New insights into executive compensation and firm performance
Evidence from a panel of “new economy” firms, 1996-2002
The Authors
Giorgio Canarella, Department of Economics and Statistics, California State University, Los Angeles, California, USA
Arman Gasparyan, College of Business and Economics, California State University, Los Angeles, California, USA
Acknowledgements
The authors gratefully acknowledge the valuable help provided by Mahmoud Nourayi. Any remaining errors are the authors' full responsibility.
Abstract
Purpose – This paper aims to examine the relation between executive compensation, firm size and firm performance on a panel of the so-called “new economy” firms in the USA over the period 1996-2002.
Design/methodology/approach – The authors use two measures of performance, total shareholder return and return on assets, and concentrate on total CEO compensation, which includes stock option compensation, as equity-based compensation practices have been prevalent in new economy firms. The estimation process uses both the feasible generalized least squares method of Parks and Kmenta and the panel corrected standard error method of Beck and Katz. These methodologies investigate error structures that do not conform to the classical ordinary least squares assumptions.
Findings – The econometric results indicate that estimates on firm size are robust to alternative specifications of the error structures. There is evidence however that the effect of firm size on CEO compensation is more significant after the stock market crash of 2000. The opposite holds true for the estimates on firm performance. In addition, estimates on firm performance are more sensitive to the estimation method and the specification of the error structures.
Research limitations/implications – The research presented in this paper is a first step in the direction of understanding the pay to performance relation in the “new economy” industries in the USA. Additional research is warranted, which should extend both the time series and the cross section aspects of the data.
Originality/value – The paper fills an important gap in the existing literature by providing rigorous econometric evidence on the pay to performance relation in the so-called “new economy” industries. The evidence provided in this paper is relevant as it complements the findings in the literature on executive compensation in the so-called “old economy” industries, which typically make up the samples of most previous studies.
Article Type:
Research paper
Keyword(s):
Business performance; Compensation; Senior management; United States of America.
Journal:
Managerial Finance
Volume:
34
Number:
8
Year:
2008
pp:
537-554
Copyright ©
Emerald Group Publishing Limited
ISSN:
0307-4358
1. Introduction
The emergence of the so-called “new economy” industries is, obviously, a rather recent phenomenon and it is not surprising that research has not produced much evidence concerning the pay to performance relations in these industries. Yet, such evidence is relevant as it complements the findings in the literature on executive compensation in the so-called “old economy” industries, which typically make up the samples of most previous studies. For example, Anderson et al. (2000) estimate a simultaneous equation model of executive compensation and firm performance for the “information technology” industry and provide evidence that the shares of both bonus and option pay increase with performance and that the pay level and the extent of incentive pay positively affect firm performance. Ittner et al. (2003) find evidence, based on 1998 and 1999 data from a proprietary sample of firms, that the determinants of equity grants are significantly different in new versus old economy firms. Murphy (2003) complements the Ittner et al. (2003) results by analyzing data over a longer time period (1992 through 2001) and document the effect of the 2000 market crash on stock-based pay in “new economy” firms. Stathopoulos et al. (2004) report that a large proportions of new economy firms in the UK, in stark contrast to US findings, issue executive stock options in-the-money and that the composition of the remuneration committee has a significant impact on the moneyness of stock options.
As amply documented by Anderson et al. (2000) and Stathopoulos et al. (2004), “new economy” firms differ from “old economy” firms in terms of the distinctiveness of their internal organization, the makeup of their control structures and internal monitoring technologies, the nature of their product markets, the composition of their assets and the characteristics of their skills requirements. In addition, Ittner et al. (2003) and Murphy (2003) report that “new economy” firms are typically smaller in size, at least in terms of sales and employees, although not in terms of market value, and their growth opportunities are on average significantly higher than “old economy” firms. Most importantly, the compensation practices in “new economy” firms rely more heavily on equity-based forms of compensation than their “old economy” counterparts. Indeed, the latter characteristic has been described as one of the main organizational innovations that have emerged in the 1990s.
Although the theoretical and empirical literature on executive compensation is fairly well developed, it is far from complete and there are still many issues worthy of continued research. This paper extends the existing executive compensation literature in several dimensions. In the first place, this paper attempts to fill an important gap in the existing literature by providing rigorous econometric evidence on the pay to performance relation in “new economy” industries in the USA using new panel data on CEO compensation in 48 firms over the period 1996-2002. The definition of “new economy” industry follows Murphy (2003). The empirical analysis focuses on total CEO compensation, which includes stock option grants, as equity-based compensation practices have been prevalent in the new economy firms. We revisit issues that have not been completely resolved in the executive compensation literature dealing with “old economy” firms (Murphy, 1986). Is the association between firm size and CEO compensation as strong and robust in “new economy” firms as it has been found in “old economy” firms? Is executive compensation responsive to changes in the firm's shareholder returns? Even if it is, is it more responsive to accounting rates of return than to shareholder returns? Both accounting and market return measures have been considered in the compensation literature on “old economy” firms, however no definite conclusion has been reached (Lambert and Larcker, 1987; Clinch, 1991). The assumption that the response of executive compensation to market return has been weaker than that to internal measures of performance, such as returns on assets, may be an important rationale for the proliferation of stock options in executive compensation packages in “new economy” firms. It is well known that agency theory suggests that in a moral hazard setting the interests of the CEO can be aligned with the preferences of the shareholders through compensation arrangements that reward the CEO for firm performance (Fama and Jensen, 1983; Jensen and Meckling, 1976). If shareholders find it to their advantage to offer such powerful incentives, and if increasing shareholder wealth increases total welfare, then it is hard to fault the growth of stock-based executive compensation. But there are reasons to doubt that both conditions hold, and for that reason to ask whether executive compensation practices are nothing but the outcome of bureaucratic extravagances they have been made out to be.
