Product differentiation ans CEO turnover

(1)

MSc Business Economics - Finance

Master’s Thesis

Product Differentiation and CEO Turnover

Author: Qian Jia

Student Number: 11086378 Supervisor: Dr Florian S. Peters August 2016

(2)

Abstract

This paper researches the sensitivity of CEO turnover to company performance under different levels of product uniqueness. It uses the new Text-Based Network Industries classification and product differentiation by Gerard Hoberg and Gordon Phillips to measure product uniqueness, and shows evidence to support the argument that turnover-performance sensitivity is lower for firms with more unique products. Poorly performing CEOs are fired less quickly when their companies' products are less similar to those of their rivals in the market.

Statement of Originality

This document is written by Student Qian Jia who declares to take full responsibility for the contents of this document.

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

CEO analysis is never an outdated topic in the corporate finance sector. More and more people focus on the implication of CEO turnover for firm value at some of the famous firms in the world. When the company underperforms, the board of directors always have to decide whether or not to replace a poor performing CEO. Although a negative relation between firm performance and CEO turnover has been well documented by the likes of Jensen and Murphy (1990), we still know relatively little about how turnover decisions are made.

There is some research that links industry to the firm performance-CEO forced turnover sensitivity. Jenter and Kanaan (2015) illustrated industry performance as a key measure behind dismissal rather than market performance. They finally found that CEOs are fired after bad firm performance even when the whole industry has performed badly. According to Parrino (1997), the evidence is consistent with arguments that poor CEOs are easier to identify and less costly to replace in industries that consist of similar firms than in heterogeneous industries. The more homogeneous the industry is, the more likely the CEO will be fired.

However, the puzzle always seems to be the difficulty to find a proper benchmark to measure the industry homogeneity. As Krishnan and Press (2003) puts, even by using the widely-used industry classification by SIC or NIC, there is still a lot to improve. Hoberg and Philips (2015) pointed out that the static classification by SIC can neither create a new product market to accommodate the innovations nor treat the within industry or across-industry relation. Further, SIC and NAICS impose transitivity even though two firms that are rivals to a third firm might not be rivals. In addition, the measurement does

(5)

not reflect the latest situation of a company.

To avoid these limitations, I use the product differentiation technique by Gerard Hoberg and Gordon Phillips to represent industry homogeneity and apply it to CEO turnover studies. They measure firm pairwise product similarity based on the 10-k report from SEC by web crawling and text parsing algorithms and then the word usage vectors from each firm generates an empirical Hotelling-like product market space where all firms reside. Later, they calculate different scores to measure how firms are related to each of their rivals and to create new industry classifications. For each year, according to the update of firms’ 10-K reports, these similarity scores are different. They name this new classification method the Text-based Network Industry Classification (TNIC). In addition, they share the data of pairwise score between firms and all their competitors in different fiscal years, mainly from 1996 to 2013 from the Hoberg-Phillips Data Library.

The goal of my research is to determine the sensitivity of CEO forced turnover rate to company performance within the new industry classification which is related to the uniqueness of products. My hypothesis is that the lower the similarity of the company to its rivals in the market, the lower the sensitivity of the CEO forced turnover to firm performance will be.

I construct a data set of US–listed companies from 1996 to 2013 and use previous one-year stock return before the CEO was announced to leave, to measure the company performance. In addition, I use three methods to measure product uniqueness for each company in different fiscal years when processing the raw data from Hoberg-Phillips Data Library: the average similarity score, number of rivals in the industry and number of close rivals in the industry. After preparing the data, I investigate the research question by

(6)

conducting a probit regression in the full data sample to check the coefficient of the interaction item related to company performance and product uniqueness and univariate regressions in subsamples within different rival proxy.

Related to the findings of Parrino (1997) that the likelihood of forced turnover increases with industry homogeneity, my research supports the argument that forced turnover-performance sensitivity is lower for firms with more unique products. In other words, the more unique products companies have, the less quickly boards will make up their minds to fire the CEO. Economically, this does have logic. When firms have more unique products, it means that there are fewer rivals in the market. Correspondingly, there are no proper benchmarks or “frame of reference” to measure the CEO performance. Further, Jenter and Kanaan (2015) show that the performance of these peer CEO performances have a huge influence on forced CEO turnover. Even if the company underperformed, it is hard to attribute this to the CEO’s failure. For the board, they need a longer period to aggregate the precise information from the market to determine whether current CEOs should be responsible for bad performance, or whether it is just due to bad luck. After that they can decide whether or not to replace the CEO.

On the one hand, as I stated above, the board are often not able to evaluate the performance of CEOs precisely and then they can’t fire the CEO easily. On the other hand, the board would like to maintain the uniqueness of products in the market and make it original. As Farrell and Whidbee (2003) put it, the board makes decisions by expected performance instead of performance itself. This kind of innovation sometimes does not reflect the short term value on profitability. So the board is willing to take a longer period to see the market reactions and then make the decision. For example, after

(7)

Steve Jobs was first fired by the board, he returned to Apple, bringing it with him the new IOS system. This time, the board gave him more time to determine the market reaction, which proved to be a great success.

My findings also have implications for present CEOs. The uniqueness of products reflects the innovation. To some extent, it is wise for executive teams to put some effort into innovation and spend more on research and development (R&D) which seems to be related to their performance.

The rest of the paper is organized as the follows. Section 2 shows the main theories in the existing literature and the relevance of my research to these theories. The research methodology is found in Section 3. Section 4 describes the data. The empirical analysis including the explanation of the main results and the robustness checks are presented in Section 5. Conclusions and implications are summarized in Section 6.

2. Literature Review

Previous studies have presented widespread concern about whether CEOs are appropriately punished for poor performance. While CEOs are more likely to be forced out if their performance is poor relative to the industry average, overall industry performance also matters.

Many scholars have valuable research related to this topic. For example, Taylor (2010) built a dynamic model to evaluate the forced CEO turnover rate and quantify the effects on shareholder value, and this model helps to explain the relation between CEO firings, tenure and profitability. Fredrickson, Hambrick and Baumrin (1988) developed a model of CEO dismissal to show that the board of directors' expectations and attributions, the

(8)

board's allegiances and values, the availability of alternative candidates for CEO, and the power of the incumbent CEO are the four social and political factors which predict the likelihood of a CEO’s dismissal. A preliminary paper by Jenter and Lewellen (2010) examined the significant relationship between firm performances on CEO turnover, accounting for performance included turnovers shortly after the start of a CEO´s tenure. Further, a CEO turnover study by Kaplan and Minton (2012) addressing key findings relating internal turnover to three components of firm stock performance. They are performance relative to industry, industry performance relative to overall market and the performance of the overall stock market.

