Stock Market Information Production and CEO Incentives
1.
Stock Market Information Production and CEO
Incentives∗
Qiang Kang † Qiao Liu ‡
University of Miami University of Hong Kong
First Draft: January 2003
This Draft: October 2004
Abstract
We examine the relation between chief executive oﬃcers’ (CEOs’) incentives and stock
market informativeness. Following Holmstrom and Tirole (1993), we hypothesize and
demonstrate both theoretically and empirically that CEO incentives tend to be stronger
when stock prices contain more information. We use probability of informed trading,
residual analyst coverage, analysts’ earnings forecast error, and dispersion of analysts’
earnings forecast as proxies for stock market informativeness, and ﬁnd that these
variables account for the cross-sectional variations in pay-performance sensitivities
well. Our results are robust to the choice of estimators, sample periods, measures
of incentives, model speciﬁcations, and estimation methods.
JEL Classiﬁcation: D80, G14, G34, J33
Keywords: Market microstructure, pay-performance sensitivity, probability of informed
trading, analysts’ earnings forecast
∗
This paper is a revision of the paper entitled “Stock Trading, Information Production and Incentive
Pay”. We appreciate comments from Ross Watts (the editor), an anonymous referee, Tim Burch, Stephen
Chiu, John Core, Craig MacKinlay, Wing Suen, James Wang, Xianming Zhou and seminar participants at
University of Hong Kong, University of Pennsylvania, and Hong Kong University of Science and Technology.
An earlier draft was completed while Kang was aﬃliated with University of Hong Kong, whose hospitality
is gratefully acknowledged. We thank Clara Vega and Rong Qi for sharing with us programming codes and
oﬀering computation suggestions. The ﬁnancial support from University of Hong Kong and University of
Miami (Kang) and the RGC Earmarked Research Grant HKU 7128/01H (Liu) are gratefully acknowledged.
All errors remain our own responsibility.
†
Corresponding author. Mailing address: Finance Department, University of Miami, P.O. Box 248094,
Coral Gables, FL 33124-6552. Phone: (305)284-8286. Fax: (305)284-4800. E-mail: q.kang@miami.edu.
‡
School of Economics and Finance, University of Hong Kong, Pokfulam, Hong Kong. Phone: (852)2859-
1059. Fax: (852)2548-1152. E-mail: qliu@hku.hk.
2.
Stock Market Information Production and CEO Incentives
Abstract
We examine the relation between chief executive oﬃcers’ (CEOs’) incentives and stock
market informativeness. Following Holmstrom and Tirole (1993), we hypothesize and
demonstrate both theoretically and empirically that CEO incentives tend to be stronger
when stock prices contain more information. We use probability of informed trading,
residual analyst coverage, analysts’ earnings forecast error, and dispersion of analysts’
earnings forecast as proxies for stock market informativeness, and ﬁnd that these
variables account for the cross-sectional variations in pay-performance sensitivities
well. Our results are robust to the choice of estimators, sample periods, measures
of incentives, model speciﬁcations, and estimation methods.
JEL Classiﬁcation: D80, G14, G34, J33
Keywords: Market microstructure, pay-performance sensitivity, probability of informed
trading, analysts’ earnings forecast
3.
1 Introduction
Market-based compensation schemes have gained huge popularity in the past two decades.
The use of stock grants and stock options, as well as other forms of equity-related incentive
pay, has become a common practice. Many researchers believe that performance-related
compensation schemes help discipline managers to act in the interest of stock investors and
that such pay arrangements are a (partial) remedy to an agency problem. However, executive
compensation has recently received a great deal of negative exposure. The consensus now
is that a large portion of what top managers get is not incentive pay, but rents extracted
by greedy CEOs from dysfunctional boards (e.g., Bertrand and Mullainathan, 2001; and
Bebchuk and Fried, 2003). Both views seem plausible. The debate thus boils down to
answering the following questions: Does incentive pay really play an eﬀective role in aligning
the interest of executives with that of shareholders? If yes, under what conditions will
performance pay be more eﬀective? If no, why is it used?
Our study is motivated by this debate. In this paper we focus on identifying factors that
drive the cross-sectional variation in CEO pay-performance sensitivity, a common measure
of CEO incentives in the compensation literature. In addition to the factors that have
been identiﬁed in previous research,1 we hypothesize and empirically demonstrate that stock
market informativeness enhances CEO pay-performance sensitivity. To better motivate our
empirical work, we develop a model that shows a negative relation between pay-performance
sensitivity and the (inverse) measure of stock price informativeness, V ar(δ|P ) (Kyle, 1985).
In this measure, δ captures the uncertainty of a ﬁrm’s value and P is the ﬁrm’s stock price.
Consistent with Holmstrom and Tirole (1993),2 we posit that incentive pay works more
1
It has been well demonstrated that there are enormous cross-sectional variations in individual incentive
measures such as pay-performance sensitivities (Hall and Liebman, 1998). A non-exhaustive list of variables
that help explain this heterogeneity includes ﬁrm size (Smith and Watts, 1992; Schaefer, 1998; and Baker and
Hall, 2002); return volatility (Garen, 1994; Aggarwal and Samwick, 1999; and Jin, 2002); growth opportunity
(Smith and Watts, 1992; and Gaver and Gaver, 1993); CEO tenure (Bertrand and Mullainathan 2001; and
Milbourn, 2002); institutional investors (Hartzell and Starks, 2003); and industry and year eﬀects (Ely, 1991;
and Murphy, 1999).
2
Holmstrom and Tirole (1993) analytically show that stock prices incorporate performance information
1
4.
eﬀectively in a more eﬃcient (or more informative) stock market.
We test this prediction using CEO compensation data drawn from Compustat’s
ExecuComp database during the period 1992–2001. Following the compensation literature
(e.g., Jensen and Murphy, 1990; Hall and Liebman, 1998; and Aggarwal and Samwick 1999),
we estimate pay-performance sensitivity by examining the empirical relation between changes
in executive ﬁrm-related wealth and changes in shareholder wealth. We use two metrics of
pay-performance sensitivity. One is stock-based pay-performance sensitivity (P P S), which
measures the change in dollar value of a CEO’s holdings of stocks and stock options per $1,000
increase in the ﬁrm’s shareholder wealth. The other is total pay-performance sensitivity
(P P S T OT ), which measures the change in dollar value of a CEO’s total ﬁrm-related
wealth per $1,000 increase in the ﬁrm’s shareholder wealth. Hall and Liebman (1998) and
Murphy (1999) ﬁnd that the former accounts for the majority of CEO incentives and can
also be measured more precisely compared to other compensation components. We thus
mainly rely on the stock-based PPS to interpret the empirical results.
We construct four proxies for stock market informativeness. The ﬁrst is probability of
informed trading (P IN ) for each ﬁrm-year observation (Easley, Kiefer, and O’Hara (1996,
1997)). P IN directly estimates the amount of information held by traders on a speciﬁc stock.
We also exploit the information contained in the analysts’ earnings forecast data to construct
three other proxies for stock market informativeness: residual analyst coverage (RESCOV ),
analysts’ earnings forecast error (F ERROR), and dispersion of analysts’ earnings forecast
(DISP ER). The four variables capture diﬀerent aspects of stock market informativeness.
We conduct various empirical tests of the hypothesis that pay-performance sensitivity
increases with stock market informativeness. We ﬁnd strong empirical support for this
hypothesis. Since P IN and RESCOV are positively related to stock market informativeness
that cannot be readily extracted from the ﬁrm’s accounting data and that the principal can use the inferred
information to design a better compensation contract. Thus, the higher the market liquidity is, the more
informative the stock price is and consequently, the more eﬀective is the performance-related pay. These
predictions have not been rigorously tested partly because it is not straightforward to construct an empirical
analysis based on Holmstrom and Tirole’s (1993) model.
2
5.
but F ERROR and DISP ER are inversely related to stock market informativeness, the
predictions go one way for the ﬁrst two variables and the other way for the last two variables.
Speciﬁcally, pay-performance sensitivity is strongly and positively related to P IN and
RESCOV , respectively. Pay-performance sensitivity is strongly and negatively related to
F ERROR and DISP ER, respectively. All of these results are also economically signiﬁcant.
For example, using the raw measure (or the cumulative distribution function) of P IN as the
proxy for market informativeness, the pay-performance sensitivity at the maximum P IN
is $15.36 (or $2.035) larger than the pay-performance sensitivity at the median P IN as a
result of $1,000 increase in the shareholder value, and the sample median of pay-performance
sensitivities is $12.85. The overall results strongly suggest that an informationally eﬃcient
stock market helps induce CEO incentives.
Our empirical results are robust to various robustness checks. In addition to showing that
the CEO compensation schemes work more eﬀectively in a more informative environment,
our results identify another important underlying economic determinant, the stock market’s
microstructure, that helps account for the cross-sectional heterogeneity in CEO incentives
even after we control for the impact of various previously identiﬁed factors. Several theoretic
models have been proposed to link the market microstructure framework with corporate
ﬁnance (Holmstrom and Tirole, 1993; and Subrahmanyam and Titman, 1999). We develop
empirical evidences showing that the market microstructure matters a lot in inducing
managerial incentives.
The paper is organized as follows. Section 2 develops the main hypothesis of the paper.
Section 3 discusses the data and the empirical method. Section 4 presents the empirical
results. Section 5 oﬀers the robustness analysis. Section 6 concludes.
3
6.
