19th of June, 2013
University of Amsterdam
Dr. F.H.M. Verbeeten MBA 2012/2013
Master thesis Accountancy & Control Semester II
CEO compensation packages:
An investigation of subjectivity in US companies
E. Stam 6039588
Total words: 14,076 Total pages: 59
2 Abstract
I investigate the use of subjectivity in determining CEO compensation across different types of firms. In contrast to objective performance measures, subjectivity cannot be easily measured and involves a lot of discretion. Based on prior literature, I predict that the use of subjectivity in CEO compensation packages increases with R&D expenses and firm size. Additionally, SGA expenses and strategic emphasis (value creation versus value appropriation) are expected to be negatively associated with subjectivity. The analyses came up with mixed results. Evidence is found for the relation between subjectivity and both SGA expenses and strategic emphasis, but the other hypotheses are rejected.
3 Table of Contents Paragraph Page 1 Introduction 4 1.1 Background 4 1.2 Research Question 4 1.3 Reading guide 5
2 Theoretical background and hypotheses 6
2.1 Agency theory 6
2.2 CEO compensation 8
2.3 Pay-performance relationship 10
2.4 Exploitation versus exploration 12
2.5 Hypotheses 15 3 Research design 19 3.1 Research methodology 19 3.2 Sample creation 20 3.2.1 Data collection 20 3.2.2 Dummy variables 22 3.2.3 Dropping observations 23
3.2.4 Log-transformed variables and Winsoring 23
3.2.5 Definition of variables 25 4 Results 30 4.1 Descriptives 30 4.2 Univariate analyses 32 4.3 Regressions 32 4.4 Robustness checks 37 5 Conclusion 42 5.1 Discussion of results 42
5.2 Suggestions for further research 43
References 44
Appendix A: Distribution of industries 46
4 Master thesis Accountancy & Control
CEO compensation packages: An investigation of subjectivity in US companies
1. Introduction
1.1 Background
In the past, research has shown that using exclusively traditional performance measures is often not adequate enough to capture firm performance. Subjective incentive bonuses and non-traditional measures of performance are often used to complement perceived weaknesses in quantitative performance measures (Bushman, Indjejikan and Smith, 1996; Gibbs, Merchant, Van der Stede and Vargus, 2004). Various aspects of non-traditional performance measures have yet been studied, with most existing literature focusing on under which specific circumstances they are applied (Bushman et al. 1996; Ittner, Larcker and Rajan, 1997). Hölmstrom (1979) found evidence that using additional performance measures can be used to improve compensation contracts, which in return allows for a more accurate judgment of employee performance. Subjectivity however, has been largely neglected in incentive compensation literature. Despite it playing an important role in current day society, it has not been studied very extensively.
It has been acknowledged in prior literature that, for most employees, it is hard to measure their individual contribution to firm value by using only objective performance measures (Bol, 2008). Yet, research on subjectivity in compensation packages has only recently begun. According to Bol (2008) most prior research has been focused on compensation contracts of employees whose individual contribution were easy to measure. This provided many insights into the design of compensation contracts (e.g. Gerhart and Milkovich, 1990) and the corresponding effects of this on employees (e.g. Gibbs et al., 2004). This paper tries to study the role of subjectivity in compensation packages by comparing compensation packages among different industries. This might provide new insights in the use of use of subjectivity in determining CEO compensation in our current society.
5 1.2 Research question
Since traditional performance measures are often complemented by non-traditional ones in current day society, it would be interesting to see what kind of effect this is having on CEO compensation in different industries. Ittner et al. (1997) already showed that firms following an innovation-oriented strategy tend to use more nonfinancial measures in their CEO incentive packages. Following this logic, R&D-focused firms should have a considerably higher ratio of non-traditional performance measures compared to more general firms focusing on cost reduction and customer satisfaction. Although this is just an example of a firm characteristic possibly influencing subjectivity in determining CEO compensation, there are bound to be more. This thesis therefore focuses on whether certain firm characteristics indeed have an effect on the use of subjectivity in compensation packages, by looking at the way compensation packages are constructed.
Therefore I would like to propose the following research question: To what extent does the use of subjectivity in CEO compensation packages differ among different types of firms?
1.3 Reading guide
This paper will provide additional information regarding the central question, and will hopefully come up with some significant results. The remainder of this paper will consist of four parts. First, a theoretical framework will be provided on relevant literature, which will result in a couple of hypotheses to test. Secondly, the research methodology will be described, to explain how and why I will undertake each step in the process. Next, the corresponding results will be discussed, after which I will conclude with some final remarks and suggestions for further research.
6 2. Theoretical framework and hypotheses
2.1 Agency theory
When we talk about CEO compensation packages and aligning its interests with that of the shareholder's, agency theory cannot be ignored. Agency theory originates from the 1960s and early 1970s, when economists began researching risk sharing among individuals, as well as groups (Eisenhardt, 1989, p. 58). In early literature it was described as a theory, dealing with the problem that arose whenever cooperating parties had different attitudes toward risk (Arrow, 1971; Wilson, 1968). Most research concerning CEO incentives have been rooted in agency theory. It tries to explain the relationship between principals on the one hand, and agents on the other. It is concerned with the difficulties arising in motivating one party (the agent), to act in the best interest of another (the principal), rather than in his or her own. Tosi et al. (2000, p. 304) argue that this is based upon several assumptions: Agents are (1) risk averse, (2) self-centered, (3) and their interests may be different from the principal's. The theory assumes herewith that both parties are trying to maximize their own utility (Roth and O'Donnell, p. 679). Typically, the CEO should act in the best interest of all the stakeholders, instead of its own. In practice however, CEOs tend to act in their interest alone, when the interests between the CEO and shareholders diverge. By providing the right incentives for an agent or incurring monitoring costs, behavioural misalignment can be limited (Jensen and Meckling, 1976).
According to Eisenhardt (1989, p. 58) agency theory focuses on two types of problems in principal-agent relationships. The first one is caused when conflicting desires or goals exist between the principal and the agent, and the principal is experiencing difficulties verifying the agent's behaviour. In this case, the principal cannot be sure whether the agent is behaving like he is supposed to. Contracting problems such as moral hazard and adverse selection are closely related to this: Moral hazard may arise when a party is willing to take greater risks than usual, because the related costs will not be (fully) incurred by this party (Hölmstrom, 1979), whereas adverse selection, sometimes called a 'lemon' problem, takes place when asymmetric information between two parties results in choosing the 'bad' products (i.e. lemons). The second type of problem arises when both the principal and the agent have different attitudes toward risk. The agent might prefer riskier actions than the principal (or vice versa), resulting from the differing risk preferences. High levels of socialization might reduce this goal incongruence (Eisenhardt, 1989).
