IMPLICATIONS OF RISING DEBT LEVELS FOR MACROECONOMIC VOLATILITY

(1)

University of Groningen Faculty of Economics and Business

Master’s Thesis International Economics and Business

IMPLICATIONS OF RISING DEBT LEVELS

FOR MACROECONOMIC VOLATILITY

Author Tom Kosse

Student ID S3189414

Email T.S.Kosse@student.rug.nl

Supervisor Dr. A.C. Steiner

Co-Assessor Prof. dr. J. de Haan

(2)

ABSTRACT

Global debt levels have hit an all-time high. This paper examines the implications of the rising debt levels for macroeconomic volatility using a panel dataset consisting of 41 heterogeneous countries and comprising the period from 1990 to 2016. Regressing consumption, investment and GDP per capita volatility over different types of non-financial debt, this paper finds government debt to have a positive linear relation with investment volatility and a U-shaped relation with consumption and GDP per capita volatility. This finding strengthens the thought that to build the fiscal buffer required to address extraordinary events, limits should be put on government debt. In addition, evidence is found for a positive association between total non-financial debt and investment volatility. While some linkages between private debt and macroeconomic volatility are found, these findings are not robust to alternative specifications. Keywords: debt; macroeconomic volatility; implications

(3)

1 INTRODUCTION ... 4

2 LITERATURE REVIEW ... 6

2.1 Household debt and macroeconomic volatility ... 6

2.2 Corporate debt and macroeconomic volatility ... 10

2.3 Government debt and macroeconomic volatility ... 12

2.4 Total non-financial debt and macroeconomic volatility... 15

3 DATA AND METHODS ... 15

3.1 Variables ... 15 3.1.1 Dependent variables ... 15 3.1.2 Explanatory variables ... 17 3.1.3 Control variables ... 17 3.2 Final sample ... 18 3.3 Regression model ... 20 3.4 Diagnostic tests ... 22

3.5 Descriptive statistics and correlations ... 24

4 EMPIRICAL RESULTS ... 26

4.1 Main estimation results ... 26

4.1.1 Results hypotheses 1A and 1B... 26

4.2 Robustness tests ... 32

4.2.1 Non-overlapping periods ... 33

4.2.2 Exclusion of country-specific crisis dummy ... 33

4.2.3 Advanced countries versus developing countries ... 34

4.2.4 Pre-crisis versus post-crisis ... 35

4.2.5 Generalized methods of moments estimation ... 35

5 CONCLUSION AND DISCUSSION ... 36

REFERENCES ... 39

(4)

1 INTRODUCTION

Over the past decades, countries all over the world have witnessed a surge in debt levels. Global total non-financial debt peaked at a record high of $175 trillion at the end of 2017, with most of the debt concentrated among advanced countries (Institute of International Finance, 2018). As shown in Figure 1, debt accumulation outpaced the growth of GDP, making global total non-financial debt reach 238% of world GDP in 2017 from 192% in 2007 and 164% in 1997. Cecchetti et al. (2011) note that the current policies and demographics make it likely that debt levels will continue to grow over the coming years.

Figure 1. Global sectoral indebtedness, as % of GDP. Source: Institute of International Finance The high and increasing debt levels raise questions about their economic impact. On the one hand, the accumulation of debt may help to smooth economic activity. With debt, households can consume even without savings or current income, corporates can invest when their sales would otherwise not allow it and fiscal authorities can play their role in stabilizing the economy. On the other hand, the consequent high indebtedness creates weaknesses in balance sheets. This suggests that there is a sense in which debt can become excessive. Or, as Cecchetti et al. (2011) put it: “Debt is a two-edged sword. Used wisely and in moderation, it clearly improves welfare. But, when it is used imprudently and in excess, the result can be disaster” (p. 1).

Since the global financial crisis, academics have paid considerable attention to the effects of high debt levels on the economy. The greater part of the literature examines debt’s impact on economic growth. Most studies indicate that there is a quadratic relationship between debt and economic growth. For example, Cecchetti et al. (2011) found evidence that while a moderate level of debt improves welfare and enhances economic growth, a level of debt (whether household, corporate or government debt) reaching over 85-90% of GDP depresses growth. While the implications of high debt levels for economic growth have been thoroughly studied in the literature, only little academic attention has been paid to the impact of debt on macroeconomic volatility. This neglect is surprising for two reasons. First, the many economic crises of the past decades have made macroeconomic volatility an important topic in analysing

0% 20% 40% 60% 80% 100%

Household Non-financial corporates Government

Global Sectoral Indebtedness, % of GDP

(5)

the determinants of economic growth. A burgeoning literature documented that macroeconomic volatility, through its links with various forms of uncertainty, has a negative impact on long-term economic growth and welfare (Loayza et al., 2007). Second, theory also suggests a possible linkage between debt and macroeconomic volatility. Higher levels of debt are believed to cause a greater change in aggregate activity for a given size of shock to the economy (Cecchetti et al., 2011). High debt levels also constrain the ability to attract additional debt to smooth economic activity (Jorda et al., 2012). In sum, there is a strong theoretical presumption that high debt levels amplify macroeconomic volatility.

Although empirical research on the subject is lacking, some studies report evidence consistent with the above presumption. Most knowledge on the relationship between debt and macroeconomic volatility comes from an OECD project which is summarized in Sutherland and Hoeller (2012). The researchers, who examined the interlinkage between debt and a broad range of macroeconomic trends, devoted part of their work to debt’s impact on volatility. Using a panel of OECD countries, they performed a modest empirical analysis to show that high levels of debt inhibit the ability of the private sector to smooth consumption and investment and of governments to cushion adverse shocks, thereby increasing macroeconomic volatility. Furthermore, Pescatori et al. (2013), who also focused on OECD countries, have found some evidence of a positive relation between debt and output volatility. The authors found that countries tend to experience a relatively higher output volatility when government debt increases above 56% of GDP.

The few papers that pay attention to the relationship between debt and macroeconomic volatility provide a simple empirical analysis on the matter. They do not consider developing countries, do not control for alternative determinants of volatility and do not examine whether there is any non-linear relation. This paper aims to fill this gap. To do so, the following research question has been developed: “What are the implications of rising debt levels for macroeconomic

volatility?”. This paper contributes to the existing literature by presenting new empirical

evidence based on a more comprehensive analysis, using a panel dataset consisting of 41 advanced and developing countries and compromising the period from 1990 to 2016. Do higher debt levels increase consumption, investment and output volatility? Are there non-linearities in this relationship, with debt increasingly rising macroeconomic volatility at very high levels? Answers to these questions have important welfare implications since a better understanding of the causes of volatility can lead to more effective government policy that directly combats the long-term, underlying causes of volatility (Spiliopoulos, 2010).

(6)

The implications of debt for macroeconomic volatility are examined using fixed effects estimation and controlling for other factors that may be associated with the level of macroeconomic volatility, namely government spending, trade openness, exchange rate volatility, inflation, real GDP per capita, average real growth rates, the political environment and crises. Moreover, the specification includes time-fixed effects to account for secular trends in the data. Household, corporate, government and total non-financial debt are measured by their respective debt-to-GDP ratio, while macroeconomic volatility is measured by the 5-year rolling standard deviation of the real growth rate of consumption, investment and GDP per capita. Each of the three volatility measures is regressed on each type of non-financial debt. The main estimation results predict government debt to have a positive linear relation with investment volatility and a U-shaped relation with consumption and GDP per capita volatility. In addition, there is some evidence pointing to a positive association between total non-financial debt and investment volatility. No robust evidence for any causal relation between private debt and macroeconomic volatility is found. Moreover, the robustness checks suggest that the effects of debt are somewhat country-specific, indicating that various factors should be evaluated when assessing the vulnerabilities arising from high indebtedness.

