Are socially responsible firms less likely to default?

(1)

Master Thesis

Are socially responsible firms less likely to default?

A study into the relation between corporate social responsibility and credit default swap

spreads

By Johannes Vlietstra

Master of Finance

Faculty of Economics and Business University of Groningen

Abstract

This study investigates the relation between corporate social responsibility and credit default swap spreads. It performs ordinary least squares regressions on a sample of 273 companies listed on the S&P500 for the time period 2008-2014. Based on existing literature a negative relation between social responsibility and credit default swap spreads is expected. The conclusion of this study is that

no relation exists between social responsibility and credit default swap spreads. It is, however, important to notice that the data sample in this study is limited and that this has a potential

influence on the significance of the results.

JEL Classification: G12, M14

Keywords: Corporate social responsibility, Credit default swap spread, Default risk.

Author: Johannes Vlietstra Student number: s1693336 Date: January 12th, 2017

(2)

2

1. Introduction

Corporate social responsibility (CSR) has become an important subject in recent years. An increasing number of companies invest in CSR and a large part of big listed companies actively provide public statements regarding CSR. A definition for CSR is corporate social behaviour that goes beyond regulatory requirements (Kitzmueller and Shimshack, 2012). Over the last years a lot of research has been performed regarding CSR. Empirical research finds in general a small, but significant, positive relation between CSR and financial performance (Margolis, Elfenbein and Walsh 2007).

Another trend in the financial world is the development of the credit default swap (CDS). The market for CDS’s has increased a lot in the past decades (Ericsson, Jacobs and Oviedo, 2009). The CDS is a derivative that acts like insurance for the buyer of a bond in the case the bond issuing party cannot pay back the notional amount. In this case the bond buyer receives back the notional amount from the CDS seller. The costs are a periodic payment to the CDS seller, this is called the CDS spread. The majority of the CDS spread is due to default risk of the bond issuing party (Longstaff, Mithal and Neis 2005). This means CDS spreads can be seen as a measurement for default risk, where a higher CDS spread means a higher default risk.

The relation between CSR and financial performance is an interesting topic that has been researched extensively. More specific, when looking at default risk, there is evidence that higher CSR leads to better credit ratings (Oikonomou, Brooks, and Pavelin 2014). Furthermore there seems to be a relation between credit ratings and CDS spreads (Hull, Predescu and White 2004). There is however no direct research into the relation between CSR and CDS spreads. This thesis is innovative because it adds to the existing literature by investigating the relation between CSR and CDS spreads. With other words, it investigates if CSR influences default risk.

(4)

4

Due to limited data availability, this thesis has, of course, restrictions. Limited data on CDS spreads is freely available. It had to be searched manually in Datastream which is time consuming. Data on CSR has another obstacle. It is only available on a yearly basis in Datastream. This data is also the only freely available responsibility data and has led to a limited dataset. Despite these early obstacles and the limited dataset the choice has been made to continue with this thesis. This is because this thesis is a chance to explore a topic that is not previously researched. With the limited data it can at minimum break new grounds for further research.

The conclusion of this study is that no significant relation is found between CSR and CDS spreads. Furthermore, adding the CSR variable to a base model for explaining CDS spreads does not increase the explanatory power of this model. This can possibly be explained by the limited dataset that is used.

(5)

5

2. Literature and hypotheses development

This chapter discusses existing research into the relation between corporate social responsibility and financial performance. It explains theory on which the relation between CSR and default risk is based. The first paragraph explains the concept of CSR. The second paragraph gives an overview of measurements for CSR, including a description of the measurement used in this thesis, the asset4 ratings. The third paragraph discusses existing research in the relation between CSR and corporate financial performance. The fourth explains the dependent variable in this thesis, the CDS spread. Paragraph five discusses existing literature into CDS spreads. Paragraph six forms a bridge between CSR and CDS spreads and formulates two hypotheses.

2.1 CSR

Corporate social responsibility is a phenomenon that has become increasingly important in the last decades (Schroder, 2014). Furthermore the amount of academic research into CSR has also increased in this period. Possible explanations for this trend are given by Bénabou and Tirole (2009). They state that this is due to the fact that CSR is becoming a normal good, information about companies is more transparent and that externalities of multinationals are difficult to deal with.

To understand what motivates companies to engage in CSR it is important to first understand what CSR is. The problem with defining CSR however is that there is not one clear definition. Almost 50 years ago Friedman (1970) argued that the only social responsibility of a company is to increase its profits. This point of view sees CSR as a way of pure profit maximising.

Another definition is given in the paper by Bénabou and Tirole (2009). They state that “A standard definition of CSR is that it is about sacrificing profits in the social interest”. Seeing CSR in this way gives it a purely philanthropically motivation.

(6)

6

The actual reason for engaging in CSR is discussed by several authors. Bénabou and Tirole (2009) state three reasons. The first is win-win, which means firms engage in CSR from strategic motivation to maximise profits. The second is delegated philanthropy, which means firms engage in CSR because this is wanted by their stakeholders. This is done from a purely philanthropic viewpoint. Stakeholders can be, among others, shareholders, customers, employees or investors. The third is insider orientated corporate philanthropy. This is just like the delegated philanthropy a reason to engage in CSR purely based on philanthropic reasons. Only now the motivation comes from the management of a firm and not from its stakeholder including shareholders.

This thesis uses the definition as used by Kitzmueller and Shimshack (2012) for CSR. This is because the motivation for the companies for conducting CSR, although interesting, is not the objective of this thesis. The goal of this thesis is to investigate if there is a relation between CSR and CDS spreads. Only the degree of engaging in CSR is important. With other words, in this thesis it only matters if companies engage in CSR, not why they engage in CSR.

2.2 Measurement of CSR

To investigate the relation between CSR and CDS spreads it is important to have a measurement for CSR. There are several ways of measuring the level of CSR a company engages in. A simple way would be directly observing a company by for example reading their annual reports with respect to CSR. In practice this happens, however availability and reliability are problematic.

A more convenient way of measurement is using third party measurement rating agencies. There are several big companies that specialize in rating other companies’ CSR behaviour. Some of these rating companies are described by Malik (2014) who performs a meta-analysis on CSR studies.

