The Dynamic Macroeconomic Effects of Sovereign Debt on Credit Supply and the Business Cycle - A Cross-Country Panel Vector Autoregression Analysis*

(1)

The Dynamic Macroeconomic Effects of Sovereign Debt on Credit Supply and

the Business Cycle - A Cross-Country Panel Vector Autoregression Analysis

* M.D.B. (Matthijs) Katz, 1877763

m.d.b.katz@student.rug.nl

Thesis for MSc Economics and MSc International Economics and Business University of Groningen, Faculty of Economics and Business

Supervisors: dr. D.J. Bezemer and dr. J.P.A.M. Jacobs Co-assessor: dr. G.H. Kuper

13-6-2016

Abstract: This paper investigates the magnitude and duration of sovereign debt shocks on

financial institutions’ lending decisions and the real economy in relatively stable economic conditions by estimating a four variable panel vector autoregression system using an unbalanced panel of 14 OECD countries from 1990 to 2007. Sovereign debt shocks have large and persistent negative effects on both credit supply and output. This finding is robust to different model specifications. The results of the effects of credit shocks on output and vice versa show more heterogeneity, as household credit shocks negatively affect growth, whereas firm credit shocks do not.

Keywords: Panel vector autoregression, sovereign debt, credit supply JEL classification: E32, E43, E44, F34

*

(2)

1 Table of contents

1. Introduction ... 2

2. Literature review ... 4

2.1 Sovereign debt ... 4

2.2 The link between sovereign debt and financial institutions ... 6

2.3 Credit and output ... 9

3. Theoretical model ... 11

4. Methodology ... 14

5. Data ... 16

6. Empirical analysis and discussion ... 22

6.1 Benchmark model ... 23

6.2 Household credit model ... 25

6.3 Firm credit model ... 27

6.4 Discussion ... 27

7. Robustness tests ... 30

7.1 Hodrick-Prescott output gap ... 30

7.2 Annual data ... 32

7.3 Alternative recursive orderings ... 32

7.4 Country groupings ... 33

7.5 Individual country models ... 35

8. Conclusion ... 37

References ... 40

Appendices ... 47

Appendix A: List of countries in sample ... 47

Appendix B: List of variables ... 47

Appendix C: List of countries with missing credit observations ... 48

Appendix D: Graphs of variables per country, per quarter ... 48

Appendix E: Granger causality tests ... 52

Appendix F: Panel VAR stability conditions ... 54

(3)

2 1. Introduction

Since the start of the European debt crisis, the issues of sovereign debt and debt crises have again received attention from researchers and practitioners. In the last decades of the twentieth century, sovereign debt crises were mostly the domain of emerging economies that suffered macroeconomic mismanagement of some sort (Reinhart and Rogoff, 2013). However, the European debt crisis that followed 2008’s global financial crisis has reminded economists that debt crises can also occur in advanced economies. In May 2016, European Union Member States continued to negotiate potential caps on the amount of sovereign debt financial institutions can hold, indicating how important the subject is for financial and economic stability (Guarascio and Strupczewski, 2016).

One issue that can make sovereign debt crises so severe is that financial institutions hold sovereign debt in the form of bonds on their balance sheets, which in turn can make these institutions run into problems when the value of these bonds fall. This in turn will impair financial institutions’ ability to supply credit to consumers and firms. Furthermore, sovereign debt crises often require governments to cut back on government spending in order to restore markets’ faith in them to bring interest rates down. Both actions negatively affect the real economy.

(4)

3

the bank-sovereign nexus, and the link between finance and growth. Compared to the standard VAR approach that Sims (1980) pioneered, using a panel VAR prevents the researcher from running out of degrees of freedom, and the extra information provided by the larger sample helps in analyzing the dynamics at the aggregate level that a number of countries share by ensuring more efficient estimates (Pool et al., 2015, Love and Zicchino, 2006, and Abrigo and Love, 2016). Moreover, this approach enables the researcher to use fixed effects to account for unobserved cross-country heterogeneity. One potential drawback is that the approach used in this paper that these cross-country differences are absorbed by the intercept, thereby implicitly assuming that all countries are structurally equivalent.

The analysis performed in this paper finds that sovereign debt shocks have persistent and relatively strong negative effects on both the real economy and the financial system during normal times: both output and the credit supply are permanently lower after a shock to the interest rate that governments pay on their debt. This finding is robust to a number of different empirical setups, subsample choices and data frequencies. The results for the effect of credit supply shocks on output and vice versa are more ambiguous. The benchmark model shows that the effect is either slightly positive or slightly negative. The model with lending to non-financial business firms is not substantially different from the benchmark model. However, an increase in the amount of credit supplied to households negatively affects the real economy. Although different specifications and subsample groupings show relatively similar effects, the models for individual countries show considerable heterogeneity in how credit and output shocks affect one another. Nevertheless, this could also be due to the relatively short time series for individual countries, which reduces the efficiency of the estimates.

(5)

4 2. Literature review

Before performing the empirical analysis, it is necessary to clarify a number of concepts and mechanisms: what is sovereign debt and how can sovereign debt crises occur, why financial institutions hold sovereign debt, how sovereign debt crises can influence the financial system’s capacity to lend, and why credit matters for output. The literature presented in this section tends to be either theoretical in nature or based on microeconometric evidence, whereas relatively little evidence has been collected at the macroeconomic level. This study builds on both strands of research by analyzing the link between sovereign debt, credit supply, and output at the aggregate level, which is necessary to investigate the macroeconomic dynamics involved.

2.1 Sovereign debt

Governments can finance their expenditure in a number of ways, for example through taxation, monetary finance, and issuing debt in the form of some sort of financial instrument. Since high taxes tend to depress aggregate demand and monetary finance can lead to high inflation, issuing debt can be regarded as the most “painless” of the three policy options (Sargent, 1982).1 To make its debt attractive to investors, governments pay interest on these instruments. Virtually every government in the world issues their debt in the form of bonds, yet there is considerable heterogeneity among countries in the amount of debt that they can shoulder. For example: in 2012 the United States and Japan had government debts levels of respectively 94.3 percent and 196 percent of gross domestic product (GDP), whereas Greece at the start of its debt crisis had a central government debt of 127 percent of GDP (World Bank, 2015). Whereas Greece plummeted into crisis, Japan has been able to steer clear of any sovereign debt crises.

The most essential form of risk in finance is that the borrower is unable to pay the lender back. This phenomenon also occurs at the level of sovereigns: it does happen that a government runs into problems and cannot pay back its debts. A sovereign default occurs when a government is either unable to honor its debt or flat out refuses to do so. A sovereign debt crisis is a financial crisis that occurs when the interest rate that governments pay on their debt increases to such an extent that it becomes more difficult to finance their expenditure

1

(6)

5

(Obstfeld and Rogoff, 1996). Although defaults do not appear to happen regularly, virtually every country has defaulted on its debt at some point in time (Reinhart and Rogoff, 2009). Moreover, a debt crisis often leads to either outright default or a restructuring of that particular government’s debt, both of which tend to negatively affect economic activity (De Haan et al., 2015, and Calvo and Coricelli, 1993). The effect can also work the other way around: business cycles can increase the probability of sovereign default (Boonman et al., 2015). Governments that mainly rely on short-term bonds to finance expenditure are generally more vulnerable to these crises. After all, most of the government’s assets are both long term and illiquid. An example would be investments made in infrastructure. This creates a maturity mismatch of the type that also makes deposit-holding banks susceptible to runs.