2 Development of Hypothesis
The Holmstrom and Tirole (1993) study (hereafter HT) is one of the ﬁrst papers to combine
market microstructure with compensation contract theory. HT show that the stock price
incorporates performance information that cannot be extracted from the ﬁrm’s current or
future proﬁt data. The amount of information contained in the stock price is useful for
structuring managerial incentives. An illiquid market makes the stock price less informative
and thus reduces the beneﬁts of stock market monitoring.
HT form stock prices in expectation of a liquidation payout π = e + θ + , where e is the
manager’s eﬀort and θ and are normally distributed mean-zero noise terms with volatilities
σθ and σ respectively. We follow HT’s notation here except that we do not report their time
subscripts.
The informed trader observes a signal of terminal value equal to e + θ + η, where η is a
mean-zero normally distributed error term with volatility ση . HT allow the informed trader
to expend resources to reduce the noise in his signal, ση . The liquidity trader submits an
order, y, which is normally distributed with mean zero and volatility σy .
Proposition 3 in HT summarizes the key comparative static results on the use of stock-
based incentives. HT show that the pay-performance sensitivity is higher when σ is lower,
or when ση is lower for exogenous reasons (e.g., increase in liquidity trade σy as a result
of reduction in ownership concentration). The HT results imply that the stock market
informativeness enhances the eﬀectiveness of incentive contracts.
The above prediction has not been rigorously tested, which might be partially due to the
diﬃculty in building an empirical analysis on the HT model.3 In the Appendix, we propose a
3
First, in the HT model, ownership concentration is the fundamental determinant of market liquidity
which further determines the level of market informativeness and hence the pay-performance sensitivity.
However, Hartzell and Starks (2003) point out that the ownership concentration could also exert its impact
on the pay-performance relation through corporate governance channels, e.g., institutional investors monitor
managers by providing incentive compensations. Second, in the HT model, stock price volatility conveys
the precision of the informed trader’s information. A more volatile stock price implies a more liquid and
informative market resulting in a better market monitoring. This aspect has been extensively studied in the
4
7.
one-period model which provides direct empirical implications on the linkage between market
microstructure and pay-performance sensitivity.
The key insight of our model, immediately drawn from Proposition 2 in the Appendix
and the fact that Vδ (1 − ρ2 ) = V ar(δ|P ), is
1
b∗ = , (1)
1 + γkV ar(δ|P )
where b is the pay-performance sensitivity, γ is the manager’s risk aversion coeﬃcient, k is
his disutility parameter of spending eﬀorts, δ stands for the uncertainty of the ﬁrm value, P
is the price of the ﬁrm’s shares, and ρ is the correlation coeﬃcient between P and δ.
As shown in Kyle (1985), V ar(δ|P ) is the inverse measure of stock price informativeness.
Vδ (Vδ +2V )
Lemma 2 in the Appendix states that V ar(δ|P ) = (N +1)Vδ +2V
, where the number of informed
traders N , decreases with the cost of collecting information µ, and increases with the
underlying uncertainty Vδ , the variance of the signals V , and the variance of liquidity trading
Vz (see Lemma 1 in the Appendix). We summarize the results as follows:
Main Results: (1) The pay-performance sensitivity decreases with the manager’s eﬀort
parameter k, the risk aversion parameter γ, and the inverse measure of stock price
informativeness, V ar(δ|P ). (2) The stock market informativeness increases with the number
of informed traders which is determined by various market microstructure parameters (µ,
Vδ , Vz , and V ).
The key prediction from our model is consistent with the key prediction from HT: the
more informative the stock price, the more eﬀective is the compensation scheme in inducing
managers’ incentives. The core hypothesis for our empirical study is:
H1: Pay-performance sensitivity increases with stock market informativeness.
empirical research on executive compensation in a diﬀerent context (i.e., studying the risk-incentive relation)
but with mixed evidences (Garen, 1994; Aggarwal and Samwick, 1999; Core and Guay, 2002; and Jin, 2002).
Third, several key variables in HT are not directly observable (e.g., the importance of liquidity trade, σy ,
and the precision of noisy signals observed by the informed trader, ση ).
5
8.
To implement the empirical test, we need to develop some sensible measures to serve as
proxies for stock market informativeness.
3 Data and Empirical Method
Our primary data set for executive compensation is the ExecuComp database for the period
1992–2001. We obtain stock return data from the CRSP Monthly Stock File and accounting
information from the Compustat Annual File. We use four proxies for stock market
informativeness and construct them from two databases. We compute a ﬁrm’s probability
of informed trading (P IN ) for a given year by using intraday trading data extracted from
the TAQ database. We calculate three other variables by using analysts’ earnings forecast
information from the I/B/E/S History Summary File.
For our empirical study, we construct a sample containing 2,507 publicly traded ﬁrms
and 17,584 CEO-ﬁrm-year observations. All monetary terms are in 1992 constant dollars.
To remove the inﬂation eﬀect from the empirical study, returns are in real terms by adjusting
nominal returns with CPI.
3.1 Measuring CEO Incentives
Because the CEO makes most major corporate decisions and exerts the greatest inﬂuence
on the ﬁrm among senior executives, we focus on incentive provisions for CEOs. We
identify CEOs by the ﬁelds “BECAMECE” and “CEOANN”. Using only “CEOANN” is
problematic, since there are many missing observations for the earlier periods of the sample
(Milbourn, 2002). We measure incentives by pay-performance sensitivity, which we deﬁne as
the dollar value change in the CEO’s ﬁrm-speciﬁc wealth per $1,000 change in the shareholder
value (Jensen and Murphy, 1990).
CEO compensation comprises several diﬀerent layers. Total current compensation (TCC)
is the sum of salary and bonus. Total direct compensation (TDC) is the sum of total current
6
9.
compensation, other annual short-term compensation, payouts from long-term incentive
plans, the Black-Scholes (1973) value of stock options granted, the value of restricted stocks
granted, and all other long-term compensation. Both TCC and TDC measure the direct
payment a CEO receives from his ﬁrm within one ﬁscal year. However, neither TCC nor
TDC considers the indirect compensation that a CEO derives from a revaluation of stocks
and stock options already granted to him in previous years.
We compute the change in value of a CEO’s stock holdings over a given year as the
beginning-of-year value of his stock holdings multiplied by the year’s stock return. It
is important to revalue the stock option portfolios. ExecuComp contains the values of
unexercised exercisable and unexercisable in-the-money options, but the database measures
only the intrinsic value of the option portfolios. We apply the Black-Scholes (1973) model to
revalue the stock option portfolios. We adopt Core and Guay’s (1999) method to compute
the hedge ratio delta (δ) of each option.4 Delta refers to how the value of a stock option
varies given a one-dollar change in value of the underlying stock. We then multiply the option
deltas by the change in the ﬁrm’s market value (or change in shareholders’ wealth), adjusting
for the percentage of stock shares represented by the options, and then add the value from
stock option exercise to obtain the change in value of the CEO’s option holdings. The change
in CEO’s total ﬁrm-speciﬁc wealth (DT OT W ) is the sum of the direct compensation and
the indirect compensation due to the revaluation of stocks and stock option holdings.
We deﬁne the change in shareholder value (V CHAN GE) over a given year as the
ﬁrm’s beginning-of-year market value times the year’s stock return. Using V CHAN GE
and DT OT W , we calculate P P S T OT as the dollar value change in the CEO’s total ﬁrm-
related wealth per $1,000 change in the shareholder value.
4
Core and Guay’s (1999) method quite accurately accounts for the economically meaningful value of the
in-the-money options, but their method ignores the value of out-of-the-money options. This may likely create
systematic bias for ﬁrms with more underwater options and for ﬁrms with options that are at-the-money or
near-the-money. This concern is partially alleviated since our results are robust to the subperiod analysis in
Section 5.2 where the sample is divided into the bubble subperiod and the post-bubble subperiod. Also, for
simplicity, some researchers use 0.7 as the δ value.
7
10.
We deﬁne the stock-based pay-performance sensitivity (P P S) as the change in dollar
value of the CEO’s existing stock-based compensation (stocks and stock options) for a $1,000
change in the shareholders’ wealth. We express the number of shares and options held by
CEOs in units of thousands and the number of shares outstanding in millions.
( CEO s Holdings of Stocks and Stock Options)it
P P Sit = . (2)
( Shareholder s W ealth)it
We can interpret the stock-based pay-performance sensitivity as the CEO’s stock
ownership percentage times 1,000 plus incentives from stock option holdings. We can do
so by using the option delta (δ):
(N umber of Options Held)it−1
P P Sit = SHROW N P Cit−1 ∗ 1000 + δ, (3)
(N umber of Shares Outstanding)it−1
where SHROW N P C is the CEO’s ownership percentage.
P P S is more appealing than P P S T OT for our empirical analysis. We can estimate
the revaluations of existing stock and stock options with relative precision compared to
measuring the impact of ﬁrm performance on the changes in total direct compensation.
Including grants of stock options and restricted stocks can be problematic since they might
serve as an ex-post mechanism like a bonus rather than an ex-ante incentive mechanism.
We therefore, focus primarily on the stock-based incentives. The analysis using P P S T OT ,
which we include as a robustness check, yields essentially the same results.
Panel A of Table 1 presents the summary statistics of the CEO’s compensation
and incentive measures. For the average (median) CEO in our sample, total current
compensation (TCC), in 1992 constant dollars, is $960,776 ($680,813). Total direct
compensation (TDC) ranges from zero to $564.17 million, with an average value of $3.3
million and a median value of $1.42 million. A CEO holds a mean of 3.33% of his ﬁrm’s stocks
with a standard deviation of 7.33%. Median CEO ownership is 0.43%, and the maximum
ownership is 99.44%. A CEO gains an average of $1.92 million from stock option revaluation
8
11.
with a standard deviation of $34.86 million. The median option revaluation is worth $0.21
million. Overall, the CEO’s total ﬁrm-related wealth (DT OT W ) has a mean value of $18.83
million and a median value of $2.3 million. We note that every compensation measure,
particularly the change in the CEO’s ﬁrm-related wealth, has a higher mean than its median
and displays a right-skewed pattern. The annual change in shareholders’ value (VCHANGE)
has a mean of $457.17 million with a median of $61.79 million.