7
According to Tosi et al. (2000, p. 304), there are at least three problems related to reducing agency costs shareholders have to deal with. Firstly, it has been proven difficult for shareholders to structure and supervise activities of top level executives. Most of their tasks involve a lot of discretion and are not (easily) programmable by the owners. Secondly, executive's knowledge about the organizational processes and business decisions is often much more extensive than that of the shareholders. The information symmetry taking place in these relationships favour the CEO in most cases. Finally, executives are authorized to use the organization's resources for whatever objective they want to achieve, which might not be aligned with the owner's interests. The challenge lies within steering the risk averse, self-centred, utility-maximizing agents to act in the interest of the principals, which is most often to increase firm value (Bloom & Milkovich, 1998).
According to Bushman, Indjejikan and Smith (1996, p. 162) agency theory also predicts that particular performance measures will be included only if they provide information about a manager's actions beyond the measures already in place. Additionally, Hölmstrom (1979) already argued that any information, imperfect information as well, can be used to improve contracts. Every piece of additional information about the performance of an agent allows for a more accurate judgment of performance (p. 89). Although Hölmstrom was not studying subjective, nonfinancial matters in particular, his results are relevant nonetheless. He provides evidence that even imperfect information in a principal-agent relationship is able to improve knowledge about performance, and in turn alleviates moral hazard.
A substantial amount of articles have been written on agency theory in the past, with authors taking all kinds of different stances toward it. Because of this, literature on agency theory has been divided into different camps (Jensen, 1983). For example, some authors (e.g. Barney and Ouchi, 1986) argue that agency theory emphasizes how capital markets can affect firms, while others make no reference to this at all (e.g. Anderson, 1985; Demski and Feltham, 1978). Therefore, the theory has been broadly applied in disciplines such as accounting, economics, finance, marketing, political science, organizational behaviour and sociology (Eisenhardt, 1989, p. 57). The paper "Agency theory: an assessment and review" by Eisenhardt (1989) delves deeper into the subject matter and tries to explain agency theory, and discuss its provided insights, by means of four questions: (1) What is agency theory? (2) What does agency theory contribute to organizational theory? (3) Is agency theory empirically valid? (4) What topics and contexts are fruitful for organizational researchers who use agency theory? Based on these questions, Eisenhardt (1989) comes to two major conclusions. Firstly, she argues that agency theory provides one with unique insights into information systems,
8
outcome uncertainty, incentives, and risk. Secondly, it is argued to be an empirically valid perspective, meaning that empirical results correlate with behaviour as measured in other contexts. This empirical validity seems to be stronger when agency theory is combined with additional perspectives.
2.2 CEO compensation
Since the research question of this master's thesis focuses on the relation between performance measures and CEO compensation, the two most important concepts are CEO compensation and the related incentives provided. In this paragraph, I will briefly discuss these two concepts for a clearer view of the research at hand.
Executive mangers are often compensated based on intricate compensation packages, composed of both base salary and variable pay components. The latter is commonly dependent on the manager's performance in a given period of time. Those corresponding incentives are based upon either financial or nonfinancial performance measures. Financial measures are often based on stock-based (e.g. stock price, market capitalization) or accounting-based performance (e.g. ROA, sales), and are typically easy to measure. Nonfinancial measures (e.g. customer satisfaction) on the other hand are subjective by nature, making it hard to evaluate a manager's performance and determine whether and to what extent he has to be compensated for this. In my model (which will be specified in more detail later), financial measures (i.e. stock return and return on assets) will be the main inputs. The amount of subjectivity will be determined by looking at the residuals.
The research conducted by Gibbs et al. (2009) focuses less on determinants of using non-traditional performance measures, but rather analyzes on what grounds certain types of performance measures are selected. The performance measure properties that are analyzed are the measure's noise, controllable risk, distortion, and manipulability. They conclude that when a measure is flawed among any of these dimensions, less weight is given to that measure for explicit incentives, supporting their claim that traditional measures are not suitable for each and every firm. Additional measures will then be used, if these can mitigate the distortion or manipulability in the previous performance measures. According to Murphy (2012, p. 24), two major factors determine what kind of incentives are created for any compensation plan: (1) how performance is measured; and (2) how compensation varies with measured performance. A typical bonus plan often includes a minimum performance target to qualify for bonus pay and a maximum bonus for superior performance, instead of compensating
9
managers linearly for their performance (Murphy, 2012, p. 34). In most cases a performance target is set, with a corresponding target bonus, which will be paid when the performance target is achieved. Subsequently upper and lower performance thresholds are set, forming a so-called "incentive zone". Within this zone, managers are paid based on their performance, whether this be linear or not. However, performance exceeding the determined thresholds is often not rewarded. This can result in a disincentive for managers to optimize performance when this threshold is reached. The table below clearly illustrates what a typical bonus plan looks like.
Figure 2.1 A typical bonus plan (Murphy, 2012, p. 34)
An important aspect of producing incentives is weighing the cost versus the value of incentive compensation. Although incentives are used to align the interests of the shareholders with that of the managers, costs should not exceed the value created by this interest alignment too much. Murphy (2012, p. 18) argues that one should distinguish between two different valuation concepts: (1) the cost to the company of granting the compensation; and (2) the value of the compensation received by the executive. This is particularly relevant for stock-based incentive pay, since executives often cannot reap the full monetary benefits immediately. For example, a company decides to compensate an executive by means of shares of restricted stock vesting in five years. In this case, the CEO would be restricted from
10
selling the stock (immediate monetary benefits) for the first five years, and the accumulated benefits (dividends plus interest) would only be received on the vesting date. The company's cost of this grant is only the market price of the share, since they could have raised money by selling it on the open market. The executive however, would obviously want to receive the monetary benefits immediately. He could then decide for himself whether he wants to invest it in company stock or spent it else wise. An important aspect for the company though, is that CEOs become more risk averse and undiversified, the more overall wealth is positively correlated with company stock prices (Murphy, 2012, p. 19). This can be seen as a positive phenomenon for a lot of companies, since the managers will become more aware of their choices when they influence their own wealth more directly.