Secondly, we evaluate the robustness of the statistical findings with respect to alternative estimation methods and stochastic error specifications. The statistical analysis employs panel data and a significant issue when dealing with panel data concerns the existence of non-spherical error structures (Greene, 2003; Wooldridge, 2002). It is well known that panel data typically exhibit group-wise heteroskedasticity, contemporaneous correlation and first-order autocorrelation (Beck and Katz, 1995). The application of ordinary least squares (OLS) to data characterized by non-spherical errors produces inefficient coefficient estimates and inconsistent standard errors. We use the insights from the econometric and statistical research on feasible generalized least squares (FGLS) and panel corrected standard errors (PCSE) models to investigate error structures that do not conform to OLS assumptions and appraise the sensitivity of the statistical findings to alternative assumptions regarding the error structure of the estimated models. We present estimates by both Parks–Kmenta (Parks, 1967; Kmenta, 1986; Kmenta and Oberhofer, 1974) and Beck–Katz (Beck and Katz, 1995, 1996) methods, and, within each method, we exploit different hypotheses. This is in contrast to the vast majority of previous studies on executive compensation, where, to our knowledge, this issue has generally been ignored.
Finally, we examine whether the downturn of the stock market in the second half of the sample period is reflected in a structural change of the pay to performance relationship. The empirical analysis extends over the period 1996-2002. This has been a time of great turbulence for many “new economy” firms, particularly as a result of the stock market bubble that exploded in the spring of 2000. This raises at least one issue that remains unresolved in the empirical literature: has the bubble burst destabilized the link between executive compensation and firm performance? The issue of parametric heterogeneity of the pay to performance relation merits additional attention since the stability of such relation undoubtedly has important consequences in terms of the explanatory power of agency theory.
The rest of the paper is organized as follows. Section 2 specifies the econometric model and discusses the relevant hypotheses. Section 3 briefly explains the details of the statistical methodology. Data sources and summary statistics are discussed in section 4. Section 5 contains the main findings. Finally, a summary is provided in section 6.
2. Econometric specification
The standard empirical model of CEO compensation contains firm size and firm performance as determinants of pay. Firm size is a measure of managerial discretion and firm performance is an indicator for managerial incentive compatibility. There is no consensus in the CEO compensation literature with respect to the suitable functional specification. Coughlan and Schmidt (1985), Hall and Liebman (1998), and Gibbons and Murphy (1992), among others, prefer the “elasticity” specification, where the level of (or the change in) the log of CEO compensation is linked to the level of (or the change in) the log of firm performance. Jensen and Murphy (1990), on the other hand, use the “sensitivity” approach, which links the level of (or the change in) CEO compensation to the level of (or the change in) firm performance. The different specifications yield quite different results, but neither the sensitivity nor the elasticity approach strictly dominates the other. Following Jensen and Murphy (1990), Ingham and Thompson, (1995) and Main et al. (1996), among others, we specify the pay to performance relationship in logarithmic first differences. This specification implies that the relationship between CEO compensation and performance is contemporaneous only, that is, a one time increase in firm performance increases compensation only within the current period. In addition, this specification removes the firm-specific fixed effects, i.e. it differences out the stable effects of time-invariant omitted factors, whether observed or unobserved, such as the diverse personal characteristics of the CEO (provided that there is no change in CEO during the sample period) that otherwise may distort the estimation of the pay to performance relationship (Murphy, 1985; Leonard, 1990; Conyon et al., 2000). Accordingly, we specify the basic econometric model as follows (see equation 1)for firm i = 1, … , N and year t = 1, … , T. Δ stands for the first difference operator (i.e. ΔX it = X it − X it−1), C it refers to CEO total compensation of firm i in year t, P it denotes a performance metric of firm i in year t and S it denotes the size of the firm, which is measured as total sales of firm i in year t. The term u it denotes a stochastic error term which reflects contemporaneous exogenous shocks to the logarithmic change in CEO compensation, whereas the term ω t captures time-specific effects. The detailed assumptions regarding u it are described below. Total compensation includes both cash and non-cash compensation. Cash compensation is defined as the sum of salary and cash bonus. Non-cash compensation, on the other hand, is composed of long-term incentive payouts, the value of restricted stock grants, the value of stock option grants and any other compensation item for the year. Stock options are valued at the grant-date using ExecuComp's modified Black and Scholes (1973) methodology. Firm performance and firm size are standard variables that have been included in prior empirical studies of executive compensation in the USA, the UK and other advanced market economies. The application of agency theory to the design of executive contracts in general predicts a positive correlation between CEO pay and some observable measure of firm performance. However, in the empirical compensation literature there is no consensus on the optimal measure of firm performance. Coughlan and Schmidt (1985) and Murphy (1985), among others, argue that firm performance should be measured by stock return because it reflects shareholders wealth. Lewellen and Huntsman (1970), Lewellen et al. (1987), Sloan (1993), Joskow and Rose (1994) and Carpenter and Sanders (2002), on the other hand, contend that accounting-based measures are more informative of the managerial contribution than market-based measures because they are less affected by the noise of the market. Although different specifications together with different measures of firm performance lead to quite different estimates of the coefficients on performance, they all reveal a statistically significant and positive relation of pay to performance. Accordingly, to avoid potential biases inherent in using either measure alone, both accounting-based and stock market-based measures are used in this study. Accounting-based performance is measured as return on assets (ROA), defined as net income before extraordinary items in fiscal year t divided by average total assets. Stock market performance is measured as return on common stock (total shareholder return (TRS)), defined as the closing price at fiscal year-end plus dividends divided by the closing price in the prior fiscal year-end.