The remainder of the paper is organized as follows. Chapter 2 reviews the existing literature to provide a solid understanding of the concepts and theories related to the research. Chapter 3 introduces and explores the data and the methods used to analyse the data. Subsequently, the empirical findings are presented and interpreted in Chapter 4. Chapter 5 provides an answer to the research question, discusses limitations and suggests recommendations for future research.

2 LITERATURE REVIEW

This chapter examines literature to identify the mechanisms through which each type of non-financial debt can affect macroeconomic volatility. Each section focuses on the theory and empirical findings concerning one type of debt and concludes with a set of hypotheses.

2.1 Household debt and macroeconomic volatility

(7)

financial crisis, the ratio remains at historically high levels and has kept growing in most developing countries.

Figure 2. Average household debt-to-GDP ratio. Source: Bank for International Settlements A wide range of household debt levels may be sustainable if debt is contracted to increase assets, that is, if net wealth is maintained (IMF, 2006). Nonetheless, several studies, which will be discussed below, indicate that simultaneous increases in debt and assets, which can leave net wealth unchanged, cause higher household consumption elasticities. Indeed, Debelle (2004) notes that while increased indebtedness, in and of itself, is not likely to be a source of shocks, the primary macroeconomic implication is that it does expose households as well as the wider economy to shocks coming from other sources. At high debt levels, there are several mechanisms that may amplify or undermine the capacity to damp shocks to income, interest rates and asset prices.

First, because debt payments represent commitments whose amount and timing can hardly be altered, households with debt find it more burdensome to maintain their loan payments through a period of lower income. Dynan and Kohn (2007) argue that reductions in income reduce the cash flow available to fund consumption proportionately more for household with a large stock of debt. As a result, shocks to income have a larger effect on consumer spending than they would have had at low debt levels. Moreover, households may be averse to have a high debt-to-income ratio, so that a decrease in income will prompt larger declines in spending among highly indebted households to attain their desired amount of debt relative to their income. This rationale is strengthened by a recent study by Baker (2018) who estimated that, even after controlling for net wealth, the elasticity of consumption with respect to income is significantly higher among households with high debt than among households with a low stock of debt. Working in the other direction, households may take on more debt to smooth consumption after an unexpected temporary drop in income, thereby reducing the impact of a shock to income. According to the permanent income hypothesis, households base their consumption on their expected long-term average income. Therefore, the hypothesis predicts that transitory changes in income only have little effect on consumption spending, since households would borrow to

0% 20% 40% 60% 80% 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

Average household debt-to-GDP ratio (1990-2016)

(8)

smooth through any shocks (Hall, 1978). However, the ability to smooth is generally only available to households with free collateral in the form of their house, who also have an existing line of home equity credit (Debelle, 2004). Households with higher levels of debt are less likely to be able to adjust by borrowing, as lenders would be less cooperative. This lack of an ability to access credit can decrease the ability of households to smooth consumption when subjected to an unanticipated negative income shock. Indeed, Baker (2018) shows that the differential household consumption elasticities with respect to income among households with varying levels of debt can largely be explained by borrowing and liquidity constraints.

In addition, Baker (2018) found that even following positive shocks to household income, households with high debt experience higher elasticities of consumption with respect to income than households with low debt, although the consumption responses are much smaller than to negative shocks. The asymmetry can be explained by the fact that households can always smooth consumption through saving when confronted with an unexpected increase in income. Similarly, at high debt levels, interest rates changes have a greater effect on total debt servicing costs than in the situation where debt levels were lower. When debts are large, a given change in interest rates has a large effect on debt service and thus a large effect on the funds available for consumption (Dynan & Kohn, 2007). Auclert (2017) estimated that the greater the interest rate sensitivity of households’ liabilities relative to that of their assets, the larger the impact on consumption. In addition, interest rate changes could have asymmetrical effects in an economy with high levels of household debt. More specifically, interest rate hikes cause consumption to contract more than cuts would cause it to expand. In the model of Auclert (2017), while borrowers adjust spending one for one with every dollar increase in the payments they must make when interest rates rise, they prefer to save an important fraction of the gains they get from low interest rates.

(9)

There also is some evidence that precautionary motives are driving the relationship between household debt and consumption behaviour. Carroll and Dunn (1997) put forward that as debt accumulates, precautionary motives make the spending of households more sensitive to uncertainty about the future. Therefore, highly indebted households would more quickly than their lower-debt peers adjust their spending in response to changes in expectations about the future path of income, house prices and interest rates. Similarly, Meier and Sprenger (2010) argue that the expectations about future income and asset values may be overly optimistic and therefore can lead households to take on more debt than appropriate. A downward revision in expected income growth or asset appreciation would encourage those households to reassess their borrowing capacity and forego consumption to pay down their debt faster. Also, a downward revision in the future inflation rate would suggest that the real value of the debt is not eroded as fast as in the past, implying higher debt in the future than previously thought. This could provide highly indebted households another incentive to pay down their debt faster (La Cava & Price, 2017).

The IMF (2017) observed that the overall trend in the household debt-to-GDP ratio is very similar to the trend in the debt-to-assets ratio, implicating that increases in debt are usually accompanied by rising leverage. With high leverage, even mild shocks can cause apparently sustainable debt to become unsustainable, and households may suddenly no longer be regarded as creditworthy. The consequent borrowing constraints imposed by financial institutions would prevent households from taking on more debt, forcing them to reduce spending to bring their leverage back to more manageable levels (Dynan, 2012). Guerrieri and Lorenzoni (2017) found that with deleveraging, households will first reduce their spending to deleverage, and once deleveraged, they continue to consume less to build up precautionary savings before restoring consumption to its original level. The sharp adjustments in the consumption pattern may also affect other parts of the economy due to linkages between sectors (Sutherland & Hoeller, 2012). In summary, as debt levels increase, households’ spending behaviour becomes progressively more sensitive to changes in (expectations about future) income, interest rates, house prices and inflation. High levels of debt amplify the impact of the initial shocks that cause these changes. High household indebtedness also limits the extent to which additional borrowing can be used to smooth consumption and makes households more likely to start deleveraging, which can bring about sharp adjustments in the consumption pattern and affect other parts of the economy. The higher the level of household debt, the bigger the change in economic activity for a given size of shock to the economy. Following this rationale, higher debt is expected to raise macroeconomic volatility. This leads to the following hypotheses:

Hypothesis 1A: Higher household debt is associated with higher macroeconomic volatility.

