The ConstantC is product of and . It is used to label the countries as market- or bank-based. Thus the Constant expresses a degree of orientation.
Below the actual Constant for each country is calculated in order to categorize them as either bank- or market based:
The table shows the Constant for all 18 countries. The higher the Constant, the more market-based the financial system. The calculations appear in Appendix 1 parallel. The dashed line indicates whether a country is more of less market-based than the average level, excl. USA. The US is excluded as its constant equals 200 and therefore is treated as an outlier causing the dashed line/the average to be misleadingly high. Countries above the dashed line are categorised as market-based. The values highlighted in red show the eight countries used in the rest of the analysis.
These are chosen, because they are the ones with most information available. Choosing countries spread more or less equally over the whole range demonstrates that bank or market-based is matter of degree.
Table 4.1 highlights the country sample used in the rest of the paper. That is, US as the most market-based country and Spain as the most bank-based country. It should be stressed that the Constant tells nothing about whether a country’s securities take the form of equity or bonds.
Table 4.1 – Categorization of countries
Constant
USA 48.47
Belgium 11.24
Finland 9.18
France 8.11
Poland 5.44
Hungary 5.3
England 4.49
Slovak Republic 3.92
Sweden 3.27
Norway 2.68
Netherlands 2.51
Austria 2.23
Greece 2.13
Germany 2.05
Portugal 1.81
Denmark 1.43
Italy 1.2
Spain 1.1
Average excl. USA 4.01
Market-based countries
Bank-based countries
43
4.2.3 Critical assessment of the country selection
The sub-chapter is meant to highlight the complications of the chosen method. Often, complications are unavoidable and the task is instead to qualify them.
The Constant described above is used to identify several countries financial structure. One could argue that more weight should have been put into the recent structure, but doing so would itself be based on pure intuition. England has changed its structure enormously during the past years (from market-based to more bank-based) and England could easily have entered the bank-based countries if more weight was put into the recent years’ structure. The implications of the chosen approach can be that we do not see a clear picture of the countries reaction to shocks. In this case, we will go back to the chosen categorization and reassess it.
DT,C does not address whether its value is high due to a high levels equity or a high level of bond financing relative to bank loans. This does not imply that equity and bonds are considered as similar financial instruments. Later on, in question C the distinction between countries who use more equity or more bond financing will be made.
4.2.4 Measurement techniques and variables
8 countries and 200 observations per variable makes it possible to conduct Ordinary Least Square (OLS) regressions and measure correlation between level data, mix of new financing, capital structure, individual countries and time. No causality is evident. Causal arguments are introduced in the analysis and discussion.
Two types of regressions are created and tested at a 5 % significance level:
Regression I supporting research question A. Three sub-regressions are created, one for bonds, one for equity and one for bank loans.
For all explanatory variables in all regression: H0: β=0, H1: β≠0
Regression II supporting research question B. Six sub-regressions are created as the mix of new financing originates from 6 different types of net-transactions: Short- and long-term bonds, equity and bank loans.
For all explanatory variables in all regression: H0: β=0, H1: β≠0 Research question C will be answered using Regression 1 and 2.
44
Regression I – presentation of the explanatory variables
The variables used for regression 1 on bond level are presented below. All sub-chapters under Regression I also use bond level to explain the method. The regressions concerning equity and bank loans follow the exact same procedure and measurement techniques. Appendix 2 shows the observations used to estimate the regressions and their underlying data source.
Table 4.2 presents an overview of the variables, i.e. only the names and components are displayed. Table 4.3 present the properties and a brief explanation of each variable.
The dependent variable and explanatory variables used in research question A for bond level. Three main explanatory variables are used: Bond - ratio which indicates the NFCs’ capital structure with respect to bond level. In addition a Time dummy is included to measure if significant changes in level happened over time. A Country dummy for each country is included to control for country specific characteristics other than capital structure. The interaction term is per definition variables created as a product of two other explanatory variables. It highlight if the combination of two explanatory variables tells a story different from the simple sum of both. As the time dummy is quoted in years and Bond - ratio is quoted in quarters, DT takes the value 1 if a quarter (t) is included in the respected year. Table 4.3 elaborates on the interpretational meaning of each variable and their properties. Similar variables are used for equity and bank loans. However, no variables are logged with regards to bank loans.