From all the managers in an organization, the CEO is arguably the most interesting to study. The working paper 'Executive compensation: where we are, and how we got there' by Murphy (2012) manages to provide us with an extensive summary of CEO compensation over the last century. The author provides us with several reasons to limit the scope of this thesis to CEOs. Firstly, CEOs are one of the highest placed managers in their firms. Their decisions are most likely to affect the organization as a whole. This necessitates the use of incentives to align the CEO's interests with that of the shareholders, making them an interesting group to study. Secondly, a large proportion of a CEO's total compensation is based on incentive-based pay. Murphy (2012, p. 17) found, based on information from the ExecuComp database, that CEO pay levels in S&P 500 firms has grown 165% in the past two decades (1992-2011). The remarkable thing is that base salary has remained mostly stable. This enormous increase in variable pay provides yet another reason to study CEO compensation packages, as to how this incentive pay is determined.
2.3 Pay-performance relationship
In the last few decades, there have been differing opinions in both academic and popular press on whether CEOs earn their pay. This has been strengthened by the increasing incentive-based compensation for executive managers, as identified by Murphy (2012, p. 17). Past research has extensively researched compensation strategy, and subsequently the relationship between pay and performance: Whether this relationship is always positive (Duffhues and Kabir, 2007); whether CEO pay is related to future performance (Cooper, Gulen and Rau, 2009); or how incentive-based pay can improve a company's competitive position (Milkovich
11
and Milkovich, 2001). Different aspects have been covered in the past few decades, collectively identifying how incentive-based pay is related to corporate performance.
For this thesis it is particularly important to know how CEO compensation packages differ amongst different companies in different industry groups. Gerhart and Milkovich (1990) focus on this subject matter by studying the extent to which organizations under similar circumstances make different decisions related to base pay, bonus pay, and eligibility for long-term incentives. They find that companies facing similar market conditions, make different decisions concerning incentive-based pay, rather than base pay. Base pay levels seem to be generally accepted, whereas variable pay differs among the observed sample of 14,000 top and middle-level managers among 200 organizations (approximate numbers). One of the main drivers behind this phenomenon is that an increase in base pay is always associated with greater production costs, affecting an organization's competitiveness. On the other hand, low base pay levels can result in poor employee attraction. Some studies argue that large companies might choose to adapt high base-pay levels to attract more applicants (Bronfenbrenner, 1956; Rynes and Barber, 1990). This enables these firms to be selective when hiring new employees. All in all, base pay does differ across organizations, but is not the main means in awarding employees. Organizations tend to differentiate themselves more with respect to pay mix, than pay level (Gerhart and Milkovich, 1990, p. 685). However, Gerhart and Milkovich's (1990) do not focus only on base pay. Their second major focus is pay mix, which is concerned with how the total compensation package is designed. As mentioned earlier, this often includes fixed and variable pay components. This pay mix is supported by two common theories in economic literature: agency theory, which we discussed before, and expectancy theory. The former describes the relationship between principals and their agents, whereas the latter argues that people will act in a certain way, based on the expected results associated with that action. Compensation packages are often designed with the idea that different degrees of emphasis on different types of objectives, influences the behaviour of a firm's employees (Gerhart and Milkovich, 1990, p. 671). Taken together, these two theories point to a positive relation between variable pay and interest alignment.
Gerhart and Milkovich (1990, pp. 683-686) conclude that pay mix, not pay level, is positively related to financial performance. This points out that pay mix has a more strategic character than pay level. Additionally, organizational differences in pay mix were larger, as well as less explained by industry, size, and financial performance.
More recent research conducted by Tosi, Werner, Katz and Gomez-Mejia (2000) tries to focuses on the determinants of CEO pay: to what extent variances in total CEO pay are
12
explained by size as well as performance. Over the last few decades CEO compensation has skyrocketed, and public press has been raising questions about the perceived fairness of this increases. At the same time, academic press have been discussing the contradictory results found in academic literature (Tosi et al., 2000, p. 302). By conducting a meta-analysis on the relationships between CEO pay and firm performance, and CEO pay and firm size, the authors hope to shed some light on the subject matter. Firm performance and firm size are arguably the most important determinants, since these have been receiving the most attention in academic literature, with studies dating back to the 1920s (Tosi et al., 2000, p. 302). The authors find that firm size is able to explain more than forty percent of the variance in total CEO pay, but firm performance accounts for less than five percent. Although their research leaves a large part of total CEO compensation variance unexplained, it does show that performance is a relatively weak determinant for CEO compensation. However, it is also argued that the objective performance criteria used (30 performance measures: e.g. return on assets-one year, net income for previous year) can contain a lot of noise, since they might only tap a small portion of a CEO's job performance requirements (Tosi et al., 2000, p. 331). This might signify the use of objective as well as subjective performance measures, which enables the firm provide a more holistic assessment of executive performance, but makes it more difficult to empirically identify a pay-performance relationship.
All in all, looking back at the literature discussed in this chapter, one could say that that pay-performance relationship literature is far from complete. Tosi et al. (2000, p. 331) also indicate that unless the CEO pay literature grows more extensive, researchers are limited to speculative arguments in explaining the relationship between CEO compensation and firm performance. However, even now it provides useful insights the way CEO compensation packages are designed.
2.4 Exploitation versus exploration
Since this thesis is for a great part about CEO incentive packages in companies adapting different strategies, it is of great importance to identify strategic differences between them. Past research in marketing often explored how resources, skills and the development of new capabilities can affect financial performance. Since the resources a firm possesses will always be limited in some way, managers will need to make decisions defining an organization's strategy. Mizik and Jacobson (2003) distinguish here between the creation and appropriation of value. They argue that these are two fundamental processes affecting financial
13
performance. Value creation (e.g. innovating, producing, and delivering products to the market) is the process of creating value for the customers. However, firms will need to appropriate this value as well, if they want to enjoy the full benefits. This involves restricting competitive forces to benefit from the value you created. As summarized by Mizik and Jacobson (2003, p. 64): "value creation influences the potential magnitude of the advantage; value appropriation influences the amount of the advantage the firm is able to capture and the length of time the advantage persists". Whenever firms are not able to restrict their competitor's ability to exploit your creation of value, these competitors will claim it for themselves (Ghemawat, 1991).
