Are CEOs Charged for Stock-Based Pay? An Instrumental Variable Analysis

Transcription

1 Are CEOs Charged for Stock-Based Pay? An Instrumental Variable Analysis Nina Baranchuk School of Management University of Texas - Dallas P.O. BOX SM31 Richardson, TX nina.baranchuk@utdallas.edu March 27, 2006 Abstract Using instrumental variables approach, this paper investigates whether an increase in stockbased compensation for CEOs is associated with a reduction in the dollar value of all other CEO compensation, as agency theory would suggest. I find that, on the contrary, larger stock-based compensation is on average accompanied by larger other components of CEO compensation. Similar results are obtained when focusing on option grants alone. The analysis of the compensation practices in corporations that have voluntarily started to include option grants into income statements shows little support for the hypothesis that accounting methods affect the use of options in CEO compensation. Acknowledgements: I am especially grateful to my dissertation committee chair Siddhartha Chib for providing valuable guidance. I also appreciate the valuable comments of Ted Day, Shane Johnson, Todd Milbourn, and seminar participants in Texas Lone Star conference in Houston, 2005.

2 I Introduction From the basic agency theory point of view, the increased role of stock-based pay in executive compensation packages that took place in the 90s is a welcome change. However, accompanied by dramatic increases in the dollar size of the compensation packages, this change has raised concerns that, instead of serving the firm s and the shareholder s interests, the current compensation practices help CEOs take money out of the firm. Critics in particular point out that according to the current accounting standards, option grants are the only form of compensation that does not have to be reported on the income statement. 1 While there is a vast existing empirical literature that focuses on the use of stock-based compensation, an analysis of other components of CEO compensation is relatively scarce. This paper attempts to shed some light on the efficiency of both stock-based and other forms of pay by evaluating the tradeoff between the dollar value of stock-based pay and all other components of CEO compensation. I find that when a CEO receives new stock-based pay, the CEO also enjoys an increase in all other compensation. The result holds even after controlling for factors that may lead to an increase in the optimal CEO compensation, such as firm size and improvements in performance. This finding is indicative of CEOs using stock-based pay to take money out of the firm. Moreover, stock-based pay is not solely responsible for unjustified increases in CEO pay, and similar inefficiencies may be present in the use of the other components of CEO compensation. Using Gompers, Ishii, and Metric (2003) governance index, I find that better governance (lower index) corresponds to lower compensation levels both in terms of stock-based and other forms of pay. There is only a fairly weak evidence that the tradeoff is more pronounced in firms with better governance. Interestingly, I find that firms with a lower index have less persistent option grant policies: smaller option grants are more likely to follow a large option grant when governance is better. This finding may be interpreted as a sign that firms with better governance are more actively involved in designing CEO compensation. Empirical analysis of the tradeoff between stock-based and all other CEO compensation is complicated by the endogeneity of the option and restricted stock grant policies and the policies on all other parts of the compensation package. Both stock-based and all other CEO compensation are affected by a host of difficult to measure or unobserved factors. These factors may include expected future investment opportunities and firm performance, risk preferences of the agents, information 1 FASB has recently changed the standards, and most firms will have to comply to the new rules starting in

3 asymmetry, and managerial ability, among others. I address the endogeneity problem using an instrumental variables approach. In particular, I use lagged stock-based compensation and four variables that are commonly used to measure agency relationship as instruments for option grants. The agency related variables are Tobin s Q, CEO share ownership, leverage, and dollar volatility of stock returns. The choice of the instruments is motivated by the observation that agency-related variables determine the optimal incentive structure of CEO compensation, but do not directly influence the dollar size of the optimal compensation package adjusted for risk. In my analysis, I rely on the sample of large publicly traded US corporations, on which the data is available from CRSP and ExecuComp. The instrumental variable models estimated in this paper capture the unobserved factors that influence both stock-based and all other CEO compensation by allowing for a non-zero correlation between the error terms in the equations for option grants and all other CEO compensation. I find the correlation to be negative and significant. This estimate indicates that the unobserved factors that affect both options and all other CEO compensation in the opposite directions, such as risk aversion and CEO talent, have more impact on the CEO compensation than the unobserved factors that affect the two variables in the same direction, such as unobserved determinants of the total compensation size. While the main results in the paper are based on panel data analysis, I investigate whether similar results are present in the cross-section by looking at the averages of stock-based and all other pay over the sample period. This approach also allows me to address concerns that current option grants are often a part of a long-term compensation plan, and therefore the charge for stockbased pay might not be contemporaneous with the stock-based pay itself. The results from the cross-sectional analysis, however, are very similar to those from the panel data analysis. While stock-based pay mainly consists of option grants, it also often includes restricted stock pay. In an important contribution, Hull and Murphy (2001) provide a theoretical analysis of a firm s optimal choice between option grants and restricted stock grants. They find that option grants are preferred when it is optimal to leave cash pay unchanged and increase the total compensation, while restricted stock is better if the cash compensation is reduced and the total compensation (adjusted for risk) remains unchanged. This analysis implies that it may be optimal to leave all other pay unchanged if stock-based pay mainly takes the form of option grants. I find that focusing on option grants alone and removing restricted stock grants produces the same evaluation of the tradeoff. The positive relationship between option grants and all other compensation is inconsistent with the 2

