Thinking about an alternative way to design CEO’s incentive compensation package? CEO’s Inside Debt : does inside debt help to mitigate CEO’s behavior of excessive risk-taking in banks?

(1)

Thinking about an alternative way to design CEO’s

incentive compensation package?

CEO’s Inside Debt

Does inside debt help to mitigate CEO’s behavior of excessive risk-taking in

banks?

Master of Science in Finance Track: Corporate Finance

Master Thesis Written by

Lei Gong 11398655 supervisor:

(2)

2

Statement of Originality

This document is written by Student: Lei Gong who declares

to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document

are original and that no sources other than those mentioned in

the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely

for the supervision of completion of the work, not for the

contents.

(3)

3

ABSTRACT

Debating about whether increasing inside debt in CEO’s compensation package helps to curb CEO’s behaviour of excessive risk-taking in banks is prevalent in academia. Because of “too big to fail” in banking, higher proportion of equity incentive plan in CEO’s compensation leads to CEO’s behaviour of excessive risk-taking. They tend to pursue short-term payoffs at the expense of the interest of debt holders. Does CEO’s inside debt help to curb CEO’s behaviour of excessive risk-taking in banks? It is a key question to figure out whether we can think about another way to design CEO’s incentive compensation package and it also helps to prevent CEO from pursuing short-term payoffs. I collect 72 banks and 109 CEOs between 2006 and 2017 and the sample size is 565 entries. Dataset is collected from CRSP/Compustat Merged, Execucomp and Compustat Fundamentals Annual. I use pooled OLS, fixed effects model to conduct this panel data study. A positive relationship between CEO’s inside debt holding and risk in banks is confirmed in my several linear regression results. The most robust result that I find is that 0.02% changes on z-score over time, on average per bank, when inside debt increase by 1% and it is statistically significant at 10% significance level. However, the effect of CEO’s inside debt on risk in banks is not very salient. The elasticity of this relationship is 0.02 (inelastic), so it is not economic significance.

(4)

4

1. Introduction

A breaking news in September 2008 that a famous investment bank- Lehman Brothers declared bankruptcy did cause a series of unexpected consequences sweeping around the world and US subprime mortgage market collapsed after that. International banking industry has been severely influenced by the collapse of US subprime mortgage market. Many economists believe it was the worst financial crisis since the Great Depression of 1930s and they blamed that banks in the United States have the behaviour of excessive risk-taking. It becomes a main factor that results in the financial crisis. At that time, many borrowers owned NINJA loans that are mortgages given to people with no income, no job and no assets. As well, this mortgage was issued with no down payment with variable interest rates called “teaser rates” that starts at low interest rate and increase over time. They make NINJA loans very hard to repay the principal of the mortgage. When borrowers noticed their home values lower than the value of mortgages, they were defaulting on their loans. It is a main reason that causes 2008 market crash and housing crisis. These excessive risk-taking by banks has many potential negative externalities for society. It is a typical example of excessive risk-taking by banks. Empirically, the increase of excessive risk-risk-taking in US banks before crisis becomes a main cause to increase the vulnerability of banking sector (DeYoung, Peng, & Yan, 2013).

The way to mitigate bank risk-taking behaviour becomes a hot-debated academic and regulatory topic in recent years. Because banks with bailout guarantees such as subsidized deposit insurance to avoid bank runs and implicit government guarantees, banks may engage in risk-shifting activities. It transfers downside risk to society. Shareholders realizes benefits from upside risk, but losses are covered by taxpayers. Troubled Asset Relief Program

(TARP) under Emergency Economic Stabilization Act of 2008 is an example of bailout guarantee. TARP is aims to purchase toxic assets and equity from financial institutions to strengthen the financial sector and cope with subprime mortgage crisis. The TARP authorizes $700 billion expenditure and many banks such as Goldman Sachs Group Inc., Morgan

Stanley, J.P. Morgan Chase & CO., etc. agreed to receive preferred stock investments from the U.S. Treasury. Apparently, this program costs taxpayers a lot to rescue banking industry.

(6)

6 Typical corporate governance mechanisms that maximizes shareholders’ benefits may lead to negative incentives in banking. There are two types of agency problems that can influence capital structure choices. One type is the distortions between firm’s equity holders and creditors. The other type is different objectives of management and shareholders. If executives aim to maximize shareholders’ equity, the interest of debtholders would be compromised. For instance, debtholders will have a high probability to facing debts default. In worst situation, debtholders are unable to receive the debt principle. Therefore, it is necessary to consider the interest of debtholders to improve regulation and health of banking industry. Recent papers highlight several key governance factors that shape the behaviour of bank risk-taking. They are effectiveness of bank boards, the structure of executives’

compensation, and the risk management system and practices employed by banks. This paper studies the relationship between structure of executives’ compensation and the behaviour of excessive risk-taking in banking. Senior executives are responsible for making important decisions every day. Traditionally, the evaluation of performance of senior executive is closely related with interest of shareholders. Therefore, in past many studies focused on option-based and performance-based equity incentives as their main research topic in studying structure of executive compensation. Usually these incentive packages are likely to encourage senior executives to pursue risky policies that yield short-term payoffs. Assessing alternative executives’ pay components is important because compensating executives with instruments that promote long-term stability is very important for policy-makers and

debtholders. In recent years, many researchers propose a concept - inside debt - that is a way to align the interest of senior executives with firms’ debtholders. It involves debt-like

instruments that are mainly deferred compensations and pension benefits. Because the payoff of inside debt depends not only on the incidence of bankruptcy but also firm value in

bankruptcy. It exposes senior executives to the same default risk with debtholders if the firm defaulted. When inside debt becoming one component of senior executives’ pay package, the risk preference of executives would converge with debtholders so higher inside debt may mitigate excessive risk-taking in banking industry.

This paper aims to answer whether inside debt owned by senior executives mitigates the behaviour of excessive risk-taking in banking industry. Is there a trend that senior executives are owning increasingly amount of inside debt in their compensation package? Understanding the relationship between inside debt and behaviour of excessive risk-taking in banking

(7)

7 industry is relevant to help policy-makers to regulate banking industry to prevent incidence of financial crisis. Also, it is also relevant to help board of directors to design incentive package for their senior executives, so the interest of debtholders and interest of shareholders can be aligned. Inside debt is an emerging topic and it attracts increasingly the number of

researchers. My study attempts to reveal the effect of inside debt on the CEO’s behaviour of excessive risk-taking in banking industry.

This paper comprises five parts. First, it is literature review that involves main theories and debates in the existing literature as well as propose my hypothesis. Second, it is methodology, data and summary statistics. In this part, I will describe the sample and methodologies that I use to test hypothesis. Third, it is empirical result that I will interpret my regression result in economics prospective. Sequent, it is robustness checks that critically assess my own results by verifying whether my regression model and data are compliant with assumptions of linear regression estimators. Finally, I will draw a conclusion.

2. Literature review

2.1. What is bank risk-taking?

Banks usually face with four type of risks and they are portfolio risk, default risk, leverage risk and market risk respectively. The portfolio risk is the volatility of asset returns from different kinds of investment activities. Empirically, the portfolio risk is measured by the book, market value of asset volatility or the ratio of risk-weighted assets to assets.