The role of firm size on CEO pay is consistent with a variety of explanations (Rosen, 1992; Kostiuk, 1990). Rosen (1992) provides a theoretical justification for the positive relation between pay and firm size. As the firm becomes larger in size the complexity in operation also increases. Company size serves as the main proxy for the complexity of managerial ability and responsibility, the relative importance of executive decisions, and also the firm's standing within the industry and in the market in general.
3. Statistical methodology
OLS can be used to estimate models created by a pooled specification, such as Equation (1). However, in the presence of non-spherical errors, such as heteroskedasticity across firms and autocorrelation within firms, OLS yields consistent but inefficient parameter estimates. Inefficiency implies that, in repeated sampling, on average, the parameter estimates in OLS will equal the true values, but the variance of the parameter estimates will not be minimized because information about the non-spherical nature of the errors is not incorporated in the estimation procedure. While efficiency implies a minimized variance of the coefficient estimate, and therefore an estimate that is more likely to have a small (“statistically significant”) probability of a Type I error, an inefficient estimator is likely to produce parameter estimates having a large probability of a Type I error.
The generalized least squares (GLS) methodology makes use of the information about the non-spherical error structures and yields unbiased and efficient parameter estimates. However, GLS assumes that the variance-covariance matrix Ω, which is used to weight the data, is known a priori. This is not always the case. Accordingly, the pooled model can be estimated by an FGLS method, which involves using an estimate of Ω in a two-stage process. The details of the estimation method are described in Parks (1967). Further discussion is in Kmenta (1986), Kmenta and Oberhofer (1974) and Greene (2003). The Kmenta–Parks method is based on GLS and can correct for temporally correlated errors and panel heteroskedasticity as well as contemporaneous correlations across units.
A brief explanation of the Kmenta–Parks methodology is as follows. Assume that in Equation (1) the random errors u it , i = 1, 2, … , N, t = 1, 2, … , T have the following stochastic structure: (see equation 2)where E(ε i t )=0, E(u i t ε j t )=0, E(ε i t ε j t )=Φ i j , E(ε i t ε j s )=0(s ≠ t), E(u i0)=0, E(u i0 u j0)=σ i j =Φ i j /(1−ρ i ρ j ). In this model the variance-covariance matrix Ω for the vector of random errors u can be expressed as (Greene, 2003): (see equation 3)where (see equation 4)
The matrix Ω is estimated by a two-stage procedure and the model parameters β k , k = 1, 2, … , p, are then estimated by FGLS. The first step in estimating Ω involves the use of OLS to estimate β k and obtaining the fitted residuals, as follows: (see equation 5)
A consistent estimator of the first-order autoregressive parameter is then obtained in the conventional manner, as follows: (see equation 6)for i = 1, 2, … , N. The Prais–Winsten transformation is then applied to remove, asymptotically, the autoregressive features of the data. That is, for i = 1, 2, … , N and t = 2, … , T (see equation 7)while for i = 1, 2, … , N and t = 1 (see equation 8)
This can be written in compact form as follows:(see equation 9)
The second step in estimating the covariance matrix Ω is to apply OLS to the transformed model, obtaining(see equation 10)from which the consistent estimator of σ i j is calculated as(see equation 11)where(see equation 12)
The resulting FGLS estimator is then obtained as(see equation 13)
Under a variety of conditions (Parks, 1967) the FGLS estimator has been shown to be consistent and efficient. For models that do not correct for autocorrelation, maximum likelihood estimates can also be obtained by iterating the described procedure to convergence (Kmenta and Oberhofer, 1974).