I will also describe theoretical background explicitly in terms of CEO forced turnover and performance, industry classification and other factors which have proved to affect CEO turnover in the next section.

2.1 CEO forced turnover and performance

Parrino (1997) focused on CEO turnover. He first argued that it is less costly and easier to identify a poor CEO in homogeneous industries. However different kinds of turnover of CEOs, including forced turnover and outside succession turn out to increase the similarity between the firms. Secondly, he argues that this evidence also showed implications of financial economics. It is predicable of the importance of incentive compensation among industries, especially the stock-based incentive compensation, which is more important in places where performance measures are not very precise. In contrast, in places where the performance measures are manageable and it is less costly to replace a CEO, the forcing contracts are more important. Finally, he argues that it is important to get precise performance measures which have a huge effect on the value of

(9)

the firm in the future. My current research question and hypotheses are related to the conclusion by Parrino.

In addition, the firm performance measure is as important as the CEO performance measure. Brickley (2003) believes that existing research has focused on the relation between firm performance measures and CEO turnover when analyzing the declining return in evaluating logit models. He argues that we are supposed to consider other less explored points, for example age-related problems, to enrich our understanding of CEO replacements and turnover.

Obviously, when companies underperform, the board will decide to take action, for example, to fire the current CEO and find a successor. Taylor (2010) mainly argues that there are three main consequences of CEO firing decisions. Firstly, the forced turnover rate is not high in practice, which means there should be large costs of turnover to fit the data. Secondly, instead of reducing the value of shareholder, large turnover is a reflection of CEO entrenchment. The entrenchment of CEOs is harmful for shareholders ex post because the boards of the firm prefer to remain some CEOs which shareholders would prefer to fire them.

When it comes to the concern of whether the punishment is properly taken in terms of the poor performance of CEOs, Eisfeldt and Kuhnen (2013) argue that in a firm where the CEO and the firm are matched based on different kinds of characteristics, complete and performance-oriented turnover can be a natural and effective result in an assignment model that is competitive. They also provided new predictions of equilibrium pay and performance of replaced managers.

(10)

1993 to 2009, and found that CEOs are much more likely to be fired from the firm because of their bad market performance, and to a lesser extent, market return. These are factors beyond the control of CEOs themselves. The performance of peer CEO performances also have a huge influence on the forced CEO turnovers. The underperformance compared to rival CEOs will increase the possibilities of firing.

According to Farrell and Whidbee (2003), it is not enough for the board to judge CEOs only on performance. Using data from five-year EPS growth rate, board pay attention to the deviation from their expectation of company performance.

2.2 Industry Homogeneity Measurement

Generally, the Standard Industrial Classification (SIC) and North American Industry Classification System (NIC) are the most acknowledged methods to conduct industry classification. However, many scholars have admitted the limitations of this method and suggested that the existing classification standard could be used in better ways.

According to Chamberlin (1993) and Hotelling (1929), product differentiation is a fundamental element to a company’s profitability as well as industrial organization. They also showed that product markets can be regarded within a spatial representation that can be explained by product differentiation. Based on this idea that product similarity is essential to industry classification, Hoberg and Philips (2015) developed new time-varying industry classifications using text-based analysis of firm product descriptions filed with the Securities and Exchange Commission.

In their research, they firstly calculate how similar each firm is to every other firm. They use web crawling and text parsing algorithms to measure firm pairwise product similarity

(11)

based on their 10-K reports. They focus on the word usage frequency vector to generate an empirical Hotelling-like product market space where all firms reside. Then, based on the similarities within a three-digit SIC code, they find 21% is a reasonable threshhold to regard the paired companies as rivals. Then they build a new measure of each company and its rivals whose pairwise score is larger than 21% and report the exceeding part as the product score of that company. Companies have different 10-K reports, so in this database, each company has different scores and its rivals in different year. This is called the text-based network industry classification (TNIC), which is analogous to a social network where each firm have a distinct set of competitors. Different industries have their own standards, and it is really important to define industry boundaries and industry competitiveness. Relative to existing industry classifications, their text-based classifications offer economically large improvements in their ability to explain differences in key characteristics such as profitability, sales growth, and market risk across industries. Compared to the original industry classification, this “unrestricted” text-based network industry classification helps capture not only within-industry heterogeneity but also product and industry change as well as cross-industry relatedness.

2.3 Other factors affecting CEO turnover

CEO turnover rate will be influenced by different kinds of factors related to the characteristics of the firm itself as well as CEOs themselves.

Companies of different sizes will also lead to different sensitivity of CEO turnover to performance. Murphy (1999) found that in small companies such as S&P Mid-Cap companies, this sensitivity is higher than that in big companies such as the S&P 500. Meanwhile, the significance of age of CEO in explaining the CEO turnover is higher in

(12)

bigger companies.

Further, CEO age is another necessary factor for the board to take into consideration when deciding the CEO turnover rate. Sometimes it is more important in explaining CEO turnover rate than the measures of firm performance. Murphy (1999) suggests that the likelihood of a CEO leaving his/her firm during the year is almost 30 percent higher when the CEO is more than 64 years old.

CEO gender also matters. Martin, Nishikawa and Williams (2009) evaluated the

influences of gender in capital market measures and found that it affects the valuation and risk of CEO appointment. For female CEOs, changes in risk following CEO

appointments are significantly lower than those for male CEOs. They support the view that the market perceives female CEOs as relatively risk averse, while firms with relatively high risk (total risk and idiosyncratic risk) are more likely to appoint female CEOs so that risk might decrease.

3. Hypothesis

My hypothesis is that turnover-performance sensitivity is lower for firms with more unique products, which means that a poor CEO is less likely to be fired quickly when the company product is more unique.

When the board begin to make up their mind, the performance measures of a company are one of the key factors to take into account. When a company has such unique products that it can’t compare with other rivals, it seems difficult to measure the company performance. As Parrino (1997) mentioned in his paper that “It tends to take longer to identify a poor CEO at a firm where the available performance measures are noisy than at

(13)

a firm where precise measures are available”, and it is hard to attribute the company’s underperformance to the CEO’s own failure. Because the board don’t have more precise information than the market which will help them to decide whether or not to replace the current CEO, it would take longer for them to recognize the situation clearly. Further, the reason may come from many areas, and this may just be due to bad luck. As a result, I believe that the more unique a company’s products are, the lower the likelihood of CEO forced turnover rate will be when the company is underperforming.

4. Methodology

The objective of the thesis is to determine the sensitivity of CEO forced turnover to a company’s performance under different levels of product uniqueness.