Table 4.2 – Short name and components of variables in Regression I
Variables Name Components
Dependent variable Bond level lag(log(Bt,C)), where B = Bond level, t = Quarter, C = Country
Explanatory
variables Bond- ratio (B/TL)
lag(log(Bt,C/ TLt,C), where TL = total external finance = sum of equity, bond and bank loan
level Time dummies (DT):
D2004, D2005, D2006, D2007, D2008, D2009
D2004=1 if T = 2004 (T≠quarters) and D2004=0 if T ≠ 2004
Same logic goes for the other time dummies.
Interaction term:
Cross time -terms (CTT)
CTT = DT * lag(log (Bt,C/ TLt,C) D2004=1 if T= 2004 and
D2004 =0 if T ≠ 2004
Same logic goes for the other time dummies.
Country dummies (DC):
DDenmark, DSpain, DEngland, DFrance,
DGermany, DUS, DFinland
DDenmark=1 if C= Denmark and DDenmark =0 if C ≠ Denmark Same logic goes for the other country
dummies.
45 Table 4.3 – Properties and brief interpretations of variables in Regression I
Name Properties and brief interpretation
Bond level Values are quoted in 2003 prices and altered to $ prices registered quarterly.
They are balance-values and they change based on accounting requirements for market value alignment, net-transaction and other technical changes. Ultimo period quotation.
Bond - ratio Same properties as for bond level. The variable is lagged and represents the primo periodical capital structure. This serves an interpretational purpose explained later. Significant coefficients tells how must log(lag(B) will change as lag (log(B/ TL) change with 1, disregarding point in time and country in question.
Time dummies The reference variable is 2003. T = years unlike the observation that are quoted in quarters (t). Significant coefficients tells how much log(lag(BT)) differs from log(lag(log(B2003)).
Country dummies The reference variable is Norway and any significant coefficient tells how much log(lag(BC) differs from log(lag(BNorway) at any point in time.
Cross time-term The reference variable in the cross time-term of 2003. Any significant coefficient tells how ∆lag (log(BT,C/TLT,C) =1 impact lag(log(BT,C) differently than in 2003.
Table 4.3 present the interpretation and properties characterizing each variable: Bond - ratio, time dummies and country dummies are referred to as constitutive variables, i.e. variables that are included in the interaction terms. The same properties and interpretational meaning exist for the sub-regressions of equity level and bank loan level.
However, no variables are logged for bank loan observations and therefore estimators for explanatory variables on bank-loans refer to the natural scale whereas the estimators for equity and bond level referred to the logged scale.
Observations for bond and equity levels are logged, as it is necessary to create normal distributed residuals. If the residuals are not normal distributed, the coefficients will not be BLUE (Gujarati, 2003, p.79) if the observations are logged. This problem does not exist for bank loans. If the observations of bank loan were logged they would not be normal distributed anymore. See Appendix 3 that illustrates the distribution of observation for bond level, equity level and bank-loan level, both logged and in the natural scale. The coefficients must be interpreted differently depending on whether the variables are logged or not. This affects the analysis and discussion throughout the paper.
46 The purpose and further use of the explanatory variables
1. The intercept captures the level that is expected in Norway, in 2003 if lag(log(B/TL)) equalled zero, i.e. if Norway used no bonds. First, this makes the intercept unintuitive and only useful together with a lag(log(B/TL)) > 0. Secondly, it does not contribute to the understanding of how economies with different capital structures reacted to the crisis, as (1) capital structure is not included in the intercept and (2) the years in which the crisis occurred is not included. Therefore, I will not spend any time interpreting its meaning in chapter 5, but the intercept estimator will be presented as it is used as a reference variable.