(7)

7

exists. Chatterji, Levine, and Toffel (2009), for instance, investigate the KLD database ratings. They find that the biggest strength of the database is to point out who the worst performing firms are with regards to CSR. The ratings are however limited when compared with more simple objective measurements. They explain this because of noise in the measurement due to subjective scoring systems. The KLD database is however not freely accessible and thus cannot be used in this thesis.

Other databases are the Bloomberg database with CSR data for over 4,000 firms and the CRD Analytics database with CSR data for over 1,000 firms. Both databases are also not freely available for the purpose of this research.

The only available database at the University of Groningen for the purpose of this thesis is the ESG Asset4 database. The ESG database is founded in 2002 and covers in-depth data on more than 4,300 companies worldwide over a period of 9 years. This database is used as the variable that describes the CSR score for a company. The way this database works is explained in more detail in chapter three. So far it seems there is no academic research available into the validity of the CSR scores. It is at least not yet addressed in literature.

2.3 Empirical research on CSR and financial performance

(8)

8

Other meta-analyses are done by Malik (2014) and Wu and Shen (2013). They conclude that on average there exists a small positive relation between CSR and financial performance.

The previous described meta-analyses all take averages of several papers that investigate CSR and overall financial performance. They do however not focus on the relation between CSR and firm risk, specifically default risk, on which this thesis focuses.

There are several well-known articles that do focus on the relation between CSR and firm risk. El Ghoul, Guedhami, Kwok and Mishra (2011) find that firms with better CSR have a lower cost of capital. They argue that the main mechanism through which this relation works is the effect of CSR on firm risk. Oikonomou, Brooks, and Pavelin (2014) investigated whether bond yield spreads and credit ratings by rating agencies are influenced by CSR. They find that in general good CSR is rewarded with better credit ratings by rating agencies and lower bond spreads. The opposite goes for bad CSR. These results are of particular interest for this thesis since CDS spreads are closely related to bond spreads and credit ratings are related to a firms’ default probability. Oikonomou, Brooks, and Pavelin (2014) again explain these results through less risk. Furthermore, Hull, Predescu and White (2004) find that credit ratings and CDS spreads are negatively related. Goss and Roberts (2011) have comparable findings regarding bank debt. They find that companies with CSR concerns pay between 7 and 18 basis points more than more responsible firms.

(9)

9

Another form of idiosyncratic risk is due to a smaller investor base because some investors exclude irresponsible firms from investing (Heinkel, Kraus and Zechner 2001, El Ghoul, Guedhami, Kwok and Mishra 2011 and Goss and Roberts 2011).

In general it seems that higher CSR ratings are associated with a lower risk for a company. It is thus expected that firms with higher CSR scores are less likely to default. If this is the case then higher CSR scores would result in lower CDS spreads.

Overall the previously described relation between CSR scores and CDS spreads seems to indicate that higher CDS scores always leads to better financial performance or CDS spreads. This is however not the case, there are some nuances to the relation. Goss and Roberts (2011) find that only the companies with the worst CSR scores pay more for bank loans but that good CSR performance is not awarded. There seems to be a threshold beyond which further improvement of CSR is not financially rewarded. Furthermore over half of the investigated literature in the meta-analysis by Margolis, Elfenbein and Walsh (2007) shows no significant relation between CSR and financial performance.

2.4 CDS spreads

The CDS spread is in this thesis the proxy for default risk and is the dependent variable on which the research is performed. Before discussing existing research into CDS spreads the mechanics of the CDS will be explained.

(10)

10

and Neis 2005), so default is the most important credit event. Restructuring is since 2009 no longer a credit event in the US. A CDS contract has a maturity, the most common maturity is five years (Ericsson, Jacobs and Oviedo 2009). During this period the CDS buyer makes periodic payments to the CDS seller until either the reference entity defaults or the contract matures. When there is no default the CDS seller only receives payment and has no obligations.

Suppose an investor holds $10 million in Apple bonds. To protect this investment against a default by Apple the investor agrees to a CDS contract with a maturity of five years with Apple as the reference entity. The bond value in the contract is $10 million and the CDS spread is 2% per year. The investor pays $200.000 to the CDS seller per year for five years. During this year the investor receives back the full $10 million from the CDS seller if Apple defaults.

The market for CDS’s is relatively young, the first CDS contract was introduced in 1995 by JP Morgan. In 1998 the CDS contract was standardized by the international swaps and derivatives association in 1998 (Gu, Liu and Hao 2016). The notional outstanding amount has grown rapidly until the financial crisis with $2 trillion in 2002 and almost $60 trillion in 2007. However after the crisis in 2009 the amount fell to $30 Trillion (Weistroffer, Speyer, Kaiser and Mayer 2009).

CDS contracts are traded over the counter, this means they are not publicly traded. They are almost exclusively traded by a network of private dealers. Furthermore this market is very concentrated with 88% of the notional amount in 2009 bought and sold by just five institutions (Weistroffer, Speyer, Kaiser and Mayer 2009). According to Longstaff, Mithal and Neis (2005) the majority of buyers of CDS’s are institutions like banks, hedge funds and security houses. Sellers are banks and insurance companies.

2.5 Empirical research on CDS spreads

(11)

11

bond spreads. This may be due to important non-default components in regular bond spreads (Blanco, Brennan and Marsh 2004). The third and most important reason of choosing CDS spreads as proxy for default risk is that there exist no extensive literature on the relation between CDS spreads and CSR. By using CDS spreads instead of regular bond spreads this thesis can break new grounds in the research into the relation between CSR and default risk.

Research into the determinants of CDS spreads has been done by Collin-Dufresne, Goldstein and Martin (2001). They test if a set of theoretical determinants of default risk have a relation with CDS spreads. These determinants are the spot rate, changes in the slope of the yield curve, changes in leverage, changes in volatility, changing in the probability or magnitude of a downward jump in firm value and changes in the business climate. They investigate a sample of 261 companies and run a series of ordinary least squares (OLS) regressions on the CDS spread with the determinants as described before as the explanatory variables. One regression is performed for each company. After this they average the results. They find limited explanatory power with an average R2 of 25%. This means 25% of the CDS variation can be explained by the variables they used and a large part is unexplained and can potentially be explained by other factors, such as a CSR rating.