Considering the current data that I can access, I aggregate the data that can measure how unique a company’s products are from the “Hoberg-Phillips Data Library”, which is a free online database1. The 10-K Text-based Network Industry Classifications (TNIC) database, lists the pairwise score between each company and its competitors in different fiscal year. A higher score indicates a higher degree of similarity and firm pairs with a higher score are closer rivals. In addition, I will collect the data related to the characteristics of CEOs from the Execucomp database, while the company’s annual fundamental information and the monthly security data related to the stock return to show a company’s performance can be obtained from CRSP COMPUSTAT. As for the CEO forced turnover rate data, I would like to thank for my thesis supervisor Florian S. Peters for sharing his dataset to support my research. Further, all of the above data can be easily merged in Stata by the exact items.

(14)

As for the empirical methodology, I conduct the probit regression as the empirical methodology. In statistics, a probit model is a type of nonlinear regression model which is specifically designed for the binary dependent variable, which means it can only take two values (0 or 1). Because a regression with a binary dependent variable Y models the probability that Y equals to 1, it makes sense to adopt a nonlinear formulation that forces the predicted values to be between 0 and 1. In addition, because cumulative probability distribution functions produce probabilities between 0 and 1, it can be used in probit regression.

The purpose of the model is to estimate the probability that an observation with particular characteristics will fall into a specific category; in addition, if estimated probabilities greater than 0.5 are treated as classifying an observation into a predicted category, this probit model is also a type of binary classification model. As a result, it can solve the problem well.

To clarify, a logistic model could also be considered. It is similar to the probit regression model except that the cumulative standard normal distribution function is replaced by the cumulative standard logistic distribution function. However, in general, they frequently produce similar results, even if there are some differences these differences are not very great. So, this may suggest that “we can basically flip a coin when it comes to deciding whether to go the logit route or the probit route”, as the one of most influential expert of Econometrics David E. Giles said. Here I just use the probit regression model.

I use two different methods of probit regression to test this hypothesis. The first is applied to the full samples and the regression equations are shown as following:

(15)

where Pr denotes probability of 𝑦_𝑖𝑡 and Φ is the Cumulative Distribution Function of the standard normal distribution.

The response dependent variable 𝑦𝑖𝑡 denotes the CEO forced turnover which is binary,

that is it can have only two possible outcomes that we will denote as 1 and 0 to represent whether the CEO will be fired or not. 𝑟_𝑖𝑡−1 and 𝑥_𝑖𝑡are the regressors which are assumed to influence the CEO turnover rate. 𝑟_𝑖𝑡−1 denotes the stock returns for one year which represent the performance. Because there is always a gap between the occurrence of CEOs being fired and the announcement of this turnover decision from the board. Here I use one-year stock returns before the CEO turnover decision was announced. This seems to explain the how“𝑟𝑖𝑡−1” can match “𝑦𝑖𝑡”. In addition, 𝑥𝑖𝑡 denotes the level of

product similarity of a company which is obtained from the “Hoberg-Phillips Industry Classification Library” to illustrate how similar a product is compared to the rivals’ products in the market. In brief, this represents the uniqueness of the company’s products. 𝑟_𝑖𝑡−1∗ 𝑥_𝑖𝑡 is an interaction which shows the influence of product uniqueness on the company’s performance. Then 𝑤𝑖 denotes control variables, such as the size of a

company, age of CEO, gender of CEO and so on.

Each coefficient, such as 𝛽₁, 𝛽₂, 𝛽₃ and 𝛽₄ , shows the sensitivity of probability of CEO forced turnover to the corresponding independent variables. For example, the probit coefficient 𝛽1is the change in the z-value associated with a unit change in 𝑥𝑖𝑡. If this 𝛽1

is positive, an increase in the product similarity score increases the z-value and thus increases the probability that 𝑦_𝑖𝑡 equals 1, which means the probability of the CEO being fired. In other words, this shows that the relation between this independent variable, product similarity and CEO forced turnover rate is positive. Otherwise, it is a negative

(16)

relation. Based on the existing research, many scholars have already documented that there is a negative relation between firm performance and the likelihood of CEO turnover. We can easily expect that 𝛽2 should be negative. So the coefficient 𝛽3 shows to what

extent the uniqueness of a company’s product influences on firm performance. In my research, 𝛽₃ is the one that I’m interested in as it shows the sensitivity of CEO forced turnover rate to the company’s performance under different levels of product uniqueness. Let’s say, if this 𝛽₃ is negative as well, it shows that the more similar product a firm has, the more negative performance will be influenced as 𝛽₂ tends to be more negative and, as a result, the probability of forced turnover rate to be 1 will increase. So I will focus on this coefficient 𝛽3 to see if the hypothesis is reasonable that the more unique the

products of a company are, the less quickly the board will make up their mind to fire the CEO.

As for the second method to continue the study, I will divide the full samples into two subsamples by measuring the level of product uniqueness. Then I will do the regression in different subgroups. The main regression equation is shows as follows:

Pr(y_it) = Φ[α + β₁r_it−1+ β₂w_i+ ε_i]

The basic idea is similar to the above. Here I do the regression within two groups that the contain the product in two levels - more unique and less unique - rather than use the interaction item. In group 1 with the product scores above median, let’s say less unique product sample set, I will obtain the coefficient 𝛽₁₁ to show the sensitivity of CEO forced turnover to the company’s performance when the company has a less unique product. In group 2 with the product score below median, which is regarded as the more unique product sample set. A similar coefficient 𝛽₁₂ can be obtained by this regression

(17)

model. Then the difference between this 𝛽₁₁ and 𝛽₁₂ obviously shows the turnover-performance sensitivity under different levels of product uniqueness. According to my hypothesis, 𝛽12 should be lower than 𝛽11, which means turnover-performance

sensitivity is lower for firms with more unique products.

Possible sources of endogeneity in this research may mainly come from different aspects. According to the current research, CEO turnover rate can be affected by a couple of factors, such as board composition (Weisbach, 1988), and CEO stock ownership (Salanick and Preffer 1980). In addition, other characteristics of CEOs may influence the turnover decision. For instance, an older CEO is more experienced than a young CEO, who may, however, turn out to be more motivated than the older one. Other endogenous characteristics include CEO gender, CEO educational background, etc. Secondly, the features of the company may also exert an impact on performance as well. These features are the size of the company, the location of the company, or the leverage ratio of the company etc.

To avoid the threat of endogeneity, all these mentioned items should theoretically have been included in the control variable in the regression. But considering the approach of the data and the applicable matched data within the whole dataset，I just use the firm size, CEO age and CEO gender as the control variable in my research.