2. The country dummies highlight country specific characteristics that determine the level, regardless of time and capital structure. It is important, because potential correlation between the country dummy and capital structure affects the estimator of lag(log(B/TL)) less than if the dummy variables were excluded (more on this below). This is the only purpose it serves. It enables us to ignore country specific characteristics in determining how different financial systems react to the crisis. Therefore, I will not spend any time interpreting its meaning in chapter 5, but the estimators will be presented, as they are used to strengthen the validity of the other estimators.
3. The time dummies, DT, show how much the level differed from the 2003 level, regardless of the country in question and regardless of the capital structure. It is used to answer research question A and therefore it is interpreted in chapter 5. It is also used as a constitutive variable for the cross time-term.
4. The bond - ratio, lag(log(B/TL)), tells to what extent the bond level is expected to increase as ratio increases, regardless of time and country specific characteristics. It therefore predicts the level based on the capital structure, but it does not predict how the level changes over time. Therefore it does not directly contribute to answering research question A and C and I will not spend any time interpreting its meaning. However, the estimator and its significance are important to report, because it is used as a constitutive variable to create the cross time-term.
5. The cross time-term uses the bond level in 2003 as a reference, i.e. it uses the intercept as a reference point. It then addresses how much the level is expected to deviate from the level in 2003 depending on its capital structure. The cross time-term is constructed as product of its constitutive variables bond-ratio and time dummies. If we expect bond-ratio and time dummies to be linearly correlated with the dependent variable log(B), then
47 the cross time-term log(B/TL)*DT provides information on whether the linear relationship suggested by the constitutive variables takes a nonlinear detour. That is, it suggests if the bond level in, let us say 2005, changes differently conditional on lag(log(B/TL)) than it would in 2003 (Cohen et al, 1983).
If the cross time-term is highly correlated with its constitutive variables, it is misleading to use it at the same time as the constitutive variables. The variable captures what happened over time and how this differed depending on the capital structure. This makes the cross time-term the most important variable in the regression and its estimators will be interpreted in chapter 5.
Statistical problems
Since the country dummy is included in order to control for multicollinearity, we recognise a certain level of this. In addition, it is also natural to find high collinear relationship between the constitutive variables (lag(log(Bt,C/TLt,C)) and DT) and their interaction term. Excluding some variables does not necessarily solve the problem, but can instead incorporate some of the effect of the omitted variables in the remaining estimators (this is the reason why country dummies were included in the first place).
Multicollinearity is not necessarily a problem, with respect to the classical OLS assumptions. However, in the event of high multicollinearity, the estimators can be difficult to measure and their standard deviation tends to be very high. This has four consequences.
Firstly, it results in higher confidence intervals (all other things equal resulting in more variables being significant). Secondly, if the variables are highly collinear, one of them tend to be insignificant in spite of the fact that the variable actually does explain the dependent variable. Thirdly, even though the individual explanatory variables can be insignificant, the overall explanation degree, R2, is usually very high. For the same reason I pay no further attention to the R2 value in this paper. Fourthly, multicollinearity results in the estimators being very sensitive towards data changes. This point is very clear in Appendix 4.1-4.3 (see definition clarification of the appendix below) (Gujarati, 2003, p. 350).
Therefore, we stand in the cross field of two countervailing arguments: One speaking in favour of including all constitutive variables and the interaction term in the same regression in order to avoid potential multicollinearity effects overtaken by another variables (Brambor et al, 2005). The opposite argument points out that including highly collinear variables can result in misleading estimators even though the overall regression is just as reliable.
48 Trials of Regression 1 on bond, equity and bank loans level showed correlation coefficients between the constitutive variables (especially the time dummies) and the interaction terms close to 1. Therefore the risk of wrong estimators caused by multicollinearity is overwhelming. To overcome such problem, one can estimate Regression I in three rounds, all with the level as the dependent variable10:
(i) Regression I.i: The first results show all variables included in 1 regression. This regression is only used to get a picture of how the estimators would be if they were extracted for the overall regression. None of these estimators are used.