(12)

12

2.6 The relation between CSR and CDS spreads

The final step in the theoretical considerations is explaining the relation between CSR and CDS spread and forming a hypothesis. As discussed in paragraph 2.3, a better CSR performance reduces the amount of firm risk as perceived by investors. Buyers and sellers of CDS contracts are for the most part large financial institutions. They can both be seen as investors in CDS contracts. Furthermore, it is plausible that they use all available information, including CSR scores, in their decision-making regarding CDS contracts. Existing research finds that investors perceive bad performing CSR firms to bear higher risks. When investors invest in CDS contracts it is hypothesised that they, among others, use CSR scores and that a lower CSR score leads to a higher perceived risk and thus a higher CDS spread. Investors in bonds think firms with bad CSR scores are more risky and thus the chance of not receiving back the bond value is bigger. They are willing to pay more for a CDS contract to protect them against this loss. This higher demand for CDS contracts leads to higher CDS prices for companies with bad CSR. This analogy leads to the following hypothesis.

Hypothesis 1: Corporate social responsibility has a negative relation with CDS spreads.

To test this hypothesis, a responsibility factor is added to the model as used by Ericsson, Jacobs and Oviedo (2009). The choice for using this model has several reasons. The first is that their model finds significant results with relatively high adjusted R2 values when explaining CDS spreads. Adding a responsibility factor to a good performing model makes more sense than adding it to an insignificant model. The second is the simplicity of the used model. It only uses three variables: leverage, volatility and the risk-free rate. Given the time and data restrictions the choice for using their model is justified. Since this thesis adds a new factor to an existing model the following hypothesis is also tested to see if the model can be improved.

(13)

13

3. Data

This chapter describes the data that is used in this thesis as well as the data sources. The first paragraph gives the sample selection. The second paragraph describes the variables. The third paragraph deals with an important aspect of the data: the frequency of which the data is available and the moment of the data selection. The fourth and last paragraph discusses the reasons and consequences of the data limitations in this study.

3.1 Sample selection

In this thesis an unbalanced panel dataset consisting of 273 companies is used with 1,911 firm year observations. All companies are listed on the S&P500. The first step of the data selection procedure is the constituent list of the S&P500 companies as of September 2016. This list can be found in Reuters Datastream. The main reason for choosing the S&P500 is the high data availability. Furthermore, most existing research on the relation between CSR and default risk focuses on the US market, this is the reason the US market is chosen in this thesis.

The next step is checking for which companies the variables are available. CSR scores can be imported for all companies in the starting sample. The same goes for leverage, volatility and the risk-free rate. CDS spreads however, are not available for all companies. They are only available for 273 companies listed on the S&P500. The time period for which they are available is 2008-2014. This leads to the final sample of 273 companies with 1,911 firm year observations. Table 1 shows the industries the companies are operating in.

Table 1

Number of firms per industry for the full sample. The industries in the table are the main industry identifiers Datastream is using.

Industry Count

Industrial 190

Utility 31

Transportation 10

Bank/Savings & Loans 9

Insurance 19

Other financial 14

(14)

14

3.2 Variables

This paragraph describes the variables used in this thesis. The dependent variable is the CDS spread. The main independent variable is the CSR score. The other independent variables are those used by Ericsson, Jacobs and Oviedo (2009), as described in chapter two. These are the leverage, the volatility and the risk-free rate.

CDS Spread

The dependent variable in this research is the CDS spread. There are two main sources of CDS data, these are Markit and Datastream. Markit has the most data available. However, using this data is not possible due to budget constraints, so Datastream is in this thesis the source of CDS spreads. It has time series of CDS spreads available, however only for a limited number of companies. The reasoning behind this limitation of companies is not known.

For each company there are spreads available with different maturities and different restructuring types. The restructuring types are full-, modified- and no restructuring. The no restructuring type is used because it is the most common, and after 2009 the only used type in the US (Weistroffer, Speyer, Kaiser and Mayer 2009). In the literature the 5 year maturity is the most investigated (Ericsson, Jacobs and Oviedo 2009). This is why in this thesis the maturity of 5 years is chosen. Datastream has time series with daily spreads available, they collect these spreads from 13 mayor CDS dealers and average the values.

CSR Score

The main independent variable in this thesis is the CSR score. This is the variable that shows the social responsibility of the company. The data source is the Asset4 ESG database in Datastream. The ESG database is founded in 2002 and covers in-depth data on more than 4,300 companies worldwide1. Over 750 data points are used to create over 280 key performance indicators. These indicators are aggregated into four main pillars: economic score, corporate governance score, social score and environmental score. All scores range from 0 to 100. Examples of indicators for scores are human rights for the social score and

1

(15)

15

emission reduction for the environmental score. The final CSR score is an equally weighted average of the four pillars.

Thomson Reuters states the CSR scores are given on a biweekly basis to assure the reflection of the most recent responsibility data and responsibility related news regarding the companies. This means the CSR scores for each company change throughout the year. However, in practice, when looking in the Datastream database the scores are the same throughout the year. They only change once per year on January the first. This implies that despite the scores change throughout the year, Datastream only receives an updated score once per year, on January the first. The changes throughout the year are not available in Datastream.

Since Thomson Reuters claims the scores are updated bi-weekly and the Reuters database only receives new scores on January the first, an important assumption has been made. The assumption is that the score published on the first of January in Datastream represents the score the company received at a maximum of two weeks ago. With other words: the CSR score is very recent and gives an accurate representation of the current social responsibility of the firm.

Leverage

Leverage can be directly imported from Datastream and is defined in equation 1.

(16)

16

Volatility

Volatility is imported from Datastream using the custom built equation 2. This is because there is no direct volatility available for import. It is an annualized historical volatility based on 180 days of stock returns.

√ ∫ ( )

Here P is the stock price of the equity.

Risk-free rate

As risk-free rate the 10 years daily United States Treasury Benchmark Bond yield is chosen. This is the same as Ericsson, Jacobs and Oviedo (2009). Datastream has daily yields available.

The previously described variables are presented in table 2. Descriptive statistics are given for each variable. The mean, median and standard deviation are shown as well as the minimum, maximum, 5th percentile value and 95th percentile value. Interestingly the mean of the CDS spread is almost twice as high as the median, indicating (a) high value(s) that strongly influences the mean. Furthermore, the CDS spread and the leverage have relatively the highest standard deviations compared to the mean. Another interesting observation is that the range of CSR scores almost reaches the full possible range of 0 through 100.