(10)

2.2 Corporate debt and macroeconomic volatility

Figure 3 shows the development of the corporate debt-to-GDP ratios for the countries in this paper’s sample. As shown, there is heterogeneity across countries in terms of both the levels and the development of the corporate debt-to-GDP ratios. During the boom that preceded the global financial crisis, corporates financed the expansion of their balance sheets primarily by borrowing, causing a considerable increase in their debt obligations (Occhino & Pescatori, 2010). The average corporate debt-to-GDP ratio in advanced countries stopped growing after the global financial crisis but remains stable at historical high standards. Corporate debt levels of developing countries are lower but keep growing and are therefore catching up.

Figure 3. Average corporate debt-to-GDP ratio. Source: Bank for International Settlements Reasons for the use of debt are the incentive effects of debt, the costliness of making claims contingent upon corporates' condition and the use of debt as a risk-sharing instrument (Bernanke & Campbell, 1988). Another explanation for the existence of debt is debt's tax-favoured status. In most tax systems, interest expenses are tax-deductible whilst dividend payments are not. This preferential treatment of interest payments encourages corporates to issue debt (Gertler et al., 1990).

Gertler and Hubbard (1993) assert that the tax bias against equity reduces the extent to which corporates insulate themselves against macroeconomic risk accompanying debt finance. The authors state that equity allows a corporate to share risks with its creditors, limiting the chance that an adverse shock will push it into financial distress. Debt cannot perfectly substitute for equity in this risk-sharing role and is, therefore, less effective than equity in insulating corporates against risks (Gertler & Hubbard, 1993). The more indebted a corporate is, the more sensitive it becomes to adverse shocks. Indeed, Davis and Stone (2004) estimated that investment falls more during a downturn when the ratio of debt to equity is high. This is consistent with empirical evidence of Mullineux et al. (2011) for European countries that while equity financing reduces investment volatility, bank financing increases it. For example, corporates with a higher reliance on debt rather than equity financing may be less able to cope

0% 50% 100% 150%

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

Average corporate debt-to-GDP ratio (1990-2016)

(11)

with adverse revenue shocks. Since debt repayments tend to be fixed, an adverse income shock will squeeze the cash flow available for investment (Jorda et al., 2012). Similarly, high indebtedness increases the sensitivity of debt servicing costs to changing interest rates.

The state of a corporate’s balance sheet matters for borrowing in normal times and especially during recessions (Jiménez et al., 2012). High levels of debt hamper the ability of corporates to raise external funds for new investments. And when they can attract debt financing, they are required to pay higher interest rates, which increases their cost of capital (Occhino & Pescatori, 2010). The inability to obtain new funds on reasonable terms is a major cost to corporates with excessive debt outstanding and can have a marked effect on their ability to smooth activity (Bernanke & Campbell, 1988). The situation in which corporates benefit little from the return to any additional investment because of debt service obligations can lead to debt overhang. Debt overhang occurs when existing debt deters new investment because most of the value generated by investments would go to the existing creditors rather than to equity holders (Myers, 1977).

The financial accelerator theory suggests that the debt overhang distortion exacerbates slowdowns in economic activity, by amplifying shocks to aggregate demand and supply (Bernanke et al., 1996). An adverse shock not only brings its standard contractionary effect, but also affects corporates’ balance sheets, thereby worsening the debt overhang distortion and further diminishing corporates’ incentives to invest (Occhino & Pescatori, 2010). This additional effect enlarges and prolongs the impact of adverse shocks on the economy. For example, Occhino and Pescatori (2010) claim that the debt overhang distortion causes the impact of a productivity shock on investment and production to double and extend for several years. A productivity shock initiating a recession would reduce the net worth of corporates. When corporates’ net worth becomes lower, they find it even more costly and difficult to raise external funds. Consequently, both their supply of goods and their demands for new capital fall, tending to worsen the recession. Another example, stressed by Bernanke and Gertler (1989), is debt-deflation. In a debt-deflation, a slower than expected inflation or an unanticipated fall in the general price level redistributes wealth from corporates to creditors. Again, the lower corporate net worth has negative effects on investments and output.

The debt overhang distortion caused by high levels of debt not only increases the vulnerability of the economy to destabilizing shocks but may also affect the structure of corporates’ investments in terms of their riskiness. Occhino and Pescatori (2010) argue that while debt overhang decreases the number of safe projects, it may increase the number of riskier investments. The rationale given is that equity holders of corporates facing debt overhang problems have an incentive to propagate risky projects because they beneﬁt from successful projects, while the creditors bear the downside risks.

(12)

(Bernanke & Campbell, 1988). This potential arises from the fact that many corporates count on being able to roll over their short-term debt as it comes due. If for some reason, the corporates’ creditors become worried about bankruptcy risk and refuse to roll over maturing debt, then these corporates would have difficulty refinancing their debt and find themselves illiquid (Bernanke, 1989). The legal proceedings initiated by possible bankruptcy would freeze the liabilities of the failing corporates. Assets that creditors previously considered to be liquid may then turn out to be illiquid, worsening the illiquidity problem of the creditors. Moreover, major bankruptcies may contribute to cash-flow problems of the corporates' suppliers and demanders. Therefore, bankruptcies among major corporates, which become more likely as debt levels increase, can contribute to a general liquidity crisis, which has a disruptive impact on the production and investment activities in the economy (Bernanke & Campbell, 1988). Another disturbing scenario is a solvency crisis. Suppose that the economy enters a recession, resulting in rising interest costs and falling revenues. Highly indebted corporates may then have trouble in servicing their debt (Bernanke, 1989). When financial problems hit, the need to meet interest payments may force management to take a very short-run perspective, leading them to retrench, cancelling even potentially profitable projects and cutting back production, employment and investment (Ruscher & Wolff, 2012). These actions would reduce total demand, thereby worsening the recession and leading to financial problems in other parts of the economy. Thus, the initial shock would be magnified by already high debt levels.

In conclusion, high levels of corporate debt may act as an automatic destabilizer, increasing the sensitivity of corporates to exogenous shocks. High corporate debt levels also limit the extent to which additional borrowing can be used to smooth investment and production and increase the probability of a liquidity or solvency crisis. The effect of additional debt is likely to be most drastic for corporates that experience debt overhang problems, suggesting that the impact of additional corporate debt on macroeconomic volatility is stronger when debt levels are already high. Therefore, the following hypotheses have been formulated:

Hypothesis 2A: Higher corporate debt is associated with higher macroeconomic volatility.

Hypothesis 2B: The positive association between corporate debt and macroeconomic volatility becomes stronger as corporate debt accumulates.

2.3 Government debt and macroeconomic volatility

(13)

countries in this paper’s sample. It is notable that especially advanced countries adopted fiscal stimulus measures of large-scale following the global financial crisis. The debt-to-GDP ratios of the developing countries have remained quite stable. Consequently, the current government debt levels of advanced countries are significantly higher than those of developing countries.