(ii) Regression I.ii: The second regression includes all variables except CCTs and CTTs.
Coefficients for the time and country dummies as well as the intercept and bond-ratio are extracted from regression 2.
(iii)Regression I.iii: The third regression includes the variable on capital structure and the cross time-term. The coefficients of CTTs are extracted.
Splitting the regression enhances the risk of one variable taking over the effect of another.
Looking at the trials of the three “partial” regressions proves that the interaction term, easily could be removed. The reason is that the additive effect of the constitutive variables provide almost the same effect of level as the interaction terms do (González & Cox, 2007), i.e.
βlog(B/TL)+ βD(T) ≈ βlog(B/TL)*D(T)11
.
The cross time-terms are so highly correlated with especially the time dummies that they work as an alternative (see appendix 3 on correlation matrixes). To use it as an alternative requires that it provide information that the constitutive variables do not. As explained above it provides very unique information. Thus, if they are significant and highly correlated with the constitutive variables, their marginal contribution to the bond level will not be added directly to the effect the constitutive variables. This paper does not seek to forecast the total value of bonds, but to identify the marginal effect of the cross time-term.
Therefore the estimator of cross time-term can be interpreted without placing any attention on the estimators of the constitutive variables. It is the most important variable in Regression 1 and it is mainly used to answer research problem C.
10 Cedric Schneider, professor in Statistics at CBS, suggested this technique
11 The illustration is a bit simplified, as the presence of logged values change the interpretation of marginal contribution coefficients.
49 Special properties of logged variables
Both the dependent variable and the explanatory variable on capital structure are logged with regards to bonds and equity, i.e. they are double log regressions. This requires cautious interpretation of marginal effects of all estimators in the regression:
The estimator of any logged explanatory variable can be read as followed: The estimators express the percentage changes in the levels when the explanatory variable changes with one percentage. Thus the estimator represents elasticity. For all explanatory variables, logged or not, their marginal effect can also be calculated in the following way:
When; β β
where ε = error term Then; β
The equation above is used extensively in the rest of the paper. Note that each variable affects B with some factor, i.e. percentage, because 10X works as a multiplier. In regressions where the dependent variable is logged any estimator provides information on the factor with which it affects B.
The factor equals a percentage impact of . Any variable rewritten to 10Xis from now on referred to as a contribution factor.
Log(B/TL) and log(E/TL) always take negative values, because we log a value < 1. Thus, the higher B/TL, the less negative or the closer to zero is log(B/TL). As lag(log(B/TL)) also affects the value of the CTT observations, its nature is important to bear in mind.
Regression II
Regression II is used to answers research question B. The interest is now the mix of new external financing and how the mix changes during the crisis as opposed to before. Net-transactions are used to measure new financing and it measures the difference between new capital obtained and liabilities redeemed each quarter (see Appendix 5 for all variables and observations)12.
12 Net-transactions are extracted at the national statistics of each country. See exact references in Appendix 5
50 Net-transactions are divided into four types:
Short-term bonds
Long-term bonds
Bank loans
Equity
For a detailed analysis of changes in the mix, 6 regressions are estimated based on ratios between short-term bonds, long-term bonds, bank loans and equity:
Table 4.4 – Dependent variable ratios of Regression II
Long-term bonds Equity Bank-loan
Short-term
bonds × × ×
Long-term
bonds × ×
Equity ×
The table states the ratios of external finance that are used as dependent variables in Regression II.
The explanatory variables purely consist of time and country dummies similar to those in Regression 1. Capital structure, as used in Regression I, is excluded, because trials showed highly insignificant coefficients for all regressions (see Appendix 7 for SAS output).
Therefore no interaction terms are included. Country dummies are again included to control for country specific characteristics, so be get a clearer picture of changes over time.
51
5 Empirical results
This chapter represents core of the paper. It seeks to answer problem statement 2 and its three research questions A, B and C:
A. How does the overall level of finance change under the crisis?
B. How does the mix of the three kinds of new finance change during the crisis?
C. How do the results of A and B differ between bank and market-based economies?
All answers are in the form of regression output. The output will be briefly analyzed as it is presented while a full analysis linked to the theory and hypotheses will be conducted in chapter 6 Discussion.
