1989
ITA1991
ESP1993
ITA1998
BEL2000
ITA1999
ESP2003
GBR1998
DEU1984
AUT1993AUS1992
ITA1989
ITA1995
BEL2002
USA1985
FRA2001
ITA2000
CAN1995
BEL1991
DEU1995NLD1995
CAN2001
ITA2002
JPN1995NOR1983
FRA1991
SWE2000
DNK2001
AUT1990
JPN2002
ITA1997
PRT1997
NOR1998
GBR1987
BEL1995BEL1988ESP1991
USA1988
JPN1989
ESP1987
DEU1986
ITA1996
SWE1992
USA1987JPN1991
USA1984
USA1998
NLD1989
BEL1992CAN1988BEL1994DEU1989DEU1992USA1992
USA1990
NLD1992
NOR1984
USA2002
FRA2002
AUS1998
ESP1992
PRT1995CAN1998BEL1984
GBR2001
AUT1992
FRA1987AUT1987
BEL2003
ESP1988
ITA1990
DNK1996GBR1992ESP1983AUT1997
ITA1992
CAN2002
SWE1999AUT1994FRA1988BEL1998
ESP1990
DNK2000
AUS1983
ITA1987
DNK1983
CAN1986
NLD1988
GBR1996
CAN1987
JPN1987
CAN1984
CAN1994CAN1990
DEU2003FRA1992
JPN1992
AUT1996NOR2000DNK1997
GBR2002
USA1996
BEL1999NLD1987
JPN1988
PRT2000
DEU1988
FRA1996DEU1996
CAN1985
DEU1987
JPN1996
NLD1990
JPN2003
USA2003
NLD1996
ITA1988
FIN1999BEL1997
AUS1999
FRA1990
PRT1994
JPN1990
FIN1982
DEU1990
FRA2003
AUT1988
NLD2003
AUT1991
NOR2001AUT2000CAN1992
FIN1991
NOR2002FIN1997
SWE1997
AUS1990
GBR2003
NOR1993
BEL1996
SWE2001NLD1985
AUS1985
NOR2003
FIN1994
BEL1982
FIN1993
FIN2000
SWE1983
SWE1982
DNK1984
DNK1982DNK1986
AUS1989
AUS1987
AUS1986
FIN1983
FIN1995
SWE1993
-2-1
01
2
GD
P G
row
th |
X
-.05 0 .05 .1Public Debt to GDP | X
coeff=2.05, robust t-stat=0.67
Figure 7: Leverage plot based on the IV estimates of column 3, Table 3
AUS1982
AUS1983AUS1984AUS1985
AUS1986
AUS1987
AUS1988AUS1989
AUS1990AUS1991AUS1992AUS1993AUS1994AUS1995AUS1996
AUS1997
AUS1998 AUS1999AUS2000AUS2001AUS2002 AUS2003AUT1982AUT1983AUT1984AUT1985AUT1986AUT1987AUT1988AUT1989AUT1990AUT1991AUT1992AUT1993AUT1994AUT1995AUT1996AUT1997AUT1998AUT1999AUT2000AUT2001AUT2002
AUT2003BEL1982
BEL1983BEL1984BEL1985
BEL1986
BEL1987
BEL1988BEL1989BEL1990
BEL1991BEL1992BEL1993BEL1994BEL1995BEL1996
BEL1997BEL1998BEL1999BEL2000BEL2001BEL2002BEL2003CAN1982CAN1983CAN1984CAN1985 CAN1986CAN1987CAN1988CAN1989CAN1990CAN1991CAN1992CAN1993CAN1994CAN1995CAN1996CAN1997CAN1998CAN1999CAN2000CAN2001CAN2002CAN2003DEU1982DEU1983DEU1984DEU1985 DEU1986DEU1987DEU1988DEU1989DEU1990DEU1991DEU1992
DEU1993DEU1994DEU1995DEU1996DEU1997DEU1998DEU1999DEU2000DEU2001DEU2002DEU2003
DNK1982
DNK1983
DNK1984
DNK1985
DNK1986DNK1987
DNK1988DNK1989
DNK1990
DNK1991DNK1992DNK1993DNK1994
DNK1995
DNK1996DNK1997DNK1998DNK1999DNK2000DNK2001DNK2002DNK2003ESP1982ESP1983ESP1984 ESP1985ESP1986ESP1987ESP1988ESP1989ESP1990ESP1991ESP1992ESP1993ESP1994ESP1995ESP1996ESP1997ESP1998ESP1999ESP2000ESP2001ESP2002ESP2003 FIN1982
FIN1983