Figure 4. Average government debt-to-GDP ratio. Source: Bank for International Settlements Imagine a deleveraging crisis, in which there is an abrupt downward revision of views about how much debt it is safe for the private sector to have, and in which this revision of views forces highly indebted households and corporates to reduce their spending. To avoid a slump, someone must increase its spending to compensate for the fact that private sector borrowers are spending less (Eggertson & Krugman, 2012). Eggertson and Krugman (2012) suggest that the public sector should borrow to fill the spending gap left by households and corporates. Indeed, Elmendorf and Mankiw (1999) note that the view held by most economists and policymakers is that expansionary fiscal measures, including measures related to increasing government debt, can stimulate aggregate demand for goods and services. Conventional analysis assumes the economy to be Keynesian in the short run, with sticky wages and prices (Elmendorf & Mankiw, 1999). An increase in aggregate demand is therefore believed to affect the utilization of the economy's factors of production, resulting in higher national income (Elmendorf & Mankiw, 1999). Particularly when private sector balance sheets are impaired and liquidity constraints bite, fiscal policy may have large multipliers, so that the final impact on national income may be much greater than the initial change in aggregate demand (DeLong & Summers, 2012). This Keynesian analysis provides a justification for deficit-financed policy of increasing government spending and cutting taxes when the economy is faced with a (possible) recession. Keynesians argue that the measure allows the economy to avoid unemployment and deflation while highly indebted private agents repair their balance sheets (Essien et al., 2016). By lessening the effects of the liquidity constraint faced by private agents, the impact of exogenous shocks on aggregate current consumption, investment and output would be alleviated, thereby reducing macroeconomic volatility (Debrun et al., 2008).

0% 20% 40% 60% 80% 100% 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

Average government debt-to-GDP ratio (1990-2016)

(14)

While some economists, among which are DeLong and Summers (2012), argue that under certain conditions regarding government bond yields, the fiscal multiplier and hysteresis effects, expansionary fiscal policy would not only be positive for the economy, but also self-financing, the dominant view is that maintaining expansionary fiscal policies requires continued borrowing resulting in an ongoing increase in debt. But the capacity of the government to borrow is limited. When a recession sets in, the ability of the government to step in depends on the stock of debt that it has already accumulated as well as what creditors perceive to be its fiscal capacity (Cecchetti et al., 2011). Therefore, highly indebted governments may become constrained both in their role as lender of last resort and in their attempt to engage in countercyclical stabilisation policies during a crisis.

Besides the financial constraints, there are several factors that reduce the effectiveness of fiscal policy in reducing macroeconomic volatility at high debt levels. The Ricardian equivalence theorem suggests that private agents save the proceeds from a debt-financed fiscal stimulus because they recognize that government debt generated through deficit spending will eventually be paid-off by increased taxes (Barro, 1989). The imposition of higher future taxes would be perceived as more tangible if indebtedness is high (Nickel & Tudyka, 2013). If it holds true that the private sector assumes that whatever is gained now will be offset by higher future taxes, a fiscal stimulus financed by higher borrowing would have no impact on aggregate demand (Barro, 1989). Röhn (2010) investigated the extent of possible offsetting private saving behaviour to fiscal policy changes and estimated that the theory holds empirically and that private saving offsets are stronger the higher the level of government debt. Thus, expansionary fiscal policy may change from being Keynesian in nature to becoming increasingly Ricardian as government indebtedness rises.

(15)

economy. High government debt levels, however, limit the capacity to engage in expansionary fiscal measures and reduce the effectiveness of such policies, thereby putting limits on governments’ ability to stabilise the economy. Thus, the higher the level of government debt, the smaller and less effective are expansionary fiscal measures and the deeper are economic downturns. This leads to the following hypotheses:

Hypothesis 3A: Higher government debt is associated with higher macroeconomic volatility.

Hypothesis 3B: The positive association between government debt and macroeconomic volatility becomes stronger as government debt accumulates.

It should be noted that testing these hypotheses may to some extent be constrained, because in the cases that an increase in government debt results from expansionary fiscal policies, which should always have a volatility-decreasing effect, the volatility-worsening effect of the consequent higher level of debt might partially be concealed.

2.4 Total non-financial debt and macroeconomic volatility

Each type of non-financial debt is expected to have a similar impact on macroeconomic volatility. This allows the creation of the following hypotheses to gain insight into the impact of the total level of non-financial debt on macroeconomic volatility as well:

Hypothesis 4A: Higher total non-financial debt is associated with higher macroeconomic volatility.

Hypothesis 4B: The positive association between total non-financial debt and macroeconomic volatility becomes stronger as total non-financial debt accumulates.

3 DATA AND METHODS

This chapter starts with describing the variables in the regression model. Subsequently, the final sample and data sources are discussed, the specifications of the regression model are presented, and diagnostic checks are performed. Lastly, the basic features of the data are presented.

3.1 Variables

3.1.1 Dependent variables

(16)

GDP (per capita) growth rate over time (Ramey & Ramey, 1995; Bejan, 2007, among others). However, Cariolle (2012) notes that it is common to use a broader definition by referring to the standard deviation of the growth rates of several macroeconomic variables. Frequently, the standard deviation of the growth rates of the following macroeconomic aggregates are considered: real consumption, real investment and real GDP (per capita) (Denizer et al., 2002; Dabla-Norris & Srivisal, 2013; Niranjan, 2017, among others). This paper chooses to follow the latter definition as it provides a more extensive view of the extent to which the macroeconomy changes quickly and unpredictably. In the empirical context, this means different models will be tested, each with its own dependent variable.

Studies also differ greatly in the way standard deviations are calculated. Fanelli (2008) notes that some studies calculate just one volatility estimate for each country for the entire sampling period, while others prefer multiple volatility estimates with each estimate pertaining to either an overlapping or non-overlapping subperiod. This paper chooses to follow the approach of Blanchard and Simon (2001) and Fanelli (2008) when calculating the standard deviations of the real growth rates of consumption, investment and GDP per capita. That is, for each variable, multiple volatility estimates are calculated at different points in time using overlapping windows. Following much of the literature on macroeconomic volatility, a 5-year window is used. Since a window of 5 years is used, the volatility estimate reported for year t is the standard deviation over years t - 4 to t. It must be noted, however, that while this paper uses windows consisting of 5 years of annual data, several other studies use 5 years of quarterly data. Due to its higher-frequency, quarterly data may have been more efficient in capturing fluctuations in the macroeconomic aggregates. However, it is not as widely available as annual data.

(17)

3.1.2 Explanatory variables

Four different explanatory variables are used to capture the separate effects of household, corporate, government and total non-financial debt. Following previous research of Sutherland and Hoeller (2012), the debt levels will be measured by their respective debt-to-GDP ratios. The aggregate sectoral debt levels have as limitation that they may mask significant heterogeneity within a sector. A high level of indebtedness at the aggregate level may not mean there is a risk to macroeconomic stability if the distribution of debt is biased toward households and corporates that have a buffer to withstand shocks and a higher payment capacity. In contrast, debt highly concentrated among households and corporates that are least able to bear it may create vulnerabilities, even if the aggregate balance sheets appear reasonably healthy. Moreover, underestimation is possible for developing countries where participation rates are low and where low indebtedness at the macro-level may coexist with high micro-level indebtedness. In this regard, using micro-level data would help identify pockets of fragility within sectors. Nonetheless, data scarcity does not allow for further disaggregation.

