MSc Finance - Corporate Finance
Master Thesis
Forced CEO turnovers and relative performance
evaluation: A product market-based analysis
Author: Muhan HU 11390891
Thesis Supervisor: Dr. Florian Peters
This document is written by Student Muhan Hu 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.
I would like to express my gratitude to my supervisor Dr. Florian Peters, who kindly provides data, the idea as well as useful advice for my thesis and being supportive during the whole procedure.
1 Abstract:
A rational board should determine CEO ability base on its performance relative to its peers according to Holmstrom (1982). Using hand-collected forced CEO turnover data from 1996 to 2015, this thesis tests whether the board of directors and eliminate systematic performance uncertainty when evaluating CEO performance. I apply a new industry classification, the text-based networking industry classification, to define the peer firms’ performance benchmark used for performance evaluation. Empirical results suggest that the board fails to entirely filter out the impact of product market common trend. Moreover, we predict that in a less concentrated market, peer performance should be better information for firm’s own performance, thus, peer return component and firm idiosyncratic return should have higher predictive power. However, increase in product market concentration does not necessarily decrease the significance of peer firm performance information.
A. Introduction ... 1
B. Literature review ... 4
a. Determinants of CEO replacement ... 4
b. Relative performance evaluation and peer selection ... 4
c. Product market competition and the text-based industry classification ... 7
C. Data and Methodology ... 9
a. Data ... 9
b. Methodology ... 11
D. Results ... 14
a. Descriptive statistics ... 14
b. Two stage regression of forced turnovers on peer and firm specific return ... 17
c. Two stage regression using TNIC HHI quintiles ... 20
d. Two stage regression using TNIC total similarity quintiles ... 22
e. Test the weak-form RPE hypothesis ... 24
E. Robustness Check ... 25
a. Using only 10 closest product market peers to compute peer benchmark return .... 26
b. Change performance variable to accounting measure: ROE ... 28
c. Use Fama-French 48 industry SIC to compute peer returns ... 30
d. Use voluntary CEO turnover as dependent variable ... 30
F. Conclusion and Discussion ... 33
1
A. Introduction
The number of CEO turnovers has increased drastically during the past few decades. Hiring and firing CEOs are essential issues in corporate governance and have triggered numerous research. The seminal study of Relative Performance Evaluation (RPE) by Holmstrom (1982) provoked numerous following empirical studies; most empirical studies of RPE focus on the application in determining executive compensation and some also focus on the relation between RPE and CEO turnovers. However, there is no consensus that whether boards eliminate systematic shocks when evaluating CEOs skills. This paper intends to investigate whether CEOs are fired due to poor management practice relative to their product market peers or weak product market performance, using the text-based network industry classification (TNIC).
Relative performance evaluation is firstly proposed by Holmstrom (1982), he applies a principal-agent model to explore moral hazard problem in teams. The model suggests that with information asymmetry, other firms’ performance can provide useful information in assessing agent firm’s own performance if they face same common uncertainties. He introduces a rational practice in corporate governance that the board of directors should utilize comparable peer firms’ information and use relative performance evaluation when making critical decisions about CEO compensation and retention decisions. Academia further makes a distinction between the weak-form relative performance evaluation and the strong-form relative performance evaluation (Albuquerque, 2009). The weak-form suggests that the peer performance should be negatively associated with the CEO dismissal probability while firms’ performance should have a positive relationship with the dismissal frequency. The strong-form, on the other hand, suggests the coefficient of systematic firm performance in the Antle and Smith (1986) model be zero. The seminal work of Holmstrom (1982) motivated much research in both executive remuneration designation and CEO retention choices. However, in respect of CEO dismissal decisions, there is a lack of consensus in empirical findings. Differences in peer group definitions and identification strategies could explain the variation among empirical results. In a recent paper on RPE in forced CEO turnovers, Jenter and Kanaan (2015) apply a two-stage model to examine the RPE hypothesis using data from 1993 to 2009. The estimated SIC peer return component in their estimation is significant in predicting CEO dismissal frequency, suggests that the board fails to
2 exclude noises from the market condition in evaluating CEO performance. However, Albuquerque (2009) suggests that the lack of consensus in current empirical studies of relative performance evaluation results from the inappropriate composition of the peer group. Therefore, this thesis intends to apply a novel classification, the Text Based Industry Classification, of the peer group to examine the relative performance evaluation in CEO retention decisions.
The main hypothesis of this thesis is whether the board can fully exclude product market-wide noises when assessing CEO performance and making CEO retention decisions. This thesis intends to use the same two-stage model of Jenter and Kanaan (2015). They use the peer group return as an effective instrument variable to cleanse the noise from the peer's common shock. More specifically, in the first stage OLS regression, peer firm return is used to predict the firm return. The linear estimation is stored as product market induced return component and the residual is kept as the firm idiosyncratic return component. The second stage probit regression uses those two return components to estimate the CEO dismissal probability. According to the RPE hypothesis, we expect the coefficient of the firm idiosyncratic return component to be negative and significant while the coefficient of the product market induced return component to be insignificant. In addition, I also try to identify whether product market structure, in terms of market concentration and product similarity, will lead to a difference when the board processing peer performance information. To investigate aforementioned hypotheses, I utilize two unique datasets. The first one contains forced CEO turnover information from 1993 to 2014, this dataset is kindly provided by Peters and Wagner (2014) and Jenter and Kanaan (2015). I manually go through web information for CEO turnovers taken place in 2015 and identify forced turnovers among all turnover observations, then extend this dataset to 2015. The second essential dataset is the TNIC dataset obtained from the Hoberg and Phillips Data Library. This classification scheme specifies a certain group of peer firms for each particular firm and scores the closeness between each firm-peer pair using product information on the firms’ 10K filings. Compared with conventional SIC measure, the TNIC is more dynamic and has a more narrowly defined peer group. Moreover, the TNIC HHI and total product similarity data are also included in the Hoberg and Phillips Data Library. Firm and stock price data are downloaded from Execucomp and CRSP respectively.
3 Using the data and identification strategy described above, my analyses find results consistent with Jenter and Kanaan (2015), which suggests that change the peer group from SIC peers to TNIC peer yields same result. In the first stage, all estimated coefficients show that product peers return is highly significant in predicting firm’s own return. It suggests that product market peer return is a relevant instrument. Nonetheless, in the second stage, both firm idiosyncratic return and product market common return components are significant. Namely, the board of directors does not distinguish the product market profitability tendency and CEO contribution when assessing firm performance. The marginal effect of the firm idiosyncratic component on non-voluntary CEO turnover is more pronounced in comparison with systematic return component. In addition, I split the sample into quintiles based on their product market concentration as well as product similarity. The regression results, in general, are in violation of RPE hypothesis as well and we fail to detect any tendency among quintiles. This indicates that product market structure does not have a material impact on how the board makes CEO retention decisions. Further investigation regarding the robustness of previous findings is done by change dependent variable, independent variable, and peer group definition. These tests further confirm our findings.