(Flannery&Rangan, 2008; Vallascas&Hagendorff, 2013; Shrieves&Dahl, 1992). Default risk is raised from investment and financing activities. There are two ways to measure default risk. One is Merton’s structural distance-to-default model that is generally studied using market value data (Gropp, Vesala, & Vulpes, 2006; Hagendorff&Vallascas, 2011). The other is Z-score that can be calculated from data in financial statements (Pathan, 2009;

Laeven&Levine, 2009; Houston, Lin, Lin, &Ma, 2010). Leverage risk is closely related to capital structure of the bank. Banks or firms need to raise capital to fund their daily operation activities or investment decisions. If the bank has a much higher amount of debts than the amount of equity, it can be regarded as high leverage risk. It is empirically measured by book capital ratios such as high-quality (Tier-1) capital or risk-adjusted capital ratios (Flannery &

(8)

8 Rangan, 2008; Gropp & Heider, 2010; Nier & Baumann, 2006). The market risk normally arises from the impact of adverse market movements on the value of banks’ on-balance and off-balance sheet positions. Accroding to Anderson & Fraser (2000), Chen, Steiner, & Whyte (2006) and Konishi & Yasuda (2004), there are two common ways to measure market risk that are stock volatility and tail risk measures like value-at-risk (VaR) based on the research of Ellul & Yerramilli (2013) and Van Bekkum (2015).

2.2. Are banks different?

Banks can be differentiated from non-financial firms by some characteristics. One cash-generating unit of banks is that they receive money from short-term deposits with high liquidity and turn them into relatively illiquid long-term loans. The quality of these loans is monitored and collected by banks privately, so it is difficult for external stakeholders to know related information in order to value the intrinsic value of their assets. Therefore, senior executives may tend to manipulate these off-balance-sheet items to increase bank value because these off-balance sheet items would not be visible on retrospective balance sheets.

Also, banks are financial institutions with relatively high level of leverage and they are protected by governments. If they are in severe financial distress or are about to file bankruptcy, government will provide them with implicit and explicit protection. Explicit protection is deposit insurance guarantees and implicit protection is emergency liquidity and capital assistance such as bailouts (Bhattacharya & Thakor, 1993). These protections from can be regarded as a put option. The value of this call option increases as banks increase the leverage ratio, but it costs very less if this call option ends up with out-of-money, so it

motivates executives of banks tend to pursue high-risk projects to boost this option value and short-term benefits from increase of share price. It exacerbates the behaviour of excessive risk-taking as indicated by Bebchuk & Spamann (2009), John, Mehran & Qian (2010) and Keeley (1990).

2.3. The relationship between inside debt and behaviour of excessive risk-taking

When it comes to the relationship between corporate governance and the excessive risk-taking behaviour by banks. There are three main factors in context of corporate governance that affect banking risk-taking. They are the board who monitor and control risks, the

(9)

9 structure of CEO pay for reducing the level of risk-taking as well as risk management and risk exposures of the bank. This study explores the topic of structuring CEO pay for mitigating the excessive risk-taking behaviour of banks.

Hagendorff and Vallascas (2011) analyse the impact of executive compensation structure on the risk decisions made by bank CEOs. In samples of acquiring U.S. bank, they use the Merton distance for defaulting their model in order to reveal that CEOs who are highly sensitive to pay-risk tends to participate in risk-inducing mergers. They also show that in large banks, the executive compensation would encourage risk-shifting activities, and shareholders tend to incentivize CEOs to accept risk-increasing investment for squeezing benefit from bondholders.

Also, they show that the high level of pay-risk sensitivity would lead to CEOs’ engagements in risk-increasing deals. Bhattacharyya and Purnanandam (2011) investigate the reason why banks tend to engage in more risky behaviour in pre-crisis phase (2007-2008 financial crisis). They investigate factors that have influences on CEO compensation and find that CEOs’ compensation depends more heavily on the short-term earnings instead of the level of stock returns. It seems like that the pattern reflects the phenomenon that executives are likely to pay more attention on enhancing short-term performance even they need to bear the excessive amount of risk. Kim, Li et al. (2016) examine how the bank CEO risk-taking incentives induced by stock option compensation would affect banks’ contribution to systematic risk by using a sample of bank-holding companies and commercial banks in the US. They show that the CEO risk-taking incentivized by option-based compensation are positively correlated with a bank’s contribution to systematic distress and crash risks. The finding is consistent with the notion that CEO option-based compensation would increase the chance of excessive engagement in risky business activities as a reaction to common risk exposure.

As shown in the study of Mehran et al. (2011), almost half of bank CEOs in the sample held option as awards vested within one year. These options held by CEOs would increase possibilities that CEOs undertake bank policies for maximizing short-term interest for their

(10)

10 equity payoffs at the cost of long-term interest. However, as Fahlenbrach et al. (2011)

indicated, bank CEOs held on to vested equity or option grants before the financial crisis bore huge wealth losses subsequently. This is contradictory with the saying that CEOs tend to pursue short-term interest.

2.4. Hypothesis

Van Bekkum (2015) reveals a negative relation between CEO and CFO inside debt, measures of subsequent market volatility and tail risk. In the study, he chooses to use market risk and tail risk measured by VaR, equity risk and expected shortfall to be dependent variables in order to measure risks taken by banks. And his study proves that there is a negative relationship between inside debt and risk.

Srivastav et al (2014) reports a negative relationship between inside debt and dividend payout ratio. Bennett et al. (2015) shows a negative relationship between inside debt and a default risk measured by market-based factors. And they use the expected default frequency to measure bank risk-taking.

In the study of Bolton et al. (2015), they show that the compulsory disclosure of inside debt held by bank CEOs in 2006 was detected positively by creditors. Higher inside debt related to lower credit default swap (CDS) spreads by using announcement effect on CDS spreads to measure bank risk-taking.

Cassell et al. (2012) find that there is a negative relation between CEO inside debt holdings, the volatility of firm stock returns in the future, R&D expenses, and the leverage level. Moreover, there is a positive relation between CEO inside debt holdings and the level of diversification and asset liquidity. They suggest that CEOs who hold large inside debt prefer investment and financial policies that are less risky.

Freund et al. (2018) examine the relation between CEO inside debt holdings and firm’s choice of sources of external capital that are debt and equity. They find that the inside debt hold by CEOs aligns managers’ and debtholders’ interests and make firms prefer debt financing.

(11)

11 Srivastav et al. (2018) check whether holding inside debt causes CEOs to undertake

acquisitions with lower risk level. They find that announcement of acquisitions published by CEOs with high inside debt incentives are related to transferring wealth from equity to debt holders. When a deal has been successfully completed, banks with CEOs who have high inside debt incentives display lower market measures of risk and higher level of loss aversion for taxpayers. If CEOs like to pursue risky acquisitions to boost their interest at the expenses of debtholders, debtholders’ interest would be compromised according to Srivastav et al. (2018)

Mohamed & Sabri (2013) studies the relationship between debt-like compensation and executives’ behaviour of excessive risk-taking. Executives who has a higher amount of inside debt holding tend to use interest rate derivatives for hedging purpose because their interest is closely related to whether debt will default. It implies that executives’ interest is aligned to debtholders’ interest.

Hypothesis: Increasing CEOs’ inside debt holdings helps to mitigate their behaviour of

excessive risk taking in banks.

Data and Methodology

3.1. Data

3.1.1. The Sample of banks

This study is about executives’ compensation and excessive risk-taking in banking industry. The data is queried from S&P Compustat Execucomp and CRSP/Compustat Merged. I follow the method that Fahlenbrach & Stulz (2011) used. I firstly restrict the query with Standard Industry Classification (SIC) codes between 6000 and 6300 from fiscal year 2006 to fiscal year 2016. It yields 132 unique firms. I exclude firms with SIC code 6282. Because this category includes companies that main operating activities are primarily provide investment information and advice on a contract or fee basis to companies and individuals about

securities and commodities.