3.1.3 Control variables

(18)

trade openness as control variable helps to capture exposure to shocks. Third, this paper includes the volatility of countries’ exchange rate to account for the effects of external shocks on domestic macroeconomic volatilities, since exchange rate changes can affect production and consumption decisions. The volatility will be calculated in a similar way as macroeconomic volatility. A fourth control variable is the inflation rate (measured by the consumer price index), because inflation may be correlated with GDP and its components when the aggregate supply curve is upward sloping. Mobarak (2005) confirms the relationship as he estimated that countries with lower inflation rates tend to exhibit less volatility. The present paper also includes real GDP per capita as a proxy for the degree of development of a country, since poorer economies may experience higher volatility. Moreover, several papers on macroeconomic volatility include the annual real growth rates of the macroeconomic aggregates whose volatility is estimated, since the growth rates may be correlated with the volatility estimates. Although the exact shape of the relationship between growth and volatility remains unclear, studies do agree about the existence of such a relationship. For instance, Imbs (2006) estimated that growth and volatility correlate positively at the sectoral level, but negatively at the country level, with faster growing economies exhibiting less variability. The latter finding is in line with Ramey and Ramey (1995) who estimated that countries with lower growth have higher volatility. Furthermore, an index of the type of political regime is included since the political environment may affect macroeconomic stability. Indeed, Mobarak’s (2005) estimated that democracy significantly reduces volatility. Likewise, Acemoglu et al. (2003) and Malik and Temple (2009) found that an important driver behind volatility arises from the effects of weak institutions, through channels such as distortionary macroeconomic policies. Lastly, this paper includes a country-specific crisis dummy since banking, currency and sovereign debt crises are typically associated by dramatic adjustment processes and can therefore be expected to be accompanied by an increase in volatility. Including the dummy for the crisis years helps to marginalize the effects of the crises episodes. Dabla-Norris and Srivisal (2013), Niranjan (2017), Kose et al. (2003) and Denizer et al. (2002) largely include similar (control) variables in their study on macroeconomic volatility, strengthening the idea that they are important determinants of macroeconomic volatility. An overview of all variables including a short description and data sources can be found in Appendix A.

3.2 Final sample

(19)

capture possible trends. A disadvantage is that it mainly covers the last business cycle, which was characterised by the Great Moderation and the global financial crisis. During this period, debt rose in most countries, while macroeconomic volatility generally declined for a variety of reasons. As such, the analysis covers a period which is somewhat atypical. Figure 5 shows the development of the average debts ratios and volatility in this paper sample, illustrating the atypical characteristic of the period examined.

Figure 5. Left panel: average volatility (measured by the 5-year rolling standard deviation of the real growth rate of consumption, investment and GDP per capita). Right panel: average debt-to-GDP.

The data used for the main analysis originates from six data sources. The National Accounts Main Aggregates Database, which is powered by the United Nations Statistics Division, presents a series of analytical national accounts tables from 1970 onwards for more than 200 countries and allowed to gather data on the real growth rates of consumption, investment and GDP. Based on these rates, the standard deviations are calculated. The database also provided data on exchange rates, allowing to calculate exchange rate volatility. The debt-to-GDP ratios are obtained from a database on credit to the non-financial sector provided by the Bank for International Settlements. The ratios capture the outstanding amount of credit provided by domestic banks, all other sectors of the economy and non-residents. Credit covers the core debt, defined as loans, debt securities, currency and deposits. Furthermore, IMF’s World Economic Outlook database, which contains selected macroeconomic data series from the statistical appendix of the World Economic Outlook report, is used to acquire data on government expenditures-to-GDP, where expenditure consists of total expense and the net acquisition of nonfinancial assets. In addition, the World Bank is consulted to gather information on real GDP per capita levels, import and export statistics, inflation, and the population growth rates. Data on the type of political regime for each country is obtained from the Polity IV dataset, which contains annual information on the level of democracy for all major countries over the period 18002016. For each year, a ‘polity score’ is determined for each country which ranges from

-0 1 2 3 4 5 6 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016

Average volatility (1990-2016)

Consumption volatility Investment volatility GDP per capita volatility

(20)

10 (full autocracy) to +10 (full democracy). Lastly, the banking crises database by Laeven and Valencia (2012) is used to determine crisis episodes to construct the country-specific crisis dummy. The database includes all systemic banking, currency, and sovereign debt crises between 1970 and 2012. The dataset does not cover the last four years of the period considered in this paper; this paper does not expect major crises to have taken place in these years.

3.3 Regression model

This paper runs a regression on panel data to uncover the impact of the explanatory variables on macroeconomic volatility, because panel data takes into consideration both time series and cross section features. With panel data, there generally are three types of regressions that are considered: a pooled model, a fixed effects model and a random effects model. Two tests are considered to determine which regression model best suits the dataset. Namely the Durbin-Wu-Hausman test, to help decide between a fixed and random effects regression and the F-test test, to help decide between a fixed effects regression and a simple OLS regression. At a significance level of 5%, which will be used in the whole paper, the null hypothesis of the Durbin-Wu-Hausman test be rejected for 21 of the 24 models in this paper, pointing towards the use of fixed effects models (Torres-Reyna, 2007). The results of these tests can be found in Appendix C. For the sake of uniformity, the fixed effects model is preferred over the random effects model in all models. Moreover, the null hypothesis (all individual effects are jointly equal to zero) of the F-tests obtained when running the fixed effects models can all be rejected, indicating that the fixed effects models outperform OLS regressions.

The fixed effects model is also theoretically preferred because it allows for different intercepts for different countries and thus controls for unobserved country characteristics. Omitted time-invariant variables that jointly influence macroeconomic volatility cannot bias the estimated coefficients in the fixed effects model as the nature of the model eliminates time-invariant variables from the regression (Kohler & Kreuter, 2009). In this way, the net effect of the explanatory variables can be studied. The disadvantage of the use of the fixed effects model is that it restricts the interpretation of the results to analysis of within-country effects.

(21)

joint tests have been performed to test if the dummies for all years are equal to zero (Appendix D). For each model, the null that the coefficients for all years are jointly equal to zero can be rejected, confirming that time-fixed effects are needed in this paper (Torres-Reyna, 2007). In addition, the debt variables will be lagged with one year to lessen potential problems of endogeneity. This decision is prompted by Denizer et al. (2002) and Sutherland and Hoeller (2012) who asserts that to avoid that changes in GDP influence both the volatility and debt variables, the debt variables should always be included with a lag. Considering the previous year’s levels of the debt-to-GDP ratios helps to minimise problems with spurious correlation, which may arise because changes in GDP during recessions can simultaneously raise volatility and debt-to-GDP levels (Sutherland & Hoeller, 2012). Since the volatility estimates are a measure of the current and previous four years, it is debatable whether a lag of one year is enough. However, the same lag is used by Fanelli (2008), who faced the same quandary on this matter. Looking at the ten control variables, which are selected based largely on Spiliopoulos (2010), Dabla-Norris and Srivisal (2013), Niranjan (2017), Kose et al. (2003) and Denizer et al. (2002), the papers in question generally do not lag the variables. Again, this paper chooses not to deviate from the mentioned literature and thus to use no lag for the control variables. As discussed, three different dependent variables are used to capture macroeconomic volatility, while four different explanatory variables are used to capture non-financial debt. Therefore, multiple regressions will be performed to determine the effect of each type of non-financial debt. More precisely, twelve regressions are performed to determine the linear relation between each type of non-financial debt and each measure of volatility while another twelve regressions are performed to determine whether there are any non-linearities.