The contributions of this research are two folds: 1) it applies most up to date hand-collected data set of forced CEO turnovers which range from 1996 to 2015. Comparing to data contain all turnovers, this data can avoid bias from including voluntary CEO turnovers to CEO dismissals; 2) this is the first study that applies TNIC to relative performance evaluation in CEO turnover decisions. This industry classification scheme provides insight about how product market matters in CEO retention decisions. More generally, it can make a connection between studies of industrial organization and corporate governance. Moreover, TNIC can better capture the dynamic of the changing business environment and is more flexible compared to conventional measures.
This thesis proceeds as follow: section B provides a comprehensive review of related literature on RPE and TNIC; section C describes the data and methodology, section D presents empirical results concerning the research question; section E reports several robustness checks to confirm previous findings, the last section includes further discussion and a conclusion of this topic.
4
B. Literature review
a. Determinants of CEO replacement
The number of CEO turnovers has increased during the past few decades while the number of CEO being fired remains at a low level. There are two main arguments about CEO turnovers as indicated in Jenter and Kanaan (2015). Firstly, firms want to hire talented CEO that can outperform the market; therefore, CEOs with relatively poor performance with respect to their peers are more likely to be fired. Secondly, firms choose the CEO who fits best to their corporate but not necessary the one with highest qualifications. Hence, as the firms evolve to different stages, their requirements for CEO also changes and the most fitted one will replace the current CEO. In addition, there is evidence that boards sometimes are reluctant to change their CEOs as the replacement costs are considerably high. The costs of CEO replacement include but not limited to entrenchment cost of the board of directors (Taylor, 2010), negative signal about board ability and the raise in the cost of capital (Dow, 2013). To be economically sound, the foreseeable benefit of CEO succession must exceed the cost of the replacement. Moreover, Garrett and Pavan (2012) proposed a dynamic theory for managerial turnovers. Their model suggests that the turnover decision varies as management’s tenure. The replacement decision would become more conservative if the CEO is in the office for a longer time. In another word, firing CEO is more likely to happen at the early stage, ceteris paribus. Among those famous theoretical frameworks of executive turnovers, the thesis has a main focus on the agency theory (Holmstrom, 1982).
b. Relative performance evaluation and peer selection
Holmstrom (1982) studies moral hazard problem in a team. He finds that, with asymmetric information, peer firms’ average performance can provide useful information about the stage of the uncertainty of the agent firm. Therefore, the efficient method for the board of directors to evaluate firm performance is to compare it with that of its peer firms’, which faces common uncertainty as the agent firm. Holmstrom and Milgrom (1987) also prove that aggregate performance of comparable peer firms in a sufficient statistic condition in providing information for the CEO of the agent firm’ action. Hence, such information should be included in evaluating CEO’s performance. Their findings suggest that an effective board should use RPE i.e. excluding market-wide
5 and industry-wide common shocks in firm performance, when assessing the CEO’s ability. The efficiency improvement comes from two aspects (Antle and Smith, 1986). Firstly, filter out common shock can better reveal the actual ability of the agent; and secondly under the assumption that the CEO is risk averse, this filtering reduces CEO’s exposure to market risk without harming his incentives. Two of the most prominent implications of the RPE theory is on CEO compensation and dismissal decisions. Recent research extensively studies whether firm filter out industry and market common shocks when design executive compensation and replacement decisions. The following section provides a systematically summary on both topics.
RPE in CEO turnovers
Holmstrom’s (1982) seminal work provoked numerous empirical research regarding the relationship between CEO turnovers and firm’s relative performance. However, these works can only provide weak, or at least, mixed evidence. After eliminating noisy reasons for CEO turnover, such as normal retirement and death, Coughlan and Schmidt (1985) find that the likelihood of CEO turnovers is associated with firm’s market performance in terms of stock return. Barro and Barro (1990) study a sample of large U.S. commercial banks from 1982 to 1987. They find that the RPE is more pronounced when assessing relative performance using stock return than using accounting measures. Moreover, they find the incumbent CEO’s performance could be an alternative benchmark in RPE. Gibbons and Murphy (1990) provide significant evidence consistent with RPE. They find that the probability of CEO dismissals decrease with the firm’s relative performance but increase with the industry and market aggregate performance. Additionally, they suggest that the board of directors is more like to assess CEO performance based on the market performance instead of industry performance. DeFond and Park (1999) adopt sample from 1988 to 1992 and find evidence in favor of RPE when the industry competition intensity is high. They suggest that the lack of consensus in previous studies might due to the inattention to the intensity of completion in the market.
Later works include more recent observations also cannot find strong evidence for RPE hypothesis. For example, Engel, Hayes and Wang (2003) study CEO turnovers from 1975 to 2000. In contrast to Barro and Barro (1990), they suggest that accounting measures
6 receive more weight in CEO replacement decision comparing to the market measure. The reason is that accounting measures are more sensitive while market performance is more variable. Kaplan and Minton (2012) applied turnover data from 1992 to 2007 and find three important factors in the board-driven CEO replacement decision. Firm relative performance, industry relative performance and the aggregate market performance all have impacts on the likelihood of CEO turnovers. The trend is stronger in the most recent period after 2000. Nonetheless, voluntary and non-voluntary turnovers show similar results regarding RPE, which indicates that some classification of turnovers might be flawed.
Jenter and Kanaan (2015) examine recent data in forced CEO turnovers from 1993 to 2009 using a hand-collected forced CEO turnover sample. Their results only provide weak evidence for RPE, namely, the board cannot fully filter exogenous industry and market-wide shock when making replacement decisions. Furthermore, they find that there are more likely to present forced CEO turnovers if the average industry stock return is low. Using the same hand-collected sample from Jenter and Kanaan (2015); Eisfeldt and Kuhnen (2013) study CEO dismissal from 1992 to 2006 under the competitive assignment model. Their findings are consistent with Jenter and Kanaan (2015) that both firms’ accounting and market returns relative to the industry are negatively related to the probability of CEO dismissal. Moreover, they argue that the failure to filter exogenous shock does not necessarily mean that the board is inefficient. Outside options which provide a better match of firm and manager can also drive CEO turnovers.
RPE in CEO compensation
In parallel with studies focus on CEO turnovers, there is extensive research investigating whether the peer performance is the benchmark in CEO compensation. Early work by Antle and Smith (1986) finds mixed evidence regarding CEO compensation from 1947 to 1977. While accounting return measures, such as ROA, shows that CEO compensation is consistent with RPE, they find market measure, such as stock return, fails to provide similar evidence. Gibbons and Murphy (1990) investigate both CEO compensation and CEO turnovers in respect of RPE. Their empirical results show that boards use RPE for designing CEO compensation package and the benchmark for performance evaluation is
7 more likely to be market aggregate performance instead of industry performance. With a particular focus on Japanese firms, Joh (1999) studies a sample of 796 firms from 1968 to 1992. He asserts that previous mixed evidence might come from strategic collusions in oligopoly. RPE based on industry performance could hamper the benefit from strategic collusive action, therefore, RPE should also base on relative industry performance. However, this is not the case in a competitive market, where the possibility of collusion is low. In an influential paper, Albuquerque (2009) suggests that the lack of consensus in recent empirical studies of RPE is due to the inappropriate composition of the peer firms. He uses firms from the same industry and with similar size as the agent firm and finds supportive evidence that firms use RPE in designing CEO compensation package.
