European macroeconomic fundamentals and their impact on crisis severity : To what extent are pre-crisis macroeconomic variables systematically related to the severity of the 2008/2009-credit crisis in European countries

(1)

European macroeconomic fundamentals and their

impact on crisis severity

To what extent are pre-crisis macroeconomic variables systematically related to the severity of the 2008/2009-credit crisis in European countries?

Bachelor Thesis by Katharina Doesburg Faculty of Economics and Business

Economics and Finance Supervised by C. van den Kwaak

(2)

1 Introduction 3

2 Theoretical Framework 4

2.1 Cross-country causes of crisis severity 4

2.2 The crisis in Europe 9

3 Methodology 12

3.1 Data description 12

3.1.1 Crisis measures 12

3.1.2 Pre-crisis factors 15

3.2 Regression model and analysis 17

3.2.1 Testing for serial correlation and heteroskedasticity 19

3.2.2 Model specification variations 20

3.2.3 Robustness checks 21

4 Analysis 22

4.1 Descriptive statistics 22

4.2 Testing for serial correlation and heteroskedasticity 23

4.3 Regression results 25

4.4 Robustness checks 29

5 Conclusion 33

Bibliography 35

Appendices 37

Appendix 1 – Country list 37

Appendix 2 – Data sources 37

Appendix 3 – Comparison of estimation techniques 39 Appendix 4 – Robustness checks: model variations 40 Appendix 5 – Robustness checks: sample variations 41

(3)

1 Introduction

The 2008/2009-credit crisis and the turmoil following have generated renewed interest in early warning indicators that could help predict a country's vulnerability to future crises (Frankel & Saravelos, 2012). As a result, a variety of literature has emerged analyzing the fundamentals and transmission mechanisms that have lead to the most severe financial crisis since the Great Depression (Brunnermeier, 2009). Authors of these studies are not only trying to understand the underlying transmission of the crisis, but are also aiming to find early warning indicators. These indicators could potentially show which countries will be most severely affected by crises in the future (Frankel & Saravelos, 2012). Due to the advanced economies nature of this crisis (Lane & Milesi-Ferretti, 2011), Europe was hit especially hard and crisis explanations for this phenomenon come in many forms. But most cross-country studies focus on a global setting, pooling countries into one group, hardly differentiating between developed and emerging economies (see for example Rose & Spiegel, 2012; Claessens, Dell'Ariccia, Igan, & Laeven, 2010; Berkmen, Gelos,

Rennhack, & Walsh, 2012).

This paper therefore adds to the existing literature by focusing on the crisis in Europe and answering the following research question: “To what extent are pre-crisis macroeconomic variables, such as government debt, private credit and house price levels, systematically related to the severity of the 2008/2009 credit crisis in European

countries?” The existence of a systematic relationship would confirm the validity of crisis explanations so far identified, deepening our understanding of the transmission

mechanisms in Europe. It would further serve as an indication of country vulnerability. Additionally, the European Union is an interesting case, due to its deep financial and economic integration.

This paper is organized as follows: Section 2 reviews the existing literature on early warning indicators with different crisis measures, including global as well as Europe focused studies. Section 3 describes the methodology of the static panel data analysis, regressing pre-crisis fundamentals on different measures of crisis severity. Section 4 analyses the data and presents the results of the regression. It further describes the robustness checks of the results. Section 5 concludes.

(4)

2 Theoretical Framework

This section first reviews a number of global cross-country studies analyzing the relationship between pre-crisis variables and crisis severity. Different macroeconomic fundamentals and different crisis measures have been researched with emphasis on trade and financial transmission channels. Next, this section turns to studies focusing on crisis severity in Europe.

2.1 Cross-country causes of crisis severity

Since the 08/09-credit crisis a variety of literature has emerged that evaluates crisis causes across different countries. Among the first to conduct a technical analysis were Rose and Spiegel (2010, 2011 and 2012) with a series of global cross-country studies of early warning signs. They investigate a total of 107 countries and more than 60

explanatory variables, but find no significant factors, neither when including nor when excluding contagion effects through US exposure (Rose & Spiegel 2010, 2012). Rose and Spiegel's (2012) initial focus is on national characteristics, such as country size, income, financial and fiscal policies, asset price appreciation, imbalances and institutions. They regress these explanatory variables on a combined measure of crisis severity. In their first analysis conducted in 2009 but published in 2012, Rose and Spiegel (2012) use a

multiple indicators multiple causes (MIMIC) model with real GDP, stock market capitalization, credit rating and the exchange rate as measurable indicators of crisis severity. They do not find any significant factors and conclude that this is due to poor measures of crisis impact. They argue that annual data up and including 2008 do not capture all crisis effects. Additionally, Rose and Spiegel (2012) concede that they do not consider contagion effects from exposure to the US, which might be another reason for their lack of findings. In a second study they therefore add measures of exposure, through financial and real channels as explanatory variables (Rose & Spiegel, 2010). But they still do not find significant results. Specifically, they investigate countries' exposure to US assets, foreign bank exposure and exports to the US. In another update, Rose and Spiegel (2011) address the concern of measuring crisis impact. But they still do not find any significant effects. In this third update, they extend the timeframe of their analysis to

(5)

include 2009 data. They also change their estimation procedure to Ordinary Least Square (OLS) estimations, as is done in similar studies. They further review their country sample and add a new measure of crisis severity, but they only corroborate their initial findings (Rose & Spiegel, 2011).

Similarly, the cross-country study conducted by Claessens, Dell'Ariccia, Igan and Laeven (2010) does not identify any significant factors either. Their approach focuses on the depth as well as the duration of the crisis, by categorizing countries into groups based on different transmission channels. The transmission channels identified are severe national imbalances, financial linkages to the US and trade linkages in general. The transmission is reflected in the timing of crisis impact, and groups are formed based on the country’s individual crisis start. As Claessens et al. (2010) see country pooling as one of the main causes of poor results in previous studies; they use this grouping approach to avoid any masking of empirical regularities. In their case, crisis severity is measured as crisis duration, income loss and change in income growth. Furthermore, a number of pre-crisis conditions such as asset price increases, the credit boom, leverage increases, lending standards and the exchange rate stability are identified to be typical

vulnerabilities leading to financial crises. As new factors relevant for the 08/09 crisis especially, Claessens et al. (2010) mention the use of sophisticated financial

intermediaries and instruments as well as the increased interconnectedness of financial markets. A special focus is put on the crisis origin from household indebtedness and the feedback loop through consumer spending. In this feedback loop, lower house values reduce consumption, which in turn reduces demand and thus leads to layoffs and unemployment, further decreasing consumption. However, despite avoiding country pooling and focusing on the increased interconnectedness of financial markets, Claessens et al. (2010) do not find any significant factors. They conclude that the crisis was a systematic and global shock not to be explained by pre-crisis macroeconomic variables.