Beck and Katz (1995, 1996) propose a less complex method, which retains OLS parameters estimates (consistent but not efficient) but replaces OLS standard errors by the so-called PCSE. Beck and Katz (1995, 1996) argue that OLS with PCSE is superior to the Kmenta FGLS approach when estimating panel models using small samples. Using Monte-Carlo experiments, Beck and Katz (1995, 1996) show that PCSE are more efficient than OLS standard errors and can be used with either OLS or Prais–Winsten (Prais and Winsten, 1954) regressions. Although the estimates of the Kmenta method might be more efficient than OLS/PCSE in situations where contemporaneous correlation and group-wise heteroskedasticity are significant, the standard errors obtained from the Kmenta method have a tendency to be too small (in finite samples). Thus, the Kmenta standard errors convey a degree of “overconfidence” and do not correctly reflect the true sampling variability of the parameter estimates. While OLS is not efficient in the presence of spherical errors, it does yield consistent estimates. OLS standard errors are inaccurate in the presence of non-spherical errors in that they do not provide correct estimates of the sampling variability of the OLS parameter estimates. Beck and Katz's PCSE are a direct extension of White's (1980) heteroskedasticity-consistent standard errors. However, since PCSE take into account the panel structure of the data, they perform even better for panel data than White's heteroskedasticity consistent standard errors do. In this method Ω = Σ ⊗ I T is an N T× N T block diagonal matrix with Σ, an N× N matrix of contemporaneous correlations along the diagonal. OLS residuals u^ i t * are used to consistently estimate the (i, j)th elements of Σ as Σ^ i,j =∑ t=1 T u^ i t * u^ j t * /T. The PCSE are then estimated by the square root of the diagonal of (X ′ X)− 1 X(Σ^ ⊗ I T )X(X ′ X)− 1 where X denotes the N T× N T matrix of stacked vectors of explanatory variables X i t .
4. Descriptive statistics
This section describes the sample, data sources and variable measurement. All data for this study are drawn from the Standard and Poor's (2004) ExecuComp database. The sample covers the period 1996-2002 and consists of firms in primary SIC codes 3570 (Computer and Office Equipment), 3571 (Electronic Computers), 3572 (Computer Storage Devices), 3576 (Computer Communication Equipment), 3577 (Computer Peripheral Equipment), 3661 (Telephone and Telegraph Apparatus), 3674 (Semiconductor and Related Devices), 4812 (Wireless Telecommunication), 4813 (Telecommunication), 5045 (Computers and Software Wholesalers), 5961 (Electronic Mail-Order Houses), 7370 (Computer Programming, Data Processing), 7371 (Computer Programming Service), 7372 (Prepackaged Software), and 7373 (Computer Integrated Systems Design) for which CEO tenure extends over the entire period. This condition is imposed to guarantee homogeneity in the pay to performance relationship and to control to some degree for human capital heterogeneity within firms. The resulting final sample is a sample of 48 (unbalanced) panels.
Descriptive statistics for relevant variables over the entire sample period are summarized in Panel A of Table I. Over the sample period of 1996-2002 total CEO compensation, which in the Execucomp database is referred to as TDC1, is on average about seven million dollars. It is interesting to note that over the sample period total compensation rose approximately by 16 per cent while net sales increased approximately by 15 per cent. Both ROA it and TRS it recorded on average a negative change. The data show a modest fall on average in ROA it (1.37 percentage-point fall) and a much more pronounced fall in TRS it (11.9 percentage-point fall).
The stock market bubble burst in the year 2000. Panels B and C in Table I report additional descriptive statistics for two sub-periods 1996-1999 and 2000-2002, respectively. In the first sub-period total CEO compensation was on average about five million dollars and increased approximately by 38 per cent. Net sales increased by approximately 27 per cent. On the other hand, the change in ROA it on average was quite modest (1.12 percentage-point rise) while the average change in TRS it was much more striking (23.74 percentage-point rise). In contrast, average total CEO compensation in the second sub-period was about nine million dollars, but on average decreased by approximately 4 per cent. The average growth in sales was also modest compared to the previous period. Both measures of firm performance recorded on average a negative change. The data show a moderate fall on average in ROA it (3.84 percentage-point fall). On the other hand, the average fall in TRS it is significantly more pronounced and prominent (47.57 percentage-point fall).
The pair-wise correlation coefficients between CEO compensation and performance measures further highlights the impact of the stock market crash of 2000 on the pay to performance relation. Over the period 1996-2002 the correlation coefficients between Δln C it and ΔROA it and between Δln C it and ΔTRS it are 0.1814 (p = 0.0020) and 0.1726 (p = 0.0033), respectively. The correlation coefficients are higher and highly significant in the first sub-period. Over the period 1997-1999, the correlation coefficients between Δln C it and ΔROA it and between Δln C it and ΔTRS it are, respectively, is 0.2742 (p = 0.0009) and 0.2266 (p = 0.0063). The correlation coefficients, however, fall dramatically and are no longer statistically significant in the second sub-period. Over the period 2000-2002, the correlation coefficients between Δln C it and ΔROA it and between Δln C it and ΔTRS it are 0.0304 (p = 0.7175) and 0.0043 (p = 0.9595), respectively.
5. Empirical results
Different measures of performance highlight different versions of Equation (1). As mentioned in section 2, in this study we employ an accounting-based measure of performance, ROA it , and a stock market-based measure, TRS it . Table II displays the results of the OLS estimation of three versions of Equation (1). In columns (i) and (ii) we report the estimate of the responsiveness of CEO compensation to each of the performance variables individually, whereas in column (iii) we present the estimates of the joint impact of the two performance variables on CEO compensation. In addition to assess the robustness of the estimates, the latter version weighs up which measure of firm performance, if any, is dominant. t-statistics are reported in parenthesis below the estimated coefficients. Also reported is the Wald χ 2 statistics and the value of the log likelihood. The Wald χ 2 statistic tests the hypothesis that all the regression parameters, except the constant, are zero (Greene, 2003). Time-specific effects in the form of yearly dummy variables are included in each regression but are not reported. Results are qualitatively unchanged if year effects are omitted. The OLS results indicate that the pay to size relation is strong and significant at the 5 per cent level, no matter which measure of financial performance is used. Moreover, the estimates on firm size are close to those found in the literature (Rosen, 1992; Murphy, 1999). However, the OLS results support the hypothesis that market-based returns rather than accounting-based returns drive the pay to performance relation. The coefficient estimates on the change in TRS are positive and statistically significant at the 5 per cent level, whereas the estimates on the change in returns on assets are statistically insignificant at the 5 per cent level.