Value creation results in societal value, explained by Mizik and Jacobson (2003) as the total difference between the utility derived from a product or service and the costs of producing it. According to Mansfield, Rapoport, Romeo, Wagner and Beardsley (1977) this societal value will then end up with three different parties on the market: The innovating firm, the customers, and other firms (including competitors, as well as non-competitors). The innovating firm will receive some of the created value as economic profit, the customers in the form of consumer surplus, and the other firms benefit from imitation and development cost savings (Mizik and Jacobson, 2003). These ratios can be influenced by trading off between value creation and value appropriation capabilities. An example by Mizik and Jacobson (2003, p. 64) illustrating variations in these ratios, is the development of the polio vaccine. Because Jonas Salk, the inventor of the vaccine, did not patent its discovery, all value was eventually claimed by the consumer group. He did not want to personally profit from this discovery and decided it was more important to have the vaccine spread as widely as possible. It has to be mentioned though, that even when there is a desire for profit, this profit is not always achieved. For example, the company that invented the CT scanner was unable to generate profits, resulting in its takeover around the same time the inventors were given the Nobel Prize in Medicine.
In their article, Mizik and Jacobson (2003) study whether an organization's strategic emphasis (value creation versus value appropriation) influences financial performance. The authors argue that many organizational resources and capabilities (e.g. technological, financial, physical) cannot be linked exclusively to either value creation or value appropriation (Mizik and Jacobson, 2003, p. 65). However, they identify advertising and R&D expenditures to have been repeatedly highlighted in academic research as key influences to the two processes.
14
On the one hand, a firm's R&D expenses have been linked to value creation, because of the improved capabilities it provides. Research and development is often used to generate superior products or improve existing processes, creating competitive advantages. According to Mansfield et al. (1977), value is created through these product and process innovations. Although value creation encompasses a wide range of activities, R&D seems to have had the most attention in prior literature on value creation (Mizik & Jacobson, p. 65). This literature has shown significant effects of R&D expenses on value creation. Research conducted by Denison (1962) for example, finds that approximately forty percent of the total increase in per capita national income could be attributed to technological change. Additionally, he shows that approximately twenty percent of this amount could be attributed to R&D. Mansfield et al. (1977) in turn estimate that R&D has a median social return of approximately 56%. Although these estimates differ amongst all the different studies conducted, more recent research by Griliches (1995) asserts that newer studies keep reporting significant social returns from R&D.
On the other hand, advertising seems to be linked to value appropriation. Mizik and Jacobson (2003) argued that value appropriation is mainly about isolating value for yourself. According to the authors, advertising is of particular importance in the appropriation process for marketing managers. However, two different views are identified with regard to advertising as an isolating mechanism. The former views advertising as anticompetitive, preventing successful imitation by competitors. The latter depicts marketing as pro-competitive, because it is supposed to serve as a means to expel the competitors' isolating mechanisms. Although these are two completely different views of the workings of marketing, both views support the claim that marketing improves a firm's appropriation capabilities. More importantly, empirical results seem to show a significantly positive relation between advertising and persistence of profits (which can be seen as a determinant for value appropriation) (Kessides, 1990; Mueller, 1990). This persistence of profits shows that firms which advertise more heavily, experience slower erosion of profits.
Although these are not the only factors influencing strategic emphasis, movements in these expenditures are probable to provide information on the emphasis on value creation and value appropriation within the firm. Mizik and Jacobson (2003) find that a strategic shift towards value appropriation is associated with increased stock returns. Additionally, they show that even in a high-technology market, with a focus on innovation and R&D, a shift towards value appropriation is viewed favourably by the market (Mizik and Jacobson, 2003, p. 74).
15 2.5 Hypotheses
One of the most important principles for forming our hypotheses is the informativeness principle, as constructed by Hölmstrom (1979). This principle states that a performance measure will only be included in an incentive scheme if it provides incremental information content over the other available measures at low cost. Since all publicly traded companies have traditional accounting and stock price performance measures available for this matter, it is reasonable to assume that companies will use additional measures when these traditional measures fail to capture a manager's performance. Most traditional performance measures typically neglect the actual effort put in by managers. Traditional performance measures mostly reflect absolute numbers, which is often influenced by many different factors. Even when the CEO seemingly makes the right choices, the final result can be in the CEO's disadvantage. For example, Ittner et al. (1997) have already shown that firms following an innovation-oriented strategy tend to use more nonfinancial measures in their CEO incentive packages. In the case of R&D firms, it is therefore safe to assume that they will use more non-traditional measures in compensation packages. After all, costs tend to be high and revenues are uncertain, making it unfair to use only traditional measures to gauge a manager's performance. Hölmstrom's (1979) principle stems from his study on the role of imperfect information in principle-agent relationships. He argued that for performance measures to be effective, they need to provide accurate, informative as well as timely indications of an individual's contributions to firm value. One can conclude from this that when quantitative performance measures prove to be effective, formula incentives are bound to be used extensively. According to Gibbs et al. (2004, p. 411) however, quantitative measures are often not perfect reflections of reality. They argue that the agent is often victim to unwarranted risk, since the performance measures are likely to include uncontrollable factors.
An important part of non-traditional performance measures is subjectivity. Although subjectivity plays an apparent role in modern day society, it has not been studied very extensively. Most studies predominantly focus on determinants influencing the use of subjectivity (Gibbs et al., 2004, p. 410). One of the main reasons behind this, is that many interesting issues arising with subjectivity are not easily tested through available quantitative datasets, because it involves qualitative concepts.
Prior research by Bushman et al. (1996) investigates the use of individual performance evaluation (IPE) in CEOs' annual incentive plans. Although IPE is not necessarily subjective in nature, Bushman et al. argue that subjectivity is an integral part of what they discuss. Just
16
like other empirical studies (e.g. Gibbs et al, 2004) on subjectivity, or non-traditional performance measures for that matter, they conclude that using exclusively traditional measures of corporate performance is often not adequate enough to capture a firm's performance (p. 189). With respect to their study, they argue that IPE is positively correlated with growth opportunities as well as product time horizon. This positive relation between IPE and corporate growth opportunities suggests that accounting and stock-based performance measures are not perfect performance measures for each firm. The authors claim to be among the first to consider non-traditional measures of performance and, more importantly, find statistically significant results to support their claims.
Hayes and Schaefer (2000) also considered the use of non-traditional measures in CEO compensation packages. In their article, they studied the relation between unobservable (to the public) performance measures in compensation packages and future performance. They argue that if compensation contracts optimally incorporate both observable and unobservable (to the public) performance measures, that variation in current compensation should predict future variation in observable performance measures. This assumption is driven by the thought that variation in current composition should be caused by measures unknown to the public, and should be reflected in future performance. They found strong evidence to support their hypothesis that unexplained variation (not related to observable performance measures) in current compensation is positively correlated with future performance. Additionally, their results indicate that when the variance of observable performance measures increases, this relation is stronger.