In sum, empirical studies found mixed results for RPE in CEO turnovers. This could result from inappropriate measures for performance (Engel, Hayes and Wang, 2003; Barro and Barro, 1990), selection of peer group (Gibbons and Murphy, 1990; Albuquerque, 2009), inattention to market competition (DeFond and Park, 1999; Joh, 1999) and errors in non-voluntary turnovers identification (Gibbons and Murphy, 1990). c. Product market competition and the text-based industry classification
Recall the agency theory asserted in Holmstrom (1982); peer performances are only informative when peer firms have same uncertainty as for the agent firm and the benefit from peer benchmarking increases in the number of peers. Empirical results on RPE could be misleading if peer groups are inappropriately defined (Gibbons and Murphy, 1990; Albuquerque, 2009). For instance, Albuquerque (2009) argues that change in equity value result from variation in firm earnings and firm’s discount rate applying dividend discount model. More specifically, either shock in revenue or cost could lead to variations in earning. In addition, as proposed by Fama and French (1992, 1993), market risk sensitivity, firm size and market-to-book ratio are factors that will simultaneously affect the firm’s discount rate. By adopting different peer selection criteria, Albuquerque (2009) finds the strongest evidence on RPE in CEO compensation when constructing peer firms that in the same industry and have similar firm size as the agent firm. It is noticeable that most previous studies of relative performance use industry classification such as SIC and NAICS. There are several limitations regarding
8 those classifications. For example, those classifications are static hence cannot reflect the most up-to-date industry composition. Further, firms with diversified business are restricted to one industry subgroup. To overcome the aforementioned limitations, Hoberg and Phillips (2010, 2016) propose a dynamic measure for industry classification, which is the Text-based Network Industry Classifications (TNIC). It uses text-based product description in the 10-K filing to define a firm’s competitors and it is updated on a yearly base. The TNIC captures more information regarding firms’ key attributes, for instance, industry-wide risk, profitability and sales growth compared with conventional static classifications. According to Hoberg and Phillips (2010, 2016), there are several advantages of TNIC compared with SIC. For instance, this classification simultaneously identifies agent firm’s similarity to its peers and the level of competition for the agent firm. According to previous research, the competition intensity within the industry plays a crucial role in RPE in respect of both CEO turnover (DeFond and Park, 1999) and CEO compensation (Kim, 1996). Thus, the TNIC arguably could be a useful tool for RPE studies.
So far, there is only limited number of research that has applied this novel classification in RPE. Nonetheless, its applications seem appealing. For instance, Hoberg and Phillips (2010) find that firms that are similar in their product market are more like to achieve merger and acquisition transactions and post-M&A profits are great if the target and the acquirer have similar product market language, especially in a competitive market. In addition, Jayaraman, Milbourn, and Seo (2015) adopted this classification for relative performance evaluation in terms of CEO compensation. While prior research on relative performance evaluation in CEO compensation are mixed and only show weak evidence, they find strong evidence that firms filter out common shock when evaluating CEO’s ability. Similarly, my intention is to apply this new classification for relative performance evaluation in terms of forced CEO turnover.
One of the limitations of TNIC could be that no industry is defined under this system; product market similarity only applies to certain company pair instead of a group of different companies. Although there are data available for Text-based Fixed Industry Classifications, Hoberg and Phillips (2016) assert that this data are subject to many restrictions. DeFond and Park (1999) suggest that level of competition within the industry has impact on frequency of CEO turnovers. Kaplan and Minton (2012) show
9 that board driven (internal) turnovers are significantly relative with the industry return relative to the overall market return. However, under the TNIC measure, there is no specific common industry or market defined; hence, it is hard to capture the competition intensity for a particular market as well as the relative performance of a specific industry performance to the overall market performance.
C. Data and Methodology
a. DataThis thesis studies whether the forced CEO turnovers of U.S. public companies during the period of 1996 to 2015 follow predictions by relative performance evaluation theory. There are three main data sources. First of all, CEO turnover data are obtained from Compustat ExecuComp and the forced turnovers data are provided by Peters and Wagner (2014) and Jenter and Kanaan (2015). Secondly, I retrieve the TNIC industrial classification data from Hoberg and Phillips Data Library. Finally, year-end stock prices are obtained from the Center for Research in Security Prices (CRSP). This section provides detailed information for all these datasets.
All CEO data from 1993 to 2016 come from Standard and Poor’s ExecuComp which contains executive information for more than 3462 companies. We identify the CEO of each fiscal year using the variable “CEOANN”. I record an occurrence of CEO turnover if the CEO for the same entity changed over two sequential years. The last year the executive serves as the CEO is recorded as the turnover year. This sample includes 41,614 firm-year observations, among which 4599 turnover events are identified. Forced CEO turnovers data from 1993 to 2014 are kindly provided by Peters and Wagner (2014) and Jenter and Kanaan (2015). As an extension, I hand collected the forced CEO turnover data for fiscal year 2015. The criterion for turnovers to be classified as non-voluntary involves several steps. Firstly, the turnover events for year 2015 are identified from ExecuComp. Secondly, manual search via Wall Street Journal is conducted to identify CEO successions in which the incumbent is reported fired, forced out, and resigned without specific reasons are classified as forced. As a third step, CEOs leaving office not due to reasons described in step 2, and at the same time are under the age of 60 are reexamined using related press releases. More details about the classification are
10 documented in Parrino (1997). This identification procedure is essential for CEO turnover studies since CEOs are merely fired publicly. This increases the forced turnover data to 1194constituting 3.2% of the total turnovers during the period of interest. The text-based industry classification, as well as firms’ TNIC Herfindahl-Hirschman Index (HHI) and total similarity measures from 1996 to 2015 can be obtained from the Hoberg-Phillips Data Library. This industry classification contains all U.S. domestic firms listed in one of the three exchanges, namely NYSE, NASDAQ and AMEX. The fact that this dataset is updated on an annual base suggests the advantage of its dynamic nature over traditional classifications. It is constructed using the text-based description of firms’ product market composition from their 10K filings on the SEC Edgar website. The TNIC assigns a score ranging from 0 to 1 to each firm-peer pair according to their closeness of product market compositions, with 1 indicating the highest similarity. More detailed descriptions of the data can be found in Hoberg and Phillips (2010 & 2016).
Security prices of sample firms are downloaded from monthly CRSP tape. Accounting measures for sample firms, such as total assets, total liabilities, and net income etc., are available from Compustat Fundamental Annual database. All preliminary data from 1993 to 2015 are obtained. As the forced CEO data is only available from 1993 to 2015 and TNIC data is from 1996 to 2015, the final dataset contains firm-year observations from 1996 to 2015. Table 1 presents details of the data processing process when different data sources are merged together. The final dataset contains 34433 firm-year observations, among which 3798 are identified as turnovers and 987 CEOs are further classified as “forced out”.