5.1 Research question A
The answer to Research question A is divided into 3 sub-chapters; one for bond level, one for equity level and one for bank loan level. For each sub-chapter Regression I (see chapter 4.2.4) is presented and then interpreted according to the significant variables and their estimators. In appendix 4 all 3 sub-regressions conducted for each type of finance are presented. Only the 2nd and 3rd sub-regressions are used to extract the final set of estimators.
The subscription t indicates a quarterly observation. The subscription T indicates that a quarterly observation is used with the purpose to estimate the level in a specific year. When quarterly observations are used to estimate yearly changes (as they are for the cross time-terms) an average of the four observations in one year is used. Thus, if a capital ratio (bond-ratio, equity-ratio and bank loan ratio) is subscribed with T it is an average of the quarterly capital ratio in year T. This quotation is used throughout the rest of the paper.
Before I present my regression results, it is illustrated how the level changes from 2003 to 2009. The illustrations below serve to ensure a more intuitive interpretation of the results
52 Graph 5.1 – Changes in bond level of NFCs
Development in the balance level of bonds: All underlying values are quoted in 2003 prices and $US. The 2nd quarter of 2003 is the reference year.
Claiming a clear trend among countries or countries divided into bank and market-based systems will be putting too much confidence into my own analysis. Not even the financial crisis stands out: For some countries (Germany, France and Finland) a decrease happens during the crisis, but such decline started off already in 2006. For other countries (Denmark and Norway) the increase continues until mid 2008. Hereafter considerable reductions in the level happen. Thus the tendency is mixed.
Graph 5.2 – Changes in NFCs’ equity level over time
Development in the balance level of equity: All underlying values are quoted in 2003 prices and $US. The 2nd quarter of 2003 is the reference year. The interpretation is similar to that of bond level.
With equity a much clearer picture stands out. All stocks tend to increase until mid 2007 where after they all fell. The 3rd quarter for 2007 is known as the time where the first signs of
50 70 90 110 130 150 170 190 210 230
2003Q2 2003Q4 2004Q2 2004Q4 2005Q2 2005Q4 2006Q2 2006Q4 2007Q2 2007Q4 2008Q2 2008Q4 2009Q2
Index 100=2003,Q2
Aggregated bond level for NFCs
Norway DK Spain England France Germany US Finland
50 100 150 200 250 300 350 400
2003Q2 2003Q4 2004Q2 2004Q4 2005Q2 2005Q4 2006Q2 2006Q4 2007Q2 2007Q4 2008Q2 2008Q4 2009Q2
Index 100 2003,Q2
Aggregated equity level for NFCs
Norway DK Spain England France Germany US Finland
53 the financial crisis occurred. Denmark and Norway stand out, but it is more a matter of higher volatility, than a different trend. Actually, all countries seem to change the level in the same direction at any given point in time. Similar trends emphasize the systematic nature of the financial crisis.
Graph 5.3 – Changes in NFCs’ bank loan level over time
Development in the balance level of bank loans: All underlying values are quoted in 2003 prices and $US. The 2nd quarter of 2003 is used as the reference year. The interpretation is similar to that of bond level.
Bank-loans follow clearer pattern than bonds, but less than equity. Disregarding Norway, all countries hold their level rather constant and signs of the financial crisis are, if any, small. It actually looks as if minor increases in loans happen in the end of 2008 followed by a weak decrease in 2009. The reasons for what appears as delayed reaction to the financial crisis will be analysed later.
50 100 150 200 250 300
2003Q2 2003Q4 2004Q2 2004Q4 2005Q2 2005Q4 2006Q2 2006Q4 2007Q2 2007Q4 2008Q2 2008Q4 2009Q2
Index 100 2003, Q3
Aggregated bank loan level for NFCs
Norway DK Spain England France Germany US Finland