But not all studies are unsuccessful in identifying significant factors. Lane and Milesi-Ferretti (2011) find several significant effects on output growth and changes in output growth during the crisis in their global cross-country study. By ranking the countries according to their crisis measure, they establish that the average growth rate identifies industrialized countries as more severely affected. Meanwhile, the change in

(6)

output growth is larger for emerging countries (Lane & Milesi-Ferretti, 2011). This is explained by larger growth rates experienced by emerging economies before the crisis. In their independent variables, Lane and Milesi-Ferretti (2011) focus on transmission

through US sub-prime mortgage securities, increases in risk-aversion, banking exposure and trade exposure. Using credit expansion, exposure to world trade and the relative size of the manufacturing sector as regressors, they capture the effects of the decline in trade and capital flows that resulted from the crisis. Additionally, Lane and Milesi-Ferretti (2011) include current account deficits and currency liabilities, as well as domestic credit and real estate prices to capture the increase in risk aversion. An exchange rate dummy for pegged exchange rates is included for trade and capital flow effects. Lane and Milesi-Ferretti (2011) find that pre-crisis per capita income, private credit growth and current account deficits have a large effect, confirming the advanced-economy nature of the crisis. They also find limited evidence that the effect of a pegged exchange rate is significant for crisis severity.

A study by Berkmen, Gelos, Rennhack and Walsh (2012) confirms some of Rose and Milesi-Ferretti's findings. Berkmen et al. (2012) focus on the characteristics of emerging and developed markets using 43 countries in their sample. Their dependent variable is the difference between actual growth and 2009 growth forecasts. They find that the financial transmission channel is most important overall, but especially

significant for emerging markets. This is confirmed by the effect of leverage, measured as the credit-to-deposit ratio and short-term debt. However, when focusing on emerging markets specifically, the trade channel also becomes relevant. Furthermore, they identify a currency mismatch in the European Union corroborating their conclusion on the significance of the financial channel. Additionally, Berkmen et al. (2012) find exchange rates, primary fiscal gaps, current account imbalances, and public debt, though this one with a counterintuitive sign, to be relevant factors. In their approach, the explanatory variables are grouped according to the transmission mechanisms: trade, financial, institutional and other. Next, two variables per group are selected for the multivariate regression based on the highest correlation with the dependent variables. Using this method, Berkmen et al. (2012) work with relatively few variables, but are able to explain two third of the differences in crisis severity across countries.

(7)

Frankel and Saravelos (2012) wrote an extensive literature review on early warning factors including, among others, the above-mentioned studies. One of their main

criticisms is that variables are selected in hindsight and only in regard to the most recent crisis. Hence these variables fit the crisis discussed, but do not necessarily have much value as predictors for future potential crises. Additionally, they find a wide variety of crisis definitions and models, of which the linear OLS regression model is the most popular. Regarding the above-mentioned studies, a common shortcoming of the different approaches is the limitation in crisis measures. Rose and Spiegel (2012) and Lane and Milesi-Ferretti (2011) both use annual data, leading to an imprecise definition of the crisis' timeframe. Meanwhile, Berkmen et al.'s (2012) crisis measure does not consider actual performance. Based on their review, Frankel and Saravelos (2012) identify the most commonly used variables and methodologies of financial crises studies since 1998, on which they build their own formal analysis. They confirm the economic significance of international reserves and the real exchange rate. Additionally, they find that pre-crisis credit growth, current account imbalances, saving rates and external and short-term debt are relevant factors in the 08/09 crisis, although their significance is not robust to all model specifications. Frankel and Saravelos (2012) add three innovations to the existing literature: First they use five instead of two or three crisis measures. They include currency markets, recourse to the International Monetary Fund and industrial production in addition to the common measures of GDP and equity markets. Second Frankel and Saravelos (2012) use leading indicators from the existing pool of significant factors, not limiting themselves to 08/09 crisis factors. Third, they consider a longer timeframe including the second quarter of 2009. Frankel and Saravelos (2012) then conduct bivariate regressions of more than 300 equations to explore among others the following variables popular in previous studies: reserves, exchange rate changes, GDP, credit, current account, equity returns, interest rates, debt composition, capital flows, external debt and income. Based on these explorations they develop their multivariate regression for an exchange market pressure index, combining crisis effects on the exchange rate with those on international reserves.

Feldkircher (2014) confirms the relevance of the trade channel and pre-crisis credit growth. He adds three additional elements in his study of cross-country vulnerabilities

(8)

and crisis transmission channels. First, Feldkircher (2014) selects crisis measures controlled for timing. Second he adds two additional measures to capture long-run effects. Third, he uses a rich data set containing over 90 potential variables. Furthermore, the Bayesian averaging technique he employs is robust to model uncertainty and allows for cross-country spillovers. In his approach Feldkircher (2014) defines the crisis period individually for each country. He further uses the cumulative loss in real output and output depth as crisis measures in addition to GDP growth. As regressors he uses averages of 2000-2006 data in the following categories: macroeconomic risk factors, external risk factors, fiscal risk, financial risk and contagion or spillover effects. He finds that the change in domestic credit is a robust factor. But in contrast to other studies, current account imbalances, fiscal soundness and foreign reserves are not significant.

Caprio, D'Apice, Ferri and Puopolo (2014) use the probability of crisis occurrence as their dependent variable in a model analyzing financial and banking indicators. They conclude that traditional banking practices result in a lower probability of being hit by the financial crisis, while capital adequacy turns out to be mostly irrelevant. Caprio et al. (2014) analyze the regressors' effect on a macro, meaning country level, and on a micro, meaning banking level, in 83 countries. The explanatory factors are categorized as financial determinants and banking indicators, including, among others, bank business model variations, funding strategy, market structure, efficiency, stability, financial integration, and regulation. All variables are calculated as 1998 – 2006 averages in order to capture their long-term evolution. Caprio et al. (2014) find that net interest margins, credit-to-deposit ratios, banking sector concentration and policy on banking restriction are statistically significant on a macro level. They confirm this finding in their micro level analysis. The inclusion of net interest margins and the credit-to-deposit ratio captures the effect of traditional banking practices. Caprio et al. (2014) explain that higher net interest margins imply an emphasis on traditional banking, reducing the use of financial innovations such as securitization. Contrarily, high credit-to-deposit ratios suggest unstable, non-deposit funding, hence indicating the extensive use of financial innovations. As both factors are significant, they ascertain that the expansion of such financial innovations has led to higher crisis probabilities and thus increased countries' vulnerability to financial crises.

(9)

2.2 The crisis in Europe

Only few studies similar to the ones described above have been conducted with a focus on crisis effects in Europe. However, Groot, Möhlmann, Garretsen and de Groot (2011) explore financial factors, trade linkages, differences in institutions and sectoral

composition as factors explaining the heterogeneous impact of the crisis in Europe. They find that sectoral composition is the most relevant. Their analysis is based on descriptive statistics using GDP and unemployment changes as crisis measures. Groot et al. (2011) check a variety of variables in each category for their correlations with the dependent variables. They include foreign assets and liabilities, a measure of housing overvaluation, governmental support to the banking sector in form of asset purchases, injections and lending guarantees, and the countries' net financial position to capture the financial channel. Net exports, current account balances, the share of exports to the US or Asia and the change in the ratio of GDP to labor costs are included for the financial channel. Meanwhile, the institutional channel is represented in government deficits and debt, tax rates and employment measures, as well as union membership. Groot et al. (2014) find little correlation with financial indicators and institutional measures. But they identify the ratio of GDP to pre-crisis labor costs as the best indicator in sectoral composition. This result is confirmed with a panel data regression of total GDP growth on the growth of value added in specific sectors from 1980 until 2003. The coefficient can be interpreted as a sector's sensitivity to the EU-wide business cycle. To further corroborate this finding, Groot et al. (2014) repeat the regression on a regional level, increasing the number of observations used and therefore the statistical significance. Their approach differs from the one applied in this paper, because in the analysis presented in section 4, the panel regression is conducted using more than one of the relevant factors identified in the literature. As a result, it is not the one-on-one relationship between crisis severity and pre-crisis factors that is investigated, as Groot et al. (2014) do by using mostly descriptive statistics. But their joint effect is also taken into consideration.