FIN1984FIN1985FIN1986
FIN1987 FIN1988
FIN1989
FIN1990FIN1991
FIN1992
FIN1993FIN1994
FIN1995
FIN1996FIN1997FIN1998FIN1999
FIN2000
FIN2001 FIN2002
FIN2003
FRA1982FRA1983
FRA1984FRA1985FRA1986FRA1987FRA1988FRA1989FRA1990FRA1991FRA1992FRA1993FRA1994FRA1995FRA1996FRA1997FRA1998FRA1999FRA2000FRA2001FRA2002FRA2003GBR1982GBR1983GBR1984 GBR1985GBR1986GBR1987GBR1988GBR1989GBR1990GBR1991GBR1992GBR1993GBR1994GBR1995GBR1996GBR1997 GBR1998GBR1999GBR2000GBR2001GBR2002GBR2003ITA1982
ITA1983ITA1984ITA1985ITA1986ITA1987ITA1988ITA1989ITA1990ITA1991ITA1992ITA1993ITA1994ITA1995ITA1996ITA1997ITA1998 ITA1999ITA2000ITA2001ITA2002ITA2003JPN1982
JPN1983JPN1984 JPN1985JPN1986JPN1987JPN1988JPN1989JPN1990JPN1991JPN1992JPN1993JPN1994JPN1995JPN1996JPN1997JPN1998JPN1999JPN2000JPN2001 JPN2002 JPN2003NLD1982NLD1983NLD1984NLD1985
NLD1986NLD1987NLD1988NLD1989NLD1990NLD1991NLD1992NLD1993
NLD1994NLD1995NLD1996NLD1997NLD1998NLD1999NLD2000NLD2001NLD2002NLD2003NOR1982NOR1983NOR1984 NOR1985NOR1986NOR1987NOR1988NOR1989NOR1990NOR1991NOR1992NOR1993
NOR1994NOR1995
NOR1996NOR1997NOR1998NOR1999NOR2000NOR2001NOR2002NOR2003
PRT1994PRT1995PRT1996PRT1997PRT1998PRT1999PRT2000PRT2001PRT2002PRT2003
SWE1982SWE1983
SWE1984SWE1985
SWE1986
SWE1987SWE1988SWE1989SWE1990SWE1991SWE1992
SWE1993
SWE1994
SWE1995
SWE1996
SWE1997SWE1998SWE1999SWE2000SWE2001
SWE2002SWE2003 USA1982USA1983USA1984USA1985USA1986USA1987USA1988USA1989USA1990USA1991USA1992USA1993
USA1994USA1995USA1996USA1997USA1998USA1999USA2000USA2001USA2002USA2003
0.0
2.0
4.0
6.0
8.1
Leve
rage
0 .01 .02 .03 .04 .05Normalized residual squared
40
Figure 8: Partial correlations between debt and growth: OLS estimates without outliers (column1, Table 10)
BEL1982
NLD2003NLD2002BEL2003
ITA1984
USA2002
NOR1990
BEL1983SWE2003JPN1994
SWE2001NLD2001
ITA1986JPN1989USA2001ITA1985
SWE2002NLD2000
BEL1984
ITA1983
GBR1992
JPN1993
ITA1987
JPN1995
JPN1984
NOR1989
BEL1986BEL1985
JPN1992
JPN1988
AUT1984JPN1982
JPN1987
GBR1991
CAN2003
JPN1990
DNK2003
USA2000
USA2003
AUT1986
ITA1982
JPN1991JPN1983
GBR2001
AUT1985
JPN1985
SWE2000BEL2002
ESP1988
JPN1986
ESP1987
CAN1983
NOR1995
GBR2002
NOR1996
ESP1985CAN1982
ESP1986SWE1990
DNK2002
ITA1988
NLD1997
GBR1999
SWE1999DNK2001NLD1996FRA1986AUS2000NLD1999
USA1983
USA1998
BEL1987
AUS2003
FRA1987AUS2002
NOR1988
DNK2000
GBR1993
DEU1986DEU1987
SWE1993
SWE1991
FRA1993SWE1997
USA1999
USA1982
NLD1995
CAN1984AUS2001
ESP1992
DNK1991
CAN2002
AUT1997
DNK1990SWE1998
GBR2000
GBR1994AUT1987
DNK1999
NOR1998
DNK1993
ESP1984
FRA1991AUT1982
FRA1992
AUS1999
JPN1996
NLD1998
BEL2001
ESP1989ESP1982