4 efficient use of option grants. Because option grants are the only form of compensation that do not have a negative effect on the reported accounting income of the firm, it may be argued that changing the accounting standards and recognizing option grants as an expense would discourage CEOs from using new option grants to take money out of the firm. To test this hypothesis, I analyze a small sample of firms that voluntarily chose to expense option grants against their income. The results show that expensing options has no significant effect on the CEO compensation in these firms. I find therefore that the data does not support the hypothesis that accounting affects option awards to CEOs. The most closely related paper is Bertrand and Mullanaithan (2001) (BM), who consider the tradeoff between option grants and all other CEO compensation. Their approach, however, is quite different. They circumvent the endogeneity problems by concentrating on the analysis of how the quality of governance affects the charge for option grants. Additionally, they focus on the firms in oil industry. They find support for the hypothesis that better governed firms charge more for option grants. While the findings of BM provide cross-firm comparisons of the charge for option grants, the results presented in this paper offer an assessment of the level of the tradeoff between option grants and all other CEO compensation for a large sample of publicly traded US corporations. There is a vast empirical literature that analyzes whether executive compensation practices are in line with theoretical recommendations; see a review by Murphy (1999). While a number of authors find that CEO incentive compensation responds to agency problems and the size of compensation is related to the CEO s abilities (for example, Mehran (1995), Core and Guay (1999), Milbourn (2003)) others argue against that (Jensen and Murphy (1990), Yermack (1995), Ofek and Yermack (2000), Hall and Murphy (2003) to name a few). The findings in this paper are consistent with stock-based pay offered when there are more agency problems. The lack of tradeoff, however, indicates that the dollar size of both stock-based and all other pay may be too high. The following section describes the data sample. Section III develops an instrumental variable panel model with censored treatment, and Section IV discusses the estimation technique. Section V presents the main estimation results as well as a number of robustness tests. Section VI discusses accounting for option grants, and section VII concludes. II Data The main source of data is ExecuComp, which provides information on executive compensation and firm characteristics for large US public companies for the years Stock price information 3

5 Table 1: List of variables. The total sample consists of 1268 firms, 8398 firm-year observations for years The source for the variables is Compustat (unless mentioned otherwise). Variable Name Year calculation method Options current (Black-Scholes value of option grants)+(stock grants) Other current (salary)+(bonus)+(other annual)+(all other) vol($ returns) current (volatility of percentage returns for the last 60 month) (market value of shares outstanding) CEO share ownership current % of outstanding stock currently owned by the CEO Tobin s Q current book value of long term debt + market value of stock divided by book value of total assets financial leverage current (book value of debt)/(book value of total assets) ROA lagged volatility of ROE divided by stock return volatility zero dividends current 1 if paid no dividends tax loss c. f. current 1 if the firm has tax loss carry-forward interest coverage current EBIT/interest (absolute value capped at 50) ln(total assets) current ln(assets) last year stock return lagged % return on common stock G current governance index of Gompers Ishii and Metric (2003) new CEO dummy current 1 when a new CEO accepts the position departing CEO dummy current 1 when a the current CEO leaves the position industry dummies current 25 industry groups based on 2-digit SIC codes is obtained from the Center for Research in Security Prices (CRSP). Finally, I use the governance index developed in Gompers, Ishii, and Metrick (2003) available from IRRC. This index is designed to be inversely related to the quality of the firm s governance. After removing the firms that have less than 4 (not necessarily consecutive) observations, I obtain a sample that contains 1268 firms during years (8398 firm-year observations). The data set used in this paper is well known, so I offer only a brief statistical description of the variables in Table 1. Many studies have observed a dramatic increase in the use of option grants for CEOs during the 90s (see, for example, Yermack (1995) and Core and Guay (1999)). A tradeoff hypothesis implies that all other forms of compensation should decrease during the same period period. Casual observation, however, reveals that that is not the case. Figure 1 shows that the average dollar values of option grants, restricted stock grants, and all other CEO compensation, all increase fairly steadily from 1993 to While there appears to be some tradeoff between option and restricted stock grants, all three variables are highly positively correlated with all pairwise correlations above This observation alone, however, is not sufficient for ruling out tradeoff between cash and 4