5. Data and Descriptive Statistics

5.1 Data Preparation

In my research, I need four pieces of data. The first is CEO forced turnover data which can be obtained from my thesis supervisor Dr. Florian Peters. The second is the measure

(18)

of product uniqueness from Hoberg-Phillips Data online Library’s Text-based Network Industry Classifications (TNIC) database. The third is related to company’s stock returns which can evaluate firm performance by using the CRSP/ COMPUSTAT merged database, both the dataset of Fundamentals Annual and Security Monthly. The last tells us CEO characteristics from Compustat Executive Compensation database – Annual Compensation dataset.

In particular, the first data piece only concludes the announcement year of CEO forced turnover which are listed from 1992 to 2014, while the Hoberg-Philips dataset only gives the product similarity score of publicly-traded firms which have a 10-k report from the US Securities and Exchange Commission (SEC) from 1996-2013 (updated version). Taking the data validity into account, I aggregate the data of listed companies of United States from 1992 to 2015 from CRSP /Compustat and the Execucomp database.

5.1.1 CEO Forced Turnover

As I mentioned above, I highly appreciate Florian Peters for sharing this dataset with me. It contains the dates of forced CEO turnovers (not voluntary ones) of all firms recorded in the ExecuComp database between 1993 and 2012. The file is a merged with an extended version of data collected by Jenter. and Kanaan. (1993-2001) and Peters. and Wagner. (2001-2010). The construction of the dataset has also benefited from turnover data generously provided to us by Greg Nini, Luke Taylor, Cami Kuhnen, Andrea Eisfeldt, and Ofer Eldar. It contains the company identifier (gvkey), the CEO identifier, the executive’s first and last name, the last fiscal year in which the executive is recorded as the CEO in ExecuComp (as indicated by the variable “CEOann” taking the value “CEO”), the announcement date of the CEO turnover (including year, month and day seperately).

(19)

The announcement date indicates the first time the turnover was mentioned in the press and thus may be several months before the actual departure of the CEO. In addition, it is the performance before the announcement date that influence the board to make the decisions. So it seems to be more reasonable to use the company performance that is one year before the announcement date instead of that before the exact leaving date. A missing announcement date indicates that the exact date could not be ascertained. The criteria for classifying a CEO turnover as forced are identical for both subsamples and are described in detail in Jenter. D and Kanaan. F (2015), and Peters. F and Wagner. A (2014). The number of forced turnovers in the attached dataset for the 1993-2009 period is slightly higher than that used in Peters. F and Wagner. A (2014) due to an upgrade of the data following the publication.

5.1.2 Product Uniqueness

This data is based on web crawling and text parsing algorithms that process the text in the business descriptions of 10-K annual filings on the SEC Edgar website from 1996 to 2013 (latest version). We can see that the similarity score can be used to identify which rivals are "nearer" rivals than others. A higher score indicates a higher degree of similarity and firm pairs with a higher score are nearer rivals. It should also be noted that the scores in this database indicate the amount by which the pairwise score exceeded the similarity threshold as 21% for being included in the TNIC database. According to Hoberg and Philips (2015), if the similarity score of a company with another is more than 21%, based on the SIC‘s three digit standard, they can be regarded as in the same industry, otherwise the two companies’ products have less similarities which cannot make them rivals. As a result, this database only shows the pairwise companies that are in the same

(20)

industries.

According to the basic explanation of how this database works, I come up with three methods to measure the product uniqueness of a company by using the raw data.

The first is to calculate the average score of each company among all the competitors in the corresponding year as the author puts in the instruction part of the TNIC database2. Then for each firm (gvkey1) in each fiscal year, we get only one average score among all the pairwise score with all its competitors. As a result, the higher average score shows the similar products a company offers to the market compared with all other rivals. The second method is to calculate the number of rivals for each company in each year. The more competitors a company has in the market, the less unique products it produces. The third method of uniqueness combines the first two measurements, taking both the number of rivals and the raw product score into consideration. The raw pairwise score has a median of 0.0448 and mean of 0.0639 (shown in the table of summary descriptive statistics). So, I choose to use 0.05 as a benchmark. For example, for a company A, I count the number of rivals only including companies for whom the pairwise score is larger than 0.05 and use this number as a measure of company A’s product uniqueness. Within this score benchmark, companies less similar should not be considered as rivals. The larger this number is, the less unique the company’s products are

5.1.3 Company performance

Generally, the decision for the board to fire a CEO is always based on the company’s previous performance. According to the current research, stock return is widely used to

2

“If you wish to control for an industry characteristic using TNIC industries, the easiest way is to use an average across related firms (a kernel-approach).” -Technical Note 4 from Hoberg-Phillips Data Library

(21)

measure the firm’s performance, which equals the sum of capital gains and dividends. Shareholders care about the stock price as well as the return on capital. It is possible that a company earns a higher return on capital but shareholders will still suffer if the market price of the stock decreases. Similarly, a lower return on capital may be accompanied with a high stock price, which shows the positive expectation from the investors in the market. From the perspective of board, returns on stock seem to be more related to their interests and benefits.

Here I will use the stock return of previous year or twelve months as a measure of company performance. Due to the gap from the announcement date and the occurrence of CEO forced turnover, these two events may happen in different fiscal years. To avoid this, it is more precise to take the twelve-month stock returns instead of yearly stock returns that are directly gathered from the Annual Fundamental dataset. I use the twelve months’ stock returns up to three-months before the exact announcement month. The initial stock price, ending stock price and dividends can also be obtained from the CRSP/ COMPUSTAT merged database – Security Monthly database. Based on the formula that stock return is the appreciation in the price plus any dividends paid, divided by the original price of the stock, we can get the stock return for each month and than calculate the returns for one-year.

The change of ROA is another method to measure the performance of a company, which shows the short term profitability of a company. So I use the accounting formula3 that ROA equals operation income minus depreciation and then divide by total assets to get the ROA for each fiscal year. To make sure that this change of ROA exactly corresponds

3

(22)

to the current CEO’s term, I will use the change of ROA in the previous year before the CEO was fired, otherwise, the performance may reflect on the successor(s).

In addition, in view of other features of company which may also influence the CEO forced turnover rate, I download the fundamental data from the CRSP/ COMPUSTAT merged database – Fundamental Annul dataset as another control variable. For instance, I will use the natural logarithm term of total asset (AT) to measure the firm size, which is also consistent with the research by Koyuncu (2010) and Raghuram (1995) that firm size is measured by the natural logarithm of annual sales.

5.1.4 CEO characteristics

CEO characteristics are mainly used as control variables in the regression model. In my research I specially use CEO age and CEO gender which can easily obtained from the Compustat Executive Compensation – Annual Compensation dataset.

Moreover, I create two dummy variables related to CEO age and CEO gender to run the regression. Generally speaking, 60 is the most-wildly-used threshold to measure age. I set the age above 60 as 1, and below 60 as 0. Similarly, I generate a gender dummy as well: the gender of males is replaced by 0, and 1 denotes all the female CEOs.