In a sovereign debt crisis, interest rates on government bonds shoot up. This can be explained by basic asset pricing theory: the asset becomes riskier, since expectations are such that the government in question might not pay back its debts, so agents demand a higher return on that particular asset (Cochrane, 2005). Therefore, the interest rate on sovereign debt provides a good indicator of the probability of sovereign default, as the higher return that the asset pays reveals that agents consider it relatively riskier.

Naturally, these government bonds have to be held by someone. Banks have traditionally held a lot of government debt because it is a relatively safe asset compared to other financial instruments, such as stocks issued by firms or derivatives. Particularly for the most developed countries, bond interest rates tend to be low. This reflects the low risk that investors assign to bonds: the probability of most developed countries suffering macroeconomic mismanagement that scares off investors is substantially smaller, for example due to central bank independence that limit a government’s possibility to finance its deficits through monetary finance, which may lead to high inflation, which in turn will erode the value of the government’s debt (Obstfeld and Rogoff, 1996, and Vegh, 2013, present a number of theoretical models that capture this phenomenon).

(7)

6

bonds. This effect entails that investors tend to prefer domestic stocks over foreign stocks, even though holding an internationally diversified portfolio might be more beneficial, since this allows investors to spread their risk over several countries. This is why Obstfeld and Rogoff (2000) call the equity home bias effect one of the six major puzzles in international macroeconomic research. Tesar and Werner (1995) find that investors in OECD countries generally hold more domestic assets compared to foreign assets, and that the transaction costs that investors incur when they hold foreign assets does not explain the home bias effect. Banks that hold larger amounts of sovereign debt can hold smaller amounts of regulatory capital, since securities issued by governments are usually considered risk free (Barth et al., 2012). For example, one reason why Euro-area banks hold a lot of sovereign debt is that the Capital Requirements Directive (CRD) allows banks to assign a zero per cent risk weight to sovereign bonds that are denominated in euros (Popov and Van Horen, 2013, and Goves et al., 2016). Furthermore, weaker banks tend to hold larger amounts of domestic sovereign debt in order to strengthen their balance sheet (Bezemer and Gardiner, 2010). However, this last point does not pose an issue in the empirical analysis since I will be looking at aggregates instead of individual financial institutions.

In terms of domestic debt, Panizza (2008) argues that the composition of who holds sovereign debt can affect default through either reputational damage or direct sanctions. If the largest share of debt is held by domestic agents in case of a default, the effects of either of these channels might be less severe than when debt is held by foreign agents. In terms of reputational effects, the worst that can happen is that the ruling party loses in the next elections (Borensztein, 2006). This is a reason why governments would want domestic agents to hold their debt. However, this home bias in holding sovereign debt can make banks’ value and solvency depend on what markets believe constitutes the sustainable level of government debt (Brunnermeier et al., 2016). As such, when domestic banks hold a lot of their government’s debt, a default can lead to domestic banking crises (Kumhof and Tanner, 2005). Furthermore, the home bias effect amplifies what some authors dub either the “deadly embrace” or the “diabolical loop” (see De Grauwe, 2013, and Shambaugh, 2012), where banks fail because the sovereign debt they hold falls in value, which in turn requires governments to provide bail-outs. This further weakens the government’s fiscal position.

2.2 The link between sovereign debt and financial institutions

(8)

7

sovereign debt exposure remains to be established. There are a number of theoretical and empirical studies that looked into this relationship. For example, Kumhof and Tanner (2005) argue that government debt plays a key role in facilitating financial intermediation: in more risky environments, it acts as a form of collateral in the sense that it increases depositors’ willingness to have their funds intermediated. Moreover, they find that exposure to government debt in national banking systems can be as high as 50 percent. In such situations, debt devaluation can lead to widespread insolvency in the banking system.

Gennaioli et al. (2013) present a theoretical model that studies the links between sovereign default and financial fragility, i.e. a situation where weak public finances cause instability in the financial sector. In their model, banks hold government bonds as liquid assets to finance future investment. If the financial sector is highly developed and a government defaults on its debts, liquidity dries up as a result and negatively affects credit supply, investment, and output. They test their theoretical model using a panel of developed and developing countries, and find that defaults are typically followed by a drop in financial activity, which in turn is followed by a decrease in output.

In another study that demonstrates the effect of sovereign debt exposure on the financial sector, De Marco (2016) argues that, at the aggregate level, loan volumes decrease and interest rates on bank loans increase due to higher sovereign debt stress. Moreover, he finds evidence that the same mechanism holds at the micro-level using an instrumental variable regression: banks cut back on the amount of credit they supply due to losses on sovereign debt portfolios. De Marco (2016) identifies two channels through which losses on sovereign debt can affect credit supply. The first is the capital channel through which banks with large sovereign debt losses will also suffer equity losses. This, in turn, requires them to deleverage and thus supply less credit. The second is the funding channel: banks that suffer losses on sovereign debt will have a hard time finding funding on the wholesale market, where they use bonds as collateral (Trichet, 2010, and Alter and Schüler, 2012). Excluding bonds as suitable collateral will require the banks in question to pay higher rates on their loans. This is a result that is also found in other theoretical and empirical studies, for example in Gertler and Kiyotaki (2010) and Popov and Van Horen (2013).

(9)

one-8

third of a per cent. Moreover, a one percent increase in government bond yields, indicating that a government has to pay a higher interest rate on its debt, increases the median bank’s lending rate by approximately 0.19 per cent. Neri (2013) uses the seemingly unrelated regression approach and finds further evidence that sovereign debt tensions affect the rates that banks demand on loans in peripheral Eurozone countries (Greece, Italy, Spain and Portugal), but not in ‘core’ countries. In this sense, banks raising the rates on loans signal that they are cutting back on the amount of credit they supply.

Further studies indicate that the effect of troubled sovereign debt exposure on credit supply exists for a wider array of banks and forms of lending. Arteta and Hale (2008) estimate a panel fixed effects regression, and find evidence that sovereign debt exposure also affects the amount of credit supplied by international banks that are operating in that particular country. This shows a different aspect of the effect that sovereign debt issues can have on credit supply: foreign banks might lend less because they update their risk assessment of a particular country. Although the mechanism at play here is different, it is nevertheless an interesting example of how financial markets’ worries about the sustainability of a government’s debt can affect credit supply. Like Altavilla et al. (2015), Arteta and Hale’s paper differs from this paper’s setup by not examining potential business cycle affects and the macroeconomic dynamics involved. Popov and Van Horen (2013) find other instances of sovereign debt exposure negatively affecting lending. Using a fixed effects panel regression on loan-level data, they find evidence that syndicated loans, which are loans provided by a group of lenders but arranged for by a single lender, by European banks that had troubled bonds on their balance sheet fell during the Euro crisis. The reduction in lending to foreign markets was especially strong.