Table 1. Sample selection process
Time period CEO-Year observations
All CEO-year observations 1993-2015 41,614
Identified CEO turnovers 1993-2015 4599 (11.5% of all observations) Forced CEO Turnovers 1993-2015 1185 (2.85% of all observations) Keep only observation from 1996-2015 1996-2015 37306
Firm fundamentals 1996-2015 37306 (0.35% unmatched)
Firm return 1996-2015 37306 (4.98% unmatched)
Peer return 1996-2015 35042
Drop observations with missing peer return 1996-2014 35042 Drop observations with missing firm return 1996-2014 34433
11 b. Methodology
There are three hypotheses this thesis tries to test. The first one is the strong-form relative performance evaluation hypothesis pioneered by Antle and Smith (1986). The second hypothesis incorporates the effects from product market structure in the strong-form RPE hypothesis. The last hypothesis is the weak-strong-form hypothesis adopting the model proposed by Holmstrom and Milgrom (1987). Detailed explanations for identification strategies for each hypothesis will be discussed in the following sub-sections.
Hypothesis 1: Strong-form RPE in CEO replacement decision
The seminal estimation model for strong-form RPE was developed by Antle and Smith (1986) to evaluate whether CEOs are compensated based on their relative performance. Similar models are applied by Bertrand and Mullainathan (2001), Garvey and Milbourn (2006) and Jenter and Kanaan (2015). This two-stage model structures as follow:
1st Stage:
ri = 𝛽0+ 𝛽1𝑝𝑒𝑒𝑟𝑟𝑒𝑡𝑖 + 𝑣𝑖 2nd Stage:
Prob(Forced CEOTurnover) = Φ(γ0+ γ1𝑟̂i+ γ2𝑣i+ ∑ γn× control variablesn n
)
The dependent variable is the forced CEO turnover dummy. Since the dependent variable is binary, in this case, linear probability model does not fit well. Hence, we use probit model for estimation, with Φ(∙) the distribution function of normal distribution. The first stage effectively uses peer performance as an instrumental variable to estimate firm performance, and then use the estimated coefficients to separate the firm total return into firm-specific return 𝑣𝑖 and industry common return 𝑟̂, where 𝑟̂ = 𝛽̂1+ 𝛽̂2× 𝑟𝑝𝑒𝑒𝑟. The rationale behinds this two-stage model is that the first stage can be considered as a market model with the market return be replaced by the peer group return. The estimated return 𝑟̂ of a particular firm is only determined by the peer group return as this model assumes all firms have the same sensitivity to their peer group return. Hence,
12 𝑟̂ can represent the return contributed by market/industry condition. The difference between firms’ actual return and the estimated return is unrelated to peer performance, thus, the residual 𝑣𝑖 is contributed by the top executive’s management skill.
In the second stage, if the RPE hypothesis is true, the industry common return should not have any predictive power over CEO turnovers (γ1 = 0) and the probability of forced CEO turnover should be negatively associated with firm idiosyncratic stock return (γ2 < 0). The major difference of this regression model and those used in previous studies lies in the computation for the variable peer return. The weighted average peer firm returns, as well as the equally weighted peer firm return, are calculated based on firm’s product market peers defined by Hoberg and Phillips (2010). Hypothesis 2: Peer and CEO idiosyncratic effects are more pronounce in competitive market
To investigate the effects of market concentration on forced CEO turnovers probability, the whole sample is split into quintiles according to their TNIC Herfindahl-Hirschman Index (HHI). This product market HHI is a firm-specific concentration measure, and it is available on Hoberg-Phillips Data Library. This measure utilizes both public firms’, as well as private firms’, information to compute the ratio, thus, it is arguably the best estimation for market concentration (Hoberg and Phillips, 2010). Lower HHI indicates a less concentrated product market, which also means more similar products and a higher degree of competition within the product market. As in a market with close substitutes, information about peer firm is more indicative to predict agent firm’s performance. In turn, CEO contributed performance can be more accurately measured and should also provide more information about CEO’s management skill. Conversely, firms with higher HHI face less competition and fewer close products in the market. Therefore, peer performance can explain little variation in firms’ own performance and CEO component of return is harder to capture in such cases. Hence, we should expect that after splitting the quintiles, less concentrated quintile’s peer performance should have a larger impact on firm performance and be more significant in predicting firm return. Namely, the peer
return coefficient β1 in the first stage regression should be decreasing as HHI measure getting larger. At the same time, the t-statistics should also decrease with the HHI quintile.
13 𝛽1,𝑞1> 𝛽1,𝑞2 > 𝛽1,𝑞3> 𝛽1,𝑞4 > 𝛽1,𝑞5
As lower HHI can predict market component of firm return more accurately, we expect that the residual for low HHI quintile in the first stage is a better estimate for CEO contributed firm performance. Hence, in the second stage probit regression, CEO
component should be decreasing as HHI measure gets larger, meanwhile the t statistics
should also decrease with the HHI quintile.
𝛾2,𝑞1 > 𝛾2,𝑞2 > 𝛾2,𝑞3 > 𝛾2,𝑞4> 𝛾2,𝑞5
The Hoberg-Phillips Data Library also provides a product market total similarity measure based on firm’s TNIC peers. Higher product total similarity indicates a large amount of close substitutes and hence market performance is a better signal for firm performance. When there are few comparable products in the market, peer firm performance provides less useful information for firm’s own performance. Therefore, in contrast to HHI quintiles, these product market total similarity quintiles should have increasing coefficients and t statistics for both peer returns in the first stage and CEO contributed component in the second stage regression.
Hypothesis 3: weak-form RPE in CEO replacement decision
In addition to empirical investigation for strong-form RPE hypothesis, this thesis also tests the weak-form RPE hypothesis. The estimation framework for weak-form RPE was originally proposed by Holmstrom and Milgrom (1987) in testing RPE in CEO compensation. Following Gibbons and Murphy (1990) and Barro and Barro (1990), in this thesis the weak-form RPE is examined using the following one-stage specification:
𝑃𝑟𝑜𝑏(𝐹𝑜𝑟𝑐𝑒𝑑 𝐶𝐸𝑂 𝑇𝑢𝑟𝑛𝑜𝑣𝑒𝑟𝑠)
= Φ (β0+ β1ri+ β2peerreti+ ∑ βncontrol variablen)
Under the weak-form RPE, firm returns should be negatively associated with the probability of CEO dismissals (β1 < 0), ceteris paribus. Conversely, peer returns should positively associate with the probability of CEO dismissals (β2 > 0), ceteris paribus. The positive relation between peer benchmark and CEO dismissal is because that when firm’s own performance is fixed, the higher the benchmark, the worse the firm’s
14 performance relative to the benchmark. Thus, if the board utilize relative performance evaluation, the worse performance relative to the benchmark should increase the dismissal probability.
Control variables
Control variables for the regression models mentioned above include CEO at retirement
age dummy and CEO with high stock ownership dummy. Previous studies found CEO’s
age is a key determinant of CEO turnovers (Murphy and Zimmerman, 1993; Weisbach, 1988). Following Jenter and Kanaan (2015), CEO at retirement age is a dummy variable, which equals to 1 if the CEO is at the age between 63 and 66; CEO with high stock
ownership is a dummy variable, which equals to 1 if the CEO owns more than 5% of the
firm’s total stock. Without controlling such variables, the estimated coefficients might be biased since they include effects from the CEO’s age and stockownership.