Table 2

Descriptive statistics for the variables in this thesis. Statistics are shown for the CDS spread, the CSR score, the leverage, the volatility and the risk-free rate. For each variable the mean, median, standard deviation, the minimum, the maximum, the fifth

percentile value and the 95th percentile value are shown.

Variable Mean Median Stdev Min Max 5th Pctl 95th Pctl

CDS Spread (%) 1.61 0.88 3.32 0.10 54.72 0.28 4.82

CSR score 74.41 84.36 23.21 2.96 97.42 24.22 96.00

Leverage (%) 46.96 43.54 44.59 0.20 1557.48 14.99 82.69

Volatility (%) 33.74 28.40 19.86 10.02 202.64 14.86 71.85

(17)

17

3.3 Further data considerations Data frequency

The variables in this thesis have different availability with regards to the frequency. Some variables have daily data, others change only once per year. Table 3 shows the variables with the available frequency.

Table 3

The variables used in this thesis with per variable the frequency for which data is available.

Variable Frequency

CDS spread Daily

CSR score Yearly

Leverage Yearly

Volatility Daily

Risk-free rate Daily

(18)

18

Data timing

A final and very important consideration is the date for which data is imported. The variables with the lowest frequency, the CSR score and the leverage both change each year on the first of January. As described in paragraph 3.2, the CSR score on the first of January gives an accurate representation of the social responsibility of the firm at that day. This is because this score is at a maximum two weeks old, as described in paragraph 3.2. It is also assumed that leverage is accurate at the first of January, because this is the day Datastream updates the leverage. The CSR score and the leverage are both accurate at January the first. The other variables are available at every day. To match the dates and frequencies of the variables, the choice has been made to import all variables at a yearly frequency with as data January the first.

3.4 Data limitation

The data in this thesis has a considerable limitation. The limitation is the fact that only seven observations through time are available for each company to perform the regression. This limitation is of such importance that it warrants some attention here. In this paragraph the reason and the effect of the data limitation and the motivation to continue with this research are discussed, in spite of the limitations of the data.

Reason of data limitation

The reason for the data limitation is data availability. Ideally data would be available at a daily frequency over several years, just like the data that is used by Collin-Dufresne, Goldstein and Martin (2001) and Ericsson, Jacobs and Oviedo (2009). This is, however, not the case. As shown in table 3 in paragraph 3.3, data for the main explanatory variable, the CSR score, is only available at a yearly frequency in Datastream. The same goes for leverage. Data for the CDS spreads, the volatility and the risk-free rate are however available at a daily level. The reason for using a data source that only has CSR data at a yearly frequency available is that it is the only option. Other data sources do have data at a daily level. These data sources were however not available while doing research for this thesis.

(19)

19

information in the daily data for some variables is not used. This choice has been made with much consideration. Other statistical methods, like MIDAS, are considered to utilise the extra information in the daily available data. The result of these considerations is simply that the daily data cannot be used and the only way to proceed is to use yearly observations for all data.

A final limitation is the time period for which data is available. The only time period for which all data is available is 2008-2014. This gives seven yearly observations.

Effect of data limitation

The effect of the limited data availability has consequences for the statistical significance of the results. T-values are calculated to test the statistical significance of the average coefficients for each variable. With the small frequency of the sample in this thesis it is possibly and perhaps likely that results are not significant. Furthermore, often performed robustness checks cannot be done. This is because the most common robustness checks like dividing the sample in time periods or company types makes the sample even smaller.

Motivation to continue

The possible relation between CDS spreads and CSR is an interesting topic. The main obstacle to do research on this topic is the unavailability of data with high frequencies. Several attempts have been done to gain access to other databases, with no result. The only possible way to continue this research is to use the limited dataset.

(20)

20

4. Methodology

This chapter describes the model and the econometric techniques used in this thesis. The first paragraph explains the main method, ordinary least squares (OLS). The second paragraph discusses the use of levels regression versus differences regression. The third paragraph discusses univariate regressions. The fourth paragraph explains the use of panel OLS.

4.1 Main OLS method

The main framework for analysing the relation between CSR score and CDS spreads is similar to the framework first used by Collin-Dufresne, Goldstein and Martin (2001), followed by Ericsson, Jacobs and Oviedo (2009) and Galil, Shapir, Amiram and Ben-Zion (2014).

This method works as follows: one equation is estimated, using OLS, for each company in the sample. Each estimated equation has a coefficient for each variable. So, for a sample of 273 companies, 273 individual regressions are performed. Hereafter the results are averaged.

The main equation in this thesis builds on the equation as used by Ericsson, Jacobs and Oviedo (2009), which is shown in equation 3. The reason for this choice is explained in paragraph 2.5. This equation regresses the CDS spread on leverage, volatility and the risk-free rate.

Here

- is the CDS Spread of firm i at time t, - is the intercept of firm i,

- is the leverage of firm i at time t, - is the volatility of firm i at time t,

(21)

21

This thesis expands equation 3 with the CSR score. The following equation is the main equation in this study.

Here all variables are the same as in equation 3 and - is the CSR score of firm i at time t and - is the coefficient for the CSR score of firm i.

This equation is estimated for each company and the results are 273 estimated coefficients and 273 estimated adjusted R2’s. Hereafter the main results can be calculated from these results. This leads to an average coefficient for the CSR score, the leverage, the volatility and the risk-free rate. Furthermore an average adjusted R2 is calculated by averaging the adjusted R2’s for each estimated regression.

The statistical significance of the coefficients is calculated using again the same method as Collin-Dufresne, Goldstein and Martin (2001). Ericsson, Jacobs and Oviedo (2009) also follow this method. This is done by using a t-statistic. The average of the 273 estimated coefficients for each variable is divided by the standard deviation over these coefficients. This standard deviation is first scaled by the square root of the sample size, √ . The equation is shown below.

√

Here

- is the t-statistic. It is a test value based on the student T-distribution to determine the significance of each factor.

- is the average coefficient of the 273 estimated coefficients for each variable.