Moreover, I manually screen firms with SIC code 6199 (non-depository institutions with finance services) and SIC code 6211 (Security brokers, dealers, and flotation companies: their main operating activities are purchase, sale, and brokerage of securities). It is necessary to screen some firms out from SIC code 6199 and 6211 because there are pure brokerage

(12)

12 houses as well as these two codes include some investment banks: American Express and Citigroup. In appendix, I list companies that are included into my sample as well as

companies that are excluded from my sample. My sample contains 94 banks that are different from the sample of Fahlenbrach & Stulz (2011) that has 98 firms in their sample. Because three companies in their sample have bankrupted and the balance sheet is not available for companies: SLM CORP and BBVA COMPASS BANCSHARES INC.

Table 1

Descriptive statistics for fiscal year 2006

The following table reports summary statistics for key variables of a sample with 94 bank holding companies and investment banks in fiscal year 2006. For each variable, the mean, median, lower quantile, higher quantile and standard deviation are calculated. COMPUSTAT Annual and COMPUSTAT Bank Annual databases are sources of the sample. Criteria applied in sample selection are described in section 2 and the list of banks in the sample are shown in the appendix. Tier 1 capital ratio shown the last row of the table below is calculated based on the Basel Accord for reporting risk-adjusted capital adequacy and is taken from the COMPUSTAT Bank database. All accounting variables in the table are in millions of dollars.

Variable Number Minimum Lower Quartile

Median Upper Quartile

Maximum Mean Standard Deviation Total Assets 94 1885.96 6295.89 14440.13 58001 1884318 141195.5 335887.9 Total Liabilities 94 1753.69 5707.46 13006.76 52848 1764535 130974 313323.4 Market capitalization 94 313.47 1222.49 2761.84 13272.96 273598.1 19272.37 45105.32 Net Income/Total Assets 94 0.03% 0.88% 1.16% 1.45% 2.91% 1.20% 0.45% Net Income/Book Equity 94 0.33% 9.88% 13.07% 16.63% 26.65% 13.44% 5.34% Book-to-Market Ratio 94 0.27 0.43 0.50 0.63 1.13 0.53 0.16 Tier 1 capital ratio 83 6.35% 8.43% 9.48% 11.09% 19.04% 9.77% 1.97%

The median of total asset $14.4 billion, and the mean value of total assets is $141.2 billion. At the end of 2006, the mean and median of market capitalization of sample banks are $19.3 billion and $2.8 billion respectively. The average net income over assets and net income over equity is 1.2% and 13.44% respectively. This table also involves a measure of capital strength for banking that is Tier 1 capital ratio. It is defined by banking core equity capital over total risk-weighted assets which include all assets that the firm holds that are systematically weighted for credit risk. The Tier 1 capital ratio is on average 9.8%.

(13)

13

3.1.2. Define key variables

(1). Bank risk-taking

There are many measures to define banks risk-taking. Roy’s (inverse) Z-score which is used by Laeven and Levine, 2009. Not relying on the market value is the notable advantage of this score. Because there is some privately held banks in my sample. Roy (1952) proposes a concept “safety first” principle in his paper. This principle describe that if executives are certain, they will maximize the expected profits by selecting a level of liabilities and assets to keep the possibility of insolvency low. By using Chebyshev’s inequality, we can define the distance to default as:

π denote the expected return of assets; εt is the random value of the return on the assets in

period t. the standard deviation is σ; d is the level of liabilities that potentially cause default. Therefore, bank insolvency is the situation where its losses (-π) are higher than its equity (E, defined as the default level). Follow Laeven and Levine’s (2009) procedure, their study substitute π with return on assets of banks and E with the value of bank equity divided by total assets that is the capital asset ratio. After that, they obtain that P(π≤ −#)≡ %('() ≤ −*)'). –CAR is interpreted as the bank’s leverage. The bank’s probability of default is:

ROA stands for return on assets and CAR stands for capital assets ratio.

'() = -./ 12345.

64/78 799./9

*)' = #:;</=

64/78 799./9

If the distribution of profits are normal distribution, the contrary of the probability of insolvency would be equal to (ROA+CAR)/σ(ROA), in which σ(ROA) is the standard deviation of ROA. Following their procedure, the inverse of the probability of insolvency is the z-score.

(14)

14 Data for ROA and CAR can be found on Compustat-Capital IQ-fundamental annual. In this database, I use ni, at, seq for net income, total assets and shareholder’s equity respectively. A higher z-score means that the bank is more stable. Because the z-score is highly skewed, I transform z-score into nature logarithm of the z-score that is close to normally distributed. However, because there are many negative values of z-score, when I transform them into nature logarithm of z-score, they generate NA in my database. i.e. there is no value for nature logarithm of a negative value of z-score. There are 12 data entries without value of standard deviation of ROA. They are BBVA COMPASS BANCSHARES INC, BEAR STEARNS

COMPANIES INC,CHITTENDEN CORP,FIRST INDIANA CORP, FIRST NIAGARA

FINANCIAL GRP,GREATER BAY BANCORP,INVESTORS FINANCIAL SVCS CP,

MAF BANCORP INC,MERCANTILE BANKSHARES CORP,SLM CORP,TD

BANKNORTH INC, WASHINGTON FEDERAL INC. They have only one-year data, so the

standard deviation of the company is not available. I drop these data entries from my sample.

(2). Inside debt

I follow Edmans and Liu (2011) methodology and inside debt holdings is defined as:

> = ?@⁄A@

?C_⁄_AC=

(DAEFGHEIEJ?K)/(FMHKNIHDMGHEF) (OM?APMIK?APM)/(D×KFRH)

Inside debt is CEO pensions and deferred compensation. Non-qualified plan is a type of tax-deferred and employer-sponsored retirement plan. And non-qualified plans are consisting of four major types plans that are deferred-compensation plans, group carve-out plans, split-dollar life insurance plans and executive bonus plans. To provide specific forms of compensation to executives, non-qualified plans are used usually. NQDC plans provide chance of tax-deferred growth, however, there are risks associated with the plan such as the risk of complete loss of the assets. The agreement between employee and employer regards to deferring a part of an employee’s annual income until a noted date in the future is an NQDC plan. And the date can be in 5 year 10 years or even in retirement based on different plans. The employee can choose the portion to defer each year from the employee’s salary, bonuses, or other forms of compensation. Deferring this compensation offers one tax advantage: the employee does not pay income tax on that portion of its compensation in the year that the employee defers it. It differs with qualified plan, NDQC stays in employer’s balance sheet and it is subject to potential loss. This NDQC generally is a promise that the employer will

(15)

15 pay you back in the future. Usually, the employer sets aside money in a trust to fulfil its future obligation. Especially, this amount of amount in the trust are still part of assets in the balance sheet. It means that creditors are able to claim this amount in the case of this employer filed bankruptcy.

DI_{stands for inside debt. The present value of accumulated pension benefits refers to as}

PENSION. And NQDC is defined as nonqualified deferred compensation in fiscal year-end balance. EI_{is equal to the value of option and stock holdings where the stock ownership}

value (STOCK) is calculated by multiplying share outstanding by the stock price. The amount of shares outstanding includes unvested stock and equity incentive plan awards. OPTIONS is calculated from the Black-Scholes value. Firm debt (DF_{) is the sum of long-term}

debt (LTDEBT) and current debt (CDEBT). The firm equity is denoted as EF_{and is}

calculated by multiplying shares outstanding by stock price.