(10)

9

that is more in line with the deadly embrace phenomenon that De Grauwe (2013) analyzes. . Acharya et al. (2014)’s theoretical model features a sovereign that provides a bailout for its troubled financial sector, which in turn increases the sovereign’s credit risk. Since the financial sector holds the sovereign’s debt as assets, this erodes the value of its balance sheet, thereby weakening the financial sector once again. The model shows that the chain of events between increased sovereign risk, decreased credit, and a fall in output might run differently, which is important to consider when performing robustness tests.

A factor that could further affect the relationship between sovereign debt and credit supply is the so-called flight-to-quality (FTQ) effect. This phenomenon occurs when investors sell off risky assets in favor of safer investments. In open economies, this would mean that in the aggregate investors sell the bonds of distressed countries, and buy those of countries that they consider safe. This implies that banks that hold a lot of safe bonds will find it easier to lend out credit: the chance that the bonds in question would seriously drop in value is rather small. This raises the question if countries that have sovereigns with “safer” debt also have relatively higher lending activity than countries with riskier debt. However, this is not something I am particularly interested in: first and foremost this paper centers on whether the financial stability of governments affects the lending channel.

2.3 Credit and output

A large body of literature linking the effects of the development of financial markets and the process of economic growth has emerged in the last twenty or so years. King and Levine (1993) argue that financial development, which is associated with higher values of financial depth, the importance of depository institutions in allocating credit, and credit supplied to non-financial and private sector firms, is associated with higher levels of real GDP growth. A mechanism through which a well-functioning financial sector can help stimulate economic growth, is removing frictions that arise in financial intermediation (De Haan et al., 2015). By reducing ex-ante information costs about investment opportunities, facilitating risk diversification, and monitoring investments, a highly developed financial sector can more efficiently intermediate funds from borrowers to lenders. Thereby idle financial resources can be channeled to aid firms in increasing their productivity.

(11)

10

Moreover, market-based financial systems, i.e. a system where agents issue and trade securities to obtain financing, seem to lead to higher economic growth than bank-based financial systems.

Numerous authors have developed theoretical models that aim to formally describe the relation between financial markets and the real economy (see Brunnermeier et al., 2013, for an overview), and how the supply of bank credit affects output in the macroeconomy. Bernanke et al. (1999) construct a discrete time dynamic stochastic general equilibrium (DSGE) model with sticky wages and prices that also incorporates the financial sector in the form of price frictions (see Woodford, 2003, and Galí, 2015, for a thorough exposition of DSGE models). Their model incorporates what they call the financial accelerator effect. This effect is the result of a number of capital market restrictions. When embodied in a standard New Keynesian DSGE model with nominal rigidities, the inclusion of capital market frictions can both amplify and propagate shocks. A fall in asset prices decreases the net worth of entrepreneurial projects, which implies that is more difficult for entrepreneurs to borrow and therefore to find funding. This decreases the volume of loans, which negatively affects economic activity, i.e. output, which further decreases asset prices etc. etc. This is the financial accelerator effect, and thus shows how financial markets can affect output. Although Bernanke et al. (1999)’s model does not capture how credit supply affects output, it does show how shocks to the financial sector can lead to economic downturns.

Other evidence that shocks to the supply of credit matter for output can be found in Pool et al. (2015). They develop and subsequently estimate a theoretical model of the macroeconomy that incorporates bank lending and credit risk to find out if an increase in credit risk increases the lending rate, which thereby decreases lending and negatively affects the real economy. The theoretical model is estimated using as a panel VAR, finding that an increase in credit risk decreases lending, which in turn negatively affects economic activity. Although the mechanism described in Pool et al. (2015)’s theoretical model is different from the one that this paper deals with, the conclusion that decreased credit negatively affects the real economy is a conclusion that also matters for this study.

(12)

11

investor takes a loan from a bank, both the loan and deposit are in that investor’s name. If he uses that loan to finance a capital investment from someone else, that loan becomes the other agent’s deposit. Because banks create purchasing power ex nihilo, they make investment possible. As such, they are a driving force behind the process of economic growth (Schumpeter, 1934, and Bertocco, 2009). When incorporating the FMC view of banking in their model, they find that changes in bank balance sheets are larger, happen faster, and have greater effects on real economic activity compared to the situation where banks are intermediaries between savers and lenders. Whether the Bernanke et al. (1999) or the Jakab and Kumhof (2015) model is more realistic is an interesting avenue for future research, but what matters for this paper is that both papers formally show how credit supply can affect output, both positively and negatively, and how lending is also driven by shocks to other variables.2

Finally, there has also been research on how different kinds of credit affect output. Bezemer et al. (2016a) find that credit extended to non-financial business firms has a positive effect on the real economy. However, the opposite is true for credit that is used in asset markets and real estate investment. In the first case, credit supports the productive side of the economy: it allows firms to invest in production factors and to become more productive. In the second case, credit does not lead to productivity gains, and thus does not stimulate economic growth. As such, the credit-to-GDP ratio grows, which Bezemer et al. (2016a) argue leads to higher volatility, financial fragility, and a higher likelihood of financial crisis. Thus, it might not be true that all kinds of credit affect output equally.

3. Theoretical model

The theory presented so far suggests the following relationship at the aggregate level: increases in the interest rate on sovereign debt indicate that financial markets worry that the sovereign will honor its debts well. Subsequently, banks curtail lending activity when the value of the debt that they hold falls, either in the form of raising the interest rate they demand or by restricting the credit supply. Moreover, the decrease in lending in the form of credit negatively affects output. To capture this negative effect on output, I base the proposed relationship on the baseline New Keynesian model as presented in Galí (2015):

2

(13)

12

𝑦_𝑡 = −1_𝜎(𝑖_𝑡− 𝐸_𝑡{𝜋_𝑡+1} − 𝑟_𝑡𝑛_{) + 𝐸}

𝑡{𝑦𝑡+1} (1)

𝜋𝑡 = 𝛽𝐸𝑡{𝜋𝑡+1} + 𝜅𝑦𝑡 (2)

where (1) is the New Keynesian IS-curve (NKIS), and (2) is the New Keynesian Philips Curve (NKPC). The NKIS is basically an Euler equation for consumption: it describes how output in an economy intertemporally adjusts to shocks in variables.3 It relates the output gap (𝑦_𝑡) to the nominal interest rate (𝑖_𝑡), expected inflation in the next period (𝐸_𝑡{𝜋𝑡+1}), the

natural rate of interest (𝑟_𝑡𝑛_{, which would be the interest rate in an economy where capital}

goods would be exchanged directly, i.e. without any sources of nominal frictions), and expected output in the next period (𝐸_𝑡{𝑦_𝑡+1}). The terms between brackets are weighted by the intertemporal substitution elasticity (𝜎). The NKPC relates current inflation to expected inflation in the next period (𝐸_𝑡{𝜋_𝑡+1}) and the output gap. Combined with the literature presented in the previous section, the standard New Keynesian model provides a theoretical underpinning for two variables that are of use in the empirical analysis: the output gap as a proxy for business cycle, and the inflation rate. Moreover, together with some of the theoretical models discussed in the literature review, such as Bernanke et al. (1999), it motivates the recursive ordering that is used in the empirical model.