(22)

22

4.2 Levels regression versus differences regression

The main equation in this thesis, as specified in equation 4, is a regression on levels. It is also possible to run a regression on differences. Existing research uses both levels and differences (Ericsson, Jacobs and Oviedo 2009 and Collin-Dufresne, Goldstein and Martin 2001). A level regression is the most common type of regressions. It regresses the level, or value, of the dependent variable to the level of explanatory variables. Differences regression means regressing the difference between observations for a variable, instead of the level of a variable. First differences regression means running the regression on the difference between the variable at time t and time t-1.

Ericsson, Jacobs and Oviedo (2009) discuss both methods. They argue that the choice for either method can be based on economical or statistical reasons. From an economic perspective it can be interesting to investigate both the variation in levels and differences. From a statistical perspective it is appropriate to use differences regression if the dependent variable and the independent variables are integrated. A unit root test can be performed to check if the variables are integrated, however with the sample Ericsson, Jacobs and Oviedo (2009) use this is not possible because among others, they have a short sample. Because of the uncertainty regarding the right regression model, they use both.

In this thesis the sample is shorter than the sample Ericsson, Jacobs and Oviedo (2009) use. This is due to limited available data. Because of this the statistical correct method cannot be determined and both the levels and the differences regression are performed. A final remark is differences are harder to explain then levels and that as a result lower R2 are expected from a regression on differences (Ericsson, Jacobs and Oviedo 2009)

The main equation for the differences regression is given below by equation 6.

(23)

23

4.3 Univariate regression

The main regression in this thesis is a multivariate regression. It is a regression which uses multiple variables as regressors. It is also possible to perform univariate regressions. These are regressions with just one explanatory variable. Ericsson, Jacobs and Oviedo (2009) use this to get a better understanding of the explanatory power of each variable. They regress the CDS spread separately to each variable used in the main equation to indicate the relative importance, or explanatory power, of each variable. This thesis also performs univariate regressions for this purpose. This is done for CSR, leverage, volatility and the risk-free rate. The equations are shown below.

The variables in equations 7 to 10 are the same as in equation 4. They are estimated both on levels and on differences regression.

4.4 Panel OLS

Panel data OLS is an econometric technique. It is one of the robustness checks used by Ericsson, Jacobs and Oviedo (2009) to verify the results from their main method. This thesis also uses panel OLS. The reason for doing this is that the main econometric method, as described in paragraph 4.1, can possibly lead to insignificant results. This is a consequence of the limited dataset.

It is possible that panel data techniques remove the problem of these possibly insignificant results. This is because the use of panel OLS has major advantages over regular regression techniques like the main method in paragraph 4.1.

(24)

24

Panel data OLS combines all data to one data set with 9,555 data points (273 companies times 35 data points). On this data one single regression is then performed with as a result a possibly higher statistical significance.

Another advantage of panel OLS over estimating separate equations for each firm, as is done in the original methodology, is that panel OLS accounts for common structures in the data. So it is accounted for if firms, on average, behave the same way with respect to a variable (Brooks 2014).

Panel data is data that has both cross-sectional and time series elements. The data in this thesis has firms as cross-sections and has different observations through time. It can thus be seen as panel data and panel data techniques can be used.

There are several forms of panel OLS. The forms are pooled-, fixed- and random effects panel OLS. The first estimation uses pooled OLS. This is the simplest form where all data is stacked as if it is one dataset. The original sample is merged into one dataset on which one regression is performed. The following equation is used for pooled panel OLS. It is the same equation as equation 4. Only now instead of estimating 273 equations only one estimation is performed.

Pooled OLS however has a big drawback. It assumes that the estimated variables are the same for each firm and are the same over time. In the main method an intercept and a coefficient for each variable is estimated for each firm. So firms differ. For example, Apple has a different intercept and a different coefficient for CSR than Amazon has. However with pooled OLS, there is just one intercept and one coefficient for each variable for the whole sample. It thus assumes there is no heterogeneity between firms and nothing changes over time. To allow for differences between firms, fixed effects OLS can be used. Fixed effects OLS has two forms: entity-fixed and time-fixed.

(25)

25

Here is a firm specific constant. This term allows for heterogeneity between firms. It can for example be the location of a firm’s headquarters or the specific product it sells. Since the current dataset has firms from different industries it is likely that entity-fixed effects are needed.

Time-fixed effect adds a constant to the equation that changes over time but is the same cross-sectional. The equation is specified in equation 13.

Here is time-varying constant that is the same for all companies. It can for example be a change in regulations that affects all companies the same. Since the current sample covers seven years where among other the financial crisis happened it is to be expected that time-fixed effects are needed.

A redundant fixed test is performed to determine if entity-fixed, time fixed, both or neither are necessary. When both entity-fixed and time fixed effects are used the equation becomes.

(26)

26

5. Results

The results of this thesis are presented in this chapter. The first paragraph gives the results from the main regression and compares them to existing literature. It relates the results to the hypotheses of this study. The second paragraph describes the results of the univariate regressions to further explain the explanatory power of each variable. The third paragraph discusses results from panel data OLS.

5.1 Main results

The main results of this thesis are presented in table 4. The results are calculated in the same way as Collin-Dufresne, Goldstein and Martin (2001) and Ericsson, Jacobs and Oviedo (2009). This is done by running one regression on each firm and averaging the results. The results are presented by first showing the estimation results without the CSR variable and then showing the result with the CSR variable included in the regression. This is done so the results can be compared.

The estimation without the CSR variable is the same estimation as Ericsson, Jacobs and Oviedo (2009) performed, the formula is specified in equation 3 in chapter four. The estimation with the CSR variable is the main equation in this thesis and is specified in equation 4 in chapter four. Results for both regressions on levels and differences are shown. It is important to notice that the sample for the regression on differences is smaller than that for levels. This is because when calculating differences there are only 6 years of data. Some variables have missing observations in Datastream, leading to only 5 years of data. 5 Years of data is not enough to perform a multiple regression with 5 variables, so these companies are dropped from the sample. This leads to 228 companies for the levels regression and 194 companies in the differences regression.