6 stock-based pay. If the total size of CEO compensation has increased during the 90s, one would expect all parts of compensation to increase, including cash compensation. The following section describes the instrumental variable approach that helps address this endogeneity issue All Other Pay 5 4 Option Grants 3 2 Restricted Stock Figure 1: Option grants and all other CEO compensation Option grants are valued using Black-Scholes method with stock return volatility over previous 60 months. All other CEO compensation is a summation of salary, bonus, other annual, long term incentive payouts, and all other total compensation. Because the distribution of both variables is skewed, natural log transformations are applied to these variables. III Modelling Option and Cash Pay Tradeoff The tradeoff between stock-based pay and other components of CEO compensation cannot be analyzed by simply regressing CEO cash compensation on the dollar value of option grants to CEOs. Such approach would produce biased estimates because various parts of the CEO compensation are determined endogenously. Moreover, the direction of the bias is not obvious and depends on what unobserved factors have more influence on CEO compensation. On one hand, the unobserved factors that increase the total size of the CEO compensation are likely to produce a positive bias in the coefficient on stock-based pay because they induce an increase in both stock-based pay and all other compensation (this point is discussed in more detail in BM). On the other hand, unobserved CEO characteristics such as CEO ability and attitude towards risk, as well as the degree of shareholders control over option grant policies may produce a negative bias in the coefficient 5

7 on option grants. To illustrate the latter, consider CEO ability. Talented CEOs, expecting a good future performance, may wish to receive a larger proportion of their compensation in the form of option or restricted stock grants. Given a sufficiently strong relationship between ability and the composition of the optimal compensation package, one would expect a positive relationship between CEO ability and a dollar value of stock-based pay and a negative relationship between CEO ability and cash compensation. Thus, failure to account fully for CEO ability may result in a downward bias of the coefficient on stock-based pay. I address the endogeneity problem using an instrumental variables (IV) approach. First, note that the tradeoff between stock-based pay and cash compensation can be identified by having either stock-based pay or cash compensation on the left hand side and the other on the right hand side. I consider both types of model specifications. Where convenient, I adopt the standard terminology from the statistical literature on instrumental variables, and refer to the variable on the left hand side as the response and the variable on the right hand side as the treatment. Instrumental variables approach then requires identifying instruments for the treatment variable. After controlling for the observed data, these instruments have to be i) significantly correlated with the treatment and ii) uncorrelated with the response. When stock-based pay is the response and all other compensation is the treatment, I use lagged value of all other compensation as an instrument. When all other compensation is the response and stock-based pay is the treatment, I identify instruments by considering variables that may affect agency problems in the firm. According to Agency Theory, such variables determine the optimal level of pay-for-performance sensitivity of a CEO compensation and do not have a direct effect on the dollar size of the CEO compensation package, beyond reimbursement for risk taking. A number of studies document that the observed pay-for-performance sensitivity of the CEO compensation is mainly defined by the stock-based parts of compensation packages, typically represented by option and restricted stock grants (Jensen and Murphy 1990). Keeping in mind the requirement that instrumental variables have to satisfy, the following five variables are chosen as instruments for option grants: (i)lagged Options - dollar value of options granted in the previous year; (ii) CEO share ownership - the percentage of the firm s shares beneficially owned by the CEO; (iii) Tobin s Q - a ratio of the market value of the firm to the book value; (iv) leverage - a ratio of book value of debt to market value of shares, and (v) vol($ returns) - a volatility (standard deviation) of dollar returns to all shareholders of the firm, calculated over 60 months. These variables are used as proxies for the agency relationship 6

8 in the firm in a number of papers (Mehran (1995), Yermack (1995), Hermalin and Wallace (2001), to name a few). CEO share ownership proxies for alignment of CEO and firm owners interests; Tobin s Q proxies for the firm s growth opportunities, which complicate monitoring and interpreting CEO s actions; leverage reflects monitoring quality: CEOs in firms with more leverage are better monitored; and vol($ returns) measures the cost of providing incentive compensation to the CEO: higher volatility means higher risk premium required by the CEO for any given option award. The choice of instruments is further discussed in Section V. It may be argued that the current accounting treatment of option grants results in an inefficiently large amount of options awarded to CEOs as a compensation for their services. According to the current standards, an option grant is the only form of compensation that is not required to be expensed against the firm s income and hence does not reflect negatively on the firm s reported earnings. Therefore, CEOs are less likely to face shareholders disapproval if they use option grants to take money out of the firm. For more than a decade, the FASB attempted to change the accounting rules to require the same treatment for all forms of compensation, including option grants. Expecting this change, a number of companies voluntarily started expensing option grants against income. If accounting practices significantly affected option grant policies, variables reflecting accounting methods could be used as instruments for option grants. The available data, however, does not show any significant relation between the use of option grants in CEO compensation and a method of accounting for option grants in the sample. The analysis of this issue is presented in Section VI. From the statistical point of view, the model where all other compensation is a response and stock-based pay is a treatment differs very little from the model where the roles of the two variables are switched. Therefore, I first discuss in detail the former model and then briefly specify the latter model. Let Options it denote stock-based pay 2 to the CEO of firm i during year t Let Other it denote the dollar value of CEO compensation excluding option and restricted stock grants. Since both Options it and Other it have skewed distributions, I use natural log transformations to these variables. To exploit the panel structure of the data and capture heterogeneity in compensation practices documented in a number of studies (see, for example, Hermalin and Weisbach (2002)), I model the constant term as well as the coefficients on covariates other than the constant term as heterogenous and use the observed data to model the coefficient heterogeneity. The response 2 In my analysis, I also include restricted stock grants into the variable that measures stock-based pay. The notation here is chosen simply for convenience. 7