D. Results
a. Descriptive statistics
The final dataset contains 34433 firm-year observations, 3798 among those are identified as CEO turnover and 987 are classified as forced CEO turnover. We generate a turnover dummy and a forced dummy, which equal to 1 if a CEO turnover or force turnover takes place respectively. Table 2 is the summary statistics of this study which presents median and mean of key company characteristics of three subgroups, including retained CEO without turnovers, voluntary CEO turnovers and forced CEO turnover. CEO is retained if the turnover dummy equals zero, CEO is left voluntarily if the turnover dummy is 1 while the forced dummy is 0, and CEO is forced out if the forced dummy is 1. All numbers are reported in million USD, except ratios. Firm characteristics reported include total assets, total liabilities, and the book value of shareholder equity, the market value of shareholder equity, net income, market return and some major accounting ratios. Total asset, total liabilities, shareholder equity and net income are directly downloaded from Compustat. ROA and ROE are calculated using net income divided by average total assets and shareholder equity respectively. The computation of market return is the cumulative 12-month rounding return of fiscal year end for firms without force turnover and cumulative 12-month rounding return of 3 months prior turnover
15 announcement for forced turnovers. The annual return is computed 3 months ahead of the turnover announcement in order to remove the effect of price movements due to market rumors about the succession plan. Return of individual peer firms is calculated as the same manner as the firm market return. The agent firm’s peer firm weighted return is the weighted average individual peer firm return, where the weight depends on their closeness in the product market using TNIC data. The equally weighted return of peers is the arithmetic return of all peers for the firm.
Table 2 Descriptive Statistics
This table shows the summary statistics of the CEO turnover dataset. The table presents key firm characteristics, accounting return, market return and CEO age by turnover outcomes. The total assets, total liabilities, book value (BV) shareholder equity, market value (MV) of shareholder equity and net income are reported in million USD. Both median and mean are reported. N represents the number of observations; the variation is mainly due to missing accounts.
Factor Retained CEO Voluntary turnovers Forced turnovers N Median Mean N Median Mean N Median Mean Total assets 30634 1817 14238 2890 1937 13570 987 1471 23246 Total liabilities 30566 977.8 11250 2882 1052 10595 983 779.8 20023 BV of equity 30634 668.7 2928 2890 705.2 2903 987 488.5 3216 MV of equity 26328 1565 7902 2513 1628 8975 851 824 6402 B/M ratio 26328 0.5 0.5 2513 0.5 0.5 851 0.6 0.4 Net Income 28257 61.8 349.7 2722 59.0 333.4 922 1.2 -86.4 ROA 28252 4.8% 4.1% 2722 4.3% 2.5% 921 0.2% -5.8% ROE 28252 11.7% 6.3% 2722 11.1% 6.1% 921 1.6% -8.2% Market return 30635 11.2% 16.8% 2890 6.3% 11.0% 987 -14.9% -12.2% Peer return 30635 10.8% 12.9% 2890 10.1% 12.0% 987 6.5% 7.2% EW peer return 30635 11.8% 13.8% 2890 11.0% 13.0% 987 7.0% 8.8% Age 30419 55.0 55.4 2462 61.0 60.7 636 54.0 54.1
As demonstrated by Table 2, the mean firm size regarding total assets suggests that firms with forced turnovers are on average larger than the other two subsamples; while the median of total assets shows the opposite. The same finding is also true for total liabilities and shareholder equity that while the median number indicates forced turnover sample firms are relatively smaller compared to the retained CEO and voluntary turnover, the average number shows the contrast. The average net income for retained CEO observations is $349.7 million and average net income for voluntary turnovers is $333.4 million; both positive. However, the average net income for forced turnovers observations is a loss of $ 86.4 million. Therefore, force turnover firms generally have more negative profit compared with firms without forced turnover.
16 Moreover, there is a clear pattern for all profitability measures (ROA, ROE, and market return) for three subsamples. In terms of ROA, the average ROA is 4.1%, 2.5% and -5.% for retained CEO, voluntary turnover CEO and forced CEO respectively. In respect of ROE, the average ROE is 6.3%, 6.1% and -8.2% for retained CEO, voluntary turnover CEO and forced CEO respectively. These two accounting returns are negatively skewed for all three groups of observations. This means there are large negative outliers in the sample. The market return is winsorized at 1% level. Since the negative market return is limited by -1, the return distribution is right skewed for all samples. The retained CEO sample has the highest market return, 16.8%, while the forced turnover sample has the lowest market return, -12.2%. The peer firm average returns are calculated using winsorized return data. The summary statistics show that equally weighted peer return is close to but slightly higher than score weighted peer return. The average weighted peer return is 10.8% for retained CEOs, 12.0% for voluntary step-down CEOs, and 7.2% for forced out CEOs. Although the peer returns also decrease among these three groups, the variation is smaller than firm’s own return. In terms of CEO characteristic, the average top executive age for retained CEO and forced out CEO are 55.4 and 54.1, which is close. The average age for voluntary turnover CEOs is 61, which is an acceptable age for retirement.
Table 3 is a correlation matrix for selected variables. Panel A presents the correlation
matrix of several firm profitability measures, which includes firm’s market return, peer firm’s weighted and equally weighted market return, firm’s accounting ratios and net income. All these variables are positively correlated. The weighted and the equally weighted peer firms’ return have the highest correlation coefficient of 0.942. Both these returns of peer firms are more correlated with firms’ market return compared with firms’ own accounting return. In terms of accounting returns, ROA is more correlated with firms’ market return compared with ROE, and ROA is also more closely connected with net income compared with ROE. Panel B shows the correlation between turnover probability and several CEO characteristics. Only the CEOs serve as a director on the board is negatively correlated with both CEO turnovers and forced CEO turnovers. While CEO at retirement age is positively correlated with all CEO turnovers, it is negatively correlated with forced CEO turnovers. This difference might because those CEOs under the retirement age are due to more strict inspection during the classification process.
17
Table 3 Correlation matrix
This table presents the correlation matrix for selected variables. Panel A reports the correlations among major firm profitability measures. Panel B shows that correlations for the retention outcomes and CEO characteristics and firm size.
Panel A. Firm profitability measures
Variables stock Firm return Weighted peer return Equally weighte d peer return
Firm ROA Firm ROE
CEO at retireme
nt age dummy Firm stock return 1
Weighted peer return 0.523 1
Equally weighted peer return 0.522 0.942 1
Firm ROA 0.198 0.097 0.096 1
Firm ROE 0.014 0.008 0.009 0.063 1
Net income 0.033 0.033 0.034 0.175 0.013 1
Panel B. CEO characteristic measures
Variables turnover CEO CEO forced out retiremenCEO at t age CEO ownership CEO sever on the board Firm market capitaliza tion
CEO turnover dummy 1
CEO forced out dummy 0.438 1
CEO at retirement age dummy 0.090 -0.038 1
CEO ownership dummy 0.027 0.018 -0.013 1
CEO sever on the board dummy -0.082 -0.055 0.010 -0.002 1
Firm market capitalization 0.006 -0.010 0.006 0.060 0.024 1
b. Two stage regression of forced turnovers on peer and firm specific return Table 4 shows results of the two-stage regression model for strong-form RPE. Panel A
demonstrates the first stage of hypothesis 1 in the empirical method section. In the first stage, the dependent variable is firm’s own stock return and the independent variable is peer return, which includes all peer firms in TNIC database. Column (1) and column (2) use the product similarity weighted peer stock return as the regressor. Column (1) shows result controlling for year fixed effect while column (2) does not control for year fixed effect. Column (3) and column (4) use equally weighted peer stock return as the regressor. Similarly, column (3) controls for year fixed effect while column (4) does not. The coefficient in the first column suggests that a 1 unit increase in weighted peer stock return is associated with a 0.845 (z-statistics of 109.080) units increase in the firm market return; similarly, the coefficient in the third column suggests that a 1 unit increase in weighted peer return is associated with a 0.816 (z-statistics of 111.029) units increase in the firm market return. This result indicates that weighted peer stock return
18 is more closely related to firm’s market return comparing to score weighted peer return. The z-statistics are higher when the regressions do not control for year fixed effect. For all regressions, the peer return variable is highly significant at 1% level. The high significance of the first stage regressions suggests that the peer return is a relevant instrumental variable and enables us to use the predicted value from the first stage results to conduct the second stage regressions.