(27)

27

Table 4

Average results for individual OLS regressions for each company in the sample. Part A shows average OLS estimates for the regression on levels. Part B shows average OLS estimates for the regression on differences. Both part A and B show in the first column the results without the CSR variable and in the second column the results with the CSR variable. For each coefficient the t-statistic is shown in parentheses below the coefficient. *, ** and *** stand for significance at the 10%, 5% and 1% level, respectively. N shows the amount of companies where a regression is performed on.

Part A: Regression on levels Part B: Regression on differences

Intercept

Without CSR With CSR Without CSR With CSR

98.06 (2.95)*** 218.80 (1.88)** -2.39 (-1.06) -5.91 (-1.21) CSR - - -1.81 (-1.22) - - -2.34 (-1.27) Leverage -0.59 (-1.10) 0.16 (0.14) -0.62 (-0.70) 2.19 (0.98) Volatility 3.98 (10.87)*** 4.18 (9.82)*** 4.05 (9.02)*** 4.08 (8.22)*** Risk-free rate -30.82 (-8.58)*** -31.58 (-6.62)*** -30.85 (-8.28)*** -34.17 (-5.82)*** Adjusted R2 0.69 0.71 0.73 0.76 N 228 228 194 194

Results of the regression with the CSR variable

The most important result is the estimated sign for the CSR coefficient. This is negative with -1.81 for the levels regression and -2.34 for the differences regression. This implies that CSR has a negative relation with CDS spreads and seems to confirm hypothesis 1. This is consistent with the expected relation as described in the literature review. However, the estimated coefficients are not statistically significant with t-statistics of -1.22 and -1.27 and thus hypothesis 1 cannot be confirmed. It is important to notice that the critical value of a t-statistic to be t-statistically significant at the 10% level is 1.282, this means that both coefficients are very close to being statistically significant. The statistical significance of the results will be further discussed in chapter six: the discussion.

(28)

28

implies that adding a CSR variable to the equation of Ericsson, Jacobs and Oviedo (2009) increases the explanatory power. This confirms hypothesis 2.

A further result is the explanatory power of the model. This is an adjusted R2 of 0.71 for the regression on levels and 0.76 for the regression on differences. This model is and extends the model of Ericsson, Jacobs and Oviedo (2009) by adding the CSR variable so it makes sense to compare the adjusted R2’s. They find adjusted R2’s of 0.60 for levels and 0.23 for differences in their regressions, which excludes the CSR variable. The adjusted R2 for levels is comparable, that for differences however is in this thesis much higher than Ericsson, Jacobs and Oviedo (2009) find. This is remarkable since explaining differences is more difficult than levels, as explained in paragraph 4.2. A possible explanation is that the sample for the regression on levels is even shorter than the original sample, since one year is removed. This can have an influence on the significance. Again this will be further discussed in chapter six.

When looking at the results for leverage, volatility and the risk-free rate, several observations can be made. First of all, the results are similar to that found by Ericsson, Jacobs and Oviedo (2009). All coefficients have the expected sign. Leverage and volatility are positive, implying a positive relation with CDS spreads which makes sense economically. The coefficient for the risk-free rate is negative, as expected from the literature. The coefficient for leverage is statistically insignificant and the coefficients for volatility and the risk-free rate are both highly significant at 1%.

Results of the regression without the CSR variable

It makes sense to describe and compare the results of the regression without the CSR score to Ericsson, Jacobs and Oviedo (2009). This is because it is the same regression, although performed with a shorter sample and a different leverage. Furthermore the main model with the CSR variable builds on this model.

(29)

29

The coefficient for leverage however, is negative and still insignificant. Possible explanations are that the leverage variable differs from that used by Ericsson, Jacobs and Oviedo (2009) and that the coefficient is insignificant. Another explanation can be the short sample. Furthermore, the coefficient for leverage has changed sign when compared to the regression with the CSR variable included. This can potentially be explained by the insignificance of the coefficient.

The adjusted R2’s are 0.69 for levels and 0.73 for differences. The 0.69 for levels is similar to the 0.60 found by Ericsson, Jacobs and Oviedo (2009). The explanatory power of the differences regression however is remarkably higher then Ericsson, Jacobs and Oviedo (2009) with 0.73 against 0.23. A possible explanation can be the short data sample. Autocorrelation in the residuals is also a possibility. Further discussion follows.

5.2 Univariate results

(30)

30

Table 5

Average estimation results for univariate regressions using OLS. CDS spreads are regressed separately against the CSR score, leverage, volatility and the risk-free rate. Part A shows the results for the regression on levels. Part B shows the results for the regression on differences. The statistical significance is shown with. *, ** and *** and stand for significance at the 10%, 5% and 1% level respectively.

Part A: Regression on levels Part B: Regression on differences

Regression using only CSR

Intercept -33.06 -7.84***

Coefficient 1.80 0.97

R2 0.20 0.26

Regression using only leverage

Intercept 126.44*** -8.53***

Coefficient -0.25*** -4.69***

R2 0.18 0.20

Regression using only volatility

Intercept -22.35*** 5.63***

Coefficient 4.45*** 5.35***

R2 0.62 0.76

Regression using only the risk-free rate

Intercept 265.42*** -16.91***

Coefficient -41.22*** -90.86***

R2 0.25 0.63

The results from the univariate regression are surprising. The most remarkable result is the sign of the CSR coefficient. Both in the levels and the differences regression the sign is positive with 1.80 and 0.97 respectively. This indicates a positive relation between CSR and CDS spreads, opposite to theory. A plausible explanation is the omitted variable argument. The coefficients are however not statistically significant.

(31)

31

their main regression. The signs and significance for the volatility and risk-free rate factors are comparable to the main findings, just as found by Ericsson, Jacobs and Oviedo (2009).

When looking at the explanatory power of each variable the CSR variable has a limited explanatory power with 0.20 for levels and 0.26 for differences. The leverage variable has limited explanatory power with 0.18 for levels and 0.20 for differences. This result differs from Ericsson, Jacobs and Oviedo (2009). They find that leverage has the highest explanatory power. Volatility has the highest explanatory power with 0.62 for levels and 0.76 for differences. This is higher than Ericsson, Jacobs and Oviedo (2009), who find 0.23 for levels and 0.14 for differences. A possible explanation is the omitted variable argument. The explanatory power of the risk-free rate is comparable to Ericsson, Jacobs and Oviedo (2009).