FRA1988
ITA1989AUS1997ESP1993
FRA1994
NOR1991AUT1983
DNK1998SWE1989
AUS1998GBR2003ESP1991PRT1996
DNK1989
CAN2001AUS1991SWE1995NOR1997ESP1983CAN1992
DEU1988DEU1985
DEU1994
GBR1998
DNK1994
PRT1997
GBR1996
ITA1991
ESP1990NOR1983
NLD1994BEL2000ITA1990FRA1985
FRA1989FRA1990
DEU1984SWE1992ITA1992
ESP1994AUT1996
PRT1998
PRT1999
CAN2000
USA1997AUT1995USA1984
GBR1997
AUS1984
DEU1996
DEU1993
FRA1996AUS1990
AUS1992
NOR1982
GBR1990NOR1999DEU1989
NOR1984
USA1996
DEU1995DNK1995
DNK1997
DNK1988
CAN1985
PRT1995
DEU1999
CAN1990
GBR1995DNK1996
DEU1998
NOR1992BEL1998
ITA2003
BEL1988USA1985AUT1993
PRT2000
USA1995
ESP1995
FRA1984
DEU1997
CAN1989SWE1994DNK1992CAN1991AUT1998FRA1995
CAN1986
DNK1982
NOR1994AUS1996AUT1994FRA1997
BEL1999
ITA1993
NLD1987
PRT2003
DEU1983
SWE1988
AUT1988
CAN1987DEU2000
NLD1985
CAN1999BEL1997
AUS1985
BEL1996
FRA1998
FRA2000
CAN1988USA1986DEU1992
PRT1994
GBR1989
DEU2002AUS1993FRA1999
DEU2001AUS1983AUS1982
PRT2001
AUS1995FRA2001AUS1994
PRT2002
CAN1998
AUT2003
DEU1990
CAN1994
NOR2000
DNK1983AUT1992
AUT1999DEU1982DEU1991
ESP1996
DEU2003
ITA1994JPN1997
BEL1995
AUS1989
AUT1990AUT1989
NLD1988NLD1986NOR2001
USA1994NOR1993CAN1997NLD1993
NOR1985
FRA2002JPN1998SWE1987
AUT2000
FRA1983
ESP1997
AUT1991
CAN1993USA1993
ESP2003
ITA2002
AUT2001
AUT2002
GBR1988
NOR2002DNK1987NOR1987
GBR1986
NLD1982
USA1992
ITA1999ITA2000
NLD1989FRA2003
USA1987NLD1992ESP1998NLD1983
ESP1999
NLD1990
CAN1996
NLD1991
SWE1986ESP2002USA1991
CAN1995
ITA2001
AUS1988AUS1987
GBR1985
ESP2000
GBR1987
USA1990
GBR1984BEL1989
FRA1982
BEL1994ESP2001USA1989USA1988
JPN1999
ITA1998
GBR1983
BEL1993
DNK1984
ITA1997
NOR2003
GBR1982
ITA1995ITA1996
SWE1985
NOR1986
BEL1992BEL1991
DNK1985
DNK1986
SWE1982
JPN2000SWE1984
BEL1990
NLD1984
SWE1983
JPN2001
JPN2002
JPN2003
-2-1
01
2
GD
P G
row
th |
X
-.2 0 .2 .4 .6Public Debt to GDP | X
coeff=-1.83, robust t-stat=4.80
Figure 9: Partial correlations between debt and growth: IV estimates without outliers (column3, Table 10)
DNK1987
BEL1986
SWE1986
DNK1990
AUS1988
AUS1997
SWE1995
AUS1982
AUT2003
BEL1987
DNK1995
AUS1984SWE1987
AUS1996SWE1990
SWE2003BEL1990
USA1993
DEU1993
FRA1993
JPN1993
ITA1983
AUS1991
NLD1993FRA1983
NOR1995
CAN1983
ITA1982
PRT2003NOR1992
SWE2002SWE1994
PRT2002FRA1982
BEL1985PRT2001
CAN2000DEU1983AUS2000JPN1983
NLD2000
SWE1988
AUS1995
CAN1989ESP1986
AUT1983
GBR1993NOR1990
CAN1997
NOR1996
NLD1984
DEU1982NLD1999AUT2001SWE1992
CAN1993DNK1994
NOR1989
NLD1983SWE1989AUS1993ITA1993DNK1993
DEU2000
SWE1985AUT1984SWE1998
GBR1997
DNK1992
NLD1982
JPN2000
SWE1991
CAN1982
NLD2001
JPN1999