9 equation is specified as Other it = Options it β 0 + X it β 1 + W it βi + ɛ 1it, (1) where X it is a k vector of covariates (excluding the constant) on the ith firm in year t, whose effect on the outcome is not heterogenous across firms, and W it is a q vector of all other covariates (including the constant), whose effect on the outcome is modeled heterogeneously. Because different models estimated in this paper use different specifications of X and W, a more precise specification of these matrices is offered later. To address the possible endogeneity problem, Options it are modelled using instrumental variables z it = (CEO share ownership it, T obin s Q it, leverage it, volatility($ returns) it ). The Tobit model specification is chosen to model Options since the observed stock-based compensation is always nonnegative and 18% of the sample observations for this variable are equal to zero. Therefore, it is assumed that { } Options it = max Xit γ 1 + W it γ i + Z it γ 2 + ɛ 2it, 0. (2) This equation will be referred to as the treatment equation. Since omitted or unobserved variables may have an effect on both Options and Other, I allow for a non-zero correlation between the error terms in equations (1) and (2): (ɛ 1it, ɛ 2it ) N(0, Σ). To introduce further notation, let matrix Σ be defined as follows, ( ) σ 2 Σ = 1 ω ω σ2 2. To complete the model specification, I assume that β i and γ i are q dimensional random vectors, whose distributions depend on a set of firm-specific covariates A i. In particular, I assume that ( ) βi = A i β 2 + b i ; γ i where A i is a 2q r matrix, and b i is a 2q 1 random vector distributed as multivariate normal with a mean vector 0 and full covariance matrix D. When all other compensation is specified as a treatment and option grants as a response, then the model can be rewritten as follows. The treatment equation is { } Options it = max Other it β 0 + X it β 1 + W it βi + ɛ 1it, 0, (3) and the response equation is Other it = X it γ 1 + W it γ i + Z it γ 2 + ɛ 2it. (4) 8

10 All distributional assumptions on the model parameters remain the same. The analysis is mainly concerned with the following two parameters: coefficient β 0, which measures the tradeoff between option grants and all other compensation, and covariance ω, which reflects the endogeneity factors that are not captured by the included control variables. Basic agency theory implies that an efficient compensation should exhibit a positive tradeoff between stock-based pay and all other CEO compensation. Since stock-based pay is risky and agents are risk-averse, the decrease in all other compensation corresponding to a $1 dollar increase in stock-based pay is expected to be smaller than $1. Therefore, if compensation exhibits the tradeoff, coefficient β 0 will lie between -1 and 0. Positive (negative) ω indicates that common unobserved factors influence option grants and all other compensation in the same (opposite) direction. Thus, covariance ω helps identify what kind of unobserved factors play a more important role in determining the endogeneity of stock-based pay. IV Model Fitting The instrumental vairables model specified in the preceding section is difficult to analyze by frequentist methods. The likelihood function is not available directly and even with simulation methods finding the maximum likelihood estimates is a challenge. For this reason I turn to Bayesian methods, which over the past ten years have gone through a period of almost revolutionary growth and development (see, for example, Chib (2001)) in large part due to the emergence of Markov Chain Monte Carlo (MCMC) simulation methods and new ways of dealing with models with latent variables (Chib (1992) and Albert and Chib (1993)) and instrumental variables models (Chib and Hamilton (2000) and Chib (2003)). In the Bayesian approach, the parameters are treated as random variables, and given a prior distribution and the observed data, the focus is on the posterior distribution of the parameters, which by Bayes theorem is proportional to the prior distribution times the likelihood function. 3 In the model, this prior-posterior calculation can be conducted by Markov Chain Monte Carlo methods even though the likelihood function is largely intractable. Evaluation of the posterior distribution of the parameters is achieved by enlarging the parameter space with suitable latent variables for the responses that are censored. Given values of the latent variables, the model can be represented in a SUR form as a continuous data hierarchical panel model. Values of the pa- 3 standard reference texts of the Bayesian approach to statistical inference include Zellner (1971) and Bernardo and Smith (1994). 9