LTDEBT is Long Term Debt- Total (DLTT) and CDEBT is Debt in Current Liabilities-Total in Compustat-Capital IQ-fundamental annual. Stock price is Price Close-Annual-Fiscal (PRCC_F) from CRSP/Compustat Merged-Fundamentals Annual. CSHO is number of outstanding shares that is csho from Compustat-Capital IQ-fundamental annual. PENSION is PENSION_VALUE_TOT--Present Value of Accumulated Pension Benefits. NQDC is DEFER_BALANCE_TOT—Total Aggregate Balance in Deferred Compensat. STOCK is calculated by SHROWN_TOT, EIP_UNEARN_NUMER, OPTION_AWARDS_FV and PRCC_F. By abovementioned variables that are from Compustat-Capital IQ-fundamental annual and CRSP/Compustat Merged-Fundamentals Annual, I can generate CEO’s inside debt from my sample. In the sample, there are some CEO who has no inside debt in their compensation. I exclude them from my sample and my sample reduces to 77 companies and 118 CEOs. Also, there is an outlier of inside debt: SOUTH FINANCIAL GROUP INC, 2009 so I drop it from my sample.

(3). Control variables

This study includes control variables to control the effect of banks’ characteristics, CEO’s characteristics and economic factors on bank’s risk-taking behaviour. Because all banks in the sample are in the US and their main operating activities also take place in the US, I do not

(16)

16 need to control country-level economic factors. But there is one effect that I need to consider, it is 2007-2008 financial crisis is covered by my sample. It is wise to control the effect of financial crisis on CEO’s behaviour of excessive risk-taking. I create a dummy variable to capture this effect- year effect that is equal to 1 if the fiscal year is 2007 and 2008. To control bank’s characteristics and CEO’s characteristics, I follow Sundaram and Yermack (2007) way to set up control variables. I include gender, CEO’s age to control CEO’s characteristics. Ln (net assets) to control the bank size. Also, I include loan loss provision ratio to control bank’s characteristics. In the banking industry, lenders earn revenues from the interest and expense they receive form lending products. A wide range of customer like small start-ups , consumers, and large enterprises would borrow from banks. So, isolating the effect of loan loss effect on behaviour of excessive risk-taking is important. Loan loss provision provide a good proxy to control this effect. Of course, it is reasonable to scale it by net interest income because in my sample the size of banks varies. Loan loss provision ratio is loan loss provision over net interest income. Expenses set aside as allowances for uncollected loans and loan payments set asides are loan loss provision. In the other words, loan loss provision will cover potential loan losses that caused by customer defaults, bad loans and renegotiated terms of a loan that lead to lower payment than previously estimation. The difference between revenues earned from assets of a bank and the costs generated by paying out liabilities is defined as net interest income

3.1.3. Summary Statistics

In this part, I am going to illustrate descriptive statistics of variables in my sample. From Graph1 to Graph 4, I visualize the dependent variable and key independent variable in histogram and boxplot.

(17)

17

Graph 2. Histogram of ln (Inside Debt)

In Graph 1, ln(Z-Score) has peaks between 2 and 4 and ln (Inside Debt) in Graph 2 has peaks between -2.5 and 0. The spread of Graph 1 is from around -3 to 6 but the left side of -3 is suspected as outliners. The spread of Graph is from -8 to 2. The symmetry of these two

graphs is the same and they are skewed left, so they are not a good fit for normal distribution.

Graph 3. Boxplot of ln (Inside Debt)

(18)

18 Graph 3 and graph 4 are boxplots that graphically depict groups of numerical data through their quartiles (A quartile is a type of quantile. The first quartile (Q1) is the number in the middle of the smallest number and the median of the dataset. The second quartile (Q2) is the median value of the dataset. Q3 is the value in the middle of the median and the highest value of the dataset). The line extending vertically is boxes (whiskers) indicating variability outside the upper and lower quartiles. Outliers are demonstrated as individual points. Because

boxplots are non-parametric, it shows variation in samples of a statistical population without making any assumptions of the underlying statistical distribution. So, it is straightforward to show degree of dispersion, skewness and outliners. I spot many outliners in my sample they are out of whiskers and they are mainly at the right side of the boxplot. Based on the position of the notch in the boxplot, these two variables have a skewness of distribution. The

distribution of ln(Z-Score) has a negative skewness and the distribution of ln (inside debt) has a positive skewness.

Graph 5. Histogram of ln(Z-Score) with trimmed dataset

(19)

19

Graph 7. Boxplot of ln(Z-Score) with trimmed dataset

Graph 8. Boxplot of ln (Inside Debt) with trimmed dataset

After removing from my dataset, the sample reduces to 72 banks and 109 CEOs. I generate new histogram and boxplot for ln (Inside Debt) and for ln (Z-Score). They are shown from Graph 5 to Graph 8. In Graph 5, ln(Z-Score) has peaks between 2.5 and 4 and ln (Inside Debt) in Graph 6 has peaks between -2.5 and -1. The spread of Graph 5 is from around 1 to 5. The spread of Graph 6 is from -4 to 2. The symmetry of these two graphs is the same and they are slightly skewed left and not a good fit for normal distribution. But compared with untrimmed dataset, trimmed dataset improves a lot in skewness and distribution. From Graph 7 and Graph 8, it is easy to conclude that many outliners are dropped from the sample. Based on the position of the notch (median) in the boxplot, these two variables have a skewness of distribution. The distribution of ln (Z-Score) has a negative skewness and the distribution of ln (inside debt) has a positive skewness.

Graph 9 shows the trend of z-score and CEO’s inside debt. I calculate the group mean of nature logarithm of z-score and inside debt for each fiscal year. The left-hand graph shows that the risk of banks in 2007 is quite good but during financial crisis the risk of banks

(20)

20 during financial crisis. But from 2009 the situation is getting better and it is an upward trend in the left graph. In the right-hand graph, it shows the change of CEO’s inside debt. There is a huge jump in the graph from a low level in 2008 to the highest level in 2009. After this jump, CEO’s inside debt stay at the relatively higher level, compared with data before 2009.

Graph 9. Trend of Z-Score and CEO’s inside debt between 2006 and 2017

In graph 10, it shows a scatterplot of inside debt and z-score. The value of x-axis is nature logarithm of inside debt and value of y-axis is nature logarithm of z-score. The grey area and red line is the 95% confidence level interval for predictions from OLS linear model. It is a flat line. In each tail, the fitted line is bent. From this graph, I suspect the OLS estimator is biased and inconsistent. In the next part of this paper, I will test in statistics way.

Graph 10. Scatterplot of Inside Debt against Z-Score

Table 2 provides summary statistics of dependent variable, its components and the key independent variable of my regression model. There are 565 records in my sample where it contains 72 banks and 109 CEOs. It is a large sample and it is convincing to provide more precise estimates of the population mean and standard deviation. The mean and median of ln(z-score) is close to each other. They are 3.08 and 3.18 respectively. Because mean and

(21)

21 median both measure central tendency, it can conclude the skewness of this variable. ln(z-score) is skewed to the left because its median is greater than its mean. Standard deviation is a good measurement to determine how the variable in the dataset deviates from its mean. If the standard deviation high, it means greater spread from the mean value. The standard deviation of ln(z-score) is 0.84. Most of observations are spread within one standard

deviations on each side of the mean. The mean and median of ln (inside debt) is also close to each other but its median is smaller than its mean. It means that ln (inside debt) is skewed to the right. And the distribution of most observation around its standard deviations has the similar situation as that for ln (z-score)

(22)