The output gap is an indicator for the business cycle that measures the difference between actual output and the so-called natural rate of output, i.e. how much an economy would produce if it were operating at full capacity (Jahan and Mahmud, 2013). Inflation matters as well: higher expected inflation in the next period will cause consumption to be shifted from the next period to the current period. Therefore, I will use the output gap as an indicator of fluctuations in real economic activity in my empirical analysis. Since the NKIS shows that inflation affects the output gap, I will also use this variable in the empirical model. The relationship between inflation and credit is more ambiguous from an economic standpoint. However, a number of theoretical and empirical studies find that higher inflation negatively affects the working of financial markets and bank lending (see for example Boyd et al., 2001, and Huybens and Smith, 1999). Furthermore, the relationship between inflation and sovereign debt is also not clear: even though higher public debt often goes hand-in-hand with higher inflation, there is no clear rationale to argue how inflation changes due to a shock to sovereign debt interest rates (Martin, 2015).

3_{Heijdra (2009) intuitively explains the mechanism behind the consumption Euler equation on pp. 116 using the}

(14)

13

The theory presented in section 2 leads us to the following hypotheses about the dynamics of sovereign debt, inflation, credit supply, and the business cycle. An increase in the interest rate that governments pay on their debt negatively affects the financial position of banks (De Marco, 2016, and Trichet, 2010). A shock to that interest rate implies that the asset is considered riskier by investors. As such, the asset loses value. Exposure to a distressed sovereign’s debt can negatively affect a bank’s capabilities to operate through losses on equity and by making it more difficult to find funding (Alter and Schüler, 2012). This in turn will lead to the bank cutting back on lending activities by providing less credit to agents. Since credit matters for the real economy, a decrease in the amount that banks supply should lead to business cycle fluctuations in the form of a negative shock to output (Bernanke et al., 1999). Finally, higher inflation has an ambiguous effect on lending activity, but does lead to a negative output gap (Galí, 2015). This implies that output is lower than its potential level, which in terms of data means that the GDP growth rate is lower than its growth trend.

To prepare for the VAR analysis to come, I present a mathematical formulation of the proposed relationship. This paper refrains from constructing a micro-founded dynamic stochastic general equilibrium model based on optimizing agents, but rather uses the empirical and theoretical literature presented in the previous section to guide the relationships in the model. As such, in individual equations the model becomes:

𝑏𝑡= 𝜑(𝐿)𝑏𝑡+ 𝜑(𝐿)𝑐𝑡+ 𝜑(𝐿)𝜋𝑡+ 𝜑(𝐿)𝑦𝑡+ 𝑒𝑡𝑠 (3)

𝜋𝑡 = 𝜑(𝐿)𝜋𝑡+ 𝜑(𝐿)𝑏𝑡+ 𝜑(𝐿)𝑐𝑡+ 𝜑(𝐿)𝑦𝑡+ 𝑒𝑡𝑖 (4)

𝑐_𝑡 = 𝜑(𝐿)𝑐_𝑡+ 𝜑(𝐿)𝑏_𝑡+ 𝜑(𝐿)𝜋_𝑡+ 𝜑(𝐿)𝑦_𝑡+ 𝑒_𝑡𝑐 ₍₅₎

𝑦_𝑡 = 𝜑(𝐿)𝑦_𝑡+ 𝜑(𝐿)𝑏_𝑡+ 𝜑(𝐿)𝑐_𝑡+ 𝜑(𝐿)𝜋_𝑡+ 𝑒_𝑡𝑏𝑐 ₍₆₎

with 𝑏_𝑡 defined as the interest rate on sovereign debt, 𝑐_𝑡 is the credit supply, 𝜋_𝑡 is the rate of inflation, 𝑦_𝑡 is the output gap, 𝜑(𝐿) is the lag polynomial, and 𝑒_𝑡 are the structural shocks hitting the economy. Starting with equation (3), these shocks are respectively sovereign debt, inflation, credit supply, and business cycle shocks. To prepare for the VAR analysis that is to come, I condense this system of equations as a structural VAR system in matrix form:

𝜞𝒀_𝒕 = 𝑩(𝑳)𝒀_𝒕+ 𝒆_𝒕 (7)

where 𝒀_𝒕 = (𝑏𝑡 𝜋𝑡 𝑐𝑡 𝑦𝑡)′ denotes a vector of endogenous variables, 𝜞 is a matrix of

(15)

variance-14

covariance matrix of the structural shocks hitting the economy. Premultiplying the structural VAR by 𝜞−𝟏_{gives us the reduced form:}

𝒀_𝒕 = 𝑨(𝑳)𝒀_𝒕+ 𝒖_𝒕 (8)

where 𝑨(𝑳)𝒀_𝒕 ≡ 𝜞−𝟏𝑩(𝑳)𝒀_𝒕, and 𝒖_𝒕 ≡ 𝜞−𝟏𝒆_𝒕. Furthermore, a number of identifying restrictions are needed to bring the model to the data and disentangle the effects of the shocks on the various variables (Lütkepohl, 2006). First, I assume that the shocks hitting the economy are orthogonal, i.e. the variance-covariance matrix’s off-diagonal elements are set to zero. Technically, this means that the covariance between different shocks is set to zero. This amounts to assuming that the shocks are unrelated and do not have a common cause: they are what Bernanke (1986) calls “’primitive’ exogenous forces.” Second, it is necessary to impose restrictions on the contemporaneity of shocks, i.e. to impose order on when shocks to one variable affect another variable. I assume that the system follows a recursive ordering, i.e. a Cholesky decomposition, where observed shocks in 𝑦₁come from 𝑦₁, observed shocks to 𝑦₂come from 𝑦₁and 𝑦₂ etc.

4. Methodology

In order to analyze the dynamic relationships between the variables of interest, I will use panel VAR methods in this study. According to Stock and Watson (2001), VARs serve four main purposes: data description, forecasting, structural inference, and policy analysis. Moreover, they provide a relatively straightforward tool to examine economic dynamics present in more than one time series. Panel VARs differ from the regular VAR methodology pioneered by Sims (1980s) due to the individual heterogeneity between cross-sectional units. Holtz-Eakin et al. (1988) were the first to use the panel VAR methodology. A useful overview of the literature on panel VARs is Canova (2013).

(16)

15

problem that VARs potentially suffer from. This entails that the researcher might run out of degrees of freedom due to the relatively short time series used in estimating a model. Using a panel VAR circumvents this problem. Moreover, the ability to use fixed effects also makes it possible to account for any potentially unobserved heterogeneity.

Panel VAR models have a number of advantages compared to regular VARs. First, the fact that the researcher is able to use more than one cross-sectional unit increases the number of degrees of freedom and leads to more efficient estimates (Abrigo and Love, 2016). As such, this method enables the researcher to circumvent the curse of dimensionality that is commonly associated with VARs and end up with more trustworthy estimates. Second, I will be able to investigate the dynamic effects that are common to all countries, which makes it possible to give more general statements about the effects that certain economic variables might have on other variables. On the other hand, this method has a drawback: when estimating a panel VAR, the researcher assumes that the cross-sectional units are not structurally different from each other. In this setup, this would mean that only the intercept is allowed to differ per country. Gambacorta et al. (2014) circumvent this problem by using a mean group estimator in the spirit of Pesaran and Smith (1995). This allows for cross-country heterogeneity. Unfortunately, this is beyond the scope of this study. Even though this is a rather strong assumption, it is quite justifiable if the cross-sectional units are similar to each other. Since all countries are advanced economies that are relatively similar, this assumption will not pose a problem.