11 rameters are then sampled from relevant conditional posterior distributions using MCMC methods, as discussed in Chib and Carlin (1999). In essence, parameters are samples from the posterior distribution, augmented with latent variables, in an iterative process: sampling latent variables conditioned on the parameters and then sampling parameters conditioned on the simulated latent variables. Within one cycle of the MCMC algorithm, conditioned on the current values of the parameters, the latent censored responses are simulated from truncated normal distributions (according to the prescription outlined in Chib (1992)). It can be shown that iterations of these cycles produce draws from the joint posterior distribution of the parameters and latent variables. The marginal posterior distribution of the various parameters can be summarized in terms of the sample means, standard deviations and quantiles of the sampled draws. Full details of the MCMC algorithm that I have developed to fit the proposed model are given in Appendix 1. In that appendix I also provide the prior distributions that I use in the analysis. The priors I have chosen are non-restrictive and allow the data to play the major role in determining the posterior distribution. The prior on the covariate coefficients is centered at zero and the marginal prior variances on these covariates are centered at ten. The results I present are robust to reasonable changes in the prior hyperparameters, as I have confirmed. V V.A Results Tradeoff Between Option Grants and Other Compensation I estimate the tradeoff between stock-based pay and all other compensation using model (1)-(2) with the following instruments and control variables: Model 1. Response: Other Treatment: Options Instruments (z it ): Lagged Options, CEO share ownership, Tobin s Q, leverage, and vol($ returns); Controls with homogeneous coefficients (X it ): (governance index) (lagged Options) ln(total assets), interest coverage, tax loss carry-forward and zero dividends indicators, last year stock returns, and new and departing CEO dummies; Controls with heterogeneous coefficients (W it ): constant and governance index (G) Covariates modelling heterogeneity (A i ): constant (A i is an identity matrix). Table 2 reports the estimates (the mean values of the posterior distributions) obtained from Model 1. Most importantly, the coefficient on Options is positive and significant. The sign of this 10

12 coefficient means that on average, CEOs are not charged for their option or restricted stock grants. Instead, an average CEO enjoys an increase in other forms of compensation in addition to new option and restricted stock grants, even after controlling for endogeneity of stock-based pay. This result contradicts the efficient compensation hypothesis of a tradeoff between stock-based and cash compensation. Moreover, if only stock-based pay were used inefficiently, the coefficient on Options would have been insignificant. The positive and significant estimate, on the other hand, implies that the average firm uses both stock-based and cash compensation inefficiently. Interestingly, the signs of the coefficients on the instruments, which were chosen for their relationship to agency costs, are consistent with basic agency theory predictions. The result suggests that, on average, more incentive (stock-based) compensation is used when agency costs are high. It is somewhat difficult to place a dollar value on the increase in cash pay corresponding to the increase in stock-based pay because of the log transformation of the variables. For the median CEO, the estimate implies that a $100 increase in stock-based pay is accompanied by approximately $7.54 additional dollars of cash compensation. Taking into account that the median size of an option grant is about $700,000, the model estimates imply that an absence of option grants would correspond to a reduction of cash compensation approximately by $52,000 (or 0.5% of the median cash compensation). The negative and significant covariance of the error terms in the response and the treatment equations implies that there are unobserved factors which affect Options and Other in the opposite directions. While it is easier to think of the factors that affect Options and Other in the same direction, the negative covariance means that these factors are probably well-captured by the included covariates (possibly such as firm size and past performance measures). This guess can be further supported by observing that the size of the coefficient on Options, which equals 0.077, is noticeably smaller than the sample correlation between Other and Options, which equals I conjecture that the following factors may affect Options and Other in the opposite directions: CEO talent, risk preferences, career concerns and the CEOs ability to control their own pay. As discussed earlier, more able CEOs also may receive a bigger proportion of their compensation in stock-based form. If the portion of compensation represented by option or restricted stock grants increases quickly enough with the CEO s ability, omitting ability would induce a negative correlation between stock-based and all other compensation. Risk aversion determines the size of reward for risk required by the CEO for a given option or restricted stock award. Therefore, risk aversion is negatively related to Other and positively related to Options. While the average risk premium is 11

13 captured by the coefficient on Options in the response equation, the differences in risk preferences among CEOs result in a negative correlation between the two error terms. Similar effect may arise due to career concerns if career concerns are positively related to the premium required by the CEO for performance-based compensation. Finally, the effect of CEOs ability to set their own pay may affect Options and Other in the opposite directions if some parts of compensation are scrutinized less by the board of directors, the firm s shareholders, and the debtholders. Table 2: Model 1. Response Equation Treatment Equation Variable response: Other response: Options Options ( ) Instruments lagged Options ( ) CEO share ownership ( ) Tobin s Q ( ) leverage ( ) vol($ returns) ( ) Controls G ( ) ( ) G (lagged Options) ( ) ( ) Error covariance ( ) IV model (1)-(2). Control variables also include constant, ln(total assets), interest coverage, tax loss carry-forward dummy, zero dividends dummy, last year stock return, restricted stock grants, new and departing CEO dummies. 95% probability interval for the posterior is in parentheses. Following BM, I include in Model 1 the product of governance index and lagged option grants to capture the effect of governance on the tradeoff. If better governance means more tradeoff, the coefficient corresponding to this variable in the response equation is expected to be positive. The estimate reported in Table 2 is indeed positive, but insignificant, and the magnitude is small compared to the coefficient on Options. This contrasts the result of BM who find that better governance significantly improves the tradeoff. The difference in the results may be due to 1) high 12