In an earlier article, the authors (Gibbs et al., 2004) argue that subjectivity can exist in several forms, which are often used in combination (p. 410). Three types are identified: a bonus, whether or not partially, based on subjective judgment of performance; the weights of quantitative measures are determined subjectively, or; whether to pay a bonus is dependent on measured performance and other relevant factors. They focus on two general questions: When do firms make greater use of subjectivity in awarding bonuses? And what are the effects of subjectivity on employee pay satisfaction and firm performance? Concerning the first question, they find that subjectivity is positively related to: the extent of long-term investments in intangibles; the extent of organizational interdependencies; the extent to which achievability of the formula bonus target is both difficult and leads to significant consequences if not met; and the presence of operating losses. With regard to the second question, they find that using subjectivity in compensation packages improves employee pay satisfaction, and leads to better firm performance when manager tenure is high. In addition,
17
the findings show that subjectivity should play an important role in complex work environments, since those require intricate job designs (Gibbs et al., 2004, p. 433). These unpredictable environments impose risk on the employee, by involving them in multiple tasks and often tough decision making. The authors note that there are only few other studies trying to relate subjectivity to financial performance and firm characteristics, acknowledging the article by Bushman et al. (1996) discussed before, as one of them. They expect that other theoretical concepts concerning subjectivity can be examined by collecting new types of variables (Gibbs et al., 2004, p. 434).
Kerssens-van Drongelen, Nixon and Pearson (2000, p. 114) argue that performance measurement will always be a difficult aspect of management, whatever aspect one is trying to measure. By putting performance measures in place, decision-making by the people affected will probably change. Additionally, performance is often tied to rewards (e.g. variable pay) and can be objectively measured. This is not so much the case with R&D. Because R&D is often part of developing new products, there is a considerable gap between incurring expenses and reaping the corresponding benefits. Although several researchers (e.a. Griffin and Page, 1993; Hauser and Zettelmeyer, 1997) have suggested that performance measurement systems should link R&D to performance, only a few support this with empirical evidence. It is acknowledged by Kerssens-van Drongelen e.a. (2000, pp. 136-138) that there is probably no simple solution to this problem, and that it will proof difficult to formulate performance measures for R&D operation. Based on the material discussed above, I develop my first hypotheses:
H1a: Companies with high R&D expenditures use high levels of subjectivity in
determining CEO compensation.
H1b: Companies with high SGA expenditures use low levels of subjectivity in
determining CEO compensation.
The next hypothesis of this thesis is based on the exploitation versus exploration ideology discussed mainly by Mizik and Jacobson (2003). They argue that a firm's strategic emphasis (i.e. value creation versus value appropriation) influences financial performance. Their results demonstrate that stock markets react favourably when this emphasis relatively shifts from value creation to value appropriation. Since R&D expenses heavily influence strategic emphasis in the current operationalisation I would expect, based on the presumption that high R&D expenditures results in more subjectivity, that a focus on value creation is positively
18
related with subjectivity. For this reason, the last hypothesis will look into whether strategic emphasis influences CEO incentive packages:
H2: The use of subjectivity in determining CEO compensation is greater among
companies focusing on value creation, compared to companies focusing on value appropriation.
For the last hypothesis, a paper by Murphy and Oyer (2001) provides the basis. They study the role of discretion in executive incentive contracts in general, and look at the trade-offs firms need to make between several types of discretion. They identify three different types of performance measure used by the firms in their model (p. 2): objective measures that imperfectly capture the performance of an individual, potentially more accurate (but non-verifiable) subjective measures, and company-wide (quantitative) performance measures. Murphy and Oyer (2001) find that using discretion in determining executive bonuses is more important as firms are larger or privately held. Additionally, their results show that the use of subjectivity increases with the importance of the executive's actions, and decreases as these measures are more easily manipulated. They note however, that subjective assessments take time to execute and involve monitoring costs (p. 36), which is in line with Hölmstrom's (1979) informativeness principle. For example, the board might have a hard time assessing a CEO's actual contribution to firm value, since they meet only a few times a year and are often dependent on data provided by the CEO himself. A performance measure should provide additional information on CEO effort at a reasonable cost. Murphy and Oyer (2001) argue that the use of subjectivity will increase as the size of top management or, more generally, firm size increases. As there are more managers, it is harder to measure an individual manager's influence on firm performance. Based on their findings that using discretion becomes more important as firms are larger, I would like to propose an additional hypothesis:
H3a: Large companies use high levels of subjectivity in determining CEO
compensation.
H3b: Small companies use low levels of subjectivity in determining CEO
19 3. Research design
3.1 Research methodology
To find answers on the hypotheses proposed in the last chapter, I will conduct a quantitative study on data collected from the Compustat and ExecuComp databases. Statistical software (i.e. Stata) will be used to determine whether the use of subjectivity in determining CEO compensation indeed differs among different industries. This study will mainly consist of two regression analyses. The first regression analysis will test the relation between CEO compensation and two accounting and stock performance measures:
CEOComp = 0 + 1 AccPerf + 2 StockPerf + (subjectivity)
With this analysis I try to understand to what extent changes in total CEO compensation (CEOComp) are explained by changes in return on assets (AccPerf) and stock return (StockPerf). I will now briefly explain the dependent and independent variables before moving on with the methodology. Firstly, CEOComp is the amount of total compensation received by CEOs in the sample's firms. This compensation is all the compensation received in a given year, including the following components: salary, bonus, other annual, total value of restricted stock granted, total value of stock options granted (using Black-Scholes), long-term incentive payouts, and all other total. This is the regression's dependent variable. The independent variables are AccPerf and StockPerf. AccPerf is used to indicate the accounting performance of the firm through return on assets. This is calculated by dividing sales by total assets. Thirdly, StockPerf indicates stock performance, and will be measured by stock return. This is calculated by taking the change in share price in a given year, and dividing this by last year's share price. Of course, dummy variables will be included as well to control for fiscal year and industry group. This will be expanded on later in this chapter.
With the abovementioned formula we can determine how much of total CEO compensation is dependent on accounting and stock performance measures. The greater the residual (), the greater part of CEO compensation is explained by variables other than accounting and stock performance measures. This means that we can identify which firms use more subjectivity in their CEO compensation packages. The following regression analysis will then help with identifying whether this subjectivity is related to different types of firms (i.e. R&D and SGA) and strategic emphasis (i.e. R&D and advertising):
20
subjectivity = f (industry group, fiscal year, R&D, SGA, strategic emphasis, size, CEO tenure, market-to-book ratio)
This regression analysis will check for differences in compensation packages between different industries, while controlling for firm size. Dummy variables are created for the different industry groups identified in the sample.