Formally, the model takes the following form:

𝒀_𝒊,𝒕 = 𝜶_𝒊+ 𝑨(𝑳)𝒀_𝒊,𝒕+ 𝒖_𝒊,𝒕 (9) Note that the only real differences with the normal VAR model presented in the previous section are the 𝑖 subscript, which indicates cross-sectional dimension and thus the number of countries in the sample, and the vector of country fixed effects 𝜶_𝒊. To estimate the fixed effects, I use the Helmert transformation, i.e. forward orthogonal deviation, which subtracts the mean of all future observations in the sample.

(17)

16

The optimal amount of lags is on Andrews and Lu’s (2001) Moment and Model Selection Criteria (MMSC) and Hansen’s (1982) J-statistic of over-identifying restrictions. The MMSC are variations on the selection criteria that are more commonly used in time series analysis and maximum likelihood-based, i.e. the Akaike information criteria (AIC), the Bayesian information criteria (BIC), and the Hannan-Quin information criteria (HQIC). Next, I calculate orthogonalized impulse response functions, where I impose a recursive ordering on the VAR as identifying restrictions in order to trace the dynamic effects that shocks to one variable have on other variables. Finally, I calculate the forecast-error variance decomposition in order to estimate how much variation in one variable is caused by shocks to the other variables.

In order to make the panel VAR system suitable for further analysis, I impose the following recursive structure on the system. First, the interest rate sovereigns pay on their debts, 𝑏_𝑡, is only contemporaneously affected by its own shocks. Second, inflation, 𝜋_𝑡, is assumed to be affected by shocks to 𝑏_𝑡 and inflation. Third, credit supply, 𝑐_𝑡, is assumed to be affected by shocks to 𝑏𝑡, 𝜋𝑡, and its own shocks, as worries about the sustainability of

government debt or an impending sovereign debt crisis, represented by shocks to 𝑏_𝑡, are assumed to negatively affect the balance sheet of financial institutions, leading to them cutting back on lending. Finally, the output gap, 𝑦_𝑡, is contemporaneously affected by all other variables. Since this relationship is based on the evidence presented in earlier theoretical and empirical research instead of on a formal theoretical model, I will test the robustness of these results by using different recursive orderings.

5. Data

(18)

17

the variables and their sources.

The panel has quarterly data, starting in 1990Q1 and ending in 2007Q4. I use 2007Q4 as a cut-off point for two reasons. First, the start of the Global Financial Crisis (GFC) around that time might cause structural breaks in the data, which in turn could affect the results. The panel VAR package that used in this study does not offer any way to solve these issues in its current version. Second, the Eurozone crisis also started around this time. Although this was a sovereign debt crisis, sovereigns only got in trouble after they had to bail out banks (De Grauwe, 2013), who in turn ran into problems due to a number of reasons, for example property bubbles in Ireland and Spain. As such, this is a different mechanism than this particular study is interested in. Even though the period after 2007Q4 show a number of developments that are of interest for this study, these two issues could affect the results. Third, this paper deals with the effects of sovereign debt shocks on the economy in normal times, not during times of economic breakdown. Including the years of the GFC might therefore severely affect any results.

The yield on government bonds with a ten year maturity in secondary markets in percentages is used as an indicator for sovereign debt sustainability (𝑏_𝑖,𝑡). These data are taken from the IMF’s International Financial Statistics. Long term bond yields are a good indicator for how sustainable institutions consider a sovereign’s debt and how markets perceive a sovereign’s default risk (De Grauwe, 2014). If markets consider one government’s debt riskier, they will demand a premium, and thus the yield on that debt will be higher. Figure 1 presents plots of the time series in groups of countries. These show that bond yields at the end of the sample period were lower than at the beginning of the sample period and fluctuate around the five percent mark. An explanation for this would be that the sample period consists of a relatively tranquil period in terms of economic fluctuations, with very little crises that had large effects. However, as can be seen from the graph, interest rates on Swedish bonds shot up during the Swedish financial crisis in the early 1990s.

As an indicator for price changes I use inflation data from the IMF’s International Financial Statistics database (𝜋𝑖,𝑡). These data measure annualized quarterly changes in the

(19)

18

Figure 1: Ten year government bond yields in percentages per quarter

Figure 2: Inflation rate in percentages per quarter

0.00 5.00 10.00 15.00 1990q1 1994q3 1999q1 2003q3 2008q1 Australia Austria Belgium Canada 0.00 5.00 10.00 15.00 1990q1 1994q3 1999q1 2003q3 2008q1 Denmark France Germany Japan 0.00 5.00 10.00 15.00 1990q1 1994q3 1999q1 2003q3 2008q1

The Netherlands New Zealand

Norway Sweden 0.00 5.00 10.00 15.00 1990q1 1994q3 1999q1 2003q3 2008q1

United Kingdom United States

-2 0 2 4 6 8 10 12 1990q1 1994q3 1999q1 2003q3 2008q1 Australia Austria Belgium Canada -2 0 2 4 6 8 10 12 1990q1 1994q3 1999q1 2003q3 2008q1 Denmark France Germany Japan -2 0 2 4 6 8 10 12 1990q1 1994q3 1999q1 2003q3 2008q1

Norway Sweden -2 0 2 4 6 8 10 12 1990q1 1994q3 1999q1 2003q3 2008q1

(20)

19

To capture aggregate fluctuations that occur during the business cycle I use the output gap (𝑦_𝑖,𝑡). Plots of the series are shown in Figure 3. Theoretically, the output gap captures how much actual output differs from potential output, which is what an economy would produce in its full employment, flexible price equilibrium. If the output gap is smaller than zero the economy is performing below trend, and thus in some sort of slump. If the output gap is larger than zero, than the economy is producing above capacity, i.e. it is booming. This could lead to inflationary pressures. Technically, the output gap is how much GDP in a particular quarter differs from that particular country’s growth trend. As such, it provides a suitable indicator for business cycle activity. The data are taken from OECD Stat.4 Since the output gap data are only reported annually, I use quadratic match average conversion methods to interpolate quarterly values.5

Figure 3: Output gap in percentages as deviation from trend per quarter

For the credit variable, I use data from the Bank of International Settlements’ (BIS) dataset on credit to the private non-financial sector. Since different kinds of credit might affect the economy in different ways (Bezemer, 2014), three credit variables are used to

4_{Readers interested in technical details of how this particular output gap indicator is calculated are referred to}

OECD (2014).

5

This can be done quite easily in EViews 9. Again, readers interested in technical details are referred to the EViews 9 user manual.