Table 4. Two stage regression for CEO dismissal on firm and product market return
The first stage uses product market peer group stock return to predict firm’s stock return. The
weighted peer firm return is the TNIC score weighted average return of peer firms and the equally weighted peer return is the arithmetic mean of peer firm returns. The second stage is
probit regression, the linear estimation (Product market induced stock return) and estimated residual (Idiosyncratic stock return) in the first stage are used to predict the non-voluntary turnover dummy. CEO of retirement age and CEO with high equity ownership are both dummy variables use as control. The panel B reports the marginal effect of the probit regression instead of coefficient. Robust z-statistics in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Panel A: First stage OLS regression
Variables (1) (2) (3) (4)
Weighted peer firm return 0.845*** 0.843***
(109.080) (112.758)
Equally weighted peer return 0.816*** 0.816***
(111.029) (114.799)
Constant 0.056*** 0.058*** 0.045*** 0.046***
(22.608) (18.970) (17.972) (14.881)
Observations 34,418 34,418 34,418 34,418
R-squared 0.277 0.284
Panel B: Second stage probit regression
Variables (1) (2) (3) (4)
Product market induced stock return -0.027*** -0.021*** -0.025*** -0.020***
(-6.462) (-6.869) (-5.910) (-6.624)
Idiosyncratic stock return -0.044*** -0.043*** -0.044*** -0.044***
(-13.368) (-13.847) (-13.319) (-13.746)
CEO of retirement age -0.017*** -0.017*** -0.017*** -0.017***
(-6.707) (-6.679) (-6.678) (-6.634)
CEO with high equity ownership 0.005*** 0.003* 0.005*** 0.003*
(2.955) (1.856) (2.918) (1.841)
Observations 34,418 34,418 34,418 34,418
19 The second stage is a probit regression which uses forced turnover dummy, which
equals 1 if there is a forced CEO turnover, as the dependent variable. Independent
variables are two distinct return components: product market induced return and firm idiosyncratic return. These two return components are obtained from the linear prediction and residual of the first stage regression. Panel B of Table 4 presents the marginal effects, that is, the change in the probability for a change in each independent variable, of the probit regressors. Consistent with strong-form RPE hypothesis, CEO contributed firm performance is negatively associated with the probability of forced CEO turnovers for all regressions, all significant at 1% level. However, the product market common return components are also negative and have strong predictive power over CEO dismissal probability for all regressions, which is in contradiction with our hypothesis. The first column suggests that a 1 unit increase in product market return component is associated with a 2.7% (z-statistics of -6.462) decrease in forced turnover probability, other things equal. Further, a 1 unit increase in CEO return component is associated with a 4.4% (z-statistics of -13.368) decrease in forced turnover probability holding other variables the same. The control variable CEO at retirement age dummy is negatively associates with the dismissal probability and significant at 1% for all regression. The CEO with high ownership dummy is positively correlated with the forced turnover; it is significant at 1% with year fixed effect and significant at 10% without year fixed effect
According to the empirical results, the strong-form RPE is violated since the market return component has some effect on force turnover probability. Namely, the boards cannot fully filter out product market common shocks when making CEO dismissal decisions. Moreover, idiosyncratic performance of firm is more like to lead to CEO dismissal compared with product market induced performance and with higher z-statistics. In terms of firm idiosyncratic return, the other three columns give consistent results that the force turnover probability decreases around 4.3% to 4.4% when firm idiosyncratic return increase by 1 unit. Nonetheless, there is a variation among four columns regarding the marginal effect of the peer firm return. The increase in the predicted product market induced return using weighted peers is associated more reduction in forced turnover probability in comparison with the predicted product market induced return using equally weighted peer firms.
20 c. Two stage regression using TNIC HHI quintiles
To examine the second hypothesis proposed in the previous section, I split the sample into quintiles according to each firm’s TNIC HHI. Firms with HHI in the bottom 20% in each year are subscribed to HHI quintile 1, while top 20% HHI firms are subscribed to HHI quintile 5. All quintiles are then used for the same empirical strategy in Table 4 column (1), i.e. use weighted peer return in the first stage and control for year fixed effect. Table 5 demonstrates the results for those HHI quintiles, column (1) shows the result for HHI quintile 1 and column (2) presents result for quintile 2 etc. All regressions control for year fixed effects.
Consistent with Table 4, the first stage regression for all quintiles are significant at 1% level. The coefficients for peer return are 1.085 (z-statistics of 54.466) for quintile 1, 1.006 (z-statistics of 48.935) for quintile 2, 0.913 (z-statistics of 48.404) for quintile 3, 0.819 (z-statistics of 42.789) for quintile 4, and 0.551 (z-statistics of 34.229) for quintile 5. There is an observable trend after splitting quintiles. The coefficient decreases with HHI quintile increases, and the same pattern also holds for the z-statistics. The intuition is straightforward as firms in a less concentrated (low HHI) product market has more competitor and many close substitute product. Hence, the market peer performance can provide more accurate information about how the overall market condition is.
However, we fail to find the expected pattern for the second stage probit regressions. The reported marginal effects for product market induced stock return are -2.4%, -2.1%, -1.7%, -3.3% and -2.8% (z-statistics of -2.210, -2.249, -2.104, -4.168 and -2.817) for HHI quintile 1 to 5, respectively. In addition, the reported marginal effects for product market induced stock return are 4.2%, 4.8%, 3.7%, 4.5% and 4.2% (zstatistics of -2.210, -2.249, -2.104, -4.168 and -2.817) for the corresponding quintiles. All estimates are significant at 5% level. These results indicate that increases in market return component and firm idiosyncratic return are both negatively associated with CEO dismissal probability. Additionally, firm idiosyncratic stock returns have a more pronounced impact on the turnover frequency. Nevertheless, for both variables, there is no detectable trend for the marginal effects, as well as for z-statistics. Therefore, we cannot find evidence in favor of the both hypothesis 1 and hypothesis 2 that market concentration plays a role in CEO dismissal frequencies.