5.3 Panel OLS results

This paragraph presents the results for the estimations using panel data OLS techniques. This is done only on levels, just like Ericsson, Jacobs and Oviedo (2009). They state that including panel OLS on differences does not increase the validity and that just doing regression on levels is sufficient. Table 6 shows the results of the panel estimations.

Table 6

Results of regressions using panel OLS. The first column shows results for pooled OLS. The second column shows the model with entity-fixed effects. The third column shows the model with both entity- and time-fixed effects, note that the risk-free rate is dropped here. This is because a model with time fixed effects cannot be estimated for data that doesn’t differ between firms like the risk-free rate. *, ** and *** a stand for significance at the 10%, 5% and 1% level respectively.

Pooled Entity-fixed Entity- and time-fixed

Intercept -46.60 -29.22 -147.67 CSR -0.89*** 0.46 0.27 Leverage 5.82*** 0.83*** 0.84*** Volatility 3.30*** 5.69*** 7.41*** Risk-free rate -26.48*** -26.50*** - Adjusted R2 0.33 0.61 0.62

(32)

32

negative, as expected. The coefficients for leverage, volatility and the risk-free rate are the same as Ericsson, Jacobs and Oviedo (2009) find. Furthermore all factors are highly statistically significant. The adjusted R2 of the pooled regression is 0.33 which is remarkably lower than the adjusted R2 of the main results which is 0.71 for the regressions on levels. This difference will be discussed in chapter six.

When including entity-fixed effects the outcome becomes different. First, the adjusted R2 when compared to the pooled model increases almost double in size to 0.61. Including entity fixed effects allows for each company in the sample to have own characteristics not yet captured in the variables. This is clearly the case; it is evident that the CSR score, leverage, volatility and the risk-free rate cannot explain all the cross company differences. There were clearly are omitted variables that are now captured in the extra term that entity-fixed effects adds to the model. The specification of this term can be found in equation 12 in paragraph 4.4. The results for a redundant fixed-effects test confirm that fixed-effects are necessary. Another remarkable result is the change of the sign of the CSR coefficient, it becomes positive. This is not expected in hypotheses one, however this coefficient is also no longer statistically significant.

The final model includes both entity-fixed effects and time-fixed effects. The risk-free rate is dropped from the equation, this is necessary to make it possible to use time fixed effects. The results are almost identical to the model with only time-fixed effects. Just as Ericsson, Jacobs and Oviedo (2009), the adjusted R2 increases slightly when including time fixed effects from 0.61 to 0.62. Apparently there is not much variation that can be explained by time-fixed effects. It is however difficult to interpret such results exactly.

(33)

33

6. Discussion

To get a better understanding of the results of this thesis it is interesting to first compare the results of the main OLS method without the CSR score with the findings by Ericsson, Jacobs and Oviedo (2009), to determine if this model is correct. This is because this thesis adds the CSR variable to this equation. For the most part the results are similar. Estimated coefficients for volatility and the risk-free rate have the same sign as Ericsson, Jacobs and Oviedo (2009) find. They are both highly significant. The sign for leverage however is different from Ericsson, Jacobs and Oviedo (2009). A possible explanation is that the coefficient is statistically insignificant, or that another type of leverage is used. Overall the model without the CSR variable for determining the CDS spreads seems to work and thus seems to be a reliable foundation for adding the CSR variable.

The result for the main regression with the CSR variable is a negative coefficient for the CSR factor for both the levels and differences regressions. This implies a negative relation between CSR and CDS spreads and seems to confirm hypothesis 1. However the coefficients are statistically insignificant, although almost statistically significant at the 10% level. Based on this result hypothesis 1 cannot be confirmed. It can at the most hint at a negative relation between CSR and CDS spreads as predicted by theory. The rest of the results for the main regression with the CSR variable are logical when compared to Ericsson, Jacobs and Oviedo (2009).

(34)

34

This is why the results from using panel OLS are of high importance. Panel OLS runs one regression on all data. It uses 9,555 data points in this regression instead of the 35 used in the main OLS method. This seems to be a solution to the limitations of the dataset. The results from the pooled OLS show a negative coefficient for the CSR variable that is highly statistically significant. This confirms hypothesis 1. The coefficients for leverage, volatility and the risk-free rate all have the same sign as Ericsson, Jacobs and Oviedo (2009), as expected. The explanatory power of the pooled OLS model is an adjusted R2 of 0.33. Furthermore adding the CSR variable to the pooled model increased the adjusted R2 with 0.02. A redundant fixed effects however shows that including fixed-effects are preferred for the panel OLS. This means the results of the pooled OLS become less usable.

The result for panel OLS with entity-fixed effects shows that the CSR coefficient becomes positive. Furthermore it is no longer statistically significant. This result shows no relation between CSR and CDS spreads and thus cannot confirm hypothesis 1. The coefficients for leverage, volatility and the risk-free rate all have still the same sign as Ericsson, Jacobs and Oviedo (2009) and are still significant at 1%. The explanatory power of the entity-fixed model does not change when adding or removing the CSR variable. This implies that adding a CSR variable to the equation as used by Ericsson, Jacobs and Oviedo (2009) does not increase the explanatory power of the model and thus hypothesis 2 cannot be confirmed. Adding time-fixed effects to the regression does not change the findings as found for the entity-fixed effects.

Overall there are some contradictory findings. The main OLS method finds a negative relation between CSR and CDS spreads that is statistically insignificant, but by a very small margin. This relation is also found by using pooled OLS, only now statistically significant. However, when using panel OLS with fixed-effects these findings disappear. The CSR coefficient is then positive and no longer statistically significant.

(35)

35

critical look it becomes apparent that with the short sample the results from the main OLS method become possibly unreliable.

The panel OLS has statistically reliable findings since it can use all data in one regression and thus has a larger sample. This means results from the panel OLS are more reliable than that of the main OLS method. This leads to the concluding insight that in this thesis the safest course of action is to base the conclusions on the results of the panel OLS. The results from the panel OLS show no relation between CSR and CDS spreads and no increase in explanatory power after adding the CSR variable. This means that both hypothesis 1 and hypothesis 2 cannot be confirmed with certainty.