14 correlation 4 between G*(lagged Options) and Options (correlation is 0.42), 2) different statistical models, 3) different data sets, and 4) different measures of governance quality used in BM. Although I cannot check the importance of the third and fourth reasons, I can investigate the role of the correlation and the model specification by estimating a regression similar to that used in BM on my data set. Hence, I next estimate the following regression: Other = β 0 Options + β 1 G lagged Options + β 2 Controls + ε, (5) where controls are constant, ln(total assets), return on assets, last year stock return, new and departing CEO dummies and year dummies. The results are reported in Table 3. Consistent with BM, the coefficient on the product of governance index and lagged Options is positive and significant. I obtain almost identical results if I use current option grants instead of lagged. Table 3: Model (5). Control variables also include constant, ln(total assets), return on assets, last year stock return, new and departing CEO dummies and year dummies. 95% probability interval for the posterior is in parentheses. Variable (response: Other) G (lagged Options) ( ) Options ( ) G ( ) Further indicator of excessive use of stock-based pay is the positive coefficient on the governance index G in the treatment equation. It implies that firms with weaker shareholder rights (larger G) offer more option and restricted stock grants. Additionally, the negative and significant coefficient coefficient on the product G (lagged Options) in the treatment equation implies that the size of option and restricted stock grants is more persistent in firms with weaker governance. It is difficult to argue on the theoretical basis that offering similar-size option grants year after year is inefficient. Nevertheless, the finding is suggestive of firms with better governance participating more actively in designing CEO compensation. As discussed earlier, the tradeoff between stock-based pay and all other compensation may be estimated not only using instruments for stock-based pay, but also using instruments for all other compensation. The model is given by equations (3) and (4) and is specified as follows: 4 Because this correlation may also affect the coefficient on Options, I estimated Model 1 excluding G*(lagged Options). The results are virtually identical to those reported in Table 2. 13

15 Model 2. Response: Options Treatment: Other Instrument (z it ): Lagged Other Controls with homogeneous coefficients (X it ): CEO share ownership, Tobin s Q, leverage, vol($ returns), ln(total assets), interest coverage, tax loss carry-forward and zero dividends indicators, last year stock returns, and new and departing CEO dummies; Controls with heterogeneous coefficients (W it ): constant and governance index Covariates modelling heterogeneity (A i ): constant (A i is an identity matrix). The results of estimating Model 2 are reported in Table 4. Qualitatively, the results resemble those presented in Table 2. The coefficient on all other compensation is positive and significant and the correlation between the two error terms is negative, but insignificant. The size of the coefficient implies that, on average, an exogenous $100 increase in all other compensation corresponds to an approximately $74.85 increase in the stock-based pay. Table 4: Model 2. Response Equation Treatment Equation Variable (response is Options) (response is Other) Other Instruments Lagged Other ( ) ( ) Controls CEO share ownership ( ) ( ) Tobin s Q ( ) ( ) leverage ( ) ( ) vol($ returns) ( ) ( ) G ( ) ( ) Covariance ω ( ) IV model (1)-(2). Control variables also include constant, ln(total assets), interest coverage, tax loss carry-forward dummy, zero dividends dummy, last year stock return, restricted stock grants, new and departing CEO dummies. 14

16 V.B Robustness Tests Looking for tradeoff between annual option grants and annual cash compensation may be misleading because many firms have option grant programs that stretch over several years. Tradeoff may be absent on a yearly basis if appropriate reduction in the cash part of compensation is made, say, at the the start of the option grant program. To address this concern, I estimate Model 1 using the data averaged over the years. If in any of the observed years cash compensation has been reduced to compensate for an increase in option grants, the averaged data should be able to capture the tradeoff. Using averages means that the data set becomes cross-sectional instead of panel. This, however, has little affect on the estimation procedure. Table 5 summarizes the results, which are very similar to those obtained from the panel data estimation of Model 1. The coefficient on Options in the response equation is positive and significant, with almost the same magnitude as in Table 2. The correlation between the two error terms is negative, although it is insignificant for the cross-sectional data, while it is significant when panel data is used. Table 5: Model 1 for cross-section. Response Equation Treatment Equation Variable (response: Other) (response: Options) Options ( ) Instruments CEO share ownership ( ) Tobin s Q ( ) leverage ( ) vol($ returns) ( ) Controls G ( ) ( ) error covariance ω ( ) IV model (1)-(2) applied to the cross-section of the firms. Control variables also include a constant, industry dummies, and averages of the following variables: ln(total assets), interest coverage, zero dividends dummy. 95% probability interval for the posterior is in parentheses. An alternative way to address the possible bias introduced by the focus on annual grants is to 15