Statements about the statistical significance of the parameter estimates in Table IV are predicated on the assumptions, among other things, that the error terms are homoskedastic and have no autocorrelation. In Table II we also report a series of diagnostic residual tests, namely:
- the modified Wald test statistic for group-wise heteroskedasticity (Greene, 2003);
- the modified Bhargava et al. Durbin Watson; and
- the Baltagi–Wu locally best invariant (LBI) test statistic (Baltagi and Wu, 1999).
The modified Wald test statistic for group-wise heteroskedasticity is χ 2 distributed with N degrees of freedom and tests the null hypothesis that σ i 2=σ 2 for i = 1, … , N where N is the number of panels. The Baltagi–Wu LBI test statistics tests the null hypothesis that ρ=0. The Durbin–Watson test for first-order autocorrelation was extended by Bhargava et al. (1982) to the case of balanced, equally spaced panel data. Baltagi and Wu (1999) modified it further to account for unbalanced panels with unequally spaced data. The p-values are not reported for either statistics. Test statistics for a first-order autoregressive process in unbalanced panels possess complex distributional properties (Baltagi and Wu, 1999). As a result, establishing critical values is computationally cumbersome and obtuse. As is often the case, however, examination of the test statistic itself is revealing. With a value of about 2.65, the Baltagi–Wu LBI test statistic would almost certainly reject the null hypothesis of no serial correlation. We conclude that the results of these tests reject the null hypothesis of no serial correlation and group-wise homoskedasticity and lend support for the use of FGLS and PCSE methods.
The results in Tables III-V demonstrate that the conclusions about the pay to performance relation change a great deal after relaxing the restrictions that errors are homoskedastic and have no autocorrelation. The estimation process allows for the most general error structure that can be considered within the data constraints – group-wise heteroskedasticity and serial autocorrelation – using both Parks-Kmenta and Beck–Katz methods. Within each method, two alternative hypotheses about the non-spherical nature of the error structure are taken into consideration. Time-specific effects are included in each regression but are not reported. Results are qualitatively unchanged if year effects are omitted. Each table presents four sets of estimates. Models 1 and 2 report the FGLS estimates based on the Parks–Kmenta method. Specifically, in model 1 we allow for group-wise heteroskedasticity (variance u i,t =σ i 2) and in model 2 we allow for both group-wise heteroskedasticity and serial correlation of order 1 with common AR(1) coefficient for all panels (variance u i,t =σ i 2 and u i,t =ρ u i,t − 1+ε i,t ). Similarly, models 3 and 4 present the PCSE estimates based on the Beck–Katz method. Model 3 displays the OLS estimates with PCSE, while model 4 presents the Prais–Winsten regression estimates with PCSE. The Prais–Winsten regression corrects for serial correlation of order 1 with common AR (1) coefficient for all panels (u i,t =ρ u i,t − 1+ε i,t ). The numbers in parenthesis below the estimated coefficients are t-statistics. In models 3 and 4 the t-statistics are based on heteroskedasticity-corrected standard errors. In models 1 and 2 the R 2 is not reported as in FGLS estimation the total sum of squares cannot be decomposed as in OLS. Specifically, R 2 computed from FGLS sum of squares is not be bounded between zero and one and does not represent the percentage of total variation in the dependent variable that is accounted by the model. Additionally, eliminating or adding variables in a FGLS model does not always decrease or increase the computed R 2 value. Instead, we report the Wald test statistics and the value of the log likelihood. The latter is not available for the PCSE models. The Wald χ 2 statistic tests the null hypothesis that all of the model parameters, other than the intercept term, are zero. In Table III we present the FGLS and PCSE results for the first version (ΔROA it only) of Equation (1). Similarly, the corresponding results for the second (ΔTRS it only) and third (ΔROA it and ΔTRS it ) versions of Equation (1) are presented in Tables IV and V. It is remarkable that the parameter estimates measuring the impact of firm size and financial performance on CEO compensation are consistently significant, unlike their OLS counterparts, across different error structure specifications. We take this as a reflection of the higher efficiency of the estimation procedures. The OLS/PCSE specifications yield standard errors that are not larger than the corresponding OLS ones. The Parks–Kmenta standard errors, on the other hand, are lower than the corresponding PCSE. In all four cases, the Wald χ2 statistic (df = 7) is significant, permitting rejection of the null hypothesis that all of the coefficients, aside from the intercept term, are zero, and the R 2 in the Beck–Katz regressions is about 15 per cent. The R 2 reported in the OLS regressions in Table IV are about of the same magnitude.