5.2 Descriptive statistics

[Insert Table 1 Here]

Tables 1 shows the descriptive statistics of all the variables that I use in this paper. From 1992 to 2014, there are 39854 CEO were fired. In my sample, there are 886 female CEOs and 38968 male CEOs and gender may be a factor that affects the decision of the board.

(23)

In terms of the age of CEO, there is a wide range from 27 to 96. The average age of CEOs is 56. In general, CEOs are not as as young.

[Insert Figure1 Figure2 Figure3 Here]

As for the uniqueness of a company’s product, besides the statistical data, I also plotted the distributions of the product uniqueness by different measurements respectively in Figure 1, Figure 2 and Figure 3. We can see that the distributions are similar to each other. The majority of companies have a small average similarity score between 0 and 0.2, and 0 to 0.1 make up the highest percentage. The average similarity score has a mean of 0.03 and the median is 0.028. In terms of the amount of rivals, the lowest number of rivals a company has is only one and the highest is 901. On average, a company has 96 rivals. Half of companies have less than 46 rivals and the rest have more than 46 rivals. The number of close rivals is less than the number of rivals. This amount is within 100 for most companies. Only a small number of firms have more than 200 close rivals. There is a trend in the product similarity score. As time passed by, the scores got smaller, which means that companies are producing more unique products.

From the company side, when it comes to its previous one-year stock return, the most underperforming company only has -0.92 but the company who had the best performance got a stock return of 4.31. The average level is 0.19 and the median is 0.12. In terms of the ROA, the difference is not so big. The smallest one-year change of ROA is -1.3, while the highest is 1.47. The median is 0. No matter which measurement is used to evaluate the company performance, both of the two variables have small standard deviations, which show that the distribution is relatively concentrated and there are less outliers. In addition, the data set has different sizes of companies. The range of the firms sizes is

(24)

from 1.49 to 14.76, which means that this dataset contains companies with different sizes and it reflects reality.

6. Empirical Analysis

6.1 Main results

6.1.1 Full sample regression with interaction items

Table 2 presents coefficient estimates for probit regressions which show the correlation between the CEO forced turnover rate and the main independent variables. Column one reflects a univariate regression between CEO forced turnover rate and one-year stock return. The coefficient of one-year stock return is -0.035 and it is significant below 1% significance level which indicates that there is a negative relation between CEO forced turnover rate and a company’s one-year stock return, which is consistent with Jensen and Murphy (1990) as well as Murphy and Zimmerman (1993) who reported that the worse the company performance, the more likelihood CEO turnover will be.

In the next nine columns, every three columns are grouped based on the measurement of product uniqueness.. We do the regression three times for each group adding different control variables as well as year fixed effects. We can see that all the coefficients of the interaction items are negative, which means that the performance-CEO forced turnover sensitivity is lower when the company has more a unique product in the market compared to its rivals. To make this clear, the more similar a company’s products seem to be, the more the negative effect will work on the one-year stock return. Based on the negative relation between the company performance and CEO forced turnover rate, the probability

(25)

of the CEO being fired increases.

It also seems that the firm size doesn’t really affect the board’s decision of keeping the current CEO or not. The coefficients of firm size are all very small and insignificant. As for the effect of CEO age, the coefficient is negative and significant, which means that if the CEO is older than 60 years old, he or she would be less likely to fired, while the likelihood of young CEOs being fired is relatively large.

Column 2 to column 4 use average similarity scores as the measurement of product uniqueness, we can see that the coefficients of the interaction item (average similarity score * one-year stock return) are all negative. This means that as the average similarity score increases by one, the coefficient of one-year stock return will be more negative to -0.1749, -0.0870 and -0.0636 respectively, which equals the coefficient of interaction item and the one of one-year stock return, for instance, -17.49 equals -0.0319 plus -0.1430. So the likelihood of CEO forced turnover rate will increase by 17.49%, 8.7% and 6.36%.

However, all of these three tests are not significant at least at the 10 % level. In my point of view, this is because the range of average similarity scores are from 0 to 1, while half of the stock returns are around 0.1 to 0.2, this interaction item doesn’t have a strong effect for the whole regression. In addition, from the robust z-statistics in brackets, we can see that it is almost close to the statistical significance level of 10%.

In the second group where the product uniqueness is measured by number of rivals in the market, we can see that the coefficient of the interaction items are all negative and strongly significant with respect to confidence levels of 5%, 1% and 5%. To begin with column 5, this shows that a one-unit decrease of one-year stock return will increase the

(26)

likelihood of CEO forced turnover by 3.24%. But when the company has more unique products, let’s say if the number of rivals in the market decreases by 1%, the total probability of the CEO bring fired will decrease from the original 3.24% to 2.81%. Adding control variables as shown in column 6, this probability also decreases from 1.71% to 1.29%. With year fixed effect, the result is similar to the previous test, as the products are more unqiue, and the performance-forced turnover sensitivity will be reduced to 1.46%.

Furthermore, columns 8 to 10 show how the performance-turnover sensitivity changes when the product uniqueness is evaluated by the number of close rivals. The coefficients of the interaction items are negative, which is consistent with the previous result. The more unique products a company has, when the company underperforms, the executive board will take longer to decide whether to fire the current CEO . Columns 8 and 9 both can prove this relation. Column 10 takes into account all the control variables and year fixed effects, a 1% increase of number of close rivals will decrease by 0.8%, which leads to the probability of CEOs being fired by the board increasing from 1.78% to 2.58% when the one-year stock return decreases by one unit. This is significant at a 1% confidence level.

6.1.2 Univarite regression on subsamples

In the second main regression, I divided the full sample into two data sets of close rivals and distant rivals. Similarly, the measurement of rival proximity is still the same as before, by average score, number of rivals and number of close rivals.

(27)

Table 3 indicates the simple correlation between one-year stock return and CEO forced turnover rate in both close rivals and distant rivals by three different measures of rival proximity. In the first two columns where I use the average similarity score to judge the uniqueness of products, the coefficient of one-year stock return in the close rival group means the average similarity score is below the median level in column 1 is less than that in the distant rival group in column 2. This illustrates that for companies with more similar products in the market, when the one-year stock return decreases by 1, the probability of CEOs being fired will increase by 3.79% while for a company with more unique products than rivals in the market, this probability will be 3.08%. In a similar way, the probability of CEOs being fired from companies with more rivals in the market is 3.66% when there is 1 unit one-year stock return decrease, which is higher than that companies with less rivals in the market (3.28%). The same story occurs when products are measured by the number of close rivals.