Hypotheses 1A, 2A, 3A and 4A expect a positive relation between respectively household, corporate, government and total non-financial debt and macroeconomic volatility. This can be characterized by the first regression equation below. It is expected that the coefficient of interest, 𝛽0, is positive and is related to an increase in the different measures of macroeconomic volatility. Hypotheses 1B, 2B, 3B and 4B expect the macroeconomic volatility-worsening effect to become stronger as debt accumulates. To find evidence for these hypotheses, the second regression equation below includes the squared term of debt to capture any non-linear effects, where it is expected that both coefficients of interest, 𝛽0 and 𝛽1,are positive.

Volatilityit = β0Debtit-1 + β1Control Variablesit + αi + δtTt + uit Volatilityit = β0Debtit-1 + β1Debt2it-1 + β2Control Variablesit + αi + δtTt + uit

(22)

dummy), αi is the unknown intercept for each country, δt is the coefficient for the binary time variables, Tt is time as binary variable, and uit is the error term.

3.4 Diagnostic tests

To have reliable estimates several assumptions must be satisfied. One of the assumptions is the absence of multicollinearity. With high multicollinearity, it is difficult to disentangle the effects of the variables involved, causing inflated standard errors of the coefficients. The most widely used approach to check for multicollinearity is the variance inflation factor (VIF). However, the VIF test, as a measure of the severity of multicollinearity, is only available for OLS regressions. Therefore, a correlation matrix is used to analyse whether there are multicollinearity issues. The correlation values between the different variables are presented are presented in Appendix E. There is no consistent argument on the level of correlation that causes multicollinearity. Farrar and Selwyn (1967) suggest that correlations that are smaller than 0.8 are acceptable, while Malhotra (2007) states that a multicollinearity problem exists when the correlation coefficient among variables is greater than 0.75. As can be seen in the correlation matrix, there is no troublesome correlation between any of the debt variables and the control variables. However, despite the low correlation, it may be argued that the crisis dummy is interrelated with the debt variables; this will be discussed in section 4.2. Moreover, the significant correlation between three control variables (real growth rates) closely approaches the threshold proposed by Malhotra (2007). Nonetheless, Allison (1999) states if the variables are only used as controls, and they are not collinear with the debt variables, there is no problem because the coefficients of the variables of interest are not affected, and the performance of the control variables as controls is not impaired.

(23)

function of the first variable (Goldberger, 1991). Allison (1999) agrees and suggest that high multicollinearity can be ignored if it is caused by the inclusion of powers of other variables. Furthermore, the models potentially exhibit heteroskedasticity. If the variance of the residuals is non-constant, then the residual variance is said to be heteroskedastic. There are graphical and non-graphical methods for detecting heteroskedasticity. A commonly used graphical method is to plot the residuals versus the fitted (predicted) values. The plot for regressing consumption volatility on household debt, which looks similar, although a little more marked, to the plots for the other regressions, is given in Appendix F. The pattern of the data points is getting wider towards the right end, which is an indication of heteroskedasticity. Confirmation for this intuition comes from a non-graphical method to test for the presence of heteroskedasticity, the modified Wald test. The test calculates a modified Wald statistic for groupwise heteroskedasticity in the residuals of a fixed effects model. Under the null hypothesis, it assumes homoskedasticity. The results for the modified Wald test can be found in Appendix G. All models have a p-value smaller than 0.05, indicating the possibility of heteroskedasticity at the 5% significance level.

Furthermore, a test is conducted to check for serial correlation. Serial correlation is a relatively common problem which occurs when time series data is influenced by its own historical values. It biases the standard errors and makes the results less efficient (Drukker, 2003). Serial correlation in panel data models can be tested for with the Wooldridge test (Wooldridge, 2002). The outcomes of the Wooldridge tests are significant, which indicates the presence of serial correlation (see Appendix H). To control for both serial correlation and heteroskedasticity, cluster-robust standard errors will be used in each regression. The cluster-robust standard errors can correct, at least partially, for the serial correlation introduced by the way volatility is measured.

(24)

Moreover, the Shapiro-Wilk test is conducted to check whether the residuals are normally distributed (see Appendix J). The null hypothesis of the test can be rejected for every variable, which indicates that the data is non-normally distributed (Hanusz et al., 2016). To check how severe the non-normality is, quintile-normal plots are constructed for the data on the volatility and debt measures (see Appendix K). The figures plot the quintiles of the residuals against the quintiles of a normal distribution and allow to check for non-normality in the extremes of the data (Torres-Reyna, 2007). As can be seen, the tails are a bit right-skewed and fat-tailed, but do not show a substantial departure from normality. Since prediction intervals are calculated based on the assumption that the residuals are normally distributed, non-normality may reduce the accuracy of the prediction intervals. But, if the distribution is not too grossly non-normal, the tests will still provide good approximations (Hanusz et al., 2016). Since the quintile-normal plots indicate the latter is the case, at least for the six main variables, the non-normality is not too much of a problem. This is endorsed by the central limit theorem, which states that for large enough sample sizes the distribution of the data will be approximately normal, regardless of the shape of the data (Ghasemi & Zahediasl, 2012).

3.5 Descriptive statistics and correlations

(25)

Similar as to previous studies examining the volatility of consumption, investment and GDP per capita, by and far the most volatile macroeconomic aggregate is investment, with consumption and GDP per capita volatility following at a proper distance. Countries standing out for high volatility levels are predominantly less developed (Chile, China, Hungary, Ireland, Russia, Saudi Arabia, Singapore and Turkey), while countries with relatively low volatility are all highly developed (Australia, Austria, Belgium, Switzerland, Germany, United Kingdom, United States and Japan). Looking at the debt variables, the mean of corporate debt appears to be highest, followed by the mean of government and household debt. This corresponds to the global non-financial debt levels shown in Figure 1. What stands out for the debt variables is that the differences between the minimum and maximum values are quite large. Remarkable is that all minimum values are found at the beginning of the 1990s, while the maximum values are found in the most recent years. Russia and Turkey had the lowest household debt levels, while the maximum household debt levels are found in Australia, Switzerland and The Netherlands. With corporate debt, Brazil stands out for its low values while Hong Kong and Ireland do for their high levels. South Korea and Saudi Arabia account for the minimum government debt levels, with the opposite applying to Greece and Japan.

The correlation matrix in Appendix D, which was assessed in the previous section to check for multicollinearity, shows that the different measures of volatility are all positively and relatively strongly associated, with the strongest correlation being that between consumption and GDP per capita volatility (.75). This suggests that consumption, investment and GDP per capita are affected quite similarly by given developments. Nonetheless, the correlations are low enough to suggest that each measure captures a slightly different feature of macroeconomic volatility.

Table 1: Descriptive statistics

Variable Obs. Mean Std. Dev. Min. Max.