-5 0 5 1990q1 1994q3 1999q1 2003q3 2008q1 Australia Austria Belgium Canada -5 0 5 1990q1 1994q3 1999q1 2003q3 2008q1 Denmark France Germany Japan -5 0 5 1990q1 1994q3 1999q1 2003q3 2008q1

Norway Sweden

-5 0 5

1990q1 1994q3 1999q1 2003q3 2008q1

(21)

20

investigate the effects of different kinds of credit supply shocks: credit to households (𝑐_𝑖,𝑡𝐻_),

credit to non-financial business firms (𝑐_𝑖,𝑡𝐹_{), and total lending to the private sector (𝑐}

𝑖,𝑡𝑇). The

data are in billions of US dollars, and includes lending by both banks and by non-bank financial institutions. Of these three variables only the total credit variable has data for the entire sample period. In order to create quarterly growth rates, I first took the natural logarithm, then took first differences, and finally multiplied all data points by 100 percent. These total credit, household credit, and firm credit data are presented respectively in Figures 4 to 6. The plots show that the only country that had a steadily growing total supply of credit was the United States, whereas others show much larger fluctuations.

Figure 4: Growth rate of total credit per quarter

-30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 Australia Austria Belgium Canada -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 Denmark France Germany Japan -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1

Norway Sweden -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1

(22)

21

Figure 5: Growth rate of credit to households per quarter

Figure 6: Growth rate of credit to non-financial business firms per quarter

-30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy Australia Austria Belgium Canada -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy Denmark France Germany Japan -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy

Norway Sweden -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy

United Kingdom United States

-30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy Australia Austria Belgium Canada -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy Denmark France Germany Japan -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy

Norway Sweden -30 -20 -10 0 10 20 1990q1 1994q3 1999q1 2003q3 2008q1 qy

(23)

22

Finally, all variables were checked for stationarity to ensure stable VAR estimates and interpretable impulse response functions. I performed panel unit roots tests for each variable. Since the panel is unbalanced, the standard Levin-Lin-Chu (2002) test will not work. Hence, use of the Im-Pesaran-Shin (2003) panel unit root test is required. All variables are demeaned, since they will also be demeaned in the VAR, and the ADF regressions use the optimal amount of lags according to the AIC. The results show that all variables are stationary.

Table 1: Summary statistics

Variables Transformation Obs. Mean Standard deviation Min. Max. 𝑏_𝑖,𝑡 - 1,008 6.014 2.251 0. 657 13.563 𝑦𝑖,𝑡 - 1,004 0.054 1.937 -6.236 5.379 𝜋_𝑖,𝑡 - 1,008 2.120 1.594 -1.383 11.522 𝑐_𝑖,𝑡𝑇 _{∆ln (𝑐} 𝑖,𝑡𝑇 ) ∗ 100 994 1.881 5.041 -27.806 17.708 𝑐_𝑖,𝑡𝐻 _{∆ln (𝑐} 𝑖,𝑡𝐻 ) ∗ 100 946 2.042 5.199 -29.110 20.099 𝑐_𝑖,𝑡𝐹 _{∆ln (𝑐} 𝑖,𝑡𝐹 ) ∗ 100 916 1.743 4.937 -27.083 16.443

Note: 𝑏𝑖,𝑡 and 𝜋𝑖,𝑡 are respectively the interest rate on government bonds and the annual

inflation growth rate, 𝑦_𝑖,𝑡 is deviation from output growth trend in percentages, 𝑐_𝑖,𝑡𝑇 is the quarterly growth rate of total credit supplied by banks and non-bank financial institutions to the private non-financial sector, and 𝑐_𝑖,𝑡𝐻 and 𝑐_𝑖,𝑡𝐹 are respectively quarterly growth rates of credit to households and non-financial business firms.

Table 2: Im, Pesaran and Shin (2003) unit root tests

Variables W t-bar statistic Average amount of lags p-Value 𝑏_𝑖,𝑡 -4.233 2.50 0.00 𝑦_𝑖,𝑡 -7.006 3.50 0.00 𝜋_𝑖,𝑡 -6.624 3.57 0.00 𝑐_𝑖,𝑡𝑇 _-21.334 _1.14 _0.00 𝑐_𝑖,𝑡𝐻 _-6.071 _1.14 _0.00 𝑐_𝑖,𝑡𝐹 _-21.4045 _0.86 _0.00

Note: 𝐻₀: all panels contain unit roots. 𝐻_𝑎: some panels are stationary. All variables are

demeaned for the test. The ADF regressions use a maximum of 6 lags according to the AIC.

6. Empirical analysis and discussion

(24)

23

total credit variable 𝑐_𝑖,𝑡𝑇, then I will discuss the model with the firm and household credit variables. All impulse response functions were computed with 95 percent confidence intervals after 1,000 Monte Carlo iterations using Gaussian approximation methods, and show the results for the next ten quarters. The number of lags used in estimating each model is based on the overall coefficient of determination, a minimized J-statistic, and the MMSC, which include the AIC, BIC, and HQIC Granger causality tests for the models are presented in Appendix E. These show that most variables Granger cause each other, with the exceptions being the total credit variable 𝑐_𝑖,𝑡𝑇, which only Granger causes the bond interest rate 𝑏𝑖,𝑡 on its

own, the firm credit variable 𝑐_𝑖,𝑡𝐹_{, which does not Granger cause any other variable on its own,}

and the household credit variable 𝑐_𝑖,𝑡𝐻, which only does not Granger cause the inflation rate 𝜋_𝑖,𝑡 on its own. However, when combined with other variables these issues disappear, indicating that, at the least, each variable is Granger caused by a combination of the other variables. Furthermore, all models were tested for stability. The graphical results of these tests, Figures F1 to F3 in Appendix F, show that all models are stable. Thus, the models are suitable for statistical inference. The estimated coefficients themselves go unreported, as is standard in the VAR literature. However, these are available on request.

6.1 Benchmark model

The orthogonalized impulse response functions for the benchmark model are presented in Figure 7. The model was estimated using two lags. I define the shocks the same way as in the theoretical model in section 3. In the graphs, the rightmost variable is the one that is first in the recursive ordering variable, and the leftmost variable is the last one. In the headers above the graph, the first term (e.g. business cycle) is the shock, whereas the term on the right (e.g. bond int. rate) is the variable being shocked.

(25)

24

interest rates permanently as well.

The rate of inflation increases permanently after a business cycle shock. A one-off innovation in the credit variable does not significantly affect inflation. An unexpected inflationary shock leads to a permanently higher rate of inflation after ten quarters. Finally, a sovereign debt shock permanently increases the inflation rate after ten quarters by about 0.1 percent.

A positive business cycle shock has a slight negative effect on credit supply after a few quarters, yet the effect goes to zero in the end. A positive credit supply shock leads to a higher amount of credit at first, yet the effect goes to zero almost immediately afterwards. An inflation shock positively affects lending slightly after a number of periods, but the effect goes to zero after ten quarters. Finally, a sovereign debt shock increases credit supply at impact, but leads to a decrease in the supply of credit shortly thereafter. This negative effect appears to be permanent.

Lastly, a positive business cycle shock temporarily increases the output gap. A shock to the credit supply makes the output gap increase permanently. However, the effect is rather small and the confidence bands are wide. As such, it is difficult to say with certainty what the effect is. An inflationary shock permanently decreases the output gap. This is in line with expectations, given the predictions of the New Keynesian IS curve as presented in equation (1): an increase in the inflation rate leads to agents intertemporally substituting consumption tomorrow for consumption today, since their money is worth less tomorrow. Finally, the output gap decreases permanently after a sovereign debt shock, leading to an economic downturn.