AUT1995
USA1983
AUT1998
DNK1985
NOR1994
DEU1999
JPN1982
DNK1988
AUT1985
ESP1989
NOR1987
AUT2002
CAN1999
CAN1996GBR1983
GBR2000
NOR1988
USA1989
SWE1984PRT1996NLD1998
ESP1982DEU2001
FRA1984
NOR1986
DNK2003
ESP1985
CAN2003AUT1989
GBR1982JPN2001
USA2000
JPN1997DEU1997
AUS2003
ITA2003
NOR1982GBR1994
DNK1989
GBR1999
BEL1993GBR1989
DNK2002
FRA2000
CAN1991NOR1997
AUT1982
ITA1984
FRA1999
USA1999
GBR1991ESP1984
NLD2002
AUS1994
FRA1994
ESP2001JPN1994
FRA1985
FRA1989
ESP1992
ESP1998
ITA1989
USA1994
BEL1992
ITA1985
DEU1998
DEU1994
ITA2001
DNK1991
DNK1998NOR1999
JPN1998
BEL2001ESP1996
USA1985NOR1998
CAN1988ESP1987
USA1982
JPN1984
GBR1988
NLD1997
BEL1989PRT1998
USA2001
ESP1991AUS2002
NOR1985GBR1985
FRA1998
JPN1989
USA1991
NLD1994DEU1984
USA1997ESP1999
AUT1990
FRA1997DEU2002
DNK1999
NOR1991
DEU1991
BEL2002
ITA2002
GBR1998
JPN2002
ITA1991
ITA1996GBR1984
ITA1992DNK2001
ITA1986
ESP2002ESP1988
JPN1985
USA1984
DEU1985
DEU1992
ESP1990
ESP1997
ITA1994
USA1986NLD1991USA1987AUS2001CAN1984
BEL2003GBR2001
GBR1990DNK1996
FRA2001
USA1992
GBR1996ESP1994
ESP2000
AUT1999
ITA1998
SWE2000
AUS1992
NLD1989USA1988
BEL1988FRA1991GBR1992
BEL2000
PRT1999BEL1991NLD1992AUT1996
AUT1992
CAN1987
USA1998USA2002ESP1993
ITA1999
ESP2003
ITA1990BEL1998DEU1996
JPN1996
DEU1989
USA1990
PRT1997
FRA1992AUT1987
BEL1994
CAN1985JPN1992
JPN1991
ITA2000
GBR1987
FRA1987
FRA1986AUT1986
AUT1993ITA1997
GBR1986
GBR1995
FRA1995
CAN1990
FRA1988
ITA1987
USA1996
NOR1984
FRA1996
NLD1986CAN2001
FRA2002JPN1995
NLD1996BEL1999DEU1995
BEL1995
JPN2003
AUS1998GBR2002ESP1995SWE1999
JPN1986
USA1995DEU2003NLD1987
JPN1988JPN1987
AUT1997NLD1988
CAN1992NOR2000
ITA1988
USA2003
NLD2003
AUT1994
NOR2002
DEU1988
NOR2001
NLD1995DNK2000CAN1998
DEU1986
BEL1983
BEL1997
DEU1987
FRA1990
CAN2002
NLD1990
CAN1994
CAN1986
JPN1990
ITA1995
BEL1984ESP1983
CAN1995
AUT1988DEU1990
FRA2003PRT2000
DNK1997
PRT1995
BEL1996GBR2003
NLD1985
DNK1983
PRT1994
NOR1983
AUS1999
AUT1991
SWE1997
AUS1983
NOR2003SWE2001
AUT2000
AUS1985
NOR1993AUS1990SWE1982
BEL1982SWE1983
DNK1984
DNK1982DNK1986
AUS1989
AUS1987
-2-1
01
2
GD
P G
row
th |
X
-.04 -.02 0 .02 .04Public Debt to GDP | X
coeff=0.033, robust t-stat=0.01
41
Figure 10: Hansen LR test for Threshold Regressions
02
46
810
LR S
tatis
tic
50 60 70 80 90 100 110 120Debt Threshold
LR_OLS LR_IV
Notes: The solid line plots Hansen (1999) LR statistics obtained with OLS estimations of Equation 7. The dashed line
plots the statistics obtained with IV estimates of the same equation. The horizontal line is the 10 percent critical value
derived by Hansen (1999).