22

Table 2

Sample summary statistics for 72 banks and 109 CEOs

The data covers fiscal year between 2006 and 2017. The following table shows summary statistics for key variables for a sample of 72 investment banks and companies held by banks for fiscal year between 2006 and 2017.Sample banks are also listed in their appendix. The data are from CompStat annual and CompStat bank annual databases and CRSP/Compustat merged. Tier 1 capital ratio is calculated according to the Basel Accord for reporting risk-adjusted capital adequacy and is taken from the Compustat Bank database. Z-score is used to measure bank risk. It is equal to the return on assets plus the capital asset ratio divided by the standard deviation of asset returns. ln(z-score) is nature logarithm of z-score. ln(z-score2_{) is nature logarithm of square of z-score. ROA is return on asset that is calculated as net}

income divided by total asset. CAR is calculated as total shareholders’ equity divided by total assets. σ(ROA) is standard deviation of ROA within each company. Inside debt is equal to present value of pension benefit plus non-qualified deferred compensation divided by current debt and long-term debt of the company and then scale this number by CEO’s value of his/her owned shares and owned options over market capitalization of this bank. Pension is present value of accumulated pension benefits from all pension plan ($). Non-qualified deferred compensation is total aggregate balance in deferred compensation plans at fiscal year end ($). Value of owned shares is production of equal to shares owned by the executives, including options that are exercisable or will become exercisable within 60 days and stock price in fiscal year end. Value of owned options is fair value of all options awarded during the year as detailed in the plan-based awards table and valuation is based upon the grant-date fair value as detailed in FAS 123R. Current debt is debt in current liabilities. Long-term debt is long term debt.

All accounting variable are measured in millions of dollars.

Variable Number Minimum _QuartileLower Median _QuartileUpper Maximum Mean _DeviationStandard

ln(z-score) 565 0.92 2.55 3.18 3.64 5.08 3.08 0.84 ROA ₅₆₅ _-0.04 _0.006 _0.008 _0.011 _0.042 _0.01 _0.01 CAR ₅₆₅ _0.03 _0.092 _0.106 _0.118 _0.189 _0.11 _0.02 σ(ROA) 565 0.0007 0.0026 0.0048 0.0086 0.0434 0.01 0.01 ln (Inside Debt) 565 -5.29 -2.37 -1.51 -0.41 1.87 -1.48 1.45 Pension 565 0 0.27 2.96 8.19 57.43 6.34 8.83 Non-qualified deferred compensation 565 0 0.17 1.19 4.51 224.87 5.87 20.93 Value of owned Shares 565 0.36 7.95 19.95 48.53 1095.65 55.03 112 Value of owned Option 565 0.00 0.00 0.00 0.68 19.01 0.71 1.76 Market Capitalization 565 129.28 1211.44 2886.96 11689.29 303681.16 17266.17 40807.95 Current Debt ₅₆₅ _0.00 _355.92 _1158.80 _4982.40 _442959.00 _22618.40 _72624.32 Long term-Debt 565 0.00 386.81 1419.73 8508.00 359180.00 18405.67 48860.11 Loan Loss Provision Ratio 565 -0.16 0.04 0.08 0.21 2.43 0.19 0.28 3.2. Empirical Models

The sample I used for this paper covers 72 entities and 11 years. It is a panel data or longitudinal data. Because of data is not available for some banks, the sample I used is an unbalanced panel. There are three common methodology to conduct panel data analysis. They are pooled OLS, random effects models and fixed effects models. Pooled OLS estimation is simply an OLS methodology run on Panel data. Because of unobserved

heterogeneity (they are correlated with the observed variables) among banks, OLS estimator is biased and not consistent. Fixed-effects and random-effects model is designed for

(23)

23 controlling unobserved heterogeneity. If the heterogeneity does not change over time and has correlation with independent variables, this time invariant component can be removed from the data by taking a first difference. The difference between fixed-effects and random-effects is that random random-effects are estimated with partial pooling, but fixed random-effects are not. Partial pooling plays an important role in studying unbalanced panel data because the effect of an entity with a few data will be partially estimated based on the more abundant data from other groups. This can be a good balance between estimating an effect by completely

pooling all entities that masks entity-level variation and estimating an effect for all groups completely separately that would generate poor estimates for small-sample entities. In the empirical result part, I will demonstrate three methodologies to conduct my panel data analysis.

Base specification model for pooled OLS linear regression:

i. Bank Risk-Takingi,t=αi+β1 inside debt holdingi,t+ β2 executives’ characteristics i,t+ β3 bank characteristics i,t + β4 economics-related control variables+ εi,t

Base specification model for fixed-effects and random-effects models:

ii. Bank Risk-Takingi,t=αi+

θt+β1 inside debt holdingi,t+ β2 executives’ characteristics i,t+ β3 bank characteristics i,t + β4 economics-related control variables+ εi,t

Subscript i and t stand for bank and year respectively. Dependent variable is risk of bank and it is measured by Z-score. Independent variables are CEO’s inside debt and control

variables. αi in the first model is the intercept and in the second model it combined with

coefficient can be interpret as individual effect for each bank. θt in second model is the time

effect. In Empirical Result, I test three scenarios: individual effect, time effect and both for fixed-effects and random-effects. αi captures effects that are specific to some panel unit but

constant over time but θt captures effects that are specific to a period but constant over panel

units. As to my research topic, θt captures might represent economic cycle effects between

2006 and 2017, αi captures bank specific effects that is time invariant over time such as

(24)

24

4. Empirical results

4.1. Pooled OLS Regression Results

The pooled Ordinary Least Squares (OLS) are summarized in Table 4. Each column reports a different OLS regression models, and each row reports a coefficient estimate, standard error in bracket, F-statistic and p-value, or other information about the regression. Also, all estimates of coefficient are controlled for robust cluttered standard error. Because in regression and time-series modelling, the assumption that error terms have same variance across all observations is usually violated. Heteroscedasticity-consistent standard errors include heteroscedastic residuals into the fitting of a model. Column (1) in Table 3 is a simple OLS regression of ln (z-score) on ln (inside debt). Many omitted variables have not been included into the model so the OLS assumption: strict exogeneity is violated. The coefficient of this simple OLS regression is statistically insignificant at 10% significance level. The regression in Column (2) includes year effect that it is a dummy variable that is 1 if the fiscal year is 2007 and 2008. The coefficient of ln (inside debt) is statistically insignificant but the coefficient of year effect is statistically significant at 1% significance level. During financial crisis 2007-2008, z-score of US banks decreases by 31%. It means that the risk in banks increased hugely during 2007-2008 financial crisis and it is consistent with our common idea about the impact of financial crisis on increasing risk in banks. In Column (3), I include some control variables into the regression model, but the coefficient of inside debt is still

statistically insignificant.

In graph 10, the scatterplot of ln (inside debt) against ln (z-score), the line is bent in each tail. To study this non-linear effect of inside debt on z-score, I create subsets of my sample to figure out whether there is a different effect of inside debt on z-score across different interquartile range. I use interquartile range of ln (inside debt) to create subsets. The first subset is minimum to first quartile, second subset is first quartile to third quartile and third subset is third quartile to maximum. They are corresponding to Column (4), Column (5) and Column (6) respectively. Their coefficients of ln (inside debt) are not statistically significant. Although they are not statistically significant, it tells a story. In the lowest quartile, the coefficient is negative, and it is a negative relationship between CEO’s inside debt and

(25)

z-25 score. In the highest quartile, the coefficient is very small comparing with the coefficient of column (3). If a CEO has a very little amount of inside debt in his/her compensation package, the bank risk might increase. If a CEO has a very high inside debt in his/her compensation package, the effect of inside debt on risk in banks is very little. The linear regression result of pooled OLS are all not statistically significant. It is due to the violation of assumptions of OLS estimator. For instance, omitted variables are not controlled in the model and it leads to the error term is correlated with the dependent variable. Also, it is a panel data and it contain time-series data. The error terms are probably autocorrelated and it leads to

heteroscedasticity. To get a more robust result, I use to fix-effects model to analyze this panel data in the following session.