3.2 Sample creation
3.2.1 Data collection
The research in this thesis will be conducted in an American setting, since data availability is most optimal for this. However, this also limits the sample to one cultural setting, which leaves open possibilities for further research. The chosen time frame spans from 1992 to 2011. There are a couple of reasons to choose this particular period of time. Firstly, the Compustat ExecuComp database has no pre-1992 data, limiting the sample to 1992 and later. Secondly, I would like to have a sample as large as possible, making it easier to get significant results. However, a lot of data was not yet available for 2012, which is why the sample has been limited to 2011. Because of these reasons I would argue that this time span is most suitable for the research at hand.
As mentioned before, I will retrieve annual compensation data from the Compustat ExecuComp database. This will include the following variables: ID number for each executive/company combination, company name, industry group description, industry group, executive ID number, full name of CEO, date became CEO, executive's age, bonus, salary and total compensation. By filtering on annual CEO flag, a dummy variable indicating whether the executive was CEO in the given year, only CEO compensation data was retrieved. Since this thesis is mostly about CEO compensation packages, the availability of ExecuComp information forms the basis of my sample. After identifying the information needed for the research at hand and setting the desired time frame, ExecuComp provided 33,518 observations over the indicated twenty years.
Additional firm information was retrieved from the regular Compustat database. This includes accounting and stock market performance measures, but also the corresponding industry group (SIC) codes, firm size and other possibly relevant factors. Some of this information will be needed to answer the research question itself, others will be used to
21
control for industry, firm size, etc. to make sure results are not only relevant for one particular type of company. By using the company codes generated in the dataset retrieved from ExecuComp, I could filter the data to be retrieved from Compustat itself. After identifying which variables I wanted to retrieve and providing the list of company codes from my ExecuComp sample, Compustat provided 47,008 observations. This dataset was subsequently consolidated with the ExecuComp dataset, by matching the data based on company ID number and fiscal year.
In addition to the variables retrieved from the databases (of which the most important ones are described in appendix A), some variables need to be computed by hand. Not all desired information could be retrieved from Compustat and ExecuComp directly. Some variables needed for the regressions to be performed (stock return, market-to-book ratio, return on assets) were not readily available. This was achieved by means of the following formulas:
Stock return: (SPt - SPt-1) / SPt-1
Market-to-book ratio: Market value per share / Book value per share Return on assets: Net income / Total assets
Abovementioned variables enable the regression analysis to be conducted and determine to what extent total CEO compensation is dependent on them. For the second regression analysis some work needed to be done with regard to its independent variables. Since the regression analyses are looking at different aspects, such as industry focus and strategic emphasis, these variables need to be explained and formulated in a formulaic way. Although variables as R&D, SGA and advertising (expenditures) seem straightforward, they need to be controlled for size before they can be used reliably in a regression analysis. Based on this assumption, variables are created by dividing those expenditure numbers by sales for each observation. By doing this, the numbers reflect the expenditures compared to the corresponding sales activity.
Next, I will discuss strategic emphasis, based on the article by Mizik and Jacobson (2003). They argue that firms are all trying to find a balance between value creation and value appropriation. To measure the shifts in strategic emphasis they use the movements in the following ratio:
22
In this formula, strategic emphasis is operationalised by dividing the sum of R&D minus advertising expenditures by assets. According to Mizik and Jacobson, R&D has had the most attention in prior literature on value creation (p.65). Significant effects of R&D expenses on value creation have been found in this literature. Value appropriation seems to be linked especially to advertising expenses, since value appropriation is mainly about isolating value for yourself. This is exactly what firms try to achieve with advertising. For that reason, value appropriation is operationalized by dividing advertising expenditures by assets. However, to achieve consistency in this thesis, sales will be used to control for size and firm activity instead. Since we already proxy for firm size with sales for SGA as well as R&D expenses, using assets for strategic emphasis will likely distort our findings.
3.2.2 Dummy variables
In the sample dummy variables are created for fiscal year, as well as industry group, to categorize data into mutually exclusive groups of observations. In the case of fiscal year, this is done to control for economy-wide effects, whereas the industry group dummy is created to control for industry-specific effects. While creating dummy variables for fiscal year is relatively straightforward, identifying industry groups requires some more discretion.
The data collected through Compustat and ExecuComp included four-digit Standard Industrial Classification (SIC) codes, identifying the observation's corresponding industry group. One can derive different kinds of information from these codes. Whereas the complete four-digit codes are related to a specific industry, the first three and the first two digits refer to respectively the industry group and major group of the company. Additionally, divisions (e.g. Agriculture, Forestry, And Fishing; Mining; Construction) are formulated and assigned a letter ranging from A to J, all encompassing a range of SIC codes (OSHA, n.d.). For this thesis, I planned to differentiate mostly at division level. However, since division D (Manufacturing) and division I (Services) both account for 74.77% and 20.86% respectively in my sample, using dummies based on industry division level will probably return very biased results. That is why I decided to create dummy variables based on a major group level instead, based on the first two digits of the firm's SIC code. This enables me to make better controls for different industries. In the sample used, 41 unique values exist for SIC major group, meaning that the firms in my sample are spread over a total of 41 different industry groups. As a result, the dataset ends up with 20 dummy variables for fiscal year, and 41 for SIC industry major groups.
23 3.2.3 Dropping observations
The data collected from the ExecuComp database consisted of 33,518 observations. However, to ensure that all observations are complete (i.e. all variables are available), observations with missing values will have to be dropped from the dataset. Although there were several variables with missing values, most missing variables were generated by the advertising, R&D and SGA expense items. Together they were responsible for dropping at least 23,170 observations from the dataset (only one of three was needed to be missing from the data), after dropping observations based on other missing variables. Altogether, dropping missing observations brought the dataset from 33,518 to 3,830 observations. This is caused mostly by missing values for advertising expense (21,870 missing values over the entire dataset of 33,518). However, I consider advertising expense to be necessary for my analysis, because a firm's strategic emphasis is based on this variable.
As will be explained further in the next paragraph, some variables are not ready to be used immediately. For some of them measures need to be taken (using log-transformed values or Winsoring), before the data is usable. For some reason, STATA created nine missing values after creating a new variable based on the natural log of total CEO compensation. This will be expanded upon in the next paragraph. Since total CEO compensation is the dependent variable in the regression analysis and taking the natural log is needed to get reliable results in this case, these observations are dropped from the data set as well. All in all, after dropping all observations with missing values the dataset consists of 3,821 observations.