The estimates on firm size are significant at the one percent or better, and are not much different in magnitude from the OLS estimates in Table II. The FGLS estimates display about the same uniformity across error specifications and measures of performance as the PCSE estimates. The PCSE estimates on firm size vary from 0.328 in model 3 (OLS/PCSE) in Table IV, to 0.229 in model 4 (Prais–Winsten/PCSE) in Table III. Conversely, the corresponding FGLS estimates vary from 0.356 in model 2 (group-wise heteroskedasticity and serial autocorrelation) in Table IV, to 0.241 in model 2 (group-wise heteroskedasticity and serial autocorrelation) in Table III.
In Tables III and IV the estimates on firm performance are consistently significant at the 5 per cent level. This contrasts with the OLS estimates. The use of Parks–Kmenta method as well as the Beck–Katz procedure accounts for uncovering the pay to performance link that OLS was unable to fully depict. The FGLS estimates are quite different from the OLS as well as the PCSE estimates. The PCSE estimates on ΔTRS it are nearly twice the size of the FGLS estimates. Within a given method, the parameter estimates on ΔTRS it are fairly uniform across each of the two error specifications. This is not true in case of ΔROA it , which lacks uniformity within a given method. In both the FGLS and PCSE case, the autocorrelation correction has the effect of increasing the size of the parameter estimates: in the FGLS case, from 0.690 to 0.868, and in the PCSE case, from 0.647 to 0.792.
The results change when the pay to performance relation is not estimated separately for each performance variable. The FGLS and PCSE estimates are displayed in Table VII. The estimates for firm size remain significant at the one per cent level and are robust to the inclusion of the additional measure of performance. Conversely, significant differences exist between the FGLS estimates and the corresponding PCSE estimates of firm performance variables. Specifically, the FGLS estimates indicate that ΔROA it dominates ΔTRS it . On the other hand, the corresponding PCSE estimates suggest that ΔTRS it dominates ΔROA it . Agency theory suggests (Holmstrom, 1979) that the strength of the relation between executive compensation and a performance measure is determined by the informativeness of the performance measure about agent behavior. In stark contrast with the FGLS estimates, the PCSE estimates are consistent with the argument that new economy firms place more weight on common stock returns than returns on assets in executive compensation contract. The PCSE findings are also consistent with Lambert and Larcker (1987) and Clinch (1991) arguments that in high growth firms accounting measures are less informative than stock market measures about firm performance because accounting measures have a tendency to lag stock market measures in reflecting investment decisions in these firms.
The inconsistency of the reported results could be due to excluded variables that other studies have found to be related to CEO compensation. For example, Jensen and Murphy (1990) and Joskow and Rose (1994) report a significant influence of past performance on current compensation. In order to test this possibility, we re-estimate the three versions of Equation (1) with the lagged change in each of the performance variables added as an additional explanatory variable. The results of this supplementary analysis are presented in Tables VI-VIII. The R 2 in the OLS/PCSE and the Prais–Winsten/PCSE regressions which are reported in Table VI is in the region of 10-12 per cent and is substantially lower than the corresponding R 2 in the OLS/PCSE and Prais–Winsten/PCSE regressions which are reported in Tables VII and IX. However, in all cases the Wald test rejects the null hypothesis that all of the regression coefficients, except for the intercept term, are not significantly different from zero. The estimates for firm size are qualitatively consistent and similar to those reported in Tables III-V. The results regarding the effect of past performance on CEO compensation, on the other hand, vary depending upon the measure of performance. Past performance measured as the previous year's change in ROA it exerts no influence on current CEO compensation; on the other hand, past performance measured as the previous year's change in TRS it has a significant influence on current CEO compensation. Including the coefficient on the previous year's change in TRS it , the total effect on current CEO compensation is as follows (t-statistics in parenthesis): 0.162 (3.14) in model 1, 0.198 (3.90) in model 2, 0.171 (2.57) in model 3, and 0.208 (3.24) in model 4. In addition, as shown in Table VIII, the inclusion of the previous year's change in TRS it substantially affects the magnitude and the statistical significance of the coefficient estimates on the current year's change in ROA it . These findings suggest that new economy firms place more weight on common stock return than ROA in executive compensation contract. In addition, they imply that the increase in the use of stock options in “new economy” firms is consistent with agency theory arguments.
An issue that is not often examined concerns the problem of heterogeneity in the pay to performance relation (Conyon et al., 2000). Is the effect of a one percentage point increase in TRS it in the year 1998, for example, the same as the effect of a corresponding increase in the year 2001? In order to test this possibility, the change in TRS it is specified to have different effects in the sub-period 1997 through 1999 and the sub-period 2000 through 2002. This subdivision of the sample period is undoubtedly arbitrary, but it seems pertinent because the two sub-periods coincide with two sharply different phases and developments in the US stock market. The first sub-period exemplifies an exceptionally well-performing stock market, especially in “new economy” stocks. The second sub-period, on the other hand, is characterized by the abrupt downturn of the stock market and the sudden burst of the stock market bubble.