In his theoretical review of pre-crisis factors, Lane (2012) describes an accumulation of vulnerabilities in Europe as the cause for the 08/09-credit crisis. He identifies a number of relevant factors: Public debt captures the European heterogeneity that exists despite the restrictions in the Stability and Growth Pact. Contrasting with the

(10)

indebtedness of countries are the generally low spreads on sovereign debt. These fueled an increase in dispersion and persistence of current account deficits that was only later perceived as risky, because it created large exposures to capital markets. A domestic credit boom added to household indebtedness and was accompanied by an increase in securitization, making financial markets vulnerable. In addition, Lane (2012) identifies these factors as most relevant in the liquidity squeeze, due to their connection to the surge in risk aversion at the offset of the crisis. According to him, the failure to tighten and the decrease in countercyclical fiscal policy have led to an increasingly vulnerable Europe. However, Lane (2012) focuses on the resulting sovereign debt crisis rather than on the initial credit crisis of 08/09. At that point in time, during the initial impact of the crisis, attention was on the banking sector. In his description of events, Lane (2012) identifies international exposure, short-term financing and a large housing sector as important transmission mechanisms from the credit to the sovereign debt crisis in Europe. Attention shifted to sovereign debt, when estimates of bank losses increased, revealing fiscal risk, and at the same time tax revenues of countries with large construction sectors decreased.

In another paper, Lane and Pels (2012) focus on pre-crisis current account imbalances as a crisis origin. They analyze the convergence mechanism of growth expectations for the current account. This mechanism is found to have a strong effect in the Euro Area (EA), because in the European Union (EU) capital flows from richer to poorer regions. This is amplified in the EA due to the common currency. The described capital flows have lead to higher growth expectations in the future for poorer regions and as a result to higher borrowing and investment today (Lane & Pels, 2012). Consequently, these countries have accumulated larger current account deficits, leading to larger

discrepancies in current account balances among European countries. The capital flows have further amplified the effect of distortions and cyclical shocks. Therefore, Lane and Pels (2012) conclude that the run-up of current account deficits has played an important role for the crisis in Europe. They expect that regulation will continue to address this issue in the future, in combination to the challenge of the common currency in the EA.

Shambaugh (2012) in turn argues that Europe actually experienced three interlinked, but individual, crises. He identifies a banking crisis, a sovereign debt crisis and a macroeconomic growth crisis that are linked through a number of channels. First,

(11)

the banking crisis resulted in liquidity strains and bank solvency problems, which spread to sovereigns as they try to support defaulting banks. Defaulting sovereigns in turn contribute to the banking crisis, as banks are large holders of government debt

(Shambaugh, 2012). Second, the slowed economic growth all over Europe reduced banks' as well as sovereigns' profitability, amplifying the problem. Third, Shambaugh (2012) argues that the introduction of the Euro in large parts of the EU has lead to cross-border contagion. According to him, the fact that currency revaluations are no longer possible, because individual currencies have been abolished, effectively reduces countries’ insulation from distressed neighbors. Furthermore, Shambaugh (2012) describes that governments’ ability to respond to crises has been limited by the elimination of monetary policy as a tool on a national level. The monetary policy conducted by the European Central Bank takes the aggregate situation of the EA into consideration and not the individual countries’. In case of geographically asymmetric shocks, which occur when countries are affected oppositely by a crisis, the unified monetary policy is thus not equally effective in all parts of the EA. According to Shambaugh (2012), this situation has amplified the discrepancies visible in the short recovery in 2009 and as a result, the crisis has had a heterogeneous impact across Europe. In addition, Shambaugh (2012) describes how the common currency has lead to a high level of financial integration and a large banking sector. In Europe, banks with a high degree of global orientation are often larger than their host countries, while supervision and regulation is still largely occurs on a national level. Shambaugh (2012) concludes that this and the fact the firms are more reliant on bank financing in Europe than in other countries, has added to cross-border contagion effects.

In conclusion, a variety of literature investigates the effect of economic variables on different crisis measures. Credit growth, current account balances and debt are found to explain crisis severity in some of the discussed studies. However, the focus has been on global cross-country analyses, pooling a large group of heterogeneous countries together. This paper however focuses on the effects of the crisis in Europe and investigates a smaller group of industrialized countries. Europe is especially interesting due to the sub group that exists in form of Euro Area.

(12)

3 Methodology

This section describes the methodology of the statistical analysis conducted. It starts with the data selection based on existing literature by first describing the three dependent variables that measure crisis intensity. The section continues with the independent variables categorized into international exposure, national account imbalances and financial channels. Next, the regression equation is presented, along with the appropriate tests for heteroskedasticity and serial correlation. Finally, the procedure of the regression is outlined and the robustness checks are described.

3.1 Data description

The sample consists of the 28 countries in the European Union plus Iceland, Norway and Switzerland. These three are part of the European Economic Area and are therefore closely linked to the EU (Lane & Pels, 2012). Croatia and Bulgaria are excluded from the analysis due to lack of data. The main regression is therefore based on 29 European countries, of which a complete list can be found in Appendix 1.

3.1.1 Crisis measures

Following the approach of existing studies, three different measures of crisis severity are investigated. The first one is GDP growth, as employed in a large number of studies according to Frankel and Saravelos (2012). In order to accommodate a precise crisis definition, quarterly data are used to calculate the percentage growth over the last period (Frankel & Saravelos, 2012). The start of the crisis is then defined in accordance with Groot et al. (2011) as two quarters of consecutive negative growth. This method identifies the start of the crisis in the first quarter of 2008. As visible in Figure 1 for selected countries, GDP started to decrease at this point in time. However, contrasting to the approach by Groot et al. (2011), the crisis period is not defined individually for each country, but the analysis homogeneously uses ten quarters for all countries. This enables the consistent application of a single timeframe in the regression analysis. In the second quarter of 2010, countries experience positive growth again, showing the slight recovery described by Shambaugh (2012) and visible in Figure 1. Hence, the effects of the

(13)

sovereign debt crisis, which led to a further decrease of GDP around 2011 (Shambaugh, 2012), are excluded from the analysis. Therefore, this research is focused on the effects of the credit crisis, while simultaneously providing a large enough sample of

observations.