21
Table 5. Two stage regression using TNIC HHI quintiles
Quintiles are split according to firms’ TNIC HHI, column 1 represents quintile 1 which contains firms with bottom 20% HHI and column 5 represents quintile 5 which contains firms with top 20% HHI. The first stage uses product market peer group stock return to predict firm’s stock return. The weighted peer firm return is the TNIC score weighted average return of peer firms and the equally weighted peer return is the arithmetic mean of peer firm returns. The second stage is probit regression, the linear estimation (Product market induced stock return) and estimated residual (Idiosyncratic stock return) in the first stage are used to predict the non-voluntary turnover dummy. CEO of retirement age and CEO with high equity ownership are both dummy variables use as control. The panel B reports the marginal effect of the probit regression instead of coefficient. Robust z-statistics in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Panel A: First stage OLS regression
Variables Lowest (1) (2) (3) (4) (5) Concentration Concentration Moderate Concentration Highest Weighted peer firm
return
1.085*** 1.006*** 0.913*** 0.819*** 0.551*** (54.466) (48.935) (48.404) (42.789) (34.229)
Equally weighted peer return
Constant 0.033*** 0.045*** 0.050*** 0.057*** 0.086***
(7.097) (8.007) (9.024) (9.944) (14.701)
Observations 6,892 6,883 6,885 6,883 6,875
R-squared 0.352 0.328 0.327 0.272 0.182
Panel B: Second stage probit regression
Variables Lowest (1) (2) (3) (4) (5) Concentration Concentration Moderate Concentration Highest Product market
induced stock return
-0.024** -0.021** -0.017** -0.033*** -0.028*** (-2.210) (-2.249) (-2.104) (-4.168) (-2.817) Idiosyncratic stock return -0.042*** -0.048*** -0.037*** -0.045*** -0.042*** (-4.888) (-6.237) (-5.937) (-7.219) (-6.959)
CEO of retirement age -0.010* -0.020*** -0.011* -0.020*** -0.021***
(-1.917) (-3.591) (-1.929) (-3.666) (-4.189)
CEO with high equity ownership
-0.005 0.007* 0.003 0.008** 0.010*** (-1.484) (1.752) (0.774) (2.316) (2.922)
Observations 6,892 6,883 6,885 6,883 6,875
22 d. Two stage regression using TNIC total similarity quintiles
In order to re-examine hypothesis 2, the whole sample was split into quintile again applying a different criterion: TNIC total similarity measure. This is also a firm-specific measure. Instead of measuring the market concentration, it captures the similarity of products a firm faces in its own product market. The total similarity measure is computed as the aggregate pairwise similarity score for a particular entity in a given year. Since higher total similarity score means they are closer substitutes, we expect the similarity quintiles to show the opposite effect as HHI quintile. Namely, as total similarity quintile increases, we expect that the marginal effect for both component in the second stage probit regression to increase. The results of these similarity quintiles regressions are reported in Table 6, with first column the quintile with least similar product market. All regressions control for year fixed effects.
Panel A presents the estimated results for the first stage. In accordance with the hypothesis, both coefficients and z-statistics are increasing in the similarity quintile. The latitude of the coefficients of peer firm return of similarity quintiles, with little deviation, mirrors those of HHI quintiles. For the lowest similarity quintile, the coefficient of peer return is 0.511(t-statistic of 35.229) and the highest similarity quintile has a coefficient of 1.094 (t-statistic of 50.596). Panel B shows estimated marginal effects of the second stage probit regressions. For the firms with least total similarity score, the marginal effects are -2.2% (t-statistic of -2.428) and -3.8% (t-statistic of -6.672) for market return and idiosyncratic return correspondingly. For the firms with least total similarity score, the marginal effects are -2.8% (t-statistic of -2.886) and -4.9% (t-statistic of -6.906) for market return and idiosyncratic return correspondingly. While all results are significant at 5% level, there is no clear tendency among those coefficients.
The quintile regressions using total similarity quintiles confirms the result from the TNIC HHI quintiles that hypothesis 2 does not hold. Although firms with less market concentration and higher product similarity increase the predictive power of peer group return on firm’s stock return, we fail to find evidence that such firms return component had consistently higher marginal effects on non-voluntary turnover probability. The failure might subject to inappropriate data processing procedure, for instance, lack of scrutiny for missing values of firm return and peer return.
23
Table 6. Two stage regression using TNIC similarity quintiles
Quintiles are split according to firms’ TNIC total similarity score, column 1 represents quintile 1 which contains firms with bottom 20% total similarity score and column 5 represents quintile 5 which contains firms with top 20% total similarity score. The first stage uses product market peer group stock return to predict firm’s stock return. The weighted peer firm return is the TNIC score weighted average return of peer firms and the equally weighted peer return is the arithmetic mean of peer firm returns. The second stage is probit regression, the linear estimation (Product market induced stock return) and estimated residual (Idiosyncratic stock
return) in the first stage are used to predict the non-voluntary turnover dummy. CEO of retirement age and CEO with high equity ownership are both dummy variables use as control. The
panel B reports the marginal effect of the probit regression instead of coefficient. Robust z-statistics in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Panel A: First stage OLS regression
Variables
(1) (2) (3) (4) (5)
Lowest
Similarity Similarity Moderate
Highest Similarity Weighted peer firm return 0.511*** 0.866*** 0.953*** 1.007*** 1.094***
(35.229) (48.803) (51.161) (52.347) (50.596)
Equally weighted peer return
Constant 0.077*** 0.046*** 0.045*** 0.046*** 0.053***
(14.209) (8.450) (8.162) (8.452) (9.585)
Observations 6,893 6,883 6,884 6,883 6,875
R-squared 0.177 0.306 0.329 0.334 0.308
Panel B: Second stage probit regression
Variables
(1) (2) (3) (4) (5)
Lowest
Similarity Similarity Moderate
Highest Similarity Product market induced stock
return
-0.022** -0.040*** -0.018** -0.017** -0.028*** (-2.428) (-4.769) (-1.986) (-2.291) (-2.886)
Idiosyncratic stock return -0.038*** -0.047*** -0.028*** -0.050*** -0.049***
(-6.672) (-6.557) (-3.927) (-8.305) (-6.906)
CEO of retirement age -0.020*** -0.014** -0.022*** -0.012** -0.012**
(-4.094) (-2.449) (-3.587) (-2.393) (-2.270)
CEO with high equity ownership
0.008*** 0.009** 0.004 0.004 -0.002 (2.632) (2.413) (1.124) (1.193) (-0.463)
Observations 6,893 6,883 6,884 6,883 6,875
24 e. Test the weak-form RPE hypothesis
The previous findings are consistent with empirical research on RPE theory that strong-form RPE does not hold. In addition, we want to use this most up to date data to investigate whether the weak-form RPE hold or not applying TNIC peers. The empirical strategy is straight forward that the forced turnover dummy is regressed on both firm own stock return and peer firms stock return.