Table 4 presented the regressions, including control variables with year fixed effects. It is obvious that the performance-CEO turnover sensitivity is higher in the close rival group than the distant rival group. Having more unique products in the market with lower average similarity score, a one-unit decease of one-year stock return will increase the likelihood of forced turnover rate by 1.7%. While for a company with more similar products offered to the market, this marginal effect could be 2.14%. When we turn to use the number of rivals of a company to measure its product uniqueness, if one-year stock return decreases by one, a CEO of the company whose products seem to be more unique will be 1.62% less likely to be fired. However, the CEO of a company which has more

(28)

close rivals and provides the market with more similar products will have 2.07% probability of being fired when the one-year stock return is reduced by one unit. The same trend happens when we measure the product uniqueness by number of close rivals. The marginal effect in both close rival groups and distant rival groups are respectively 1.89% and 1.67% with companies’ underperformance.

From the above, no matter which regression I use to test the sensitivity between the company performance and CEO forced turnover rate under different levels of product uniqueness, the estimates can still prove the hypothesis that turnover-performance sensitivity is lower for firms with more unique products as well as backing up the conclusion of Parrino R (1997) that the likelihoods of forced turnover increase with industry homogeneity.

6.2 Robustness Checks

6.2.1 One-year change of ROA

Considering that stock return is not the only measurement of company performance, I replace one-year return with another variable, one-year change of ROA, to present a company’s performance so that we can verify if the result changes. To clarify, in order to avoid the mismatch between the CEO and the company’s performance, which means to make sure the performance can exactly reflect the CEO’s term instead of the successor’s, I use the previous one-year change of ROA which equals the difference of last year’s ROA and that of two years before. Here I still take the multinomial probit regression with the interaction item to see how the uniqueness of company products could influence companies’ performance as well as CEO forced turnover rate.

(29)

Table 5 presents the correlation between the one-year change of ROA and the CEO forced turnover rate using different measurements of product uniqueness. The first column still has the same negative relation as we obtained before. A 1 unit increase in the one-year stock return will decrease the probability of CEOs being fired by 5.97%. Compared to the one-year stock return, this coefficient is smaller, which shows a more obvious negative relation between the change in ROA and CEO forced turnover rate.

6.2.2 Dummy variable to define one-year stock return

Jenter and Lewellen(2010) plotted the relation between the probability of CEO turnover rate and company performance decile in their paper. We see that within the lower part of the performance, the curve is steeper and sharp, while the curve of probability of CEO turnover is smoother in the rest of the performance decile part. So I use the dummy variable to represent the company performance to make the relation clear. I define the lower 20% of one-year stock return as 1 which shows that stock return is low and the rest are set to 0. By replacing the one-year stock return with the dummy variable, I still do the similar multinomial probit regression as the last one.

Table 6 illustrates the performance-forced turnover sensitivity in different measures of product uniqueness by using one-year stock return as the dummy variable. Because 1 denotes the lower one-year stock return, from table 6 we can see that the coefficients of this dummy variable are negative which means the relation between that and the CEO forced turnover is negative, which is consistent with previous studies. The coefficient of

(30)

interaction items in these regressions are all positive and show that the more similar products a company has, when the company underperforms, the probability of CEOs being fired is increased.

7. Conclusion and Implication

There has been a quite lively discussion about corporate governance since the early eighties which started with a paper by Jensen and Murphy (1990) on CEO compensation and CEO turnover. The paper makes an argument that incentives for CEOs are way too small compared to the value increase of the company brought by CEOs. Since then, thousands of papers have been written about CEO turnover. And lots of laws passed, including the Sarbanes–Oxley Act. These have helped to improve corporate governance. The idea is that these regulations would increase the strength of independence of the boards, enabling then to fire CEOs for poor performance. From all these reports, we can expect that the frequency of CEO dismissal should have increased because these laws have become much stricter and much more attention has been focused on good governance. Thus we could expect that CEO turnover would increase over the last twenty years. But it hasn’t actually increased as much as expected.

This paper studies the sensitivity of CEO forced turnover rate to company performance under different levels of product uniqueness. I use data on listed companies of the United States from 1996 to 2014 to do a probit regression. The findings show that the likelihood of forced turnover increases with similarity of companies’ products when the company underperforms. Based on the innovated product differentiation database by Gerard Hoberg and Gordon Philips., I proved that the forced turnover-performance sensitivity is lower for firms with more unique products, which means that a poor CEO is less likely to

(31)

be quickly fired when the company's product is more unique.

This also indicates the economic implications. On the one hand, there is a limit to firing the CEO, which is the ability to benchmark CEO performance. Even if the board is really independent and strong enough, sometimes it is hard for them to make a decision to fire the CEO. This is because it is hard to evaluate their performance due to the lack of proper benchmarks. On the other hand, product differentiation has increased over time. The score data from Hoberg and Philips also indicates that over time uniqueness of firms’ products has increased. This would make it harder for boards to evaluate performance of a CEO. As for a companies with more unique products, it is more difficult to evaluate the quality of the company’s performance, in that there are not enough market participants as reference points for the board to get precise information from the market to evaluate the performance of the CEO, so it is tough to figure out the exact factors that lead to the underperformance of the company. Sometimes a company’s underperformance is not the CEO’s failure and the current CEO is not supposed to be responsible for it. Then it will take longer to decide whether or not to fire a CEO when a company doesn’t have a good performance. In addition, to some extent, the board would like to retain the innovation and product uniqueness and they give the CEO more time to examine the market reaction instead of the performance alone.

As a result, there were two trends, one is that the board becomes more independent and stronger which would lead to more firing. However, also, product differentiation has become stronger and greater. The same board has a harder time evaluating CEOs. These two forces balance each other. Therefore, we don’t see more firing today than we saw ten or fifteen years ago. Product differentiation has placed some limits on board decisions or

(32)

governance in general.

However, there are still some ways in which my research which could be improved. For example, this variable of similarity score is still a choice which can generate endogeneity. Maybe there is another variable that can influence this sensitivity of forced turnover to performance. At present, it seems that this expected variable plays a role in the similarity score and then indirectly affects the independent variable. In addition, the board’s unwillingness to fire CEOs who have unique products created in their tenure could be proved in a more convincing and empirical way.

(33)

References

Brickley J.A.(2003). “Empirical Research on CEO Turnover and Firm-Performance: A Discussion”, The Journal of Accounting & Economics, vol.36(1-3),pp.227-233

Eisfeldt A.L. and . Kuhnen C.M,. (2013) “CEO turnover in a competitive assignment framework”, Journal of Financial Economics: pp.351–372

Farrell K.A. and Whidbee D.A (2003) “Impact of firm performance expectations on CEO turnover and replacement decisions”, Journal of Accounting and Economics vol.36, pp.165-196

Fredrickson J W., Hambrick D.C. and Baumrin,S.(1988). “A Model of CEO Dismissal”, The Academy of Management Review: Vol. 13, No. 2, pp. 255-270

Hoberg, G. and Philips G. (2015). “Product Market Synergies and Competition in Mergers and Acquisitions: A Text-Based Analysis”, Review of Financial Studies 23 (10), 3773-3811.