The mean and median values of the stock-based incentive measure (P P S) as deﬁned in
equation (3) are, respectively, $40.79 and $12.85 per $1,000 change in the shareholders’ value.
The P P S measure has a standard deviation of $74.59 with a maximum value of $994.45. The
total pay-performance sensitivity (P P S T OT ) has a mean of $46.51, which is higher than
P P S, but it also has a much larger standard deviation ($81.92). The median of P P S T OT
is $13.09 and is almost the same as the median of P P S. The stock-based incentive does not
deviate much from the total incentive, corroborating ﬁndings in prior studies that changes
in the value of stock and stock option holdings account for the majority of pay-performance
sensitivities (Hall and Liebman, 1998; and Murphy, 1999).
3.2 Measuring Stock Market Informativeness
3.2.1 Probability of Informed Trading (PIN)
Easley, Kiefer and O’Hara (1996, 1997) develop and use the P IN variable to measure the
probability of informed trading in the stock market. The measure is based on the market
microstructure model introduced in Easley and O’Hara (1992), where trades can come from
liquidity traders or from informed traders. The literature has ﬁrmly established the P IN
variable as a good measure of stock price informativeness in various settings.5 P IN directly
estimates the amount of information held by traders on a speciﬁc stock, and it is a good
5
See, e.g., Easley, Kiefer, and O’Hara (1996, 1997); Easley, Kiefer, O’Hara, and Paperman (1996); Easley,
O’Hara, and Srinivas (1998); Easley, Hvidkjaer, and O’Hara (2002); Chen, Goldstein, and Jiang (2003); and
Vega (2004).
9
12.
empirical variable to test our hypothesis that stock prices have an active role in enhancing
pay-performance sensitivity.
Our description of the model and how we construct the P IN measure are as follows.
There are three types of players in the game, liquidity traders, informed traders, and market
makers. The arrival rate of liquidity traders who submit buy orders is b and the arrival rate of
liquidity traders who submit sell orders is s. Every day, the probability that an information
event occurs is α, in which case the probability of bad news is δ and the probability of good
news is (1 − δ). If an information event occurs, the arrival rate of informed traders is µ.
Informed traders submit a sell order if they get bad news and a buy order if they get good
news. Thus, on a day without information events, which occurs with probability (1 − α), the
arrival rate of a buy order will be b and the arrival rate of a sell order will be s. On a day
with a bad information event (with probability αδ), the arrival rate of a buy order will be
b and the arrival rate of a sell order will be s + µ. On a day with a good information event
(with probability α(1 − δ)), the arrival rate of a buy order will be b + µ and the arrival rate
of a sell order will be s. The likelihood function for a single trading day is:
( b )B − s ( s )S
− − b ( b)
B
− s +µ ( s + µ)
S
L(θ|B, S) = (1 − α)e eb
+ αδe e
B! S! B! S!
( b + µ)B − s ( s )S
+α(1 − δ)e− b +µ e , (4)
B! S!
where θ = ( b , s , α, δ, µ), B is the number of buy orders, and S is the number of sell orders
in a single trading day.6
Using the number of buy and sell orders in every trading day within a given year,
M = (Bt , St )T , and assuming cross-trading day independence, we can estimate the
t=1
6
The trade direction is inferred from intraday data based on the algorithm proposed in Lee and
Ready (1991).
10
13.
parameters of the model ( b , s , α, δ, µ) by maximizing the following likelihood function:
t=T
L(θ|M ) = L(θ|Bt , St ). (5)
t=1
Thus, we estimate P IN by dividing the estimated arrival rate of informed trades by the
estimated arrival rate of all trades:
αµ
P IN = . (6)
αµ + b + s
We maximize the likelihood function in equation (5) for the parameter space θ and then
calculate P IN for the period 1993-2001. There are 8,456 ﬁrm-year observations after we
merge the TAQ database with the ExecuComp database.
Panel A of Table 2 presents summary statistics of P IN . In our sample, the average
P IN is 0.163 with a standard deviation of 0.051. The median value of P IN is 0.157. The
maximum P IN is 0.797 with a minimum of zero. The average sample property of our
P IN estimates is consistent with the sample proprety identiﬁed in previous studies (see, e.g,
Easley, Hvidkjaer, and O’Hara, 2002).
3.2.2 Measures Based on the Analysts’ Forecast Data
Since information-based trading can only be inferred, we supplement the P IN measure with
three other variables using the analysts’ earnings forecast data from the I/B/E/S History
Summary File: number of forecasts, forecast errors, and dispersion of forecasts. For each
ﬁrm in the ExecuComp database, as the raw measure we use the summary information on
analysts’ ﬁscal year 1 earnings estimates in the ninth month of a ﬁscal year.7
We control for the inﬂuence of size and other variables on the number of analyst forecasts
by using residual analyst coverage (RESCOV ), which we obtain by regressing the log value
7
We also conduct the analysis by using summary information on each of the other month’s ﬁscal year
1 earnings estimates within that ﬁscal year. The summary information is stable across months and is very
close to the annual average value. Those analyses all yield qualitatively similar results.
11
14.
of one plus the number of analyst forecasts against ﬁrm size, market-to-book ratio, turnover
ratio, Nasdaq dummy, leverage ratio, industry dummies and year dummies.8 As in Hong,
Lim, and Stein (2000), if no I/B/E/S value is available within that ﬁscal year (i.e., ﬁrms
in the ExecuComp database are not matched with ﬁrms in the I/B/E/S database), we set
the number of coverage to zero. Since I/B/E/S does not include some investment banks
and hence their analysts, in unreported analysis we also drop observations with no I/B/E/S
values and the results are qualitatively similar.
Panel A of Table 2 presents summary statistics of the residual coverage that we obtain
from such a regression. RESCOV has mean of 0.023 and median of 0.264. The standard
deviation of RESCOV is 1.134, with a minimum of -3.28 and a maximum of 2.57.
In addition to RESCOV , we construct two other variables, analysts’ forecast error
(F ERROR) and dispersion in analysts’ earnings forecasts (DISP ER), as proxies for stock
market informativeness. We compute F ERROR as the absolute value of the diﬀerence
between mean earnings forecasts and actual earnings, divided by the absolute value of
actual earnings. When a ﬁrm’s information production process is intense and eﬀective,
analysts following the ﬁrm’s stock are more likely to agree with one another on the ﬁrm’s
earnings prospect. F ERROR measures the intensity of a ﬁrm’s information production.
We merge I/B/E/S/ with ExecuComp and delete the observations with zero actual earnings
since F ERROR would be undeﬁned. This results in a sample of 9,420 observations over the
period 1992-2001. Panel A of Table 2 presents summary statistics for F ERROR. It has a
mean (median) of 0.444 (0.066) and a standard deviation of 5.014.
Akinkya and Gift (1985), Diether et al. (2002), and Johnson (2004) show that the
dispersions of analysts’ earnings forecasts is a proxy for the information environment in
which various traders trade. A larger analyst forecast dispersion implies less intensive
information production or a less transparent information environment and vice versa. We
8
For brevity, we do not report the results of the analyst coverage regression. The most signiﬁcant variable
to explain analyst coverage is ﬁrm size, followed by turnover, Nasdaq dummy and leverage. Details are
available upon request.
12
15.
compute DISP ER as the standard deviation of earnings forecasts scaled by the absolute
value of the mean earnings forecast and times 100. If the mean earnings forecast is zero, we
exclude the stock from the sample.9 Combining I/B/E/S/ with ExecuComp and deleting
the observations with zero mean earnings forecast, we obtain a sample of 9,161 observations
for the period 1992-2001. Panel A of Table 2 presents summary statistics of DISP ER. Its
mean (median) is 12.116 (3.175) and its standard deviation is 66.754.
Panel B of Table 2 presents the bivariate Pearson correlation coeﬃcients and their
signiﬁcance levels of the four market informativeness variables. After merging the four
variables into a balanced panel, the number of observations drops to 4,460. In this balanced
panel, the correlation between P IN and RESCOV is -0.016 and is not signiﬁcant. The
correlation between P IN and F ERROR is 0.018 and is not signiﬁcant either. The
correlation between P IN and DISP ER is 0.040 and is signiﬁcant at the 1% level. RESCOV
and F ERROR are almost uncorrelated (the correlation coeﬃcient is 2e-5). The correlation
between RESCOV and DISP ER is 0.032 and is signiﬁcant at the 5% level. The correlation
between F ERROR and DISP ER is 0.016 and is not signiﬁcant. The piecewise correlations
among the four market informativeness proxies in an unbalanced panel have a similar pattern
and are not reported for brevity. Overall, the four variables have no or small correlations
among each other, indicating that they capture diﬀerent aspects of a ﬁrm’s stock market
informativeness.
3.3 Control Variables
Panel B of Table 1 summarizes the control variables we use in our empirical analysis. We
obtain stock return data from CRSP. We calculate the market value (M KT V AL) at each
year-end. Our sample comprises a range of ﬁrms with an average market capitalization of
$4.29 billion and a median of $0.86 billion. The smallest ﬁrm has only $6,100 in market
9
We have also assigned the observations with zero mean earnings forecast to the highest dispersion decile
(90% percentile). Doing so does not qualitatively change our main results.