The FGLS and PCSE results are displayed in Table IX using the current and previous year's change in TRS it and the current change in ROA it . The previous year's change in ROA it is not included for two reasons. First, as in the results in Table VIII clearly suggest, the coefficient estimates on the previous year's change in ROA it are insignificant. Second, this insignificance maintains even when the previous year's change in ROA it is specified to have different effects in the two sub-periods. These findings are not reported for economy of space but are available from the authors. On the whole, the results are robust to the change in estimation method, the difference being mostly in the standard errors. The estimates on firm size in either sub-period are not much different in magnitude from the estimates presented in the previous tables. These estimates show only a trivial variability between the first and the second sub-periods, which corroborates their robustness to vastly different market conditions. The estimates in the second sub-period are however more significant than those in the first sub-period. The estimates on the current and previous year's change in TRS it in the two sub-periods are also relatively close, but their significance varies with the estimation method. In models 1 and 2 the coefficient estimates on the current change in TRS it in the first sub-period are slightly higher and more significant than those in the second sub-period. Conversely, the coefficient estimates on the previous year's change in TRS it in the second sub-period are more significant than those in the first sub-period. In models 3 and 4 the coefficient estimates on the current change in TRS it in the first sub-period are also higher and more significant than those in the second sub-period, while the coefficient estimates on the previous year's change in TRS it are insignificant in both sub-periods. What are vastly different from the estimates in the previous tables, on the other hand, are the estimates on the current change in ROA it in the first and the second sub-periods. These conclusions are uniformly confirmed by a series of Wald tests on the equality of the coefficients in the two sub-periods. The p-values of these tests are presented in Table X and indicate that only the response of total CEO compensation to ROA is significantly different in the two sub-periods, while it is relatively unaffected by changes in market conditions.
6. Summary-conclusions
This study focuses on two novel aspects of the empirics of the pay to performance relation: the existence of patterns of heteroskedasticity and autocorrelation and the presence of parameter heterogeneity. Group-wise heteroskedasticity and autocorrelation are second-order issues, as they affect the efficiency of estimates. However, in relatively short panels these differences can have noticeable effects on estimates, producing small sample biases. In fact, our findings indicate that autocorrelation and heteroskedasticity can significantly bias the OLS estimates.
The estimation process exploits both the cross section and the time series dimensions of the data and applies two methods that explicitly take care of the non-spherical nature of the random disturbances, the Parks–Kmenta FGLS method and the Beck–Katz PCSE method. The empirical analysis is applied to a panel of 48 new economy firms between 1996 and 2002. CEO compensation in “new economy” firms is positively significantly related to firm size and firm performance. The econometric results indicate that estimates on firm size are robust to alternative specifications of the error structure. There is however some evidence that the effect of firm size on total CEO compensation is more significant in the second sub-period than in the first one. The opposite holds for the estimates on firm performance. We find that both measures of firm performance are more significant in the first sub-period than in the second. Estimates on firm performance are also more sensitive to the estimation method, the specification of random error structures, and the time period. Mixed evidence on parameter homogeneity in the pay to performance relation is also reported. We find that in the sub-period immediately preceding the stock market crash both ROA and TRS affect total CEO compensation. The response of CEO compensation to performance measures, however, shifts to TRS only in the sub-period following the crash. The application of the Wald tests of equality fails to reject the hypothesis of stability with respect to the coefficient estimates on firm size and TRS, but not with respect to the coefficient estimates on return of assets.
Table I.Descriptive statistics
Table II.Total CEO compensation and firm performance: OLS estimates
Table III.Total CEO compensation and ROA
Table IV.Total CEO compensation and TRS
Table V.Total CEO compensation, ROA and TRS
Table VI.Total CEO compensation, current and previous year's returns on assets
Table VII.Total CEO compensation and current and previous year's TRS
Table VIII.Total CEO compensation, current and previous year's returns on assets and current and previous year's TRS
Table IX.Total CEO compensation and TRS: inter-temporal heterogeneity
Table X.
p-values of the Wald tests of coefficients equality
(see equation 1)
(see equation 2)
(see equation 3)
(see equation 4)
(see equation 5)
(see equation 6)
(see equation 7)
(see equation 8)
(see equation 9)
(see equation 10)
(see equation 11)
(see equation 12)
(see equation 13)
References
Anderson, M., Banker, R., Ravindran, S. (2000), "Executive compensation in the information technology industry", Management Science, Vol. 46 No.4, .
Baltagi, B.H., Wu, P.X. (1999), "Unequally spaced panel data regressions with Ar(1) disturbances", Econometric Theory, Vol. 15 No.6, pp.814-23.
Bhargava, A., Franzini, L., Narendranathan, W. (1982), "Serial correlation and the fixed effects model", Review of Economic Studies, Vol. 49 No.4, pp.533-49.
Black, F., Scholes, M. (1973), "The pricing of options and corporate liabilities", Journal of Political Economy, Vol. 81 No.3, pp.637-54.
Beck, N., Katz, J.N. (1995), "What to do (and not to do) with time series cross-section data", American Political Science Review, Vol. 89 No.3, pp.634-47.
Beck, N., Katz, J.N. (1996), "Nuisance vs. substance: specifying and estimating time series cross-section models", Political Analysis, Vol. 6 No.1, pp.1-34.
Carpenter, M.A., Sanders, W.G. (2002), "Top management team compensation: the missing link between CEO pay and firm performance?", Strategic Management Journal, Vol. 23 No.4, pp.367-75.
Clinch, G. (1991), "Employee compensation and firms' research and development activity", Journal of Accounting Research, Vol. 29 No.1, pp.59-78.
Conyon, M.J., Peck, S.I., Sadler, G.V. (2000), "Econometric modeling of UK executive compensation", Managerial Finance, Vol. 26 No.9, pp.3-20.