42
Figure 11: Coefficients in OLS threshold regressions
-4-3
-2-1
01
dGR
/dD
ebt
50 60 70 80 90 100 110 120Debt Threshold
Debt<Threshold
-4-3
-2-1
01
dGR
/dD
ebt
50 60 70 80 90 100 110 120Debt Threshold
Debt>Threshold
OLS Estimates
Notes: The solid lines plot the coefficients at different thresholds and the dotted lines are 95% confidence intervals.
43
Figure 12: Difference in coefficients in OLS threshold regressions
-.50
.51
Deb
t Bel
ow-D
ebt A
bove
50 60 70 80 90 100 110 120Debt Threshold
Notes: The solid line plots(
∂GROWTH∂Debt
|Debt < Threshold)−
(∂GROWTH
∂Debt|Debt > Threshold
). Since ∂GROWTH
∂Debtis
always negative, a positive value indicates that the negative effect of debt on growth is larger above the threshold. The dotted
line plots the p-value for the null hypothesis that(
∂GROWTH∂Debt
|Debt < Threshold)−(
∂GROWTH∂Debt
|Debt > Threshold)= 0.
The horizontal line is at 0.05. Therefore, whenever the dotted line is above the horizontal line the difference between
coefficients is not different form zero at the 5 percent confidence level.
44
Figure 13: Coefficients in IV threshold regressions
-50
-25
025
50
dGR
/dD
ebt
50 60 70 80 90 100 110 120Debt Threshold
Debt<Threshold
-50
-25
025
50
dGR
/dD
ebt
50 60 70 80 90 100 110 120Debt Threshold
Debt>Threshold
IV Estimates
Notes: The solid lines plot the coefficients at different thresholds and the dotted lines are 95% confidence intervals. In the
figure, we bounded the coefficients and standard errors at (-50, 50).
45
C Additional Tables
Table C.16: Summary Statistics
Mean Std. Dev. Min Max N. Obs.