Figure 2: Evolution of GDP for selected countries

Source: DataStream The second crisis measure chosen is the level of unemployment, following research such as the one by Groot et al. (2011). Unemployment is a lagged variable and increases in unemployment usually follow decreases in GDP with some delay (Groot et al., 2011). Therefore, following the procedure described above, two quarters of consecutive

increases identify the appropriate crisis timeframe. As a result, the third quarter of 2008 is identified as the starting point of the crisis based on unemployment. Figure 2 clearly shows the increase in unemployment at this point in time for the same selection of

(14)

is also measured over ten quarters, even though some countries already experience earlier decreases in unemployment. Additionally, it is worth mentioning that the short recovery in 2009 described by Shambaugh (2012) is not as evident in unemployment, as shown in Figure 2, as it is in GDP.

Figure 2: Evolution of unemployment for selected countries

Source: DataStream

The third crisis measure is the percentage change in consumption, because it captures consumer sentiment (Lane & Milesi-Ferretti, 2011). To be consistent in the time period chosen, consumption data are gathered for the same timeframe as chosen for GDP growth. However, the start of the crisis is not as clearly identifiable in consumption as it was the case for GDP growth. The two consecutive quarters drop is visible, but the initial impact of the crisis differs significantly per country, and a general starting point is

difficult to pinpoint. Additionally, the slight recovery in 2009 is less present in the consumption data.

(15)

3.1.2 Pre-crisis factors

The literature identifies three relevant transmission channels of crises in general and the 08/09-credit crisis especially: international exposure, national account imbalances and financial channels (Berkmen et al., 2012; Lane & Milesi-Ferretti, 2011). In order to capture pre-crisis levels of the different explanatory variables, data from the third quarter of 2004 until the fourth quarter of 2006 are used. Based on GDP growth, this timeframe shows positive growth and avoids early manifestations of the 08/09-credit crisis as visible in Figure 1. Consequently, this timeframe is applied to all regressions, despite the time lag present in the data of unemployment.

In order to measure international exposure, studies such as the one by Berkmen et al. (2012) include imports and/or exports as explanatory factors (Also see Claessens et al., 2010; Groot et al., 2011; and Lane & Milesi-Ferretti, 2011). The start of the recession greatly decreased export demands (Groot et al., 2011). It is therefore reasonable to

assume that countries with large export sectors were affected more severely (Groot et al., 2011). Consequently, Lane and Milesi-Ferretti (2011) find that their measure of trade openness, namely the ratio of exports and imports to GDP, has a significant effect on output. Berkmen et al. (2012) additionally consider trade composition and trade direction, but only find trade openness to be relevant, using the same measure as Lane and Milesi-Ferretti. The analysis following in section 4 evaluates the ratio of exports plus imports to GDP, as well as net exports as a percentage of GDP as relevant factors.

In addition to trade, the current account balance measures international exposure and its build-up is thought to have contributed to the crisis in Europe (Berkmen et al., 2012; Feldkircher, 2014; Frankel & Saravelos, 2012; Lane & Pels, 2012). Large current account imbalances are visible within the EU due to financial liberalization, whereas the EU as a whole towards the rest of the world is nearly balanced (Lane & Pels, 2012). Current account balances measure the net capital flow of a country in addition to its trade flows (Mankiw & Taylor, 2008). Capital flows were an important factor during the crisis, when they suddenly reversed (Lane & Milesi-Ferretti, 2011). It is therefore reasonable to assume that countries depending on large capital inflows were hit more severely by the crisis. In the analysis, the current account balance is evaluated as a percentage of GDP (Frankel & Saravelos, 2012). However, the current account balance and net exports are

(16)

closely related (Mankiw & Taylor, 2008), a fact which is explored in the alternative model specifications described in subsection 3.2. Including both factors in the regression would thus lead to biased estimations due to multicollinearity (Stock and Watson, 2012). Therefore, the current account is considered as an alternative measure of trade.

Other factors found relevant in previously mentioned studies are national account imbalances, such as the build up in government debt resulting from repeated government deficits (Berkmen et al., 2012; Feldkircher, 2014; Lane 2012). As described by

Shambaugh (2012), the credit crisis in Europe was followed in short succession by a sovereign debt and a macroeconomic growth crisis. The three crises are closely connected through governmental imbalances, although the imbalances are thought to have contributed more severely to the sovereign debt crisis than to the banking crisis (Shambaugh, 2012). Nevertheless, government debt and government deficits both as percentages of GDP are considered as explanatory variables.

In addition, low interest rates before the crisis have fueled domestic credit build-ups, which are part of the financial transmission channel and are captured in bank lending to the private sector (Feldkircher, 2014; Lane & Milesi-Ferretti, 2011). This measure includes loans, purchases of non-equity securities and trade credit. It therefore represents debt of private enterprises as well as households. Consequently, household debt, though potentially relevant due to the expansion in mortgages and homeownership, is not included as a separate variable in the analysis. Bank lending to the private sector is an appropriate measure for credit build-ups in general and it is included as a percentage of GDP. Furthermore, national savings may have played a role in crisis initiation. But savings are implicitly incorporated due to their role in net capital flows as measured by net exports. This connection is made through the national accounts identity (Mankiw & Taylor, 2008).

Another cause of the 08/09-credit crisis may have been a rise in asset prices, more specifically pre-crisis developments in house prices and in the stock market. House prices have strongly increased before the crisis, and the resulting bubble is one of the main focus points in the literature (Claessens et al., 2010; Groot et al., 2011; Rose & Spiegel, 2012). Their surge has fueled a construction boom that has led to a more severe crisis in countries, where the government relies on taxes paid by the construction sector, such as

(17)

Ireland (Lane, 2012). Therefore the house price index and the price index of construction work in housing may both be used to capture this effect. However, the construction price index measures the costs of input materials in construction and is therefore only remotely related to house prices. As data on housing prices are sufficiently available, the

construction price index is excluded from the analysis. The level of house prices in form of an index is included as an explanatory variable. Furthermore, to evaluate asset prices more generally, countries’ stock market capitalization as a percentage of GDP is added to the regression. This measure represents the described surge in asset prices and captures investors’ sentiment. Investors’ sentiment may be a relevant factor due to the sudden increase in risk aversion during the crisis (Lane & Milesi-Ferretti, 2011).

Finally, as the crisis in Europe started in the banking sector (Shambaugh, 2012), a measure of financial sector concentration is included. This paper follows Caprio et al. (2014) in using the ratio of the asset value of the three largest banks to the total asset value of the financial sector as calculated by Beck, Demirgüc-Kunt and Levine (2000). As this measure is only available as an annual measure, the gaps are filled with the help of linear interpolation. Furthermore, the level of foreign reserves is a common factor found in cross-country studies capturing the financial transmission channel (Frankel & Saravelos, 2012; Berkmen et al., 2012). However, as these studies look at global country samples, they differ significantly from this analysis. In Europe foreign reserves play a reduced role for most countries since the introduction of the Euro and are therefore excluded (Lane, 2012). A detailed overview of data sources can be found in Appendix 2 and this paper now turns to the model specifications of the regression.

3.2 Regression model and analysis

This section first presents the regression model with the initial specification of variables. Next, the Wooldridge test for serial correlation and the modified Wald test for

i,t-14

+µ

i,t

(EU1)

Where Y is quarterly GDP growth in the main analysis and unemployment or consumption growth in the robustness checks. The index "i" represents the country and "t" specifies the time period. The error term is denoted by µi,t, which contains the

unobserved country fixed effects ui and the observation specific error terms ei,t, such that

µ

i,t = ui +ei,t .