13
16.
capitalization and the largest one has $427.22 billion. We use the log value of market
capitalization (SIZE) as one measure of ﬁrm size. The mean (median) value of SIZE is
$6.91 million ($6.76 million) with a standard deviation of $1.6 million. The ﬁrm’s annualized
percentage stock return (AN N RET ) averages at 22.45% (median value at 11.54%). The
annualized percentage volatility of stock returns (AN N V OL), which we compute by using
the past ﬁve years of monthly stock returns, has a mean value of 39.44% and a median value
of 35.23%.
We extract ﬁrm-speciﬁc accounting data from Compustat. To measure a ﬁrm’s growth
opportunities, we calculate Tobin’s q as the ratio of the market value of assets to the book
value of assets. We obtain the market value of assets as the book value of assets (data 6)
plus the market value of common equity (data 25 times data 199) less the book value of
common equity (data 60) and balance sheet deferred taxes (data 74). Table 1 shows that
the average Tobin’s q of the ﬁrms in our sample is 2.16 with a standard deviation of 2.75.
The median is 1.48. Tobin’s q is as low as 0.28 and as high as 105.09.
Other control variables are CEO tenure, industry dummies, and year dummies. We
calculate CEO tenure as the number of years the executive has been the CEO of this ﬁrm
as of the compensation year in ExecuComp. The mean and median of the CEO tenure are
7.68 and 5.5 years, respectively. The maximum tenure in our sample is 54.92 years. Finally,
we construct 20 industry dummies based on the 2-digit SIC code from CRSP.
3.4 Econometric Speciﬁcations
There are two basic econometric methods in empirical studies of incentives. One is to directly
regress the incentive measure against a set of explanatory variables. This approach makes
it straightforward to interpret results and also provides the ﬂexibility to control for factors
that may aﬀect the pay-performance sensitivity. Following this approach, we specify the
14
17.
basic model as:
P P Si,t = γ0 + γ1 ∗ IN F OVi,t−1 + γ2 ∗ sizei,t−1 + γ3 ∗ annvoli,t−1 + γ4 ∗ T obinqi,t−1
+γ5 ∗ tenurei,t + γk ∗ dummiesi,t + εi,t , (7)
k≥6
where P P S refers to the computed pay-performance sensitivity as speciﬁed in equation (3);
IN F OV stands for the set of market informativeness variables, i.e., P IN , RESCOV ,
F ERROR and DISP ER; and dummies are industry dummies and year dummies.
In equation (7), we control for several empirically relevant variables: ﬁrm size, return
volatility, growth opportunities, CEO tenure, industry eﬀects, and year eﬀects. It is an
empirical regularity that the pay-performance sensitivity is related to the ﬁrm size (Jensen
and Murphy, 1990; Schaefer, 1998; and Baker and Hall, 2002). Aggarwal and Samwick (1999),
Core and Guay (2002), Jin (2002), among others, ﬁnd that the return variability plays a
key role in explaining some of the heterogeneity in pay-performance sensitivities. Smith
and Watts (1992) and Gaver and Gaver (1993) report that a ﬁrm’s growth opportunities
have a signiﬁcant impact on managerial incentives. Bertrand and Mullainathan (2001),
Milbourn (2002), and Bebchuk and Fried (2003) argue in diﬀerent aspects that CEO tenure
is related to CEO incentives. Many studies document the presence of industry eﬀects and
year eﬀects in variations in managerial incentives (Ely, 1991; and Murphy, 1999).
In equation (7), the coeﬃcient γ1 captures the inﬂuence of market informativeness on
incentives and is our main parameter. We test our hypothesis by examining the sign and
signiﬁcance of the estimated γ1 . The economic signiﬁcance of market informativeness on
P P S is easy to interpret in this setting.
Another speciﬁcation, adopted in Jensen and Murphy (1990), and Aggarwal and
Samwick (1999), among others, uses the CEO’s compensation level as the dependent variable.
15
18.
Based on the approach, we specify our model as:
DT OT Wi,t = α + vchangei,t ∗ (β0 + β1 IN F OVi,t−1 + β2 sizei,t−1 + β3 annvoli,t + β4 T obinqi,t−1
+β5 tenurei,t + βd dummiesi,t ) + β6 ∗ IN F OVi,t−1 + β7 ∗ sizei,t−1 + β8 ∗ annvoli,t
+β9 ∗ T obinqi,t−1 + β10 ∗ tenurei,t + βk ∗ dummiesi,t + εi,t , (8)
where DOT OWi,t denotes the changes in the CEO’s ﬁrm-related wealth while employed by
ﬁrm i in year t.10
To address cross-sectional variations in the pay-performance sensitivity, in equation (8)
we interact the shareholder dollar return (V CHAN GE) with ﬁrm size, return volatility,
growth opportunities, CEO tenure, and industry and year dummies. To test the relation
between the pay-performance sensitivity and the market informativeness, we interact the
variable IN F OV with V CHAN GE to control for any heterogeneity in the pay-performance
sensitivity beyond that attributable to ﬁrm size, return variability, growth opportunity, CEO
tenure, industry eﬀect, and year eﬀect. The coeﬃcient β1 associated with this interaction
term is the parameter of interest. If the heterogeneity of PPS does not depend on the market
informativeness, the estimated β1 will not be signiﬁcantly diﬀerent from zero. By including
the control variables in equation (8), we ensure that our estimates of the interaction term
coeﬃcients are not aﬀected by any relation between the control variables and the level of
compensation that may happen to exist in the cross section.
If we use DT OT W as the dependent variable, we do not need to measure incentives
10
We can also replace DT OT W in equation (8) with the total direct compensation or its diﬀerent
components or their annual changes. The breakdown of T DC helps us study how its diﬀerent
components, e.g., salary, bonus, and stock options grants, respectively, respond to the ﬁrm performance
(see Albuquerque (2004)). This approach is less appealing in our context for two reasons. First, our goal is
to examine the relation between stock market informativeness and CEO incentives. Having the empirically
appropriate incentive measures is the key, but it is still unclear whether salary, bonus, stock or stock option
grants should be counted as ex-post bonus or ex-ante incentive mechanisms. Second, it has been shown that
CEO incentives are mainly provided through revaluations of existing stocks and stock options holdings (Hall
and Liebman, 1998; and Murphy 1999). Even if salary, bonus, stock or stock option grants can be treated
as the ex-ante incentive mechanisms, their individual eﬀects on the overall CEO incentives are small.
16
19.
directly. This approach is appealing when the proposed pay-performance sensitivity measures
are very noisy or when it is diﬃcult to measure incentives directly. However, the approach is
less straightforward. We can only calculate the imputed pay-performance sensitivity based
on the estimated coeﬃcients of those interaction terms, which makes it relatively diﬃcult to
interpret the economic signiﬁcance of variables of interest. For example, from equation (8),
we infer PPS as β0 + β1 IN F OVi,t−1 + β2 sizei,t−1 + β3 annvoli,t + β4 T obinqi,t−1 + β5 tenurei,t +
βd dummiesi,t for a CEO of ﬁrm i in period t. In this formula, the level of PPS and the
contribution of IN F OV to PPS are not immediately clear. The imputed PPS proﬁle is a
linear function of the variables IN F OV , size, annvol, T obinq, tenure, and dummies. β0
measures the intercept of the PPS proﬁle. β1 , β2 , β3 , β4 , and β5 capture the respective
responses of the proﬁle to changes in the ﬁve variables.
Another pitfall of this econometric speciﬁcation is that it is still questionable whether
some components of the total CEO compensation should be included. For example, do
salary, bonus, and the grants of stocks and stock options in the current ﬁscal year serve as
incentive mechanisms for future CEO performance?
3.5 Testable Hypotheses and Estimation Methods
As noted earlier, the four IN F OV variables, P IN , RESCOV , F ERROR and DISP ER,
capture the market informativeness from diﬀerent perspectives. A larger value of P IN (or
RESCOV ) corresponds to a more informative market and a larger value of F ERROR (or
DISP ER) corresponds to a less informative market. The testing of our core hypothesis
comprises several hypothesis tests, depending on which IN F OV variable and which
econometric speciﬁcation is used.
Speciﬁcally, with the ﬁrst econometric speciﬁcation in equation (7), we have
H0’: γ1 = 0, versus
H1’: γ1 > 0 if the IN F OV variable is P IN or RESCOV ; or versus
17
20.
H1”: γ1 < 0 if the IN F OV variable is F ERROR or DISP ER.
We replace γ1 in the above hypotheses with β1 if we use the second econometric speciﬁcation
in equation (8).
Given the apparent right skewness in the CEO compensation and incentive data (as
shown in Table 1), we follow the literature and estimate median regressions to control for
the possible inﬂuence of outliers. We obtain standard errors of the estimates by using
bootstrapping methods.
As a robustness check, we replace each raw measure of the four market informativeness
proxies with its cumulative distribution function (CDF) in both model speciﬁcations. As
Aggarwal and Samwick (1999) argue, the use of CDF helps to make the highly non-
homogeneous data homogeneous and the regression results economically more sensible. It
is easier to transform the estimated coeﬃcient values into pay-performance sensitivities at
any percentile of the distribution of the proxies. Thus, the CDFs give the regression results
more immediate economic intuition than the raw measures.