(36)

36

7. Conclusion

This thesis investigates the relation between corporate social responsibility (CSR) and CDS spreads. This is done by investigating a sample of 273 companies. For each company the CSR score, the CDS spread, the leverage, the volatility and the risk-free rate are imported. This is done on a yearly basis for the period 2008-2014. This study follows the methodology first used by Collin-Dufresne, Goldstein and Martin (2001) and followed by Ericsson, Jacobs and Oviedo (2009). Using an average of OLS regression results for each company in the sample as well as panel data techniques two hypotheses are tested. Hypothesis 1 is shown below.

Hypothesis 1: Corporate social responsibility has a negative relation with CDS spreads.

There are mixed findings regarding hypothesis 1. The result from the main OLS is a negative relation between the CSR variable and the CDS spread. This result however is statistically insignificant, although by a very small margin. Furthermore the reliability of the average OLS method is questionable at a minimum because of the limited dataset. Based on the main OLS method hypothesis 1 cannot be confirmed.

The results from pooled panel OLS show again a negative relation between CSR and CDS spreads that is statistically significant. This seems to confirm hypothesis 1. Further testing however shows that fixed effects are necessary. When adding fixed effects the CSR coefficient becomes positive and statistically insignificant. This result does again not confirm hypothesis 1.

Compared with the main OLS method the panel OLS is statistically more reliable. Overall it can thus not be concluded that there exists a negative relation between CSR and CDS spreads since the fixed-effects panel OLS shows no relation between CSR and CDS spreads. The conclusion is that hypothesis 1 gets rejected.

Hypothesis 2 is the second tested hypothesis and is shown below.

(37)

37

The findings with regards to hypothesis 2 are more consistent. In the average OLS model adding the CSR variable increases the explanatory power with 0.02. This result is however statistically unreliable. In the pooled OLS model the explanatory power increases also with 0.02. The fixed effects model however does not get a higher explanatory power when adding the CSR variable to the base model. Since the fixed effects model is necessary and the panel OLS is statistically reliable it cannot be concluded that adding a CSR variable to the determinants for CDS spreads increases the explanatory power. Hypothesis 2 gets rejected.

In summary the main conclusion of this thesis is that with the data in this study it cannot be proven that there is a relation between CSR and CDS spreads. Furthermore it cannot be proven that adding the CSR variable increases the explanatory power of the main model. Hypothesis 1 and hypothesis 2 both get rejected.

It is important to take a critical look at this thesis. It is the first direct research into the relation between CSR and CDS spreads, so there is no existing literature on this topic that can be directly used. There is however extensive research into the relation between CSR and financial performance as well as into CDS spreads. The literature review in this thesis combines this to form grounded hypotheses as well as a solid foundation for the methodology. The data used in this thesis however, is limited. This is because availability. The consequence of this limited data shows in the results. Possible insignificance makes the results hard to interpret en leads to the fact that results of the main method have to be seen with caution. This is a missed chance, since the main method is also the main method that is used in existing literature and ideally conclusions would be based on the main method. The conclusions of this thesis are still solid. They are based on a combination of results from the main method and the more statistically reliable panel OLS.

(38)

38

(39)

39

Literature

Bénabou, R., Tirole, J., 2010. Individual and corporate social responsibility. Economica 77, 1-19.

Blanco, R., Brennan, S., Marsh, I. W., 2004. An empirical analysis of the dynamic relation between investment grade bonds and credit default swaps. Bank of England.

Brooks, C., 2014. Introductory econometrics for finance. Cambridge university press.

Chatterji, A. K., Levine, D. I., Toffel, M. W., 2009. How well do social ratings actually measure corporate social responsibility?. Journal of Economics & Management Strategy 18, 125-169.

Collin‐Dufresne, P., Goldstein, R. S., Martin, J. S., 2001. The determinants of credit spread changes. The Journal of Finance 56, 2177-2207.

El Ghoul, S., Guedhami, O., Kwok, C. C., Mishra, D. R., 2011. Does corporate social responsibility affect the cost of capital? Journal of Banking & Finance 35, 2388-2406.

Ericsson, J., Jacobs, K., Oviedo, R., 2009. The determinants of credit default swap premia. Journal of Financial and Quantitative Analysis 44, 109-132.

Friedman, M., 1970. The Social Responsibility of Business is to increase its Profits. The New York Times Magazine, September 13, 1970.

Galil, K., Shapir, O. M., Amiram, D., Ben-Zion, U., 2014. The determinants of CDS spreads. Journal of Banking & Finance 41, 271-282.

(40)

40

Goss, A., Roberts, G. S., 2011. The impact of corporate social responsibility on the cost of bank loans. Journal of Banking & Finance 35, 1794-1810.

Gu, W., Liu, Y., Hao, R. 2016. Attenuated Model of Pricing Credit Default Swap under the Fractional Brownian Motion Environment. Journal of Mathematical Finance 6, 247.

Heinkel, R., Kraus, A., Zechner, J., 2001. The effect of green investment on corporate behaviour. Journal of financial and quantitative analysis 36, 431-449.

Hsiao, C., 2014. Analysis of panel data. Cambridge University Press No. 54.

Hull, J., Predescu, M., White, A. 2004. The relationship between credit default swap spreads, bond yields, and credit rating announcements. Journal of Banking & Finance 28, 2789-2811.

Kitzmueller, M., Shimshack, J., 2012. Economic perspectives on corporate social responsibility. Journal of Economic Literature 50, 51-84.

Longstaff, F. A., Mithal, S., Neis, E., 2005. Corporate yield spreads: Default risk or liquidity? New evidence from the credit default swap market. The Journal of Finance 60, 2213-2253.

Malik, M., 2014. Value-enhancing capabilities of CSR: a brief review of contemporary literature. Journal of Business Ethics 127, 419-438.

Margolis, J. D., Elfenbein, H. A., Walsh, J. P., 2007. Does it pay to be good? A meta-analysis and redirection of research on the relation between corporate social and financial

performance. Ann Arbor 1001, 48109-1234.

(41)

41

Schroder, M., 2014. Financial effects of corporate social responsibility: a literature review. Journal of Sustainable Finance and Investments 4, 337-350.

Weistroffer, C., Speyer, B., Kaiser, S., Mayer, T. 2009. Credit default swaps. Deutsche bank research.

Are socially responsible firms less likely to default?

Master Thesis