Average 5-yr GDP per capita growth overall 2.02 1.16 -2.95 6.44 N = 391between 0.34 1.5 2.66 n = 17within 1.12 -2.99 5.8 T = 23
Debt/GDP overall 0.74 0.3 0.19 1.9 N = 476between 0.24 0.43 1.2 n = 17within 0.19 0.21 1.52 T = 28
Log Initial GDP PC overall 10.21 0.25 9.27 10.84 N = 476between 0.19 9.66 10.53 n = 17within 0.17 9.82 10.55 T = 28
National Gross Savings overall 22.23 4.59 10.61 39.9 N = 476between 3.74 16.13 30.13 n = 17within 2.81 11.68 32.59 T = 28
Population Growth overall 0.50% 0.42% -0.34% 2.10% N = 459between 0.34% 0.18% 1.31% n = 17within 0.25% -0.10% 1.56% T = 27
Schooling overall 8.95 1.88 3.78 12.05 N = 476between 1.82 4.97 11.85 n = 17within 0.64 6.34 10.04 T = 28
Trade openness overall 64.16 30.8 16.11 170.53 N = 476between 30.16 22.41 139.85 n = 17within 9.52 44.3 95.94 T = 28
Inflation rate overall 3.89 3.72 -9.63 28.78 N = 476between 1.7 1.01 8.78 n = 17within 3.34 -8.18 23.88 T = 28
Dependency ratio overall 50.05 3.29 43.12 58.38 N = 476between 2.7 46.26 55.06 n = 17within 1.98 46.24 59.05 T = 28
Banking Crisis overall 0.16 0.36 0 1 N = 476between 0.11 0.04 0.36 n = 17within 0.35 -0.2 1.12 T = 28
Liquid liabilities/GDP overall 0.75 0.38 0 2.42 N = 476between 0.34 0.43 1.93 n = 17within 0.2 -0.04 1.46 T = 28
VE overall 0.17% 0.97% -3.41% 7.50% N = 464between 0.26% -0.02% 0.84% n = 17within 0.94% -3.67% 7.24% T-bar = 27.29
Foreign Currency debt overall 0.11 0.14 0 0.66 N = 467between 0.13 0 0.52 n = 17within 0.05 -0.09 0.28 T-bar = 27.47
Effective Exchange rate overall -1.47 10.65 -34.67 43.02 N = 476(deviation form country average) between 6.16 -12.12 14.12 n = 17
within 8.82 -29.4 27.42 T = 28
Debt_UP/GDP overall 0.61 0.30 0.10 1.95 N = 467between 0.26 0.2 1.12 n = 17within 0.16 0.07 1.46 T-bar = 27.47
TPGROWTH overall 0.31 0.29 -0.01 1.77 N = 391between 0.28 0.03 1.08 n = 17within 0.12 -0.42 0.99 T-bar = 23
Notes: VE and Population Growth are reported in percentage here but used in levels in the regressions.
46
Table C.17: Description and sources of the variables used in the paper
Variable Description and sources
Average 5-yr GDP per capita growth Average five-year growth of real GDP per capita from period t+1 to pe-riod t+6. The original data are from the Penn World Tables. We used thedataset provided with the Cecchetti et al. (2011) paper on the BIS websiteat: http://www.bis.org/publ/work352.htm
Log Initial GDP PC Same source as aboveBanking Crisis Dummy variable that tables a value one in country-years with a banking crisis.
The original data are from Carmen Reinhart. We obtained the data from:http://www.bis.org/publ/work352.htm
Debt/GDP Public debt as a share of GDP based on national accounts (flow of funds)data. We obtained the data from: http://www.bis.org/publ/work352.htm
Debt_UP/GDP Public debt as a share of GDP. The data are from Panizza (2008)Dependency ratio Population aged 0-15 and older than 64 over population aged 15-64. The
original data are from the World Bank’s WDI. We obtained the data from:http://www.bis.org/publ/work352.htm
Inflation rate CPI inflation. The original data are from the World Bank’s WDI. We obtainedthe data from: http://www.bis.org/publ/work352.htm
Liquid liabilities/GDP Liquid liability as a share of GDP (proxy for financial depth). The orig-inal data are from the World Bank’s WDI. We obtained the data from:http://www.bis.org/publ/work352.htm
National Gross Savings The original data are from the IMF World Economic Outlook. We obtainedthe data from: http://www.bis.org/publ/work352.htm
Population Growth The original data are from the World Bank’s WDI. We obtained the datafrom: http://www.bis.org/publ/work352.htm
Schooling Average years spent in school of population aged 15 and older. The Orig-inal data are from Barro and Lee (2010). We obtained the data from:http://www.bis.org/publ/work352.htm
Trade openness Total trade as a share of GDP. The original data are from the Penn WorldTables. We obtained the data from: http://www.bis.org/publ/work352.htm
VE Valuation effect. This variable is described in great details in Section 2 of thepaper.
Effective Exchange rate Real effective exchange rate (deviation from country mean). The data arefrom: http://www.bis.org/publ/work352.htm
Foreign Currency debt Foreign currency public debt over total public debt. The sources are the sameas those for VE.
TPGROWTH Trading Partners Growth. The construction of this variable is described ingreat detail in Section 5. The original data are from Panizza and Jaimovich(2007) and the Correlates of War Project’s Trade Data Set (Barbieri, Keshkand Pollins, 2012, 2009).
47