Table 3 Results Pooled OLS Linear Regression

Excessive risk-taking in banks and CEO’s inside debt.

This table presents OLS linear regressions result of CEO’s inside debt on excessive risk-taking in banks for 72 banks and 109 CEOs. The fiscal year are between 2006 and 2017. Dependent variable is nature logarithm of z-score. Year-effect is indicator that is equal to 1 if fiscal year is equal to 2007 or 2008. MALE is 1 if the CEO is male. Age is CEO’s age in the corresponding fiscal year. Size is the bank’s log of total assets. Loan loss provision ratio is the ratio of the bank’s loan loss provisions to net interest income. (1) It is pooled OLS regression of ln (z-score) on ln (inside debt). (2) It is pooled OLS regression of ln (z-score) and ln (inside debt) and control year-effect. (3) it is pooled OLS regression of ln (z-score) on ln (inside debt) with control variables. (4) it is pooled OLS regression of ln (z-score) on ln (inside debt) with control variable by using dataset covering minimum to first quartile. (5) it is pooled OLS regression of ln (z-score) on ln (inside debt) with control variable by using dataset covering first quartile to third quartile. (6) It is pooled OLS regression of ln (z-score) on ln (inside debt) with control variable by using dataset covering third quartile to third maximum. All standard error in the brackets are HAC standard errors. *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively

Variable OLS (1) OLS (2) OLS (3) OLS (4) OLS (5) OLS (6) Inside Debt 0.044 (0.024) 0.023 (0.052) 0.021 (0.06) -0.06 (0.10) 0.119 (0.12) 0.01 (0.29) Ln (net assets) 0.013 (0.05) (0.08) 0.04 0.021 (0.05) (0.085) -0.08 Loan Loss provision ratio -0.949*** (0.15) -0.80* (0.31) -0.970*** (0.20) -1.05*** (0.16) MALE 0.72*** (0.11) 0.76*** (0.17) 0.589*** (0.13) 0.70** (0.15) Age 0.004 (0.014) 0.01 (0.017) -0.004 (0.020) -0.003 (0.017) Year-effect -0.31*** (0.09) -0.171* (0.994) -0.378* (0.178) -0.045 (0.135) -0.14 (0.244) Clustered Standard

Errors? YES YES YES YES YES YES F-statistics _3.33 _7.30 _11.73 _3.96 _4.44 _5.13 P-value _0.069 _0.0007 ₀ _0.001 ₀ _0.0002 Degree of freedom 563 562 476 93 234 136 Number of banks 77 77 77 44 60 35 Number of Obs. ₅₆₅ ₅₆₅ ₅₆₅ ₁₄₂ ₂₈₀ ₁₄₃ Adjusted-R2 0.006 0.022 0.117 0.15 0.08 0.13

(26)

26

4.2 Fixed-Effects and Random-Effects Regression Results

The fixed-effects are summarized in Table 4. Each column reports a different regression, and each row reports a coefficient estimate, standard error, F-statistic and p-value, or other information about the regression. The standard error in fixed-effect model can be adjust to robust by controlling for heteroskedasticity. I follow -T.S. Breusch & A.R. Pagan (1979), A

Simple Test for Heteroscedasticity and Random Coefficient Variation; R. Koenker (1981), A Note on Studentizing a Test for Heteroscedasticity; - their methodology to performs the

Breusch-Pagan test against heteroskedasticity. The Breusch-Pagan test is developed to test whether there is a heteroscedasticity by evaluating the relationship between variance of error terms and independent variables. If there are too many variances of error term can be

explained by independent variables, we can reject its null hypothesis. The null hypothesis is presence of homoscedasticity and the alternative hypothesis is presence of heteroskedasticity. The BP value is 425 and p-value of it is close to 0. It means that our model has presence of heteroskedasticity. Standard errors that are valid if uit is potentially heteroskedastic and

potentially correlated over time within an entity are referred to as heteroskedasticity-and-autocorrelation-consistent (HAC) standard errors. So, the standard errors in Table 4 are adjusted to HAC standard errors.

From column (1) to column (3) in Table 4, they are Least Square Dummy Variable model and control for the fixed effects of banks, year and CEO respectively. Estimates of Column (1) and (3) are statistically significant but estimate of Column (2) is not statistically

significant. The positive relationship between risk in banks and CEO’s inside debt is confirmed statistically.

The column (1) in Table 4 presents results for the LSDV regression of ln (z-score) on ln (inside debt) with controlling bank specific effects. The coefficient on the ln (inside debt) is positive: according to this estimate, increasing CEO’s inside debt reduce risk in banks. Statistical interpretation of Column (1) and (3) is: At 1% significance level, the coefficient of ln (inside debt) indicates 0.049% changes on ln (z-score) overtime, on average per bank, when ln (inside debt) increase by 1%. Because dependent variable and independent variable are nature logarithm of the variable, 0.049 is also the elasticity of ln (z-score) with respect to X. An elastic variable is defined as the elasticity value is greater than 1 and inelastic variable is defined as the elasticity value is less than 1. In this case, the elasticity value is far smaller

(27)

27 than 1. It means that the relationship is not very economic significance. Compared with the simple regression (1) in the Table 3, the regression adjusted R-Squared jump from 0.6% to 99.8% when fixed effects are included. The bank effects account for a large amount of the variation in the data.

Coefficient from LSDV model (2) that I create 11 dummy variables to control time effect (2006-2017) is not statistically significant. By comparing this coefficient with coefficient from LSDV (1) and LSDV (2), I can conclude that time effect plays an important role in changing ln (z-score).

Model (4) is fixed-effects model and it controls individual effect. Unlike model (1), (2) and (3), it does not contain dummy variables. So, the adjusted R-squared reduces significantly from around 99% to 8%. The coefficient of ln (inside debt) is statistically significant at 1% significance level. The coefficient of ln (inside debt) indicates 0.038% changes on ln (z-score) overtime, on average per bank, when ln (inside debt) increase by 1%. 0.038 is the elasticity of ln (z-score) with respect to X and the relationship is inelastic. Also, the effect of ln (inside debt) on ln (z-score) in model (4) is smaller than that in model (1) and model (3). Model (5) is fixed model-effects model and it controls both individual effect and time effect. Although adjusted R-squared is slightly higher than that of model (4), the coefficient

decreases and standard error increases comparing with model (4). The coefficient is 0.02 and it is statistically significant at 10% significance level.

Model (6) contains CEO dummy variables and control variables. It controls time effect, but it does not control individual effect. The coefficient of ln (inside debt) is not statistically

significant. Model (7) contains CEO dummy variables and control variables. It controls individual effect, but it does not control time effect. The coefficient is statistically significant at 1% significance level. Model (8) is also statistically significant and it contains CEO dummy variable, controlling individual effect and controlling time effect. From this group of models (model (6); model (7); model (8)), I observe that when time effect is included in the specification model, the coefficient will either become statistical insignificance or the standard error become larger. This finding is consistent with the finding from the previous regression result of OLS.

(28)

28 From model (9) to model (11), they are random-effects models, and control individual effect, time effect as well as both respectively. Model (9) is statistically significant at 1%

significance level, but model (10) and model (11) are not statistically significant. The coefficient from model (9) is 0.037 and it is statistically significant at 1% significance level. Interpretation of the coefficients in random effects models is tricky because they contain both the within-entity and between-entity effects. In my case, the coefficient represents that the average effect of ln (inside debt) over ln (z-score) is 0.037% when ln (inside debt) changes across time and between countries by 1%. 0.037 is the elasticity of ln (z-score) with respect to X. The result is similar with result from model (1) and this relationship is inelastic. Their adjusted R-squared are also in a low level at the range from 5% to 20%. It means that fixed-effects and random-fixed-effects model can only explain lower than 20% of variation in ln (z-score) by the variation of ln (inside debt).