3.2.4 Log-transformed variables and Winsorising
There are certain variables that might not be usable directly from Compustat and ExecuComp, because of high kurtosis and skewness values that can cause variables to distort a study's findings. Kurtosis and skewness both tell something about the normal distribution of the variable. A normal distribution is very important in statistical analyses, since in most cases methods based on normal theory provide statistically significant results, even when the sample is not normally distributed in reality. Additionally, there is a very strong connection between the numbers of observations in a sample and the normality of a distribution. Large samples can often be treated as normally distributed, even when the distribution itself is not normal. This is why it is important that the most important variables in the sample comply to
24
certain requirements, that the skewness and kurtosis variables are within agreeable boundaries.
Skewness indicates to what extent the probability distribution of a variable leans to one side of the mean. Negative skewness, or left-tailed, skewness indicates that there are relatively few low values in the data and the probability distribution is concentrated to the right. Positive skewness, or right-tailed, skewness indicates the exact opposite that there are relatively few high values in the data. In the case of a normal distribution, always fifty percent of the observations are concentrated at one side of the mean, and fifty percent of the observations at the other side. Kurtosis, on the other hand, provides insight into the occurrence of peaks within the sample distribution. High values of kurtosis practically mean that there is a greater chance of extreme outcomes, compared to a normal distribution. A kurtosis value of 3 is considered normal for a normal distribution. For both skewness and kurtosis I would like to have a value of a maximum of five.
The most important variables that I control for high kurtosis and skewness values are total CEO compensation and total assets, my proxy for firm size. It is important for the statistical analyses that the independent variable approaches a normal distribution. For total assets this is important, because it is one of the few variables (together with total CEO compensation) that can be difficult to put in relation to other variables (ranges from very small to very large). Most of the other variables included in the data set are already controlled for size, making their relative values more comparable to each other. The table below provide summary information on both the total CEO compensation and total assets variables.
Table 3.1 Summary statistics variables
Variable Mean Std. Dev. Skewness Kurtosis
Total compensation 6524.789 15092.47 20.74662 691.8925
Total assets 9222.188 30503.63 7.512311 77.7645
As can be seen here, both total CEO compensation and total assets are suffering from high skewness and kurtosis values. Especially the kurtosis value of total CEO compensation is extremely high. As of now, using these variables will not provide an accurate representation of the truth. A common method to solve this problem is to use log-transformed variables in our analyses, meaning that natural logs of the variable are used, instead of the original values. The major advantage of this method is that the regression coefficients of other variables will now show proportionate effects on total CEO compensation, instead of a dollar value
25
specification. Given our mixed sample of large as well as small companies, this specification would be an issue for our analyses. If the sample would have been composed of generally large firms, a dollar value specification could have been appropriate (Core, Holthausen and Larcker, 1999, p.402). After creating the new log-transformed variables for the two variables, the following values came forward:
Table 3.2 Summary statistics log-transformed variables
Variable Mean Std. Dev. Skewness Kurtosis
Total compensation 8.084025 1.466426 -4.331292 45.60354
Total assets 7.438012 1.689522 .5496503 3.019996
Although using log-transformed variables fixed the problem for total assets, the problem persists (to a lesser extent) for total compensation. Skewness is barely an issue within the preset boundaries, while the value for kurtosis is still way too high. In order to make sure that these values come within reasonable boundaries, Winsorising will be conducted as well on total CEO compensation. Winsoring is the act of limiting extreme values in statistical data to reduce the number of outliers. In this case, the highest and lowest 1% of observations will be equalized to respectively the 1% and 99% observations. After doing this, the skewness and kurtosis values of total CEO compensation are well within the desired boundaries, with values of respectively -0.1005549 and 2.680655.
3.2.5 Definition of variables
In this paragraph I will briefly discuss the descriptives of the variables retrieved from Compustat and ExecuComp. This will provide an overview of the raw variables (not yet log-transformed and/or winsorised) and provides some clarity with respect to the range of values belonging to them.
It is important to note that the non-ratio variables (i.e. total CEO compensation, R&D expenses, SGA expenses and total assets) are formulated in thousands of dollars. The summary statistics of the most important variables are displayed in table 3.3. This concerns the raw data as retrieved from the Compustat and ExecuComp databases. Let me start with discussing total CEO compensation first. What immediately stands out is the low minimum value of total CEO compensation. This value (.001) would mean that the CEO in question was only compensated one thousand dollars in total in the given year, probably less as Stata
26
does not provide numbers with more than three decimals for minimum values. When we take a look at the more detailed descriptive statistics, one can see that this is the case for several cases (at least the four lowest values in the sample). They seem to be the few exceptions in the sample, since the 1% percentile shows a value of 254.937. These low values could be explained by compensation packages consisting of mainly equity arrangements which only provide financial advantages in later years. The highest value found in total CEO compensation is also an exceptional high value, since the second highest value is not even half of this (293,097.3). This seems to be the case for each variable with extreme values. Mostly it is caused by a small group of outliers, after which the values quickly increase or decrease towards more acceptable values. These extreme values for total CEO compensation have been eliminated by Winsorising and log-transforming for the actual regression.
Table 3.3 Summary statistics variables
Variable Mean Std. Dev. Min Median Max
Total CEO compensation 6524.789 15092.47 .001 3417.276 600347.4
Return on assets .0391656 .1817804 -5.868446 .0596776 .4995103 Stock return .8858754 46.19808 -.97 .03 2854.77 Market-to-book ratio 3.791981 5.545797 -54.58 2.68 122.92 R&D expenses 373.8105 1090.27 1.2 55.857 12183 SGA expenses 1694.992 3962.456 6.818 367 34663 Strategic emphasis -.0625605 .1444427 -3.938345 -.0341019 .3267625 Total assets 9222.188 6.928123 22.235 1456.365 479921 CEO tenure 6.715519 6.928123 -13 5 39
Return on assets seems to have an extremely low minimum value, given its positive mean value. Whereas most observations seem to have slightly positive results, and top values do not exceed .5, the lower values are somewhat more extreme. The lowest 1% of the sample observations have a return on assets ranging from -5.87 to -.51. This lowest extreme value seems to belong to a firm with only one observation in the dataset, possibly indicating a very short-lived company. The second lowest ROA in the dataset is not even half of this number (-2.908), indicating much more healthy figures (although still loss-making). Although these figures are quite extreme, they do only cover the lowest 1% percentile of the sample. In addition, the mean and standard deviation seem normal.