Consumption volatility 1097 2.17 1.61 .30 8.31

Investment volatility 1097 5.59 2.88 1.28 13.75 GDP per capita volatility 1097 2.24 1.54 .38 7.44

Household debt 941 48.16 28.11 1.30 117.40

Corporate debt 937 76.86 35.80 12.60 166.40

Government debt 873 57.42 33.54 5.40 152.60

Total non-financial debt 783 187.13 74.77 42.30 389.80 Government spending 1015 37.27 12.24 13.54 59.97

Inflation 1100 4.02 4.17 -1.14 19.82

Trade openness 1105 83.49 67.85 17.51 382.84

Exchange rate volatility 1098 .18 .63 0 5.22

GDP per capita 1107 27.70 19.54 .76 87.48

Consumption growth 1099 3.16 3.32 -6.40 12.20

Investment growth 1099 3.50 7.03 -15.40 18.60

GDP per capita growth 1099 2.21 3.19 -7.83 10.59

Polity index 1105 7.48 4.71 -10 10

Crisis dummy 1107 .13 .34 0 1

(26)

The same conclusion holds for the different debt variables, with the strongest correlation being that between corporate and total non-financial debt (.85). The moderate correlation between the debt variables confirms that each debt variable has different underlying mechanisms and therefore should be examined separately.

Contrary to what may be expected given the hypotheses in this paper, there is a negative correlation between the different types of debt and the measures of volatility. However, negative correlation does not mean there cannot be a positive causal relation. As discussed in section 3.2, the negative correlation can be explained by the atypical period examined, in which volatility decline for a variety of reasons unrelated to debt.

4 EMPIRICAL RESULTS

This section starts with discussing the empirical results and linking them to the hypotheses. Subsequently, several checks are performed to assess the robustness of the results.

4.1 Main estimation results 4.1.1 Results hypotheses 1A and 1B

As for the link between household debt and macroeconomic volatility, the results of executing the equations as a regression are tabulated in Table 2 below. The volatility measures of consumption, investment and GDP per capita are the endogenous variables and are regressed on household debt and the control variables. The overall significance of the fixed effects models can be determined with the F-test. As can be seen, the null hypothesis of the F-tests can be rejected, indicating that the models are significant. Furthermore, the number of observations in each regression is 860. This is lower than the number of observations in Table 1, where each involved variable has 941 observations at minimum. The drop is a consequence of the lagged debt variables and the missing values in the dataset.

(27)

Looking at the link between household debt and the different measures of macroeconomic volatility, little empirical evidence is found. From the first two rows, it can be noted that the coefficients of L1_Household debt and L1_Household debt2 remain insignificant in all different models, indicating that there is an absence of a positive and statistically significant causality between household debt and macroeconomic volatility. Hence, hypotheses 1A and 1B cannot be confirmed, thereby undermining the arguments made previously about how high household debt levels increase the sensitivity of household to shocks. A driver of the insignificant results may be that the accumulation of debt helps households to smooth through shocks, as predicted by the permanent income hypotheses. This accumulation effect may offset and conceal the adverse effects of the consequent high household indebtedness in the regression.

4.1.2 Results hypotheses 2A and 2B

The estimation results for the models with corporate debt are depicted in Table 3. Each model is highly significant with, again, moderate R-squared values. Just like in Table 2, the only Table 2: Regressions of measures of macroeconomic volatility on household debt

(1) (2) (3) (4) (5) (6) L1_Household debt .0059 (.0082) .0097 (.0167) .0010 (.0168) -.0200 (.0545) -.0014 (.0092) .0137 (.0201) L1_Household debt2 -.00003 (.0001) .0002 (.0004) -.0001 (.0001) Government spending .0206 (.0139) .0218* (.0129) .0327 (.0473) .0260 (.0561) -.0253 (.0180) -.0205 (.0185) Inflation .0299 (.0243) .0317 (.0244) -.0359 (.0653) -.0458 (.0679) -.0037 (.0361) .0034 (.0381) Trade Openness .0053 (.0042) .0051 (.0042) .0007 (.0155) .0014 (.0159) .0080 (.0074) .0075 (.0073) Exchange rate volatility .0673

(.1137) .0689 (.1134) .3677 (.3130) .3592 (.3102) .0671 (.0997) .0732 (.0991) GDP per capita -.0626 (.0392) -.0611 (.0376) -.0131 (.0813) -.0212 (.0892) .0281 (.0400) .0340 (.0384) Consumption growth -.0262 (.0250) -.0258 (.0250) -.0915 (.0618) -.0934 (.0631) -.0453 (.0297) -.0439 (.0298) Investment growth -.0103 (.0116) -.0101 (.0117) -.0423* (.0222) -.0434** (.0212) -.0046 (.0092) -.0039 (.0091) GDP per capita growth -.0300

(.0309) -.0292 (.0303) .0615 (.0649) .0576 (.0662) -.0229 (.0278) -.0200 (.0276) Polity index .0920* (.0532) .0924 (.0532) .0495 (.0862) .0475 (.0847) .0877* (.0507) .0892* (.0512) Crisis dummy .2598 (.2120) .2622* (.2144) 1.4489*** (.4293) 1.4359*** (.4383) .8120*** (.2196) .8214*** (.2172) F-statistic 123.52*** 139.16*** 28.62*** 35.69*** 75.18*** 73.29*** R-squared .2230 .2232 .2556 .2568 .3517 .3542

(28)

significant control variables are Investment growth, which is found to have a negative effect on investment volatility, and Crisis dummy, which is found to have a positive effect on investment and GDP per capita volatility.

Hypothesis 2A predicts that higher corporate debts levels lead to higher macroeconomic volatility. Models (7), (9) and (11) test this hypothesis by examining the linear effects of corporate debt on respectively consumption, investment and GDP per capita volatility. As shown in the first row, the coefficient of the L1_Corporate debt variable is positive as expected and significant at the 5% level in model (7), but insignificant in models (9) and (11). Thus, hypothesis 2A can only partially be confirmed: higher values of corporate debt are found to be associated with higher values of consumption volatility but not with higher values of investment and GDP per capita volatility. The magnitude of the coefficient of the corporate debt variable in model (7) indicates that for each unit (= percentage point) increase in the corporate-debt-to-Table 3: Regressions of measures of macroeconomic volatility on corporate debt

(7) (8) (9) (10) (11) (12) L1_Corporate debt .0107*** (.0039) .0315 (.0188) .0103 (.0111) -.0446 (.0394 .0027 (.0046) .0099 (.0192) L1_Corporate debt2 _-.0001 (.0001) .0003 (.0002) -.00004 (.0001) Government spending .0240 (.0154) .0296* (.0163) .0255 (.0484) .0108 (.0525) -.0270 (.0191) -.0250 (.0209) Inflation .0280 (.0254) .0317 (.0262) -.0317 (.0146) -.0415 (.0657) -.0025 (.0367) -.0012 (.0384) Trade Openness .0031 (.0041) .0044 (.0049) -.0018 (.0146) -.0051 (.0147) .0075 (.0072) .0080 (.0071) Exchange rate volatility .0857

(.1233) .0833 (.1211) .3883 (.3104) .3946 (.3314) .0728 (.1018) .0720 (.0993) GDP per capita -.0683* (.0376) -.0607 (.0382) -.0258 (.0781) -.0458 (.0810) .0242 (.0421) .0268 (.0406) Consumption growth -.0332 (.0239) -.0322 (.0225) -.0977 (.0620) -.1002 (.0609) -.0457 (.0288) -.0454 (.0285) Investment growth -.0114 (.0111) -.0104 (.0111) -.0480 (.0211) -.0507** (.0205) -.0045 (.0097) -.0042 (.0097) GDP per capita growth -.0183

(.0820) -.0167 (.0269) .0731 (.0639) .0689 (.0639) -.0204 (.0270) -.0199 (.0269) Polity index .0820 (.0586) .0786 (.0641) .0416 (.0928) .0507 (.0973) .0852 (.0525) .0840 (.0551) Crisis dummy .1306 (.1612) .0794 (.1630) 1.2723 (.4841) 1.4069*** (.4703) .7575*** (.2148) .7399*** (.2300) F-statistic 583.29*** 270.44*** 84.17*** 82.16*** 84.22*** 80.13*** R-squared .2381 .2437 .2613 .2689 .3517 .3523

(29)

GDP ratio, consumption volatility increases by 0.01. Although the economic significance may appear small, this is not necessarily the case when considering the small mean of consumption volatility (2.17).