In this session, I test many models: LSDV, Fixed-effects and random-effects model. Their results are very similar. If the coefficient is statistically significant, it has a positive effect on ln(z-score). It means that increasing a CEO’s inside debt holding reduces risk of the bank. However, they are inelastic, and coefficients are in a range of 0.02 to 0.05. i.e. they are not very economic significant, although they are statistically significant. In these three types of model, time effect plays an important role in this relationship. When time effect is controlled in the model, the standard error of the coefficient of ln (inside debt) increase. i.e. Time invariant effect in time effect has an impact on the consistency and biasness of the coefficient.

(29)

29

Table 4 Result of Fixed-Effects Regression

Linear regression of risk in banks and CEO’s inside debt.

This table presents fixed-effects linear panel data regressions result of CEO’s inside debt on excessive risk-taking in banks for 77 banks and 118 CEOs. The fiscal year are between 2006 and 2017. Dependent variable is nature logarithm of z-score. Year-effect is indicator that is equal to 1 if fiscal year is equal to 2007 or 2008. Bank-effect is controlled by creating dummy variables for each bank. CEO-effect is controlled by creating dummy variables for each CEO. (1) It is least square dummy variable (LSDV) model and I created 76 dummy variables to control effect of every bank’s time invariant components on risk in banks. (2) it is also LSDV model, but I create 12 dummy variables to control year effect from 2006 to 2017. (3) It is LSDV model, but I control every CEO individual specific time invariant component by creating dummy variables for every CEOs. (4) It is fixed-effects model with control variables. It controls individual effect. (5) It is fixed-effects model with control variables. It controls both time effect and individual effect. (6) It is fixed effects model with CEO dummy variables and control variables. It controls only time effect. (7) It is a fixed effects model with CEO dummy variables and control variables. But it controls individual effect. (8) it is a fixed-effects model with control variables and CEO dummy variables. It controls both individual effect and time effect. (9) It is random effect model with control variables and it controls individual effect. (10) It is random effect model with control variables and it controls time effect. (11) It is random effect model with control variables and it controls both individual effect and time effect.

*, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively

Dependent variable: nature logarithm of z-score.

Variable LSDV (1) LSDV (2) LSDV (3) Fixed-effects (4) Fixed- effects (5) Fixed-effects (6) Fixed-effects (7) Fixed- effects (8) Random- effects (9) Random- effects (10) Random- effects (11) Inside Debt 0.049*** (0.007) 0.003 (0.024) 0.049*** (0.007) 0.038*** (0.01) 0.02* (0.009) (0.04) -0.04 0.04*** (0.011) 0.026* (0.01) 0.037*** (0.46) 0.035 (0.051) 0.028 (0.060) Male -0.18*** (0.02) -0.10** (0.03) 0.81*** (0.23) -0.2*** (0.03) -0.11** (0.04) -0.17*** (0.021) 0.680*** (0.103) -0.13 (0.559) Age 0.0009 (0.003) 0.002 (0.003) 0.012 (0.012) -0.002 (0.004) 0 (0.003) 0.001*** (0.003) 0.005 (0.014) 0.001 (0.016) ln (net assets) -0.008 (0.05) -0.081 (0.065) -0.008 (0.080) (0.05) -0.025 -0.02 (0.054) -0.002 (0.043) 0.017 (0.050) -0.024 (0.057) Loan loss provision rate -0.17*** (0.042) -0.193*** (0.057) -0.9*** (0.169) -0.2*** (0.051) -0.23 (0.067) -0.176*** (0.042) -0.974*** (0.156) -0.198 (0.215)

CEO-Effect NO NO YES NO NO YES YES YES NO NO NO

Individual-effect YES NO NO YES YES NO YES YES YES NO YES

Time-effect NO YES NO NO YES YES NO YES NO YES YES

Degree of freedom 492 552 455 418 407 362 336 325 477 477 477 Number of banks 77 77 77 60 60 60 60 60 60 60 60 Number of Obs. 565 565 565 483 483 483 483 483 483 483 483 Adjusted-R2 0.998 0.934 0.9983 0.08 0.11 0.61 0.24 0.12 0.20 0.11 0.05

(30)

30

5. Robustness checks

5.1. Specification test: Pooled OLS or Fixed-effects or random effects model?

To figure out whether there is a problem of model misspecification, I use

Dubin-Wu-Hausman (DWH) to test for endogenous regressors (independent variables) in the regression model (fixed-effects and random-effects model), based on the methodology from Hausman,

J.A. (1978), Specification tests in econometrics; Hausman, J.A./Taylor, W.E. (1981), Panel data and unobservable individual effects; Wooldridge, Jeffrey M. (2010), Econometric Analysis of Cross Section and Panel Data, 2nd ed; The null hypothesis is that the preferred

model is random effects and the alternate hypothesis is that the model is fixed effects. i.e. this test aims to find out whether there is a correlation between the unique errors and independent variables in the model and the null hypothesis is that there is no correlation between two. i.e. Durbin -Wu- Hausman test is to evaluate the endogeneity of fixed-effects and random-effects model. To conduct Hausman test, I pair two models: model (4) and model (9) into a group. Because the coefficient of ln (inside debt) in these two models are statistically significant.

Table 5 Specification test: Hausman Test

This table presents fixed-effects linear panel data regressions result of CEO’s inside debt on excessive risk-taking in banks for 77 banks and 118 CEOs. The fiscal year are between 2006 and 2017. (4) It is fixed-effects model with control variables. It controls individual effect. (9) It is random effect model with control variables and it controls individual effect. Hausman test is conducted to these two models. The Chi-square value is shown in the row and p-value in the bracket.

Variable Fixed-effects (4) Random- effects (9) Inside Debt 0.038*** (0.01) 0.037*** (0.46) Male -0.18*** (0.02) -0.17*** (0.021) Age 0.0009 (0.003) 0.001*** (0.003) ln (net assets) -0.008 (0.05) -0.002 (0.043)

Loan loss provision rate -0.17***

(0.042) -0.176*** (0.042) CEO-Effect NO NO Individual-effect NO YES Time-effect NO NO Hausman Test 1.31 (0.93) 1.31 (0.93) Degree of freedom 5 5 Number of banks 60 60 Number of Obs. 483 483 Adjusted-R2 0.08 0.20

(31)

31 The result shows that the p-value is 0.93 so the null hypothesis cannot be rejected. It means that the null hypothesis: preferred model is random effects cannot be rejected and random- effects model is preferable.

After knowing random-effects model is a better fit for my sample, it is time to compare random-effects model with pooled OLS model. I follow the methodology fromBreusch, T. S./Pagan, A. R. (1980) The Lagrange multiplier test and its applications to model

specification in econometrics to analyze this comparison. In a panel model, αi captures the

individual effects and θt captures the time effects. The null hypothesis of BP test is that the

variance of the θt is equal to zero. It means that every individual has the same intercept, so the

pooled regression is a better choice.

Table 6 Breusch-Pagan Lagrange Multiplier Test

This table presents fixed-effects linear panel data regressions result of CEO’s inside debt on excessive risk-taking in banks for 77 banks and 118 CEOs. The fiscal year are between 2006 and 2017. (3) it is pooled OLS regression of ln (z-score) on ln (inside debt) with control variables. (9) It is random effect model with control variables and it controls individual effect. Breusch-Pagan Lagrange Multiplier test is conducted to test null hypothesis: every individual has the same intercept. The Chi-square value is shown in the row and p-value in the bracket.