The explanatory variables in the basis specification (EU1) are the following pre-crisis factors:

Trade: country's reliance on trade as the ratio of net exports to GDP

Gov: national account imbalances measured as government deficits as percentage of GDP Private: private debt measured as bank lending to

the private sector as a percentage of GDP Market: asset bubble build-up measured as the

ratio of market capitalization to GDP House: house price bubble measured as the level

of housing price index

Fin: financial sector concentration measured as the ratio of the asset value of the three largest banks to the total asset value of the financial sector

(19)

At first equation (EU1) is estimated with quarterly GDP growth as the dependent variable on the entire sample. The fixed effects estimator is applied, because it controls for

omitted variable bias caused by the unobserved country-level effects captured in the error term µi,t (Stock & Watson, 2012). However, serial correlation in the idiosyncratic errors

can result in biased standard errors and may cause estimates to be less efficient (Stock& Watson, 2012). Additionally, the fixed effects estimator, based on OLS, assumes that errors are independently and identically distributed (Stock & Watson, 2012). This

assumption may be violated in cross-section panel data as analyzed in this paper, leading to the presence of groupwise heteroskedasticity (Baum, 2001). Therefore, the Wooldridge test for serial correlation and the modified Wald statistic for heteroskedasticity are

calculated to test whether the assumptions for the fixed effects model are satisfied and if clustered (robust) standard errors have to be implemented.

3.2.1 Testing for serial correlation and heteroskedasticity

Based on the estimation of specification (EU1), the model is tested for serial correlation using the Wooldridge test as discussed by Wooldridge (2002). This test requires few assumptions about the nature of country-level effects and applies clustered standard errors, which means that it is robust to conditional heteroskedasticity. Additionally, Drukker (2003) shows that it has good size and power properties, making it applicable for the panel setting discussed in this paper. The Wooldridge test uses the first-difference of the regression equation to estimate the coefficients and to calculate the observation specific error term

Δ

êi,t. By taking the first-differences, the unobserved country effect ui

is eliminated from the equation (Drukker, 2003). The procedure centers on the assumption that in the absence of serial correlation in the residual µi,t, the correlation

ρ(Δet, Δet-1) = -0.5 (Drukker, 2003). In the next step, the estimated residual

Δ

êi,t is

regressed on its lag and the Wald test statistic is calculated to test whether the resulting coefficient indeed equals -0.5. Rejecting the null hypothesis therefore indicates that serial correlation is present in the error term.

Next, the modified Wald test statistic for heteroskedasticity as described by Baum (2001) and based on the theory developed by Greene (2000) is calculated for regression equation (EU1). This statistic can be used to test the null hypothesis that the variance of

(20)

the cross-sectional unit residuals equals the overall residuals’ variance, hence H0: σ2i = σ2

(Baum, 2001), which is true for homoskedastic error terms. The modified Wald statistic is calculated using the estimates of the cross-sectional unit variances and it is Chi Square distributed under the null hypothesis. However, Baum (2001) points out that the test’s power is low for fixed effect estimates in panel data with a small number of observations per cross section. Therefore, the results will need to be viewed with caution. However, jointly, both tests will guide the choice of standard errors in the regression analysis following and the test results are reported in section 4.2.

3.2.2 Model specification variations

After the initial evaluation in regard to standard errors, the model is analyzed for the inclusion of the different variables in the regression. First, (EU1) is estimated as presented above. Next, variations of the model specification are considered in order to evaluate the robustness of the results and to justify the inclusion of specific variables in the regression. In specification (EU2), net exports are replaced with the sum of exports and imports. The sum of exports and imports is a measure used for trade openness in the literature (see for example Lane & Milesi-Ferretti, 2011). Next, in specification (EU3) current account balances are used instead of net exports. Current accounts and net exports are closely related, which would lead to bias in the estimation due to multicollinearity if both were included. However, as current account balances are commonly used in the literature and sometimes found to be significant (see for example Berkmen et al., 2012; Lane & Milesi-Ferretti, 2003) they are considered in this paper as an alternative trade measure. In specification (EU4), government deficits are replaced by government debt, in order to evaluate which one better captures national account imbalances. Finally, in specification (EU5) house prices are dropped. The house price index is not available for the entire sample and its exclusion increases the number of observations for the

(21)

Subsequently, the regression is repeated for the reduced sample of Euro Area (EA) countries and for the reduced sample excluding financial centers using the

i,t-16

+µ

i,t

(EU1b)

This is due to the lag in unemployment and is consistently applied to specifications (EU2) – (EU5). Meanwhile, consumption growth is regressed in the form of equation (EU1). First, specifications (EU1) – (EU5) are repeated to confirm the appropriateness of the variables included. Finally, the regressions are repeated for the EA and excluding financial centers using unemployment and private consumption growth as the dependent variables.

In addition, the Arellano-Bond generalized method of moments (GMM) estimator (Arellano & Bond, 1991) was considered as an alternative estimation technique. This estimator includes the first lag of the dependent variable as a regressor, which is by construction correlated with the unobserved country level effects captured in the error term µi,t. The Arellano-Bond GMM estimator therefore uses first-differences to avoid the

resulting bias and is thus specifically suited to address serial correlation. Using all past information of Yt as additional instruments the Arellano Bond GMM estimator is

unbiased and efficient for dynamic panel data estimations. However, when estimating equation (EU1) and its variations across crisis measures, the only coefficient to be found consistently significant was the one for the lag of the dependent variable. This indicates

(22)

the presence of short-run effects, while the intention of this paper is to focus on the long-run systematic relationship of macroeconomic variables with crisis severity. Additionally, there was the risk of weak instruments in the Arellano-Bond estimator, as the number of instruments was consistently higher than the number of groups included in the regression. It was therefore decided to estimate the model using the fixed effects estimator with robust standard errors instead.

Finally, a balanced panel was considered to corroborate the results. However, sufficient data were only available for a balanced panel of 18 countries. Iceland, Greece and a number of other countries would have been excluded from this analysis. Since Iceland and Greece are among the countries that were hit especially hard by the 08/09 crisis and the number of observations would have been reduced even further, the

regression reported remains based on the unbalanced panel of the 29 countries as describe above. Additionally, the analysis of the balanced panel did not yield results that differed significantly from the outcome of the unbalanced panel. A small comparison of (EU1) estimation results using the Arellano-Bond GMM estimator and the balanced panel with GDP as a crisis measure can be found in Appendix 3.

4 Analysis

This section first analyses the descriptive statistics of all variables included. Next it presents the results of the Wooldridge test and the modified Wald statistic. The regression analysis of specifications (EU1) – (EU5) on GDP growth for the complete sample

follows and is repeated for the EA and for the reduced sample excluding financial centers. This section finishes with robustness checks by repeating the entire analysis for the alternative measures of crisis severity.

4.1 Descriptive statistics

The descriptive statistics presented in Table 1 are generated across time and countries using the timeframes as described above. They are thought to give a general idea of the data used and a few factors are worth pointing out in Table 1:

(23)

First, the average current account is lower than the average net export balance. This indicates that net capital flows must have been negative for most countries. Therefore, as the collapse in trade during the crisis is expected to have affected countries with large pre-crisis export sectors more severely, the reversal of capital flows may have reduced this effect, at least on average.