Hoberg, G. and Philips G. (2015). “Text-Based Network Industries and Endogenous Product Differentiation”, Journal of Political Economy, Accepted for Publication

Huson M.R, Parrino R and Starks L.T (2001). “Internal Monitoring Mechanisms and CEO Turnover: A Long-Term Perspective”, The Journal of Finance: Vol. 56, No. 6, pp. 2265-2297

Jensen, M.C. and Murphy, K.J. (1990). “Performance pay and top-management incentives”. Journal of Political Economy 98, 225-264. 

(34)

Jenter D. and Kanaan F. (2015) “CEO Turnover and Relative Performance Evaluation”, The Journal of Finance: vol. lxx

Jenter D. and Lewellen K.(2010) “Performance – Induced CEO Turnover”, preliminary paper.

Kaplan, S.N. and Minton B.A.(2012) “How was CEO Turnover Changed”, International Review of Finance, vol.12, pp.57-87

Martin A.D, Nishikawa T and Williams M.A. (2009). “CEO Gender: Effects on Valuation and Risk”, Quarterly Journal of Finance and Accounting Vol. 48, No. 3, pp. 23-40

Murphy K.J. (1999). “Executive compensation”, In: O. Ashenfelter and D. Card (Eds.)”, Handbook of Labor Economics. Elsevier, Amsterdam, pp. 2485–2563.

Parrino, R. (1997). “CEO turnover and outside succession: A cross-sectional Analysis”, Journal of Financial Economics: pp165-197

Taylor, L.A,.(2010) “Why Are CEOs Rarely Fired? Evidence from Structural Estimation”. The Journal of Finance: Vol. 65, No. 6, pp. 2051-2087

(35)

Figure 1

This figure draws a histogram of all the average similarity score for each company in different year. And it is drawn as fractions so that the sum of the height of the bars equal to 1.

Figure 2

This figure draws a histogram of all the number of rivals for each company in different year. And it is drawn as fractions so that the sum of the height of the bars equal to 1.

(36)

Figure 3

This figure draws a histogram of all the number of close rivals for each company in different year. And it is drawn as fractions so that the sum of the height of the bars equal to 1.

Table 1 Descriptive Statistics

This table shows the descriptive statistics of all the aggregated variables. For each variable, it is described by the minimum value, mean, median, maximum value, standard deviation and the number of observations. All the values are reported with two decimal fractions.

Variables Min Mean Median Max SD N

CEO Forced Turnover 0.00 0.03 0.00 1.00 0.16 39837

Average Similarity Score 0.00 0.03 0.03 0.82 0.03 31483

Number of Rivals 1.00 95.64 46.00 901.00 133.68 31483

Number of Close Rivals 1.00 41.89 10.00 700.00 95.44 26815

One-year Stock Return -0.92 0.19 0.12 4.31 0.57 35513

One-year change of ROA -1.30 0.00 0.00 1.47 0.07 35549

Firm size 1.49 7.58 7.45 14.76 1.77 38677

(37)

Table 2

This table reports the result of probit regression where the dependent variable is the CEO forced turnover rate.1 means the CEO was fired in that year while 0 shows there is no turnover that year. Independent variables include one-year stock return, three measurement of product uniqueness (average similarity score, number of rivals and number of close rivals) and the three interaction items between the stock return and product uniqueness. Additional control variables include firm size, CEO age (1 denotes CEO who is older than 60) and CEO gender (1represents male CEO, while 0 stands for female CEO). From column2 to column10, every three columns are grouped based on the measurement of product uniqueness.Robust z-statistics are prested in brackets, *** p<0.01, ** p<0.05, * p<0.1

Dependent Variable: CEO forced turnover dummy

Measure of rival proximity Average similarity score Number of rivals Number of close rivals

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

One-year stock return -0.0350*** -0.0319*** -0.0190*** -0.0192*** -0.0324*** -0.0171*** -0.0176*** -0.0337*** -0.0172*** -0.0178*** [-12.53] [-5.67] [-4.73] [-4.99] [-8.75] [-5.60] [-5.35] [-9.09] [-5.88] [-5.61] Avgerage similarity score -0.0395 0.0039 0.0035

[-0.68] [0.11] [0.11]

Average similairy score * 1yr stock return

-0.1430 -0.0680 -0.0444 [-0.98] [-0.76] [-0.61] Firm size 0.0002 -0.0001 0.0002 -0.0003 0.0006 0.0002 [0.54] [-0.31] [0.50] [-0.68] [1.41] [0.41] Age>60 -0.0069*** -0.0064*** -0.0069*** -0.0064*** -0.0068*** -0.0063*** [-5.25] [-5.31] [-5.27] [-5.32] [-4.81] [-4.90] Male -0.0053 -0.0016 -0.0052 -0.0015 -0.0010 0.00198 [-1.38] [-0.50] [-1.36] [-0.48] [-0.25] [0.60] Number of rivals -0.0000 -0.0001 0.0005 [-0.04] [-0.23] [1.20]

Number of rivals * 1yr stock return

-0.0043** -0.0042*** -0.0030**

[-2.07] [-2.70] [-2.04]

Number of close rivals -0.0021 -0.0025*** -0.0017*

[-1.514] [-2.578] [-1.805]

Number of close rivals * 1yr stock return

-0.0090** -0.0109*** -0.0080***

[-2.40] [-3.45] [-2.70]

Observations 36,232 29,617 28,638 27,080 29,617 28,638 27,080 25,137 24,311 22,990

(38)

Table 3

This table reports the result of the univariate probit regression where the dependent variable is the CEO forced turnover rate.1 means the CEO was fired in that year while 0 shows there is no turnover that year. One-year stock return in the only independent variable. Every two columns are grouped based on the measurement of product uniqueness. And in each group, it is divided by the rival proximity. Robust Z-statistics are presented in brackets. *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels respectively.

Table 4

This table is aligned with Table 3, but including control variables. Additional control variables include firm size, CEO age (1 denotes CEO who is older than 60) and CEO gender (1represents male CEO, while 0 stands for female CEO). Every two columns are grouped based on the measurement of product uniqueness. And in each group, it is divided by the rival proximity. Robust Z-statistics are presented in brackets. *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels respectively.