17 look at the changes in the determinants of the optimal incentive pay instead of the levels of these determinants. Core and Guay (1999) argue that firms use option grants to restore incentives of the CEO when these incentives deviate from the optimal. This suggests that option grants may be related to the changes in share ownership, leverage, Tobin s Q, and volatility, rather than their levels. I next test this hypothesis by estimating the following model: Model 3. Response: Other Treatment: Options Instruments (z it ): Lagged Options (CEO share ownership), (Tobin s Q), (leverage), vol($ returns) Controls with homogeneous coefficients (X it ): CEO share ownership, ln(total assets), interest coverage, tax loss carry-forward and zero dividends indicators, last year stock returns, and new and departing CEO dummies; Controls with heterogeneous coefficients (W it ): constant and governance index Covariates modelling heterogeneity (A i ): constant (A i is an identity matrix). The results of estimating Model 3 are reported in Table 6. The coefficient on Options is still positive and significant, although the size is smaller than the one obtained from Model 1. While three of the instruments (lagged Options, (CEO share ownership), and (leverage)) are significant, the positive sign of the coefficient on (CEO share ownership) contradicts the basic agency theory implication that an increase in share ownership reduces the need for additional incentive pay. As in Model 1, the covariance between the two error terms in this model is estimated to be negative and significant. Stock-based pay variable used in this paper consists two parts: option grants and restricted stock grants. Option grants, however, typically face more criticism than restricted stock grants because it is easier to neglect accounting for options as an expense. Whether this bias against option grants is justified can be analyzed by evaluating the tradeoff between cash compensation and the option grants part 5 of stock-based pay. Column (1) in Table 7 reports the results of estimating Model 1 using just option grants instead of the sum of option and restricted stock grants, and instead including restricted stock grants in the list of covariates. As in the original version of Model 1, the estimate of the coefficient on Options is positive and significant, and its size is similar to that 5 In spite of the recent increase in the use of restricted stocks, only a few companies in the sample incorporate restricted stock grants into the CEO compensation: 79% of observations on restricted stock grants are zero. This large percentage of zero observations on restricted stock grants makes this variable largely uninformative as either a treatment or response 16

18 Table 6: Model 3. Response Equation Treatment Equation Variable (response: Other) (response: Options) Options ( ) Instruments lagged Options ( ) CEO share ownership ( ) Tobin s Q ( ) leverage ( ) vol($ returns) ( ) Controls G ( ) ( ) G (lagged Options) ( ) ( ) error covariance ω ( ) Control variables also include a constant, CEO share ownership, ln(total assets), interest coverage, zero dividends dummy, last year stock returns, last year ROA, industry and year dummies. 95% probability interval for the posterior is in parentheses. reported in Table 2. This result is not consistent with option grants being used more inefficiently than restricted stock grants. As in Model 1, the correlation between the two error terms is negative and significant. Black-Scholes valuation method used so far in this paper to value option grants assumes zero transaction costs. Arguably, this assumption is not valid in the context of option grants to undiversified executives who often face restrictions on trading in the firm s stock. Thus, Black-Scholes method may overstate the true value of option grants to the executives. Because Black-Scholes method is a fairly good proxy for the cost of option grants to the firm, one would expect fewer option grants when there is a larger discrepancy between the Black-Scholes valuation (cost to the firm) and the executive valuation of option grants. This implies that using Black-Scholes valuation may induce a spurious finding of tradeoff between option grants and all other compensation. Thus, using a more precise valuation method would only strengthen the conclusion reached in the 17