To save space, the forecast error variance decomposition (FEVD) results are presented in Table G1 in Appendix G. These show the FEVDs after a maximum of 20 quarters.6 Table G1 shows that after 20 quarters, approximately 87.4 percent of the unexplained variation in bond yields can be explained by sovereign debt shocks. Business cycle shocks can explain about 9.1 percent of the variation, and the rest is explained by credit supply shocks (0.4 percent) and inflation shocks (approximately 3.1 percent). Sovereign debt and business cycle shocks can explain respectively 14.5 and 21.9 percent of the variation in the inflation rate after 20 quarters. Most of the variation is explained by inflation itself (approximately 63.6 percent), with credit supply shock explaining very little variation (0.008 percent). Of the

6_{I experimented with longer time horizons for the forecast error variance decompositions to find how the}

(26)

25

variables affecting credit supply, the only shock that has a sizable effect apart from credit itself is a sovereign debt shock, which explains approximately 5.8 percent of the unexplained variation. Finally, the results show that sovereign debt shocks explain a very large part of the variation in the output gap over a long time horizon: after four quarters sovereign debt shocks explain only 1.4 percent of the variation in the output gap. However, after 20 quarters, this percentage increases to 55.6. Remarkably enough, sovereign debt shocks explain more of the unexplained variation in the output gap than business cycle shocks, which explain approximately 40.1 percent of the variation after 20 periods. Inflation and credit supply shocks explain respectively approximately 3.8 and 0.4 of the unexplained variation of the output gap

Figure 7: Impulse response function for benchmark panel VAR (impulse : response)

6.2 Household credit model

The orthogonalized impulse response functions for the model with the household credit variable are presented in Figure 8. The model was estimated with three lags. In general, the results found by the benchmark panel VAR model remain intact. However, there are two main differences. First, a business cycle shock has a clearer significant negative effect on household credit, i.e. the confidence intervals are smaller. Second, a household credit supply shock

0 .2 .4 .6 -.05 0 .05 .1 -.15 -.1 -.05 0 .05 -.4 -.2 0 -.4 -.2 0 .2 .4 -2 0 2 4 6 -.2 0 .2 .4 .6 -.5 0 .5 1 0 .1 .2 .3 -.05 0 .05 0 .2 .4 .6 0 .1 .2 0 .1 .2 .3 -.1 -.05 0 -.05 0 .05 .1 .15 .2 .3 .4 .5 0 5 10 0 5 10 0 5 10 0 5 10

Business cycle : Output gap

Credit supply : Output gap

Inflation : Output gap

Sovereign debt : Output gap

Business cycle : Credit supply

Credit supply : Credit supply

Inflation : Credit supply

Sovereign debt : Credit supply

Business cycle : Inflation

Credit supply : Inflation

Inflation : Inflation

Sovereign debt : Inflation

Business cycle : Bond int. rate

Credit supply : Bond int. rate

Inflation : Bond int. rate

Sovereign debt : Bond int. rate

(27)

26

causes the output gap to slightly increase at impact. However, the effect becomes negative shortly thereafter, and is still negative after ten quarters.

The FEVD results are presented in Table G2 in Appendix G. These show that after 20 periods, approximately 83.9 percent of the unexplained variation in bond yields can be explained by sovereign debt shocks. Business cycle shocks can explain 8.7 percent of the variation, and smaller parts are explained by household credit shocks (1.1 percent) and inflation shocks (6.7 percent). Sovereign debt and business cycle shocks can explain respectively 18.1 and 15.2 percent of the variation in the inflation rate, whereas household credit shocks account for 1.1 percent. Of the variables affecting credit supplied to households, sovereign debt shocks explain approximately 5 percent of the unexplained variation, and inflation and business cycle shocks respectively explain around 2.1 and 1.5 percent. Finally, household credit and inflation shocks account for respectively 3.5 and 2.5 percent of the variation in the output gap, whereas sovereign debt shocks cause 51 percent of the unexplained variation. As such, the FEVD results for the household credit model are broadly in line with those of the benchmark model. The largest difference is that household credit shocks tend to explain larger parts of the variation in variables than total credit supply shocks, although the magnitude of these effects is still rather small.

Figure 8: Impulse response function for household credit panel VAR (impulse : response)

0 .2 .4 -.15 -.1 -.05 0 .05 -.1 -.05 0 .05 .1 -.3 -.2 -.1 0 -.5 0 .5 0 2 4 6 -.5 0 .5 1 -1 -.5 0 .5 1 -.1 0 .1 .2 .3 -.1 -.05 0 .05 0 .2 .4 .6 0 .1 .2 .3 0 .05 .1 .15 .2 -.1 -.05 0 .05 -.1 0 .1 .2 .1 .2 .3 .4 .5 0 5 10 0 5 10 0 5 10 0 5 10

Household cred. : Output gap

Business cycle : Household cred.

Household cred. : Household cred.

Inflation : Household cred.

Sovereign debt : Household cred.

Household cred : Inflation

Household cred : Bond int. rate

(28)

27 6.3 Firm credit model

The orthogonalized impulse response functions for the model with the household credit variable are presented in Figure 9. The model was estimated with two lags. The results show that the firm credit model is virtually indistinguishable from the benchmark model. As such, the main results of sovereign debt shocks negatively affecting both credit supply and the output gap remains intact. Like in the benchmark model, a credit supply shock slightly increases the output gap on impact, but thereafter the confidence intervals are too wide to tell what is going on.

The results of the FEVDs are presented in Table G3 in Appendix G. The table shows that after 20 quarters, approximately 93.3 percent of the unexplained variation in bond yields can be explained by sovereign debt shocks, whereas business cycle shocks explain about 4.5 percent of the variation. The inflation and firm credit shocks explain respectively 1.9 and 0.5 percent of the variation. In explaining the variation in the inflation rate, sovereign debt and business cycle shocks explain most of the variation apart from inflation shocks, with sovereign debt shocks causing 15.2 percent and business cycle shocks causing 20 percent of the variation. Like in the benchmark model, sovereign debt shocks are the only kind of shocks that substantially affect the credit supply: these explain 5.9 percent of the variation after 20 quarters. Finally, just as in the benchmark model, sovereign debt shocks explain a sizeable part of the variation in the output gap after 20 quarters: these explain around 57.2 percent of the variation. Business cycle shocks only account for 39 percent. Again, quite similarly to the results found in the benchmark model, inflation and credit supply shocks explain respectively approximately 3.3 and 0.5 of the unexplained variation of the output gap. Essentially, the impulse response functions and FEVDs show that there is very little difference between the benchmark model and the model with firm credit supply as a variable.7

6.4 Discussion

Even though this paper is the first to use panel VAR methods to analyze sovereign debt issues, the results found so far are generally in line with previous research. The impulse response functions show that bond interest rates increase after both a business cycle and inflation shock. The inflation shock raises the premium that investors demand on holding ten year bonds, and therefore drives up the yield they demand (De Haan et al., 2015). This is in

7_{To make sure the results were not driven by any potential outliers, the models were also estimated without the}

(29)

28

line with earlier findings that inflation is an important driver of sovereign bond yields (Constantini et al., 2014). The effect of the business cycle shock might seem counterintuitive at first, since one would expect that an increase in the rate of economic growth would make investors worry less about the sustainability of government debt. However, previous empirical research has found that bond yields are determined differently per country: Jaramillo and Zhang (2013) find that expected real GDP growth decreases bond interest rates in emerging economies, but increases them in advanced economies.