Panel B: Regressions including control variables

Depedent Variable CEO forced turnover dummy

Measure of rival

proximity Average similarity score Number of rivals Number of close rivals Close rivals Distant rivals Close rivals Distant rivals Close rivals Distant rivals (1) (2) (3) (4) (5) (6) One-year stock return -0.0214*** -0.0170*** -0.0207*** -0.0162*** -0.0189*** -0.0167*** [-7.88] [-4.12] [-8.20] [-4.15] [-6.77] [-4.33] Firm size -0.0002 -0.0002 0.0002 -0.0013** -0.0002 -0.0002 [-0.55] [-0.37] [0.63] [-2.39] [-0.44] [-0.32] Age>60 -0.0063*** -0.0061*** -0.0055*** -0.0068*** -0.0059*** -0.0065*** [-4.83] [-3.24] [-4.25] [-4.04] [-4.77] [-3.25] Male 0.0006 -0.0033 0.0029 -0.0065 -0.0033 0.0021 [0.17] [-0.65] [0.87] [-1.42] [-0.88] [0.43] Observations 18,093 12,927 18,006 13,699 21,642 10,063

Year Fixed Effects Yes Yes Yes Yes Yes Yes

Panel A:Univariate Regressions

Depedent Variable CEO forced turnover dummy

Measure of rival

proximity: Average similarity score Number of rivals Number of close rivals Close rivals Distant rivals Close rivals Distant rivals Close rivals Distant rivals (1) (2) (3) (4) (5) (6)

One-year stock return -0.0379*** -0.0308*** -0.0366*** -0.0328*** -0.0356*** -0.0339***

[-11.69] [-7.02] [-11.06] [-6.93] [-11.88] [-6.22]

Observations 21,334 14,898 21,159 15,073 23,860 12,372

(39)

Table 5

This table reports the result of probit regression where the dependent variable is the CEO forced turnover rate.1 means the CEO was fired in that year while 0 shows there is no turnover that year. Independent variables include one-year change of ROA, three measurement of product uniqueness (average similarity score, number of rivals and number of close rivals) and the three interaction items between the one-year change of ROA and product uniqueness . Additional control variables include firm size, CEO age (1 denotes CEO who is older than 60) and CEO gender (1represents male CEO, while 0 stands for female CEO). From column2 to column10, every three columns are grouped based on the measurement of product uniqueness. Robust Z-statistics are presented in brackets. *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels respectively.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

One-year Change of ROA -0.0597*** -0.0387* -0.0170 -0.0226 -0.0544*** -0.0206 -0.0227 -0.0577*** -0.0248* -0.0274** [-5.12] [-1.84] [-1.09] [-1.59] [-2.73] [-1.29] [-1.55] [-3.48] [-1.87] [-2.20] Average similarity score -0.0320 0.0148 0.0177

[-0.567] [0.455] [0.656]

Average similarity score * 1yr Change of ROA

-0.7170* -0.6000** -0.4620* [-1.78] [-2.09] [-1.89] Firm size 0.0002 -0.0002 0.0001 -0.0004 0.0007 0.0002 [0.29] [-0.33] [0.20] [-0.74] [1.27] [0.45] Age>60 -0.0088*** -0.0079*** -0.0087*** -0.0078*** -0.0090*** -0.0082*** [-5.28] [-5.13] [-5.23] [-5.09] [-4.85] [-4.78] Male -0.0090* -0.0049 -0.0093* -0.0050 -0.0045 -0.0006 [-1.85] [-1.14] [-1.89] [-1.17] [-0.83] [-0.14] Number of rivals 0.0007 0.0006 0.0014*** [0.97] [1.30] [2.99]

Number of rivals * 1yr Change of ROA

-0.0074 -0.0135* -0.0117

[-0.61] [-1.66] [-1.58]

Number of close rivals -0.0014 -0.0012 -0.0003

[-1.07] [-1.50] [-0.46]

Number of close rivals * 1yr Change of ROA

-0.0183 -0.0303*** -0.0251**

[-1.04] [-2.63] [-2.32]

Observations 35,549 29,209 28,253 26,679 29,209 28,253 26,679 24,735 23,933 22,601

(40)

Table 6

This table reports the result of probit regression where the dependent variable is the CEO forced turnover rate.1 means the CEO was fired in that year while 0 shows there is no turnover that year. Independent variables include lower one-year stock return which is a dummy variable (1 represents that it is in the lower 20% of the whole return sample, 0 shows that it is beyond the 20% standard of the lower stock return), three measurement of product uniqueness (average similarity score, number of rivals and number of close rivals) and the three interaction items between the lower stock return and product uniqueness. Additional control variables include firm size, CEO age (1 denotes CEO who is older than 60) and CEO gender (1represents male CEO, while 0 stands for female CEO). From column2 to column10, every three columns are grouped based on the measurement of product uniqueness. Robust Z-statistics are presented in brackets. *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels respectively.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Lower one-year stock return 0.0475*** 0.0415*** 0.0254*** 0.0260*** 0.0457*** 0.0274*** 0.0274*** 0.0502*** 0.0292*** 0.0287*** [21.23] [10.35] [8.04] [9.14] [14.85] [11.12] [11.71] [16.80] [12.13] [12.63] Average similarity score -0.0858 -0.0343 -0.0179

[-1.27] [-0.70] [-0.45]

Average similarity score * Lower 1yr stock return

0.1460** 0.0924* 0.0507 [2.09] [1.78] [1.20] Firm size 0.0008* 0.0004 0.0008* 0.0002 0.0012*** 0.0007* [1.88] [1.03] [1.76] [0.60] [2.69] [1.70] Age>60 -0.0076*** -0.0069*** -0.0076*** -0.0069*** -0.0075*** -0.0069*** [-5.33] [-5.24] [-5.29] [-5.22] [-4.78] [-4.76] Male -0.0044 -0.0008 -0.0046 -0.0009 -0.0002 0.0028 [-1.12] [-0.25] [-1.15] [-0.26] [-0.05] [0.81] Number of rivals -0.0002 -0.0003 0.0006 [-0.27] [-0.60] [1.11]

Number of rivals * Lower 1yr stock return

0.0028** 0.0021** 0.0010

[2.15] [2.27] [1.16]

Number of close rivals -0.0016 -0.0024** -0.0013

[-1.04] [-2.50] [-1.43]

Number of close rivals * Lower 1yr stock return

0.0032 0.0046*** 0.0023

[1.52] [3.03] [1.64]

Observations 36,232 29,617 28,638 27,080 29,617 28,638 27,080 25,137 24,311 22,990

Product differentiation ans CEO turnover

MSc Business Economics - Finance

Master’s Thesis

Product Differentiation and CEO Turnover

Table of Contents

Table 1 Descriptive Statistics