Coughlan, A.T., Schmidt, R.M. (1985), "Executive compensation, management turnover, and firm performance", Journal of Accounting and Economics, Vol. 7 No.1-3, pp.43-66.
Fama, E.F., Jensen, M.C. (1983), "Agency problems and residual claims", Journal of Law and Economics, Vol. 26 No.2, pp.327-50.
Gibbons, R., Murphy, K.J. (1992), "Optimal incentive contracts in the presence of career concerns: theory and evidence", Journal of Political Economy, Vol. 100 No.3, pp.468-505.
Greene, W.H. (2003), Econometric Analysis, 5th ed., Prentice Hall, Englewood Cliffs, NJ, .
Hall, B., Liebman, J. (1998), "Are CEOs really paid like bureaucrats?", Quarterly Journal of Economics, Vol. 113 No.3, pp.653-91.
Holmstrom, B. (1979), "Moral hazard and observability", Bell Journal of Economics, Vol. 10 No.1, pp.74-91.
Ingham, H., Thompson, S. (1995), "Mutuality, performance and executive compensation", Oxford Bulletin of Economics and Statistics, Vol. 57 No.3, pp.295-308.
Ittner, C.D., Lambert, R.A., Larcker, D.F. (2003), "The structure and performance consequences of equity grants to employees of new economy firms", Journal of Accounting and Economics, Vol. 34 No.1-2-3, pp.89-127.
Jensen, M., Meckling, W. (1976), "Theory of the firm: managerial behavior, agency costs and ownership structure", Journal of Financial Economics, Vol. 3 No.4, pp.305-60.
Jensen, M.C., Murphy, K.J. (1990), "Performance pay and top-management incentives", Journal of Political Economy, Vol. 98 No.2, pp.225-64.
Joskow, P., Rose, N. (1994), "CEO pay and firm performance: dynamics, asymmetries, and alternative performance measures", NBER working paper no. 4976, .
Kmenta, J. (1986), Elements of Econometrics, 2nd ed., Macmillan, New York, NY, .
Kmenta, J., Oberhofer, W. (1974), "A general procedure for obtaining maximum likelihood estimates in generalized regression models", Econometrica, Vol. 42 No.3, pp.579-92.
Kostiuk, P. (1990), "Firm size and executive compensation", Journal of Human Resources, Vol. 25 No.1, pp.90-105.
Lambert, R., Larcker, D. (1987), "An analysis of the use of accounting and market measures of performance in executive compensation contracts", Journal of Accounting Research, Vol. 25 No.0, pp.85-129.
Leonard, J. (1990), "Executive pay and firm performance", Industrial and Labor Relations Review, Vol. 43 No.3, pp.13S-29S.
Lewellen, W., Huntsman, B. (1970), "Managerial pay and corporate performance", American Economic Review, Vol. 60 No.4, pp.710-20.
Lewellen, W., Loderer, C., Martin, K. (1987), "Executive compensation and executive incentive problems: an empirical analysis", Journal of Accounting and Economics, Vol. 9 No.3, pp.287-310.
Main, B., Bruce, A., Buck, T. (1996), "Total board remuneration and company performance", Economic Journal, Vol. 106 No.439, pp.1627-44.
Murphy, K.J. (1985), "Corporate performance and managerial remuneration: an empirical analysis", Journal of Accounting and Finance, Vol. 7 No.1-2-3, pp.11-42.
Murphy, K.J. (1986), "Incentives, learning, and compensation: a theoretical and empirical investigation of managerial labor contracts", Rand Journal of Economics, Vol. 17 No.1, pp.59-76.
Murphy, K.J. (1999), "Executive compensation", in Ashenfelter, O., Card, D. (Eds),Handbook of Labor Economics, North-Holland, Vol. 3 pp.2485-563.
Murphy, K.J. (2003), "Stock-based pay in new economy firms", Journal of Accounting and Economics, Vol. 34 No.1-2-3, pp.129-47.
Parks, R.W. (1967), "Efficient estimation of a system of regression equations when disturbances are both serially and contemporaneously correlated", Journal of the American Statistical Association, Vol. 62 No.2, pp.500-9.
Prais, S.J., Winsten, C.B. (1954), "Trend estimators and serial correlation", Chicago, IL, discussion paper no. 383, Cowles Commission, .
Rosen, S. (1992), "Contracts and the market for executives", in Wein, L., Wijkander, H. (Eds),Contract Economics, Blackwell Publishers, London, pp.181-211.
Sloan, R. (1993), "Accounting earnings and top executive compensation", Journal of Accounting and Economics, Vol. 16 No.1-2-3, pp.55-100.
Standard and Poor's (2004), Standard & Poor's Execucomp: The Executive Compensation Database, McGraw-Hill, Englewood Cliffs, NJ, .
Stathopoulos, K., Espenlaub, S., Walker, M. (2004), "UK executive compensation practices: new economy vs. old economy", Journal of Management Accounting Research, Vol. 16 pp.57-92.
White, H. (1980), "A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity", Econometrica, Vol. 48 No.4, pp.817-38.
Wooldridge, J.M. (2002), Econometric Analysis of Cross Section and Panel Data, MIT Press, Cambridge, .
Corresponding author
Giorgio Canarella can be contacted at: gcanare@calstatela.edu