Second, both market capitalization and bank lending to the private sector show a large standard deviation relative to their averages. Luxembourg, Switzerland, the United Kingdom and Iceland all have market capitalizations that exceed 100 percent in the analyzed time period. Furthermore, they display the largest values for bank lending to the private sector in addition to Austria and Malta. Except for Austria, this is not surprising, as the named countries are identified as financial centers. Their possible role as outliers is explored as such in the regression analysis in the following section.

Table 1: Summary of descriptive statistics

(1) (2) (3) (4) (5)

VARIABLES N Mean SD Min Max

GDP 290 -0.00447 0.0459 -0.250 0.136 Unemployment 290 0.0803 0.0359 0.0170 0.203 Private Consumption 290 0.0000624 0.0595 -0.318 0.204 Net Exports 290 0.00112 0.0962 -0.256 0.331 Imports+ Exports 290 1.082 0.503 0.481 3.113 Current Account 289 -0.00478 0.0238 -0.0772 0.0495 Government Debt 290 0.540 0.531 0.0364 4.286 Government Deficit 290 -0.00607 0.0159 -0.0667 0.0369 Bank Lending to Priv. Sec. 290 0.849 0.782 0.000850 3.195 Market Capitalization 284 0.659 0.496 0.0482 2.653 House Price index 242 0.833 0.143 0.315 1.014 Fin. Sec. Concentration 289 0.765 0.169 0.334 1

Notes: Data for GDP and private consumption span the timeframe Q1 2008 until Q2 2010.

Unemployment data are collected between Q3 2008 and Q4 2010. All other variables are from Q3 2004 until Q4 2007. GDP and private consumptions are quarterly growth rates. Net Exports, Imports+ Exports, Current Account, Government debt and deficit, Bank lending to the private sector are as percentages of GDP. Financial sector concentration is the ratio of the asset value of a countries three largest banks to the asset value of its total financial sector.

(24)

Third, government debt has a large maximum compared to its average. This value is produced by Iceland, where government debt soared before the crisis. Next, data

availability plays a role in the regression specifications used for the analysis. As visible in Table 1, house prices have the least number of observations, with five countries missing data entirely. This leads to regression specification EU5, where house prices are excluded in order to evaluate whether the increase in observations alters the results.

Finally, the large standard deviations for the sum of imports and exports,

government debt, market capitalization and bank lending to the private sector, relative to their means, reflect the heterogeneity of countries. Despite the high degree of economic integration in the EU and the common currency in the EA, countries still differ

significantly in macroeconomic fundamentals and fiscal policy approaches.

4.2 Testing for serial correlation and heteroskedasticity

To start the regression analysis, the Wooldridge test for serial correlation in the first order is calculated for the estimation of equation (EU1). The test is repeated for all three crisis measures and the results are presented in Table 2.

As visible in Table 2, the null hypothesis of no first order autocorrelation is only rejected in the case of unemployment as the dependent variable. This confirms the use of robust standard errors for at least this crisis measure.

Table 2: Wooldridge test

H0: no first order autocorrelation

DEPENDENT VARIABLE F (1, 24) Prob.> F

GDP 2.314 0.1412

Unemployment 35.291 0.0000

Private Consumption 2.672 0.1152

Notes: The test is conducted by regressing net exports (% of GDP), government deficit (% of

GDP), bank lending to the private sector (% of GDP), market capitalization (% of GDP) house price index, and financial sector concentration on quarterly GDP growth, unemployment and private consumption growth respectively. Data for GDP and private consumption spans the timeframe Q1 2008 until Q2 2010. Unemployment data are collected between Q3 2008 and Q4 2010. All other variables are from Q3 2004 until Q4 2007. GDP and private consumptions are quarterly growth rates.

(25)

Next, the modified Wald statistics is calculated to test for heteroskedasticity, which, if present would lead to bias in the standard errors. The test is applied to the regression outcome of model specification (EU1) implementing homoskedastic standard errors. Table 3 presents the test statistics and the appropriate probabilities, clearly showing that group wise heteroskedasticity is present in the regression for all three crisis measures. This confirms the decision to use robust standard errors across the entire analysis as presented in the following section.

Table 3: Modified Wald statistic H0: σ2i = σ2

DEPENDENT VARIABLE χ2 (25) Prob.> χ2

GDP 6072.90 0.0000

Unemployment 686.80 0.0000

Private Consumption 5252.81 0.0000

Notes: The test is conducted after regressing net exports (% of GDP), government deficit (% of

GDP), bank lending to the private sector (% of GDP), market capitalization (% of GDP) house price index, and financial sector concentration on quarterly GDP growth, unemployment and private consumption growth respectively. Fixed effects estimator with homoskedastic standard errors. Data for GDP and private consumption spans the timeframe Q1 2008 until Q2 2010. Unemployment data are collected between Q3 2008 and Q4 2010. All other variables are from Q3 2004 until Q4 2007. GDP and private consumptions are quarterly growth rates.

4.3 Regression results

The variables used in model specification (EU1) were selected based on existing literature. However, the variety of variables considered in previous research, the

possibility of multicollinearity among them and the number of observations available led to the model specifications (EU2) – (EU5). They explore the success of different models in capturing the desired effects and Table 4 presents the estimation output.

Specification (EU2) uses the ratio of the sum of imports and exports to GDP as a measure of trade openness instead of net exports as a percentage of GDP. Similarly, specification (EU3) replaces net exports with the current account balance as a percentage of GDP. However, as visible in Table 4, none of the three pre-crisis trade measures is statistically significant in predicting GDP growth during the crisis.

(26)

Table 4: Regression output using GDP as crisis measure – model variations

(1) (2) (3) (4) (5)

VARIABLES EU1 EU2 EU3 EU4 EU5

Net Exports -0.339 -0.357 -0.270

(0.219) (0.229) (0.183)

Government Deficit -0.672* -0.687* -0.630 -0.688**

(0.354) (0.369) (0.377) (0.297)

Bank Lending to Priv. Sec. 0.0169 0.0385 0.0241 0.0211 0.00330 (0.0383) (0.0495) (0.0404) (0.0387) (0.0352) Market Capitalization -0.00500 0.00579 -0.00214 0.000393 0.0233

(0.0193) (0.0178) (0.0158) (0.0203) (0.0199) House Price index 0.0640 0.0793 0.0749 0.0518

(0.0423) (0.0563) (0.0472) (0.0386)

Fin. Sec. Concentration 0.0128 0.0204 0.0157 0.0257 -0.0234 (0.0213) (0.0260) (0.0237) (0.0245) (0.0226) Imports+ Exports -0.00433 (0.0549) Current Account -0.536 (0.534) Government Debt -0.0173 (0.0108) Constant -0.0853* -0.128** -0.107** -0.0798** -0.00744 (0.0447) (0.0597) (0.0447) (0.0375) (0.0424) Observations 235 235 234 235 283 R-squared 0.125 0.064 0.082 0.097 0.069 Number of countries 25 25 25 25 29

Notes. Dependent variable for all specifications is quarterly GDP growth from Q1 2008 until Q2

2010. Fixed effects estimations with robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 denote significance levels. Net Exports, Imports+ Exports, Current Account, Government debt and deficit, Bank lending to the private sector are as a percentage of GDP. Financial sector concentration is the ratio of the asset value of a countries three largest banks to the asset value of its total financial sector. Independent variables span Q3 2004 until Q4 2007.