This brings us to the stock-related variables, stock return and market-to-book ratio. The mean values are .8858754 and 3.791981 respectively. These values would mean that the
27
average firm in the dataset enjoys a 88.6% increase in share price in a given year, while the market price is 3.8 times higher than the book value of the share. Although there is not much wrong with this data (stock return might be a little high), the data is heavily influenced by very extreme values in the dataset: -.97/2854.77 and -54.58/122.92 respectively as minimum/maximum values for the two variables. To make sure that these extreme values will not influence the results of the regression, I decided to Winsorise both stock return and market-to-book ratio, by replacing the highest and lowest 1% of observations with the values of respectively the 1% and 99% observations. Doing this resulted in much more reliable data, as shown in table 4.2.
Table 3.4 Summary statistics Winsorised variables
Variable Mean Std. Dev Min Max
Stock return .0977519 .5417109 -.79 2.6
Market-to-book ratio 3.729448 3.588988 -3.19 21.06
Next are the R&D and SGA expenses, which are pretty straightforward to read, without any extreme values likely to distort the results. The present minimum and maximum values are caused by the huge variety of companies with regard to firm size. As mentioned before, all numbers here should be read in thousands of dollars. The same is the case for total assets, which is relatively easy to understand. The only difference with respect to R&D and SGA expenses is that log-transformed total assets will be used in the regression (raw data is displayed in table 3.3).
Strategic emphasis requires some more discretion to read however. Earlier, I discussed the work of Mizik and Jacobson (2003) on value creation versus value appropriation. The strategic emphasis variable displays the relationship between R&D and advertising expense, with negative and positive numbers meaning a relative focus on R&D (value creation) and advertising (value appropriation) respectively.
The last variable to be discussed here is CEO tenure. The descriptive statistics brought up some interesting information. There is nothing special about either the mean values, standard deviations or maximum value, but the minimum value draws attention. A value of -13 would mean that the corresponding CEO would have experience of negative thirteen years. The 1% percentile still has a value of -4. It seems that former CEOs (e.g. founders) leave the company after a while and return again later, which could possibly reset the date of becoming CEO variable in the database. Since CEO tenure is calculated by subtracting the year
28
someone became CEO by fiscal year, this can result in weird data such as this. To fix this problem, I will replace all cases of negative CEO tenure with the average for the main regression. With all the negative values replaced, these are the new descriptives:
Table 3.5 Summary statistics amended CEO tenure
Variable Mean Std. Dev Min Max
CEO tenure 6.946525 6.729997 0 39
Please note that in this paragraph I have been discussing the raw data retrieved from Compustat and ExecuComp, not the Winsorised and log-transformed variables I will use in the regression analysis (i.e. total CEO compensation, total assets, stock return, and market-to-book ratio). These will be discussed in the next chapter concerning results. Compustat definitions of the variables can be found in table 3.6 on the next page.
29
Table 3.6 Definition of variables
Variable Database item Definition provided by database
Total CEO compensation TDC1 Total compensation for the individual year, comprised of the following: Salary, Bonus, Other Annual, Total Value of Restricted Stock Granted, Total Value of Stock Options Granted (using Black-Scholes), Long-Term Incentive Payouts, and All Other Total.
Market value per share PRCC_F Price Close - Annual - Fiscal
Book value per share BKVLPS For companies, Book Value Per Share is based on fiscal year-end data and represents Common Equity Liquidation Value (CEQL) divided by Common Shares Outstanding (CSHO).
Net income NI This item represents the fiscal period income or loss reported by a company after subtracting expenses and losses from all revenues and gains.
Total assets AT This item represents the total assets/liabilities of a company at a point in time. If the company does not report a useable amount, this data item will be left blank
Total sales SALE This item represents gross sales (the amount of actual billings to customers for regular sales completed during the period) reduced by cash discounts, trade discounts, and returned sales and allowances for which credit is given to customers, for each operating segment. Differences, which exist between the data as reported by the company and the Compustat definition, will be indicated by a footnote.
Advertising expenses XAD This item represents the cost of advertising media (i.e., radio, television, and periodicals) and promotional expenses.
R&D expenses XRD This item represents all costs incurred during the year that relate to the development of new products or services.
SGA expenses XSGA This item represents all commercial expenses of operation (i.e., expenses not directly related to product production) incurred in the regular course of business pertaining to the securing of operating income.
30 4. Results
4.1 Descriptives
The sample consists of 3,821 observations over a twenty-year period, with firms spread over a total of 41 different industries. The following table provides summary statistics on the study's most important variables, as they will be used in the regressions (log-transformed and Winsorized where needed):
Table 4.1 Summary statistics variables
Variable Mean Std. Dev. Min Median Max
Total CEO compensation 8.133642 1.079098 5.541017 8.136599 10.61905
Return on assets .0391656 .1817804 -5.868446 .0596776 .4995103
Stock return .0977519 .5417109 -.79 .03 2.6
Market-to-book ratio 3.729448 3.588988 -3.19 2.68 21.06
R&D expenses / sales .0927764 .1359541 .0001809 .0559735 3.938345
SGA expenses / sales .359935 .23361 .0136187 .3222848 5.216127
Strategic emphasis -.0625605 .1444427 -3.938345 -.0341019 .3267625
Total assets 7.438012 1.689522 3.101668 7.283699 13.08138
CEO tenure 6.946525 6.729997 0 5 39
In this paragraph I will discuss the economic details concerning the variables presented above, instead of their statistical relevance as done in the last chapter. Firstly, total CEO compensation and total assets are now somewhat harder to interpret. Log transformed data better reflect a normal distribution, often resulting in better results, but the numbers cannot be interpreted as economic values anymore.
Next we have return on assets, which indicates how profitable a company is relative to its total assets. This performance indicator shows how efficiently firms can turn investments into profits. The higher this number, the more profit is made relative to total assets (which is my proxy for firm size). This means that the average firm in our sample converts approximately 3.9% of its assets' worth into profit. The minimum value in the sample is an extreme one, namely -5.87. A value as low as this means that the firm made a loss of approximately 6 times (587%) the worth of its assets, which is a very bad sign. The data ranges from -587% to a positive 50%. Given a mean of 3.9%, this shows that the low minimum value is indeed an extreme case, since the mean value is positive.