Figure 9: Impulse response function for firm credit panel VAR (impulse : response)

The inflation rate shows a strong increase after a business cycle shock. This is in line with the theoretical literature (e.g. New Keynesian macroeconomic models) and empirical evidence (Bolt and Van Els, 2000) that economic booms tend to increase the inflation rate. Furthermore, the inflation rate also increases after a sovereign debt shock. This is in line with earlier literature. Since inflation expectations are an important determinant of bond interest rates, the reverse being true was to be expected (Christensen et al., 2010, and Barr and Campbell, 1997.

The response of credit supply to inflation shock is similar in all three models: all show a positive effect. The relationship between credit and inflation is an ambiguous one, which is also something that is found in earlier research, e.g. Martin (2015) and Liviatan (1985).

0 .1 .2 .3 .4 -.05 0 .05 .1 -.15 -.1 -.05 0 .05 -.4 -.2 0 -1 -.5 0 .5 -2 0 2 4 6 -.5 0 .5 -.5 0 .5 1 0 .1 .2 .3 -.05 0 .05 0 .2 .4 .6 0 .1 .2 -.1 0 .1 .2 -.1 -.05 0 .05 -.05 0 .05 .1 .2 .3 .4 .5 0 5 10 0 5 10 0 5 10 0 5 10

Firm credit : Output gap

Business cycle : Firm credit

Firm credit : Firm credit

Inflation : Firm credit

Sovereign debt : Firm credit

Firm credit : Inflation

Firm credit : Bond int. rate

(30)

29

Moreover, the positive effect found here is rather short-lived. None of the impulse response functions show a strong reaction of credit supply to a business cycle shock. This is in line with empirical evidence that demand factors do not strongly affect financial institutions’ lending decisions: banks and financial markets continue to extend credit despite a small upturn or downturn in the real economy (Everaert et al., 2015). A sovereign debt shock increases credit supply at impact, but leads to a decrease in the supply of credit shortly thereafter. This negative effect is highly significant, and is highly persistent: after ten years the effect has still not died out. This is in line with earlier empirical evidence on the effect of sovereign debt shocks on lending by financial institutions, e.g. Albertazzi et al. (2014), Altavilla et al. (2015), De Marco (2016), Gennaioli et al. (2014), and Guerrieri et al. (2012).

Finally, the response of the output gap to shocks is broadly the same in all three models. The effect of an inflation shock on the output gap is negative on impact, but the confidence intervals are relatively wide for most of the response period: it only becomes significant again after approximately nine quarters. After that time, the effect is negative again. This is in line with the results found in New Keynesian models a la Galí (2015): a higher inflation rate incentivizes consumers to intertemporally substitute consumption tomorrow for consumption today, leading to a loss in future output. Sovereign debt shocks have a very strong and persistent negative effect on the output gap in all three models. This is line with earlier effects on sovereign debt issues on economic growth (e.g. De Haan et al., 2015, and Calvo and Coricelli, 1993). However, no earlier research has been done on the effects of bond interest rate shocks themselves on growth, setting the models presented in this section apart from other papers on sovereign debt and growth dynamics.

(31)

30 7. Robustness tests

To ensure that the results found so far are robust to different specifications of the model I perform a number of robustness checks using a different indicator for the output gap, changing the frequency of the variables, experimenting with different recursive orderings, and estimating the model for both subgroups of the sample and individual countries.8

7.1 Hodrick-Prescott output gap

In order to be certain that the benchmark panel VAR’s results are not affected by the use of the OECD output gap data, the model is estimated once again after applying the Hodrick-Prescott (HP) filter to quarterly GDP data, thereby calculating an alternative output gap.

The HP filter separates the trend and cyclical parts of a time series (Hodrick and Prescott, 1997, and Ravn and Uhlig, 2002). When applied to GDP data, the cyclical parts represent the output gap. First, I take the natural logarithm of real GDP, and then I use a filter parameter value of 1600 for quarterly data, as is standard in the literature. Appendix D shows series of the OECD and HP output gaps per country. The plots show that, although the series do tend to move in the same direction, the HP output gap tends to be more volatile than the OECD output gap. However, the HP filter is not without its problems: it suffers from the so-called endpoint problem, where the cyclical component of the first and last values in a time series are poorly approximated (Mise et al., 2005). Moreover, it can also introduce spurious fluctuations after filtering (Enders, 2010). However, since the HP output gap is merely used as a check to test the robustness of the benchmark panel VAR and not the focal point of this paper’s analysis, this can be taken for granted.

Figure 10 shows the impulse response functions for the panel VAR model with the HP output gap, and Figure F4 in Appendix F shows the model’s stability graph. The model is estimated using two lags. The most important differences with the benchmark model are that a business cycle shock leads to an increase in the credit supply after one quarter, and that a sovereign debt shock increases the output gap at first before it becomes negative after approximately five quarters. Furthermore, a credit supply increases the output gap on impact, but this effect dies out after five quarters. Thus, qualitatively the effects found in the HP output gap model similar to those in the OECD output gap model. The model was also

8_{In order to save space, I only present the results with the total credit variable, as there were no substantial}

(32)

31

estimated with the household credit and firm credit variables. To save space, the orthogonalized impulse response functions for these models are available on request. Just as in the benchmark model, the model with firm credit was virtually identical to the model with total credit supply. The only differences in the model with household credit are the complete absence of any significant effect of household credit shocks on the output gap and of business cycle shocks on credit supplied to households.

The FEVD results are presented in Table G4 in Appendix G. The largest differences with the benchmark model are that the only shocks that substantially affect bond interest rates are sovereign debt shocks, which account for 97.6 percent of all variation. Furthermore, 77.2 of the variation in the inflation rate are caused by inflation shocks. Business cycle and sovereign debt shocks account for respectively 14.1 and 8.4 percent of the variation. The error variance decomposition of credit supply is roughly in line with the results found in the benchmark model. Finally, whereas sovereign debt shocks accounted for 55.6 percent of the variation in the output gap in the benchmark model, this drops to 9.1 percent in the HP output gap model. The effects of inflation and credit supply shocks on the output gap are negligible.

Figure 10: Impulse response function for panel VAR with Hodrick-Prescott filter output gap (impulse : response) 0 .2 .4 .6 -.05 0 .05 .1 -.1 -.05 0 .05 -.1 0 .1 .2 -.4 -.2 0 .2 .4 -2 0 2 4 -.5 0 .5 1 -.5 0 .5 1 0 .05 .1 .15 .2 -.05 0 .05 .1 0 .2 .4 .6 0 .05 .1 .15 .2 -.05 0 .05 .1 -.1 -.05 0 .05 -.05 0 .05 .1 .2 .3 .4 .5 0 5 10 0 5 10 0 5 10 0 5 10