4 Empirical Results
4.1 Using PPS as the Dependent Variable
Our main empirical results are based on the econometric speciﬁcation in equation (7),
where we use CEO incentive measure as the dependent variable. Table 3 presents the
median regression results of the stock-based pay-performance sensitivity P P S, deﬁned in
equation (3), by using P IN , CDF (P IN ), RESCOV , and CDF (RESCOV ) as the measures
for stock market informativeness. In all regressions we control for ﬁrm size, return volatility,
growth opportunity, CEO tenure, industry dummy, and year dummy. For brevity we do not
report coeﬃcient estimates for the industry and year dummies.
Columns 1 and 2 show how the CEO’s incentives respond to the probability of informed
18
21.
trading (P IN and its CDF). The estimated coeﬃcients of the PIN measure (γ1 ) in both
columns are positive and statistically signiﬁcant at the 1% signiﬁcance level. γ1 is 23.999
in Column 1 where we use the raw measure of P IN . The diﬀerence in P P S between any
two P IN levels is calculated as the diﬀerence in the two P IN values multiplied by γ1 .
For example, since the maximum, median, and minimum P IN are 0.797, 0.157, and 0,
respectively, the pay-performance sensitivity at the maximum P IN is $15.36 higher than
that at the median P IN as a result of $1,000 increase in shareholder value, all else equal. The
incentive with the median P IN is $3.77 larger than the incentive with the minimum P IN .
The economic magnitude of P IN on pay-performance sensitivity is signiﬁcant, considering
that the mean (median) stock-based pay-performance sensitivity in our sample is $40.787
($12.845).
Column 2 further illustrates the economic signiﬁcance of P IN . Using CDF (P IN ) as
the explanatory variable, the coeﬃcient γ1 is 4.07. The PPS at the maximum P IN , i.e.,
CDF (P IN )=1, is $2.035 larger than that at the median PIN, i.e., CDF (P IN )=0.5, per
$1,000 increase in the shareholder value.
Columns 3 and 4 show how the CEO’s incentives react to the residual analyst coverage
(RESCOV and CDF (RESCOV )). The estimated coeﬃcients γ1 are 0.414 for RESCOV
and 1.599 for CDF (RESCOV ), respectively. Both are statistically signiﬁcant at the 1%
signiﬁcance level, lending support to our hypothesis. We note that the economic signiﬁcance
of RESCOV is much smaller than that of P IN , which might be due to the fact that we
obtain RESCOV as the residual from the regression of the raw analyst coverage against a set
of ﬁrm-speciﬁc variables. RESCOV is likely a noisy proxy for market price informativeness.
Using CDF (RESCOV ) partially alleviates the concern. Column 4 shows that after we use
the CDF of residual coverage as the explanatory variable, the economic magnitude of the
analyst coverage on stock-based PPS almost quadruples.
Table 3 shows that the estimated coeﬃcients of control variables are all statistically
signiﬁcant and consistent in both sign and magnitude across the four regressions. The
19
22.
coeﬃcient of the lagged SIZE, γ2 , is around -3.12 to -3.98, reaﬃrming that the pay-
performance sensitivity is negatively and strongly correlated with the ﬁrm size even after
we take into account the last three years of observations (1999-2001). The coeﬃcient of
AN N V OL, γ3 , ranges from 18.06 to 20.14, implying a positive risk-incentive relation. Our
results are consistent with the model prediction of Prendergast (2002) as well as the empirical
evidence in Core and Guay (2002). However, we note that the empirical evidence on the
risk-incentive relation is mixed, and the economic rationale is still subject to a heated debate
(see the survey in Prendergast (2002)).
The coeﬃcient of the lagged T obinq, γ4 , is about 1.44 to 1.63. There is a clearly positive
relation between growth opportunities and CEO incentives, corroborating the evidences
reported in Smith and Watts (1992) and consistent with the hypothesis that a higher
incentive level is needed when CEO eﬀort becomes more valuable. There is also a positive
relation between CEO tenure and PPS as the estimated coeﬃcient of tenure is approximately
0.94 to 1.22.
4.2 Using DTOTW as the Dependent Variable
Table 4 reports median regression results from the four models in which we use the lagged
values of P IN , CDF (P IN ), RESCOV , and CDF (RESCOV ) separately as proxies for
stock market informativeness. The coeﬃcients of the interaction terms are related to
incentives. All the coeﬃcients of the interaction terms are statistically signiﬁcant at the
1% level. The coeﬃcients are consistent in both sign and magnitude with the corresponding
estimates from Table 3, where we use P P S as the dependent variable. The coeﬃcient of
V CHAN GE, β0 , ranges from 10.65 to 15.79. The incentive is strongly and negatively
related to SIZE and β2 is around -2.10. The incentive is strongly and positively related to
AN N V OL (β3 around 18 to 24), T obinq (β4 around 0.10 to 0.43), and T enure (β5 around
0.58 to 0.78).
The parameter of interest, β1 , is signiﬁcant and positive across all regressions. Its estimate
20
23.
is 16.96 for P IN , 1.91 for CDF (P IN ), 0.39 for RESCOV , and 1.36 for CDF (RESCOV ),
respectively. The t-statistics of the estimates are large enough to readily reject the null
hypothesis in favor of the alternative hypothesis that β1 > 0. The p-values of the estimated
coeﬃcient β1 are zero for all four regressions. Both the t-statistics and the p-values lend
strong support to the claim that β1 > 0, clearly pointing to a strong and positive relation
between CEO incentives and the two market informativeness proxies, the probability of
informed trading and the residual analyst coverage.
Table 4 also reports the estimates of the coeﬃcients that link the control variables to the
level of CEO pay. Firm size is strongly and positively related to the total compensation. This
ﬁnding, together with the ﬁnding of the negative relation between ﬁrm size and incentive,
might have diﬀerent implications.11 The results suggest that CEOs of larger companies are
able to extract rents from shareholders in the form of higher compensation levels independent
of ﬁrm performance.12 The positive size eﬀect on the compensation level is also consistent
with the evidence that larger ﬁrms require more talented managers who are more highly
compensated (Smith and Watts, 1992) and who are consequently expected to be wealthier
(Baker and Hall, 2002). The negative size eﬀect on the PPS can be partially explained by
the fact that a CEO in a larger ﬁrm has diﬃculty in acquiring a large percentage of shares
either directly or indirectly through stock and option grants (Demsetz and Lehn, 1985; and
Baker and Hall, 2002).
Like Aggarwal and Samwick (1999), we also ﬁnd a strong positive relation between return
volatility and the level of CEO compensation, although the principal-agent theory does not
give a clear prediction on whether the expected level of compensation is increasing with
the ﬁrm stock return volatility. We conjecture that as the volatility of the underlying asset
increases, the value of option portfolios held by a CEO increases, leading to a higher level of
total compensation, all else equal.
11
We do not explore these implications in detail as they are beyond the scope of this study.
12
We thank the referee for suggesting this explanation to us.
21
24.
We ﬁnd that the estimated coeﬃcients of T obinq are positive and signiﬁcant in all
regressions, which indicates a positive relation between the growth opportunity and the
compensation level. Together with a positive relation between the growth opportunity and
incentives, this ﬁnding is again consistent with Smith and Watts (1992). That is, a higher
growth opportunity increases the marginal value of CEO eﬀort. Therefore higher incentive
and compensation levels are needed, holding all else ﬁxed.
CEO tenure is positively related to both the incentive and compensation levels, which can
be subject to diﬀerent interpretations. On the one hand, Bertrand and Mullainathan (2001)
and Bebchuk and Fried (2003) argue that the CEO has a lot of inﬂuence in setting his
compensation contract. CEO tenure might measure the extent to which a CEO is entrenched
and is therefore better able to appropriate money, including stock-related pay, for himself.
On the other hand, CEO tenure is argued to be a proxy for CEO reputation and experience,
which helps to improve incentives as uncertainty about a CEO’s ability is resolved over time
(Milbourn, 2002). A longer tenure also captures the accumulation of stock holdings and
hence a higher level of compensation, other things equal. A third interpretation is that
CEO tenure is a proxy for the agency cost. As CEOs approach retirement, increased equity
incentives can be used to counteract potential horizon problems (Jensen and Murphy, 1990).
The relation between the two market informativeness proxies and the compensation level
is mixed. Although it is positive, P IN is not signiﬁcantly related to the CEO compensation.
RESCOV is strongly and negatively related to the CEO compensation, a ﬁnding for which
we do not have a good explanation.
4.3 Further Analysis with Alternative Price Informativeness
Measures
As noted earlier, by using the analyst earnings forecast database we can construct two other
variables, analysts’ forecast error (F ERROR) and dispersion in analysts’ earnings forecasts
22
25.
(DISP ER), as our proxies for stock price informativeness. The larger the forecast error
measure or the dispersion measure is, the less eﬃcient is the information production of
a ﬁrm’s stock and hence the lower is the stock price informativeness. When we test the
hypothesis that CEO incentives respond positively to the stock price informativeness, we
expect to ﬁnd a signiﬁcant negative correlation between the stock-based PPS and F ERROR
or DISP ER. The results are reported in Table 5.
In Table 5, Columns 1 and 2 present the median regressions results of the stock-
based PPS when we use F ERROR and its cumulative distribution function as the market
informativeness measure, respectively. The parameter of interest, γ1 , is -0.096 for F ERROR
and signiﬁcant at the 10% level. γ1 is -2.964 for CDF (F ERROR) and signiﬁcant at
the 1% level. Columns 3 and 4 report the results when we use DISP ER and its
cumulative distribution function as the market informativeness measure, respectively. γ1
is estimated respectively, at -0.004 for DISP ER, signiﬁcant at the 5% level, and -4.418 for
CDF (DISP ER), signiﬁcant at the 1% level. Consistent with the results when we use PIN
and residual analyst coverage as the market informativeness measures, the signiﬁcant and
negative estimates of γ1 in Table 5 support the hypothesis that CEO incentives do respond
positively to the stock price informativeness.