Variable OLS (4) Random- effects (9) Inside Debt 0.021 (0.06) 0.037*** (0.46) Male 0.72*** (0.11) -0.17*** (0.021) Age 0.004 (0.014) 0.001*** (0.003) ln (net assets) 0.013 (0.05) -0.002 (0.043)

(0.15) -0.176*** (0.042) CEO-Effect NO NO Individual-effect NO YES Time-effect NO NO BP-LM Test 1160*** (0) 1160*** (0) Degree of freedom 1 1 Number of banks 60 60 Number of Obs. 483 483 Adjusted-R2 0.08 0.20

The p-value shows that the null hypothesis can be rejected at 1% significance level, so every individual does not have the same intercept. i.e. the pooled OLS is not a better choice

(32)

32 Although the result from Hausman test shows that random effects is a better model for my dataset, the choice between fixed-effects and random-effects can be simply decided by Hausman test. The key way to distinguish them is the assumption. That is whether the individual specific effects are uncorrelated with the independent variables. In my specification model, the independent variable is ln (inside debt) and individual specific effects are gender, education, personality traits etc. It is plausible to decisively conclude that they are highly correlated or uncorrelated. Although random effects approach does not create variables to the model representing individuals effect, it models the correlations structure of the error terms. It leads all within individual observations to be correlated. Also, I want to make inferences on an entire population and banks in my dataset represent only a sample from the population. Based on evidence above, random-effects model is indeed a better fit for my study.

5.2. Is time effect needed?

In the session of empirical results, I find that when time effect is included into the

specification model, the standard error of the coefficient of ln (inside debt) increases. I follow the medology from Baltagi, B. H./Li, Q. (1990) A lagrange multiplier test for the error

components model with incomplete panels to test whether time effect is needed in the

specification model. I test model (4) and model (9) in Table 4. The results show that the null hypothesis can be rejected at 1% significance level. i.e. time fixed-effects is needed in the specification models. Therefore, it is safe to choose the coefficient from models with controlling time fixed-effect to be the most robust result for my study. In table 4, specification model (5) is the most robust one in my tested specification models. The coefficient of ln (inside debt) is statistically significant at 10% significance level. The coefficient of ln (inside debt) indicates 0.02% changes on ln (z-score) overtime, on average per bank, when ln (inside debt) increase by 1%.

(33)

33 Table 7 Breusch-Pagan Lagrange Multiplier Test

This table presents fixed-effects linear panel data regressions result of CEO’s inside debt on excessive risk-taking in banks for 77 banks and 118 CEOs. The fiscal year are between 2006 and 2017. (4) It is fixed-effects model with control variables. It controls individual effect. (9) It is random effect model with control variables and it controls individual effect. Breusch-Pagan Lagrange Multiplier Test is conducted to these two models. The Chi-square value is shown in the row and p-value in the bracket.

Variable Fixed-effects (4) Random- effects (9) Inside Debt 0.038*** (0.01) 0.037*** (0.46) Male -0.18*** (0.02) -0.17*** (0.021) Age 0.0009 (0.003) 0.001*** (0.003) ln (net assets) -0.008 (0.05) -0.002 (0.043)

(0.042) -0.176*** (0.042) CEO-Effect NO NO Individual-effect NO YES Time-effect NO NO

BP Lagrange Multiplier Test 8.09***

(0.004) 8.09*** (0.004) Degree of freedom 1 1 Number of banks 60 60 Number of Obs. 483 483 Adjusted-R2 0.08 0.20 5.3. Stationarity of ln (z-score)

In time series data, the value of dependent variable in one period typically is correlated with its value in the next period. The correlation of a series with its own lagged values is called autocorrelation or serial correlation. If a regressor has a stochastic trend (has a unit root), then the OLS estimator of its coefficient and its OLS t-statistic can have nonstandard distributions, even in large samples. There are three common problem if stochastic trend is present.

Problem 1: Autoregressive coefficients that are biased toward zero; problem 2: Non-normal distribution of t-statistics; problem 3: spurious regression. In my case, problem 2 and problem 3 are crucial for my study. Because non-normal distribution means that hypothesis tests cannot be conducted as usual. Also, stochastic trends can lead two time series to appear related when they are not. i.e. the relationship of ln (z-score) and ln (inside debt) can be spurious.

(Augmented) Dickey-Fuller test is a formal statistical procedure to test the hypothesis that there is a stochastic trend in the series against the alternative that there is no trend. I follow

(34)

-34 A. Banerjee, J. J. Dolado, J. W. Galbraith, and D. F. Hendry (1993): Cointegration, Error

Correction, and the Econometric Analysis of Non-Stationary Data; S. E. Said and D. A.

Dickey (1984): Testing for Unit Roots in Autoregressive-Moving Average Models of

Unknown Order- methodology to compute the augmented Dickey-Fuller test. In my test, I set

the lag order is equal to 2 to calculate the test statistic. The result of Dickey-Fuller is equal to -7.38 and p-value is equal to 0.01. It means that the null hypothesis can be rejected, so there is no unit root in ln (z-score). i.e. the relationship between ln (z-score) and ln (inside debt) is not spurious and the standard errors are robust.

6. Link to previous studies and Conclusion

Is there an alternative way to design CEO’s incentive package to reduce risk in banks? Does CEO’s inside debt holding help to reduce their behaviour of excessive risk-taking? My finding confirms this relationship between CEO’s inside debt and risk in banks. It is a positive effect of increasing of CEO’s inside debt on risk in banks. The most robust result that I find is that 0.02% changes on ln (z-score) over time, on average per bank, when ln (inside debt) increase by 1% and it is statistically significant at 10% significance level. This result is selected by comparing many specification models that are pooled OLS estimator, fixed-effects model estimator and random-effects model estimator. As well, this result is tested by many robustness checks such as Hausman test, Breusch-Pagan Lagrange Multiplier test for time effect and augmented Dickey-Fuller test for unit root testing etc. Although this coefficient suggests a positive relationship between CEO’s inside debt holding and z-score, this relationship is inelastic. It means that this relationship is not very economically

significant. By increasing CEO’s inside debt holding as a main tactic to reduce their behavior of excessive risk-taking does not have a salient effect. Therefore, increasing CEO’s inside debt holding can help to reduce risk in bank. However, it is not very economically

significant.

My result is consistent with finding (Srivastav, Armitage & Hagendorff, 2014). They also think that the increase of CEO’s inside debt holding can help to address risk-shifting concerns (they think the increase of bank payout ratio consume most liquid, cash etc. to shareholders, and firms to do not have enough liquid assets to repay debts in the future) and it aligns the interests of CEOs with creditors. Nishikawa et al. (2017) also find that there is a negative relationship between CEO inside debt holding and their behavior of excessive

(35)

risk-35 taking and they propose increase CEO’s inside debt holding could be an effective way to curb the excessive risk-taking behavior by CEOs. Furthermore, my result is consistent with

finding (van Bekkum, 2016) and he find increasing debt-like compensation (inside debt) mitigate bank risk. It also encourages CEO to make conservative investment decision so excessive risk-taking behavior can be curb. Curatola et al (2017) also document the negative relationship between inside debt and bank risk that they used credit spread of credit default swap contracts to measure risk.

Thinking about an alternative way to design CEO’s incentive compensation package? CEO’s Inside Debt : does inside debt help to mitigate CEO’s behavior of excessive risk-taking in banks?