Next, in specification (EU4), government deficits as a percentage of GDP are replaced by government debt as a percentage of GDP. As visible in Table 4, the

coefficient of government deficits is significant at the 10 percent level in specifications (EU1) and (EU2). Repeating the regression with both variables included does not alter this significance and therefore government deficits are chosen as the appropriate measure for national imbalances. House prices are excluded from the regression in specification

(27)

(EU5), which increases the number of observations. Nevertheless, the increase does not change the results, except for the fact that the coefficient of government deficits is now significant at the 5 percent level. It does however not increase the explanatory value of the regression, as the R-square measure is smaller for (EU5) than for (EU1).

To sum up, using GDP as a crisis measure and analyzing the entire European country sample, only government deficits stand out as a pre-crisis factor explaining crisis severity. This means that larger pre-crisis deficits have led to a more severe crisis

manifestation in GDP. However, the explanatory power of the model is low as shown in the R-squared (0.125) in Table 4, and the joint hypothesis, testing whether all the

coefficients in the model are different from zero, is only rejected at the 10 percent level (F=2.37, Prob. > F= 0.0608) in the case of specification (EU1).

This paper now turns to the comparison across samples. Regression (EU1) is repeated for the EA and excluding financial centers. The results are shown in Table 5. When

considering the Euro Area none of the coefficients is significant, even though the joint hypothesis of all coefficients being equal to zero is rejected at the 1 percent level (F= 6.42, Prob. > F= 0.0014). This implies that EA countries, despite not being able to use monetary policy as an instrument, do not show any different relationships between pre-crisis macroeconomic fundamentals and pre-crisis severity. However, the ability to employ monetary policy would only have become relevant in dealing with the crisis once it had manifested. This fact would not be captured in the regression analysis presented, as this paper considers a lag of 14 periods between the pre-crisis factors and the crisis itself. Additionally, the significance of government deficits is not confirmed for the EA.

The results differ when the regression is repeated excluding financial centers. Here, net exports and market capitalization are significant factors in addition to government deficits as visible in Table 5. As expected, net exports display a negative sign, meaning that a larger pre-crisis export levels would have resulted in lower GDP growth. This can be explained by the reduction in trade during the crisis, leading to lower GDP for countries dependent on exports. As pointed out above, the group of financial centers produces most of the outlier values in the data. This may explain the increase in predictive power of the regression as visible in the higher R-square (0.205). However,

(28)

this result will need to be considered with caution as the number of observations is reduced compared to the full sample.

Table 5: Regression output using GDP as crisis measure – sample variations

(2) (3) VARIABLES EA excl. FC Net Exports -0.357 -0.772** (0.247) (0.298) Government Deficit -0.783 -1.168** (0.529) (0.537) Bank Lending to Priv. Sec. 0.0306 0.0333 (0.0541) (0.0525)

Market Capitalization -0.00751 0.0540*

(0.0334) (0.0282)

House Price index 0.0877 0.0350

(0.0610) (0.0475)

Fin. Sec. Concentration 0.0289 0.0181

(0.0225) (0.0199) Constant -0.138*** -0.113** (0.0364) (0.0423) Observations 162 157 R-squared 0.142 0.205 Number of countries 17 17

Notes. Dependent variable for all specifications is quarterly GDP growth from Q1 2008 until Q2

2010. Fixed effects estimations with robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 denote significance levels. Net Exports, Imports+ Exports, Current Account, Government debt and deficit, Bank lending to the private sector are as a percentage of GDP. Financial sector concentration is the ratio of the asset value of a countries three largest banks to the asset value of its total financial sector. Independent variables span Q3 2004 until Q4 2007. Country samples are defined in Appendix 1.

In conclusion, the only significant pre-crisis factor standing out as robust across model and sample specifications, in predicting GDP growth during the crisis, is government deficits as a percentage of GDP. This factor has further been found to be significant by Berkmen et al. (2012). However, this paper cannot confirm the relevance of any of the other pre-crisis factors considered in the literature. It is noteworthy, that none of the considered trade measures is found to have an effect, even though current account balances appear to be significant in several studies (Berkmen et al., 2012; Frankel &

(29)

Saravelos, 2012; Lane & Milesi-Ferretti, 2011), while the trade channel in general is further confirmed as a significant factor by Feldkircher (2014). Especially intra-European trade plays an important role for most of the countries included in this analysis. However, the studies cited above consider the trade channel on a global level, and towards the rest of the world, Europe’s current account is nearly balanced. Therefore, dependence on trade may have been an important factor that spread the crisis around the world; within Europe there were other channels that influenced crisis severity.

4.4 Robustness checks

In order to check the robustness of the findings presented above, the analysis is repeated for the alternative crisis measures of private consumption and unemployment. There are few significant coefficients to be found when repeating the regression of model

specification (EU1) – (EU5) using private consumption as the dependent variable. While a full output table can be found in Appendix 4, it is worth pointing out that the

significance of government deficits, as a predictor, is not confirmed in any of the

specifications. (EU1) does not find a single relevant pre-crisis factor. Single coefficients are found to be significant in (EU2) and (EU5). However, they are not confirmed in the alternative model specifications. Overall, the explanatory value of the regressions as measured by R-squared is low with R-squares below 0.04, for the crisis measure of private consumption. Nevertheless, the joint hypothesis that all coefficients are equal to zero is rejected at the 5 percent level for all specifications except in (EU5).

Table 6 reports the regression output of specifications (EU1b) – (EU5b), where unemployment is used as the dependent variable. The outcome differs from the

previously reported results. In all specifications except in (EU5b), market capitalization and house prices are significant at the 1 percent level. However, the coefficient of market capitalization shows a counterintuitive sign, indicating that a larger pre-crisis stock market has led to lower crisis unemployment and thus a less severe crisis. In specification (EU5b), where house prices are dropped, government deficits and bank lending to the private sector are significant at a 5 percent level. Additionally, financial sector

concentration becomes relevant at the 10 percent level. However, specification (EU5b) shows a lower R-squared, while the other specification show a reasonably high

(30)

explanatory power with R-squared levels above 0.6. Furthermore, specification (EU5b) is not reasonable in this setting, as house prices are significant and should therefore not be excluded.