Table 7. Weak-form relative performance evaluation
The dependent variable of the probit regression in this table is the forced turnover dummy. The independent variable is firm stock return and peer firm return. The weighted peer firm return is the TNIC score weighted average return of peer firms and the equally weighted peer return is the arithmetic mean of peer firm returns. CEO of retirement age and CEO with high equity ownership are both dummy variables use as control. The panel B reports the marginal effect of the probit regression instead of coefficient. Robust z-statistics in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Variables (1) (2) (3) (4)
Firm stock return -0.043*** -0.043*** -0.044*** -0.044***
(-13.105) (-18.449) (-13.278) (-18.865)
Weighted peer firm return 0.012*** 0.017***
(3.719) (6.172)
Equally weighted peer return 0.015*** 0.019***
(4.238) (7.047)
CEO of retirement age -0.017*** -0.017*** -0.017*** -0.017***
(-6.707) (-6.732) (-6.680) (-6.707)
CEO with high equity ownership
0.005*** 0.003* 0.005*** 0.003* (2.959) (1.865) (2.923) (1.853)
Observations 34,433 34,433 34,433 34,433
Year FE YES NO YES NO
Column (1) and (2) use the weighted peer return as regressor while column (3) and (4) use the equally weighted peer return as a regressor. For all columns, the firm stock return is negatively correlated with the dismissal frequency and peer firm stock return is positively correlated with force turnover probability. All coefficients are significant at 1% level. This result provides evidence for weak-form relative performance evaluation. The first column indicates that as the firm’s own stock return increase by 1 unit the dismissal probability decreases by 4.3% (z-statistics of -13.105). This effect is
25 approximately the same as the impact of firm idiosyncratic return using the two-stage regression model presented in Table 4. Conversely, the coefficient peer return shows an opposite sign as in Table 4. One unit increase in the stock return of peer group is associated with a 1.2% increase in CEO dismissal probability holding firm market return constant. The intuition is that when firm return is fixed, better performance of peer firms means poorer relative performance of the firm itself, thus, top executives are more likely be forced out.
The previous analyses of relative performance evaluation using product market peer firms are, in general, in accordance with earlier papers. We find that by changing the industry benchmark from conventional SIC industry peer performance to TNIC product market peer performance, weak-form relative performance evaluation hypothesis holds while strong-form relative performance evaluation hypothesis is violated. This suggests that when the board making the CEO retention decision; they cannot completely distinguish product market systematic shock from the firm’s performance. Thus, some CEOs are fired due to bad luck instead of poor management skill. Additionally, there is no clear tendency on the coefficients after splitting quintiles based on both TNIC HHI and product market total similarity. Namely, product market structure has little influence on CEO turnover decisions. The robustness of this conclusion will be tested in the next section and some possible explanation for why boards sometimes fail to assess the CEO’s own ability will be discussed.
E. Robustness Check
Several robustness tests are presented in this section in order to validate the aforementioned conclusion of the research question. Firstly, we argue that boards have limited attention and make decision using salient information. Hence, we replace the peer return variable using only the weighted average return of the top 10 closest peer firms in the market. Secondly, stock market price might deviate from the firm’s underlying value due to irrationality of investors, thus, we apply the same empirical strategy in Table 4 for accounting return ROA. Thirdly, the dependent variable changes from the forced turnover dummy to the voluntary turnover dummy to investigate the different impact of peer return on this group. Lastly, we try to replicate Jenter and Kanaan (2015) using the Fama-French 48 industry classification to define peer return.
26 a. Using only 10 closest product market peers to compute peer benchmark
return
To keep the 10 closest peer firms, each firm’s peer firms have been ranked according to their TNIC similarity scores, only firms ranked top 10 are kept for the following weighted average peer group return computation, the weight for each firm is also base only those 10 peers. For firms with less than 10 peers in the original dataset, all peer firms available are used to compute the peer group return. Table 8 shows estimation outcome using the same methodology in Table 4 column (1) and (2), in addition, the sample are divided into HHI quintiles to re-examine the hypothesis 2 about market structure’s impact on CEO retention decisions. Table 8 shows the estimated outcome using the new peer return. The result, in general, is in line with Table 4 and Table 5. In the first stage, all peer returns have strong predictive power over firm return and the impact increase with competition intensity. In the second stage, idiosyncratic return components are all negatively associated with CEO dismissal probability while market common returns are not significant at 5% level. This indicates the board successfully filters out market shocks when making CEO replacement decision. It is notable that the top 10 peers weighted average stock return are less correlated with firm stock return compared with using all peers weighted average return as the independent variable. The coefficient is 0.728 (z-statistics of 113.945) in Table 8 column (1) and 0.845 (z-statistics of 109.080) in the corresponding column in Table 4.
In respect of the second stage regression, the reported marginal effects remain close to those in Table 4. For the TNIC HHI quintiles, coefficients for peer benchmark return in the first stage decease gradually form 0.815 (z-statistics of 55.252) to 0.522 (z-statistics 35.325) as the HHI quintile increases, the z-statistics also decease with the HHI quintiles. In the second stage, both coefficients for firm idiosyncratic and market-wide factors fluctuate with no tendency. The quintile with moderate market concentration has lowest marginal effects on turnover probability regarding both firm idiosyncratic component and market systematic component. Hence, regardless with the difference in predictive power of the peer return on firm return; the empirical results are in favor of our previous finding.
27
Table 8. Use only 10 closest peers’ return as performance benchmark
The first stage uses 10 closest peers’ average stock return to predict firm’s stock return. The
weighted peer firm return is the TNIC score weighted average return of peer firms. The second
stage is probit regression, the linear estimation (Product market induced stock return) and estimated residual (Idiosyncratic stock return) in the first stage are used to predict the non-voluntary turnover dummy. CEO of retirement age and CEO with high equity ownership are both dummy variables use as control. The panel B reports the marginal effect of the probit regression instead of coefficient. Robust z-statistics in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Panel A: First stage OLS regression
Variables
(1) (2) (3) (4) (5) (6) (7) Lowest
Concentration Concentration Highest Peer firm return 0.728*** 0.727*** 0.815*** 0.801*** 0.754*** 0.731*** 0.522*** (113.945) (118.466) (55.252) (50.147) (48.322) (44.284) (35.325) Constant 0.054*** 0.054*** 0.033*** 0.047*** 0.054*** 0.055*** 0.082*** (21.908) (19.443) (7.233) (8.497) (9.769) (9.755) (14.128) Observations 34,418 34,418 6,892 6,883 6,885 6,883 6,875 R-squared 0.295 0.359 0.339 0.327 0.286 0.191
Panel B: Second stage probit regression
Variables
(1) (2) (3) (4) (5) (6) (7) Lowest
Concentration Concentration Highest Product market induced stock return -0.027*** -0.022*** -0.026*** -0.024*** -0.017** -0.034*** -0.027*** (-6.868) (-7.206) (-2.669) (-2.748) (-2.253) (-4.369) (-2.878) Idiosyncratic stock return -0.044*** -0.043*** -0.042*** -0.047*** -0.037*** -0.045*** -0.042*** (-13.176) (-13.584) (-4.606) (-6.078) (-5.916) (-7.224) (-6.966) CEO of retirement age -0.017*** -0.017*** -0.010* -0.020*** -0.011* -0.020*** -0.021*** (-6.693) (-6.662) (-1.927) (-3.550) (-1.932) (-3.662) (-4.185) CEO with high equity ownership 0.005*** 0.003* -0.005 0.007* 0.003 0.008** 0.010*** (2.969) (1.839) (-1.416) (1.769) (0.805) (2.317) (2.913) Observations 34,418 34,418 6,892 6,883 6,885 6,883 6,875