The estimated γ1 is very small, although statistically signiﬁcant, if we use the raw
informativeness measures, F ERROR and DISP ER. This ﬁnding might cast doubt on
their economic signiﬁcance. Since both variables are highly right-skewed and fat-tailed,
the estimates from the cumulative density functions of F ERROR and DISP ER are more
economically meaningful. Using CDF (F ERROR) (CDF (DISP ER)) as proxies for the
price informativeness, we ﬁnd that the pay-performance sensitivity at the maximum forecast
error level (or the maximum analyst forecast dispersion level) is $1.482 ($2.136) lower
than the pay-performance sensitivity at the median forecast error level (or the median
analyst forecast dispersion level) per $1,000 change in the shareholder value. The economic
magnitude is by no means small.
23
26.
Consistent with the results from Sections 4.1 and 4.2, the coeﬃcients of control variables
are all statistically signiﬁcant and have the expected signs. Incentive is negatively related to
ﬁrm size and is positively related to return volatility, growth opportunity, and CEO tenure.
5 Robustness Analysis
5.1 Using Alternative Pay-Performance Sensitivity Measures
P P S considers only changes in the value of CEO stock and stock option holdings. The
CEO’s direct compensation from salary and bonus, new stock and stock option grants, long-
term incentive pay and other annual compensation might also matter. P P S T OT , deﬁned
as the dollar value change in the CEO’s total ﬁrm-related wealth per $1,000 change in the
shareholder value, measures total CEO incentives because it is derived from both the direct
and indirect stock-based CEO compensation. We repeat our analysis using P P S T OT as the
dependent variable in equation (7). We use CDFs of the four market price informativeness
proxies and report the median regression results in Table 6.
The results in Table 6 are consistent with those results using P P S. The estimates
for the parameter of interest, γ1 , are 5.283 for CDF (P IN ), 1.463 for CDF (RESCOV ),
-3.072 for CDF (F ERROR), and -4.169 for CDF (DISP ER), respectively. All estimates
are statistically signiﬁcant at the 1% level. Similar to when P P S is used as the incentive
measure, we ﬁnd that the coeﬃcient estimates using P P S T OT are economically signiﬁcant.
In Table 6 when we use P P S T OT as the incentive measure, the signs and magnitude
of the other parameter estimates are comparable to the corresponding signs and magnitude
of the parameter estimates reported in Tables 3 and 5, where we use P P S as the incentive
measure. These results, from another perspective, substantiate the ﬁnding of Hall and
Liebman (1998), among others, that changes in the value of CEO stock-based compensations
contribute to nearly all of the pay-performance sensitivity, and that the CEO direct
24
27.
compensation has little impact on the pay-performance sensitivity. This result alleviates our
concerns on how to compute T DC, e.g., should we also include future direct compensations
into T DC, since ﬁrm performance might have an impact on the future pay levels? Should
new stock and option grants be treated as ex-ante incentives or ex-post bonuses? Table 6
suggests that the results are not sensitive to such choices.
Table 6 further shows that incentive is related negatively to ﬁrm size and positively to
return volatility, growth opportunities, and CEO tenure.
5.2 Subperiod Analysis
We perform a subperiod analysis for three reasons: (1) The U.S. stock market bubble burst
in 2000, rendering many executives’ options virtually worthless. The way in which we
compute the changes in value of stock option portfolios might create bias for ﬁrms with
more out-of-money options, and for ﬁrms with options that are at-the-money or near-the-
money. Comparing the results before and after the bubble burst is thus useful and interesting.
(2) We might rationally suspect that a ﬁrm’s incentive provision policies changed after the
bubble burst, which might directly aﬀect our results. (3) Most of the prior research uses
data up to 1999. By conducting a subperiod analysis, we are able to compare our analysis
with the prior research. We can also check the stability of the relation between the incentive
and the variables of interest.
To conduct the subperiod analysis, we split our sample into two subsamples, the bubble
period 1992-1999 and the post-bubble period 2000-2001. Using the stock-based PPS as the
dependent variable and the CDF transformations of the four price informativeness measures
as the main explanatory variables, we apply median regressions to the two subperiods. Table
7 presents the results of this subperiod analysis.
Columns (1)-(4) and Columns (1’)-(4’) in Table 7 report the results for the 1992-1999
period and the 2000-2001 period, respectively. There is very little diﬀerence in the results
25
28.
across the two periods. For each coeﬃcient of the control variables, the corresponding
estimates from the two subperiods have the same signs. Most of the estimates have similar
magnitudes and signiﬁcance levels across the two subperiods. For either subperiod, the
incentive is negatively related to ﬁrm size and positively related to return volatility, growth
opportunities, and CEO tenure.
The corresponding estimates of the parameter γ1 are consistent with each other across the
two subperiods, and for the full sample period. When we use CDF (RESCOV ) to measure
the market price informativeness, γ1 is 1.561 for 1992-1999, 1.944 for 2000-2001 and 1.599
for the full sample period 1992-2001. All three results are signiﬁcant at the 5% level. When
we use CDF (F ERROR), γ1 is -2.322 for 1992-1999, -2.965 for 2000-2001, -2.964 for 1992-
2001, and all are signiﬁcant at the 1% level. When we use CDF (DISP ER), γ1 is -3.427 for
1992-1999, -3.949 for 2000-2001, and -4.418 for 1992-2001. Again, all results are signiﬁcant
at the 1% level. When we use CDF (P IN ) as the market price informativeness measure, the
estimates of γ1 increase sharply from 1.464 for the ﬁrst subperiod to 9.208 for the second
subperiod. Both estimates are signiﬁcant at the 5% level. The estimated γ1 for CDF (P IN )
is 4.070 for the whole sample 1992-2001. Consistent with the full-sample analysis, the
subperiod analysis brings about the same results, that is, CEO incentives increase with
market informativeness. This relation is generally stable across the two subperiods despite
the stock market bubble burst in 2000. The incentive enhancement eﬀect due to the market
informativeness seems to strengthen slightly after the bubble burst, which is obvious if we
use the CDF (P IN ) as our proxy for market informativeness.
5.3 Other Robustness Checks
We examine two other types of robustness: robustness of control variables, and robustness
of estimation methods.
In addition to market capitalization, net sales and total assets are other commonly used
variables to measure ﬁrm size. We replace market capitalization with either net sales or total
26
29.
assets in our econometric models. We also include the squared ﬁrm size proxies to control for
possible nonlinearities in the data. All these alternative speciﬁcations yield similar results
on the relation between CEO incentives and market informativeness.
In unreported median regressions we also control for other factors that may be related to
CEO incentives, such as the ratio of capital over sales, R&D expense over capital, and the
investment-capital ratio (Jin, 2002). These variables are likely to capture the value of CEO
eﬀort. Including those control variables does not change our main results qualitatively.
Following Core and Guay (1999), we try the log value of CEO tenure in the regression
model to capture the possible concave relation between CEO incentives and CEO tenure.
Again, our main results do not change qualitatively.
Apart from the median regressions on which we build our empirical analysis, we adopt
other types of model estimation methods. We perform OLS with robust standard errors,
CEO ﬁxed-eﬀects OLS (Aggarwal and Samwick, 1999), and robust regression (Hall and
Liebman, 1998).13 The CEO ﬁxed-eﬀects OLS controls for all diﬀerences in the average level
of incentives across executives in the sample. Having CEO ﬁxed eﬀects helps alleviate the
potential “unobserved endogeneity” problem if those unobservables tend to stay constant
over time. All these regression methods generate qualitatively similar results. For brevity
we do not report these results.
6 Summary and Conclusion
In this paper we examine the relation between chief executive oﬃcers’ incentives and stock
market informativeness. Following Holmstrom and Tirole (1993), we develop a model that
links market microstructure to CEO incentives. We empirically investigate whether stock
market informativeness has a signiﬁcant impact on CEO pay-performance sensitivity.
13
A robust regression begins by screening out and eliminating gross outliers based on OLS results, and
then iteratively performs weighted regressions on the remaining observations until the maximum change in
weights falls below a pre-set tolerance level, say, 1%.
27
30.
Using probability of informed trading, residual analyst coverage, analysts’ earnings
forecast error, and dispersion of analysts’ earnings forecasts as four proxies for the market
informativeness, we conduct various empirical tests and justify the key prediction that CEO
pay-performance sensitivity increases with market informativeness. As predicted, we ﬁnd a
positive relation between probability of informed trading (or residual analyst coverage) and
CEO pay-performance sensitivity, and a negative relation between analysts’ earnings forecast
error (or dispersion of analysts’ earnings forecasts) and CEO pay-performance sensitivity.
Our results are robust to alternative estimators, measures of incentives, sample periods,
model speciﬁcations, and estimation methods. Our results suggest that incentive pay is more
eﬀective in a more informative stock market. Thus, this paper not only provides strong
empirical support for the propositions in Holmstrom and Tirole (1993), but also contributes
empirically to the research area that combines the market microstructure literature with the
corporate ﬁnance literature.
28
31.
Appendix: A Model Linking Market Microstructure to Executive
Compensations
We introduce a single-period model that illustrates the link between stock trading and
the eﬀectiveness of market-based executive compensation.
At the initial point of time 0, a publicly held ﬁrm is established with a random terminal
payoﬀ at time 1: r = e + δ, where e is the eﬀort level that the potential manager would
privately choose and e cannot be contracted on. δ is a zero-mean, normally distributed
random variable with variances, Vδ . The shares are issued on the ﬁrm’s future cash ﬂow.