Table 6: Regression output using unemployment as a crisis measure – model variations

(1) (2) (3) (4) (5)

VARIABLES EU1b EU2b EU3b EU4b EU5b

Net Exports -0.00740 -0.0117 -0.0128

(0.0179) (0.0174) (0.0232)

Government Deficit 0.0785 0.0807 0.0753 0.142**

(0.0747) (0.0731) (0.0768) (0.0542)

Bank Lending to Priv. Sec. 0.0390 0.0395 0.0391 0.0399 0.0819** (0.0237) (0.0235) (0.0233) (0.0234) (0.0331) Market Capitalization -0.0530*** -0.0543*** -0.0526*** -0.0488*** -0.00638 (0.0115) (0.0121) (0.0114) (0.0124) (0.0194) House Price index 0.184*** 0.180*** 0.186*** 0.183***

(0.0268) (0.0260) (0.0268) (0.0270)

Fin. Sec. Concentration -0.0204 -0.0195 -0.0201 -0.0202 -0.0680* (0.0248) (0.0251) (0.0248) (0.0247) (0.0370) Imports+ Exports 0.0117 (0.0202) Current Account 0.00842 (0.0631) Government Debt -0.00562 (0.00354) Constant -0.0588 -0.0680 -0.0609 -0.0592 0.0660 (0.0372) (0.0408) (0.0371) (0.0360) (0.0434) Observations 235 235 234 235 283 R-squared 0.610 0.611 0.608 0.609 0.310 Number of countries 25 25 25 25 29

Notes. Dependent variable for all specifications is unemployment using data from Q3 2008 until

Q4 2010. Fixed effects estimations with robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 denote significance levels. Net Exports, Imports+ Exports, Current Account, Government debt and deficit, Bank lending to the private sector are as a percentage of GDP. Financial sector concentration is the ratio of the asset value of a countries three largest banks to the asset value of its total financial sector. Independent variables span Q3 2004 until Q4 2007.

These results are notable in so far that previous cross-country studies have explored the role of market capitalization in explaining crisis severity, but have not found it to be significant. House prices however have been identified as a crisis cause in Europe (Lane,

(31)

2012), because of countries’ dependence on tax revenues from housing construction. Furthermore, their effect on unemployment was to be expected, because work in construction was reduced as a result of plummeting house prices.

To sum up, the results of the robustness checks so far do not confirm the significance of pre-crisis government deficits as an explanatory variable for crisis severity as found in the above-described regression using GDP growth. But they do confirm the validity of variable choice in model specification (EU1), which is used in the sample comparison. In all cases except one, namely (EU5) using private consumption as the dependent variable, none of the trade measures considered is significant. Furthermore, government debt does not stand out as a more appropriate measure for national account imbalances in comparison to government deficits.

Next, the regression analysis is repeated for the different sample specifications using the alternative crisis measures. When regressing on private consumption, there is no single coefficient significant and robust across the sample variations. As a matter of fact, when using sample specification EA, none of the coefficients is significant. Meanwhile, when reducing the sample to exclude financial centers, only market capitalization is significant at a 5 percent level. Furthermore, the statistical power of the results is weak, with R-squares below 0.06. The output for the regressions using private consumption can be found in Appendix 5.

In Table 7, the output for the regression using unemployment as the dependent variable is presented. It can be seen that market capitalization and house prices are significant at the 1 percent level across the sample specifications. Additionally, bank lending to the private sector is significant at the 5 percent level, although it is not when considering the entire sample as above in Table 6 under the heading (EU1b). Its

significance could be explained by the fact that, in Europe, firms’ reliance on bank financing, which is included in bank lending to the private sector, is thought to have been a contagion factor during the crisis (Shambaugh, 2012). However, the variable is not found to be robust across sample specifications or crisis measures. In addition, when excluding financial centers, the significance of government deficits is confirmed. This

(32)

regression shows the highest explanatory power in the analysis with a R-squared measure of 0.69. It further corroborates the outlier role of financial centers.

Table 7: Regression output using unemployment as a crisis measure – sample variations (1) (2) VARIABLES EA excl. FC Net Exports -0.0201 -0.0333 (0.0153) (0.0381) Government Deficit 0.0280 0.186* (0.0752) (0.0905) Bank Lending to Priv. Sec. 0.0536** 0.0972**

(0.0231) (0.0344) Market Capitalization -0.0581*** -0.0622***

(0.0120) (0.0176)

House Price index 0.192*** 0.172***

(0.0223) (0.0287) Fin. Sec. Concentration -0.0176 -0.00112 (0.0236) (0.0202) Constant -0.0829** -0.0789** (0.0341) (0.0297) Observations 162 157 R-squared 0.669 0.692 Number of countries 17 17

Notes. Dependent variable for all specifications is unemployment from Q3 2008 until Q4 2010.

Fixed effects estimations with robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 denote significance levels. Net Exports, Imports+ Exports, Current Account, Government debt and deficit, Bank lending to the private sector are as a percentage of GDP. Financial sector concentration is the ratio of the asset value of a countries three largest banks to the asset value of its total financial sector. Independent variables span Q3 2004 until Q4 2007. Country samples are defined in Appendix 1.

To conclude the robustness checks, none of the pre-crisis variables considered is robust across the three different crisis measures. However, house prices and market

capitalization, although this one with a counterintuitive sing, are significant and robust across model and sample variations, when regressing on unemployment. Furthermore, the regressions on unemployment show the highest explanatory power in R-squared.

Unemployment therefore appears to be a suitable measure of crisis severity, whereas private consumption may not capture crisis severity sufficiently. This is confirmed by the

(33)

fact that private consumption data do not show a clear two consecutive quarters drop as described above and used in order to identify the start of the crisis. Furthermore, as neither the use of the Arellano-Bond estimator nor the application of a perfectly balanced panel find any significant results that are robust across model specifications, one must conclude that the chosen pre-crisis variables do not explain crisis severity in Europe.

5 Conclusion

This paper analyses the manifestation of the 2008/2009-credit crisis across Europe and attempts to identify pre-crisis macroeconomic factors that predict crisis severity. However, none of the factors considered is significant and robust across model

specification and crisis measures. Nevertheless, pre-crisis government deficits are found to be significant in predicting GDP growth during the crisis. Additionally, house prices and market capitalization, though this one with a counterintuitive sign, are found to have significant coefficients when regressed on unemployment.

Despite the consideration of alternative model specifications and estimation techniques, there are limitations to this study: First, compared to the existing literature only a limited number of variables is considered. While these are selected carefully, it does not compare with the rich data sets included in cross-country studies such as the one by Frankel and Saravelos (2012) or by Feldkircher (2014). Therefore, the possibility of omitted variable bias in the estimates exists and warrants further investigation.

Second, as this paper focuses on the crisis in Europe, the small number of relevant countries limits the number of observations included in the analysis. This issue is addressed in the use of panel data, but may nevertheless be a contributing factor. The different results found when excluding financial centers support this explanation, as the financial centers produce most of the outlying observations.

Third, in the analysis conducted the crisis period is defined homogeneously, starting in the first quarter of 2008. This is done to aid a smooth application of the data, however, in reality the two consecutive quarters drop in GDP that guides this definition occurs at different points in time for different countries. Additionally, not all countries included in

(34)

the sample experience negative GDP growth for ten quarters. This may cause the measures of crisis severity to be less accurate than in the above-described cross-country studies, which frequently use single observation points such as the maximum decrease of GDP or its cumulative loss (Feldkircher, 2014; Frankel & Saravelos, 2012).

Finally, the lack of significant pre-crisis factors confirms the conclusion of Claessens et al. (2010), that the 08/09-credit crisis was a systematic and global shock. Due to country heterogeneity, the crisis did not manifest in a systematic way but affected every country individually. This explains the fact that there is no systematic relationship between pre-crisis factors and crisis severity found in this paper.