Moving to hypothesis 2B, which predicts that the macroeconomic volatility-worsening effect of rising corporate debt becomes greater as corporate debt accumulates, no significant effect can be found. The first two rows of models (8), (10) and (12) are all insignificant, suggesting that there is no non-linear relationship between corporate debt and macroeconomic volatility. Because the coefficients of L1_Corporate debt2 remain insignificant in all models, this paper does not find evidence for hypothesis 2B to be supported. This contrasts with the theoretical arguments of existing literature that the impact of additional debt is likely to be most drastic for corporates that experience debt overhang problems, which suggested that the impact of additional corporate debt on macroeconomic volatility would be stronger when debt levels are already high.

Hypotheses 3A and 3B are tested in a similar way as the previous hypotheses. Therefore, it follows that the models are again highly significant with only a modest percentage of the variation in the dependent variables that can be explained by the models. Several control variables are statistically significant: government spending is found to have a positive effect on consumption volatility, investment growth is found to have a negative effect on investment volatility, GDP per capita growth is found to have a negative effect on GDP per capita volatility, a shift towards a more democratic political environment is found to have a positive effect on GDP per capita volatility, and crises are found to have a positive effect on all types of volatility. These results confirm and complement those shown in Table 2 and 3.

(30)

debt and consumption and GDP per capita volatility is found. In model (14), the linear term is negative while the squared term is positive, suggesting that government debt is associated with lower consumption volatility at low debt levels, but with higher consumption volatility at high debt levels. More precisely, the predicted turning point where the effect of government debt on consumption volatility becomes negative is a debt-to-GDP ratio of 65.57%. Thus, although the coefficient of the squared term seems small, it does provide a realistic turning point, calculated by dividing the unrounded coefficient on the linear term over twice the absolute value of the unrounded coefficient on the squared term. In model (18), while the linear component of how GDP per capita volatility changes as government debt changes is insignificant, the significant, positive squared term predicts a U-shaped relation.

Although government debt is associated with lower consumption and GDP per capita volatility at low debt levels, thereby contradicting hypothesis 3B, the expected direction of changes in the slope as the value of government debt changes (the slope becoming more positive as Table 4: Regressions of measures of macroeconomic volatility on government debt

(13) (14) (15) (16) (17) (18) L1_Government debt .0040 (.0051) -.0440*** (.0003) .0408*** (.0094) .0140 (.0266) .0209*** (.0044) -.0148 (.0118) L1_Government debt2 _.0003*** (.0001) .0002 (.0002) .0002*** (.0001) Government spending .0342* (.0200) .0529** (.0199) .0051 (.0510) .0155 (.0497) -.0224 (.0204) -.0085 (.0208) Inflation .0018 (.0292) -.0011 (.0276) -.0463 (.0664) -.0479 (.0668) -.0104 (.0349) -.0125 (.0351) Trade Openness .0122 (.0091) .0143 (.0075) -.0136 (.0181) -.0125 (.0176) .0083 (.0119) .0099 (.0106) Exchange rate volatility .0910

(.1561) .0570 (.1511) .6658* (.3352) .6468* (.3528) .1663 (.1103) .1410 (.1154) GDP per capita -.0631 (.0396) -.0301 (.0322) -.0186 (.0895) -.0002 (.0923) .0079 (.0399) .0324 (.0344) Consumption growth -.0196 (.0361) -.0152 (.0343) -.0062 (.0550) -.0038 (.0538) .0140 (.0315) .0172 (.0278) Investment growth -.0145 (.0124) -.0136 (.0123) -.0491** (.0192) -.0487** (.0195) -.0155 (.0100) -.0148 (.0101) GDP per capita growth -.0480

(.0292) -.0507* (.0278) -.0201 (.0638) -.0216 (.0636) -.0816** (.0363) -.0836** (.0345) Polity index .0659 (.0690) .0719 (.0582) .0383 (.0722) .0417 (.0673) .0913** (.0391) .0958** (.0369) Crisis dummy .4384* (.2297) .4650** (.2123) 1.528*** (.4311) 1.5424*** (.4297) .9034*** (.2247) .9232*** (.2151) F-statistic 36.38*** 594.72*** 131.59*** 166.07*** 116.37*** 54.02*** R-squared .2400 .2981 .3117 .3158 .3957 .4218

(31)

government debt increases) is in line with hypothesis 3B. The finding that at low levels of government debt, changes in the level of debt are associated with lower subsequent consumption and GDP per capita volatility, may be explained by debt being issued to engage in countercyclical policies. Since changes in debt resulting from expansionary fiscal measures always reduce volatility, this may conceal the adverse effects of the consequent high government debt in the regression. However, at higher levels of debt, volatility rises, suggesting that high government debt levels do constrain governments’ ability to stabilise the economy.

The estimation results for the equations with total non-financial debt are depicted in Table 5. First, it should be noted that most models in this paper that include total non-financial debt lack an overall model F-statistic because those models have a small number of clusters, but large number of (time) variables, resulting in exhausted degrees of freedom and a failure of the F-test. Illustrative, when not using time-fixed effects or cluster-robust standard errors, the models turn highly significant. Nonetheless, the failure of the overall model test does not invalidate the variables' coefficients or standard errors. Besides, if it were defined, it would test the null hypothesis that all model coefficients are zero. Given that most of the model coefficients are time variables, the hypothesis of all model coefficients being zero is of little interest anyway. Before elaborating on the impact of total non-financial debt on the different measures of macroeconomic volatility, it is interesting to note that, again, several alternative determinants of volatility are found. The results for the control variables are consistent with those found in the previous models. However, contrary to the previous models, now we do find a significant (positive) association between exchange rate volatility and investment volatility and a significant (negative) association between the level of GDP per capita and consumption volatility.

Hypothesis 4A predicts that higher total non-financial debts levels lead to higher macroeconomic volatility. Models (19), (21) and (23) test this hypothesis by examining the linear effects of total non-financial debt on respectively consumption, investment and GDP per capita volatility. As shown in the first row, the coefficient of the total non-financial debt variable is positive as expected and significant at the 5% level in each of the three models. Thus, hypothesis 4A can be confirmed: higher values of total non-financial debt are found to be associated with higher values of macroeconomic volatility. The impact of total non-financial debt is of a similar size as that of corporate debt on consumption volatility and government debt on investment and GDP per capita volatility. For each percentage point increase in the total non-financial debt-to-GDP ratio, consumption, investment and GDP per capita volatility are expected to increase by respectively 0.01, 0.03 and 0.01.

IMPLICATIONS OF RISING DEBT LEVELS FOR MACROECONOMIC VOLATILITY