At stage 1, the ﬁrm owner (the principal) hires one manager (the agent). The owner
writes a compensation contract on two performance measures, – the ﬁrm’s terminal payoﬀ
r, and stock price P :
W = a + b1 r + b2 P, (A. 1)
where a represents the ﬁxed salary. b1 and b2 capture the sensitivities of the manager’s
compensation relative to the ﬁrm’s terminal payoﬀ, r and its stock price, P , respectively.
Given the compensation contract, the manager chooses an eﬀort level e ∈ [0, ∞), which is
not observable.
At stage 2, the stock market opens. Stock market participants can observe an informative
yet non-contractible signal on the ﬁrm’s future value δ, at a cost. We assume that there is an
endogenous number of N investors who choose to do so. A potential investor will search for
the private signal on δ only if the expected value of doing so exceeds her reservation value µ.
The costly signal acquired by investor i (if she so chooses) is δ + i , where i is i.i.d. with a
mean of zero and a variance of V . Both the informed and uninformed traders submit their
order ﬂows to the market maker. Informed trader i submits a market order that is linear in
her signal, β(δ + i ). We assume that the total liquidity demand in the market is z and z
is a zero-mean, normally distributed variable with variance Vz . Since there are N informed
traders and z liquidity demand, the total order ﬂow observed by the market maker is given
by ω = N βδ + N (β i ) + z. The competitive market maker, given the aggregate order ﬂow
i=1
ω, sets a price such that P = E [r|ω]. In order to obtain analytically tractable solutions, we
ignore the manager’s compensation W in the price function. Given that W in our model is
linear in both r and P , including it in the price function does not change the information
˜
content of the stock price. The stock price derived from this price function is informationally
equivalent to that from the more general price function speciﬁcation, P = E [(r − W )|ω].
At time 1, the payoﬀ is realized, the incentive contract is honored, and the ﬁrm is
liquidated. The resulting liquidation proceeds are distributed between the manager and
the principal.
All agents are risk-neutral except the manager. The manager’s preference is represented
by a negative exponential utility function over her compensation W with the (absolute) risk
aversion coeﬃcient γ. Her cost of choosing the eﬀort e is denoted as C(e) = 1 ke2 . The cost
2
is measured in money and is independent of the manager’s wealth. The manager’s evaluation
of the normally distributed income W , given her choice of eﬀort e, can then be represented
29
32.
in the certainty equivalent measure as follows:
γ
U (W, e) = E(W ) − V ar(W ) − C(e). (A. 2)
2
Based on the model set-up, we can solve a rational-expectation equilibrium in which the
players in the real sector (the principal and the manager) use the information contained
in the stock price to make decisions, and both the real sector and the stock market reach
equilibrium at the same time.
We begin with the stock market equilibrium. With z liquidity demand, the total order
ﬂow observed by the market maker is ω = N βδ + N (β i ) + z. The market maker sets a
i=1
linear price schedule of the form P = e + λω (Kyle, 1985). Using standard techniques, we
1
−2 1 N Vδ2 (Vδ +V )
obtain the equilibrium value of λ as λ = Vz Γ 2 , where Γ = [(N +1)Vδ +2V ]2
.
1
Vδ2 (Vδ +V )Vz2
The expected proﬁt of an informed trader is then given by ER = 1 . A
Γ 2 [(N +1)Vδ +2V ]2
potential trader will make an eﬀort to search for the private signal if and only if the expected
proﬁt from doing so exceeds her reservation value µ. Thus, the equilibrium number of
informed traders N , is determined by
1
Vδ2 (Vδ + V )Vz2
1 = µ. (A. 3)
Γ 2 [(N + 1)Vδ + 2V ]2
Using the implicit function theorem, we obtain
Lemma 1 The number of informed traders N , (1) decreases as the investors’ reservation
value µ on becoming informed increases, (2) increases as the volatility of the ﬁrm’s cash ﬂow
Vδ increases, and (3) increases as the volatility of the liquidity trading Vz increases.
Proof. Omitted and available from the authors. Q.E.D.
We also have
Lemma 2 The (inverse) informativeness of the stock price P is given by
Vδ (Vδ + 2V )
V ar(δ|P ) = . (A. 4)
(N + 1)Vδ + 2V
Proof. Given the linear price schedule P = e + λω, we have V ar(δ|P ) = V ar(δ|λω).
Thus,
Cov(δ, λω)2 Vδ (Vδ + 2V )
V ar(δ|P ) = Vδ − = . (A. 5)
V ar(λω) (N + 1)Vδ + 2V
Q.E.D.
Equation (A. 4) shows that as N increases, the conditional variance V ar(δ|P ) decreases
monotonically. From equation (A. 4), we have:
30
33.
Proposition 1 As the number of informed traders increases, the stock market becomes more
informative.
We can now solve the optimal compensation contract. We set up the principal’s problem
as follows:
max E(r − W )
a,b1 ,b2
γ
s.t. max E(W ) − V ar(W ) − c(e) ≥ W (A. 6)
e 2
where W is the reservation utility to the manager.
The manager’s compensation, as speciﬁed in (A. 1), is a linear function of both the
realized payoﬀ, r, and the stock price P . Rather than dealing with W = a + b1 r + b2 P
˜ ˜
directly, the analysis of contracting becomes analytically much easier if we transform the
wage function into the following equivalent normalized function:
ˆ ˜ ˆ
W = a + b1 r + b2 P , (A. 7)
ˆ
where a = a + b2 E(P ) = a + b2 e∗ and P = P − E(P ) = P − e∗ . Note that equation (A. 7) is a
ˆ
linear transformation based on hypothesized equilibrium values. One beneﬁt of dealing with
ˆ ˆ
P rather than P is that P is a zero-mean and normally distributed variable. Since the mean
ˆ
of P is zero, it is much easier to analytically illuminate its information role in the manager’s
ˆ
contracting process. At equilibrium, P is informationally equivalent to P . That is, we have
ˆ ˆ
V ar(P ) = V ar(P ) and Corr(δ, P ) = Corr(δ, P ).
ˆ
With W = a + b1 r + b2 P , we can rewrite the principal’s problem as
ˆ
max (1 − b1 )e − a
ˆ
a,b1 ,b2
ˆ
s.t. b1 = ke
b2 γ
1 b2 γ ˆ ˆ ke2
a = W − b1 e +
ˆ V ar(δ) + 2 V ar(P ) + γb1 b2 Cov(δ, P ) + (A. 8)
2 2 2
The ﬁrst-order conditions yield
1 ˆ b1
− b1 γV ar(δ) − b2 γCov(δ, P ) − = 0, (A. 9)
k k
and
ˆ ˆ
−γb2 V ar(P ) − γb1 Cov(δ, P ) = 0. (A. 10)
1 1
ˆ ˆ
Plugging Cov(δ, P ) = ρVδ2 V ar(P ) 2 into equation (A. 10), we have
1
ρVδ2
b2 = −b1 . (A. 11)
ˆ 1
V ar(P ) 2
31
34.
1
ρVδ2
Plugging b2 = −b1 ˆ 1
back into equation (A. 9), we obtain
V ar(P ) 2
b1 = (1 + γkVδ (1 − ρ2 ))−1 , (A. 12)
N Vδ 1
where ρ = [ (N +1)Vδ +2V ] 2 .
Given equation (A. 11), we can rewrite the transformed compensation function in (A. 7)
as: 1
ρVδ2 ˆ
W = a + b1 (˜ −
ˆ r P ), (A. 13)
ˆ 1
V ar(P ) 2
1
ρVδ2 ˆ
where r −
˜ ˆ 1
P is an aggregate measure built on two performance instruments. b1 then
V ar(P ) 2
captures the sensitivity of incentive pay to the aggregate performance measure. Immediately,
we obtain:
Proposition 2 In the rational expectations equilibrium, the optimal pay-performance
sensitivity, b∗ , is
1
1
b∗ =
1 . (A. 14)
1 + γkVδ (1 − ρ2 )
∂b1 ∂b1 ∂b1
From equation (A. 14), we have ∂k
< 0, ∂γ
< 0, and ∂Vδ (1−ρ2 )
< 0.
∂b1
Since Vδ (1 − ρ2 ) = V ar(δ|P ), we have ∂V ar(δ|P )
< 0. V ar(δ|P ) captures the
informativeness of stock price (Kyle 1985).
Further, we show
Proposition 3 (1) The more costly it is to collect information about ﬁrm performance, the
less sensitive is pay to performance ( ∂b1 < 0).
∂µ
(2) Given the number of informed traders N , the more volatile the cash ﬂow, the less sensitive
∂b
is pay to performance ( ∂V1 < 0).
δ
(3) Given the number of informed traders N , the more dispersed are the informed investors’
opinions (or the less informative of the performance signal), the less sensitive is pay to
∂b
performance ( ∂V1 < 0).
Proof. Omitted and available from the authors. Q.E.D.
32
35.
References
[1] Aggarwal, Rajesh K., and Andrew A. Samwick, 1999. The other side of the trade-oﬀ:
The impact of risk on executive compensation. Journal of Political Economy 107, 65–
105.
[2] Akinkya, Bipin, and Michael Gift, 1985. Dispersion of ﬁnancial analysts’ earnings
forecasts and the (option model) implied standard deviations of stock returns, Journal
of Finance, 40, 1353-1365.
[3] Albuquerque, Ana, 2004. Who are your peers? A study of relative performance
evaluations. Working paper, University of Rochester.