19 previous section. To verify that, I next estimate Model 1 using an alternative valuation method developed by Hall and Murphy (2002), which takes into account the additional risk premium required by executives who have a limited ability to diversify away idiosyncratic risks. As in the previous robustness test, I focus on option grants only and include restricted stock grants in the list of covariates. As expected, the results reported in Column (2) of Table 7, are similar to those obtained using Black-Scholes valuation method, with the coefficient on Options larger than that reported in Table 2. The informativeness of the results presented so far crucially depends on the validity of the instruments. As mentioned in Section 2, valid instruments have to be, first, significantly related to the treatment variable and, second, insignificantly related to the error term in the response equation. As Table 2 shows, three out of five instruments used in Model 1 satisfy the first requirement. The second requirement on instruments is more difficult to verify. Since there is no direct test, I analyze the results of Model 1 for robustness to removing some of the instruments. The CEOs that own the firm s shares receive firm-related income not only in the form of compensation but also in the form of stock returns. The income obtained from stock returns therefore contributes to compensating the CEO for the appropriate level of effort. Moreover, CEOs with higher share ownership benefit less from any given increase in the compensation, since increases in compensation negatively affect the value of the firm s shares. Therefore, CEO share ownership may be negatively related to the reported total CEO compensation and thus may violate the second requirement that instruments must satisfy. To address this critique, I modify Model 1 by excluding CEO stock ownership from the list of instruments and instead using them as controls in both response and treatment equations. The results are reported in Column (3) of Table 7. The coefficient on Options is positive and significant and covariance between the two error terms ω is negative and significant. Both parameters are similar to those produced by Model 1 (as reported in Table 2). Similarly, I address the possible critique of leverage by excluding it from the list of instruments and including it as a control in both equations. The results, reported in Column (4) of Table 7, are qualitatively similar to those produced by Model 1. It may be argued that CEOs in firms with more volatile performance require higher fixed salary to compensate for the additional risk imposed by the volatile performance to their careers (Chen and Jiang (2003)). While I control for career concerns by including probability of turnover in both response and treatment equations, this may not be sufficient. To analyze whether career concerns affect the results, I estimate a modification of Model 1 where Tobin s Q and vol($ returns) are used 18

20 as controls and not as instruments. The results are reported in Column (5) of Table 7. It shows similar results to those obtained by fitting Model 1: coefficient on Options is positive, covariance ω is negative, and both parameters are significant and have approximately the same values as in Model 1. Table 7: Robustness tests. (1) (2) (3) (4) (5) Main Equation Options CEO share ownership (0.003) (0.013) (0.007) (0.014) (0.015) (0.0002) Tobin s Q leverage (0.005) (0.425) vol($ returns) (0.147) Treatment Equation Lagged Options (0.045) (0.016) (0.049) (0.005) (0.062) CEO share ownership (0.001) (0.001) (0.001) (0.001) (0.001) Tobin s Q (0.023) (0.021) (0.029) (0.028) (0.029) leverage (0.326) (0.236) (0.398) (0.320) (0.481) vol($ returns) (0.523) (0.535) (0.617) (0.658) (0.741) error covariance matrix Σ ω (0.019) (0.113) (0.093) (0.182) (0.093) IV model (1)-(2). Control variables also include constant, ln(total assets), interest coverage, tax loss carry-forward dummy, zero dividends dummy, last year stock return, restricted stock grants, new and departing CEO indicators, and controls for time and industry. Standard deviation of the posterior is in parentheses. It is useful to note that the variables that capture agency problems may also be correlated with the efficiency of CEO compensation. If governance index is not sufficient to control for compensation efficiency, then using agency-related variables as instruments would produce biased estimates. This critique, however, relies on the assumption that CEO compensation is inefficient in some firms, which is in line with the estimation results. 19

21 VI Does Accounting for Option Grants Matter? During the last decade, FASB has been arguing that the current accounting standard that does not require companies to expense employee option grants is misleading shareholders. 6. Expecting a change in accounting rules, a number of companies recently announced that they will voluntarily start reflecting employee option grants on their income statements. If accounting affects option award policies, then the recent events should shed some light on the use of option grants as well as tradeoff between option grants and other parts of compensation packages. In this section, I present descriptive statistics for the data sample, which at this point does not detect any significant relationship between the method of accounting for option grants and composition of CEO compensation. Information about firms that announced plans to expense option grants is obtained from the testimony of Chairman of the Board Robert Herz, issued on April 20, This testimony contains a list of 483 companies that made the announcement before February 12, Along with company names, the testimony provides the date of the first announcement, the year when the practice is adopted, industry classification and market capitalization information. Since 28 companies in the list are private, market capitalization data is available for 455 firms. The testimony does not contain information on compensation. For 122 companies, however, I am able to obtain this information from ExecuComp. Since the last year available in ExecuComp is 2002, the data on pre- and postannouncement compensation is available only for the 53 companies that made the announcement before Out of these 53 companies, 50 made the announcement in Therefore, to see whether the announcement affected CEO compensation structure, I compare the change in compensation in year 2002 for these 50 firms to that for the control group. If the accounting treatment of option grants affects composition of executive compensation, I would expect to observe a change in the size of option grants following a change in accounting. Specifically, I would expect the firms that start reporting option grants on the balance sheet to reduce option grants and increase restricted stock grants. Most of the announcements concerning the change in accounting came during or after 2002, although two companies made the announcement as early as 1995 and eight more followed by year The total sample of 483 firms (treatment group of firms) represents a variety of industries, and even includes 5 internet firms: Amazon, Netflix, Interactive Corporation, Rae systems, and 6 Potential economic effects of changing the accounting standards as well as explanations for the fierce opposition on the part of corporations against the new rules are discussed in Guay, Kothari, and Sloan (2003) 20