[4] Baker, George P., and Brian J. Hall, 2004. CEO incentives and ﬁrm size. Journal of
Labor Economcs, forthcoming.
[5] Bebchuk, Lucian Arye, and Jesse M. Fried, 2003. Executive compensation as an agency
problem. Journal of Economnic Perspectives, 17, 71-92.
[6] Bertrand, Marianne, and Sendhil Mullainathan, 2001. Are CEOs rewarded for luck?
The ones without principals do. Quarterly Journal of Economics 116, 901–932.
[7] Bushman, Robert M., and Abbie J. Smith, 2001. Financial accounting information and
corporate governance. Journal of Accounting and Economics 32, 237–333.
[8] Chen, Qi, Itay Goldstein, and Wei Jiang, 2003. Price informativeness and invetment
sensitivity to stock prices. Working paper, Columbia University.
[9] Core, John, and Wayne Guay, 1999. The use of equity grants to manage optimal equity
incentive levels. Journal of Accounting and Economics 28, 151–184.
[10] Core, John, and Wayne Guay, 2002. The other side of the trade-oﬀ: The impact of
risk on executive compensation: A revised comment. Working paper, University of
Pennsylavania.
[11] Demsetz, Harold, and Kenneth Lehn, 1985. The structure of corporate ownership:
Causes and consequences. Journal of Political Economy 93, 1155–1177.
[12] Diether, Karl B., Christopher J. Malloy, and Anna Scherbina, 2002. Diﬀerence of
opionion and the cross section of stock returns. Journal of Finance 57, 2113-2141.
[13] Easley, David, Soeren Hvidkjaer, and Maureen O’Hara, 2002. Is information risk a
determinant of asset returns? Journal of Finance, 57, 2185-2221.
[14] Easley, David, Nicholas Kiefer, Maureen O’Hara, and Joseph Paperman, 1996. Liquidity,
information, and less-frequently traded stocks, Journal of Finance, 51, 1405-1436.
[15] Easley, David, Nicholas Kiefer, Maureen O’Hara, and Joseph Paperman, 1996. Cream-
skimming or proﬁt-sharing? the curious role of purchased order ﬂow, Journal of Finance,
51, 811-833.
33
36.
[16] Easley, David, Nicholas Kiefer, Maureen O’Hara, and Joseph Paperman, 1997. The
informative content of the trading process, Journal of Empirical Finance, 4, 159-186.
[17] Easley, David, Maureen O’Hara, and P.S. Srinivas, 1998. Option volumn and stock
prices: evidence on where informed traders trade, Journal of Finance, 53, 431-465.
[18] Ely, Kirsten M., 1991. Interindustry diﬀerences in the relation between compensation
and ﬁrm performance variables. Journal of Accounting Research, 29, 37–58.
[19] Garen, John, E., 1994, Executive compensation and principal-agent theory. Journal of
Political Economy 102, 1175–1199.
[20] Gaver, Jennifer J., and Kenneth M. Gaver, 1993. Additional evidence on the
association betweenn the investment opoetunity set and coporate ﬁnancing, dividend,
and compensation policies. Journal of Accounting and Economics 16, 125-160.
[21] Hall, Brian J., and Jeﬀrey B. Liebman, 1998. Are CEOs really paid like bureaucrats?.
Quarterly Journal of Economics 113, 653–691.
[22] Hartzell, Jay, and Laura Starks, 2004. Insitutional investors and executive
compensation. Journal of Finance, forthcoming.
[23] Holmstrom, Bengt, and Jean Tirole, 1993. Market liquidity and performance
monitoring. Journal of Political Economy 101, 678–709.
[24] Hong, Harrison, Terrence Lim, and Jeremy Stein, 2000. Bad news travels slowly: Size,
analyst coverage, and the proﬁtability of momentum strategies. Journal of Finance 55,
265–295.
[25] Jensen Michael C., and Kevin J. Murphy, 1990. Performance pay and top-management
incentives. Journal of Political Economy 98, 225–262.
[26] Jin, Li, 2002. CEO compensation, diversiﬁcation, and incentives. Journal of Financial
Economics 66, 29–63.
[27] Johnson, Timothy C. 2004. Forecast dispersion and the cross section of expected returns.
Journal of Finance 59, 1957-1978.
[28] Kyle, Albert S., 1985. Continuous auctions and insider trading. Econometrica 57, 1315–
1335.
[29] Lee, Charles M. and Mark J. Ready, 1991. Inferring trade direction from intraday data.
Journal of Finance, 46, 733-746.
[30] Milbourn, Todd T., 2002. CEO reputation and stock-based compensation. Journal of
Financial Economics, forthcoming.
[31] Murphy, Kevin J., 1999. Executive compensation. In: Orley Ashenfelter and David Card
(Eds.), Handbook of Labor Economics, Vol 3, North Holland.
34
37.
[32] Prendergast, Canice, 2002. The tenuous trade-oﬀ of risk and incentives. Journal of
Political Economy 110, 1071–1102.
[33] Schaefer, Scott, 1998. The dependence of pay-performance sensitivity on the size of the
ﬁrm. Review of Economics and Statistics 3, 436–443.
[34] Smith, Cliﬀord, and Ross Watts, 1992. The investment opportunity set and corporate
ﬁnancing, dividend, and ﬁnancing policies. Journal of Financial Economics 32, 262–292.
[35] Subrahmanyam, Avanidhar, and Sheritan Titman, 1999. The going-public decision and
the development of ﬁnancial markets. Journal of Finance 54, 1045–1082.
[36] Vega, Clara, 2004. Are investors overconﬁdent? Working paper, University of Rochester.
35
38.
Table 1 – Summary Statistics: CEO Compensation/Incentive Measures and Firm /CEO
Characteristics
Total current compensation (TCC) is the sum of a CEO's annual salary and bonus. Total direct
compensation (TDC) is the sum of TCC, other annual short-term compensation, payouts from
log-term incentive plans, the Black-Scholes (1973) value of stock options granted, the value of
restricted stocks granted, and all other long-term compensation. SHROWNPC is the CEO’s
ownership percentage at year-end. We compute the annual change in a CEO's total firm-related
wealth (DTOTW) as the sum of TDC and the changes in value of the CEO's existing stocks and
stock options. The change in value of stock holdings is computed as the beginning-of-year value
of CEO’s stock holdings multiplied by the current year’s stock return. The change in the value of
stock options (DOPTV) equals the product of option deltas and the change in the firm’s market
value, adjusting for the percentage of stock shares represented by the options, and plus the value
from stock option exercises. TCC, TDC, and DTOTW are expressed in thousands of dollars. We
calculate the change in the firm's market value (VCHANGE) as the percentage stock returns
times the market capitalization of equity at last year-end (in millions of dollars). The stock-based
pay-performance sensitivity (PPS) is the change in value of the CEO’s stock-based holdings
(stocks and stock options) divided by the change in the firm’s stock market value. We define the
total pay-performance sensitivity (PPS_TOT) as the dollar value change in the CEO's total firm-
related wealth per $1,000 change in the shareholder value. Size is the log of MKTVAL, which is
the fiscal-year-end equity market value (in millions of dollars). Stock return (ANNRET) is the
annualized percentage return in real terms. We compute the annualized percentage volatility of
stock returns (ANNVOL) using the past five years of monthly stock return data. We calculate
Tobin's Q as the ratio of the market value of assets to the book value of assets, where the market
value of assets is defined as the book value of assets (data 6) plus the market value of common
equity (data 25 times data 199) less the book value of common equity (data 60) and balance sheet
deferred taxes (data 74). We compute TENURE as the CEO’s tenure as of year t. Nobs is the
number of non-missing observations. All monetary variables are in 1992 constant dollars.
Variables Mean Std Dev Median Min Max Nobs
Panel A: Compensation and incentive measures
TCC ($K) 960.776 1,348.354 680.813 0 91,634.85 17,584
TDC ($K) 3,299.286 9,899.765 1,424.882 0 564,171.19 17,051
SHROWNPC (%) 3.331 7.326 0.428 0 99.445 16,708
DOPTV($K) 1,920.610 34,856.17 211.803 -1,509,144 1,641,605 14,123
DTOTW ($K) 18,834.77 497,626.39 2,296.758 -22,056,001 35,166,453 17,348
VCHANGE ($M) 457.171 5,541.287 61.788 -257,239.5 182,126.2 16,944
PPS 40.787 74.590 12.845 0 994.446 16,769
PPS_TOT 46.506 81.916 13.093 0 997.479 16,904
Panel B: Control Variables
MKTVAL ($M) 4,292.293 15,270.39 864.983 6.10e-3 427,220.76 17,369
SIZE ($M) 6.909 1.601 6.763 -5.100 12.965 17,369
ANNRET (%) 22.451 95.466 11.542 -96.654 7,010.12 17,349
ANNVOL (%) 39.438 18.845 35.227 10.279 347.80 16,890
TOBIN’s Q 2.159 2.747 1.480 0.279 105.090 17,307
TENURE 7.682 7.409 5.500 0.083 54.917 16,028
36
39.
Table 2 - Summary Statistics of Stock Price Informativeness Measures
Panel A reports summary statistics of the four stock price informativeness variables used in our
empirical analysis: probability of informed trading (PIN) per Easley, Kiefer and O’Hara (1992,
1996, and 1997), residual analyst coverage (RESCOV), analyst forecast error (FERROR), and
forecast dispersion among analysts (DISPER). RESCOV is the residual from regressing the log
value of one plus the number of analysts’ earnings forecast against firm size, market-to-book ratio,
turnover ratio, Nasdaq dummy, leverage ratio, industry dummies, and year dummies. We
compute FERROR as the absolute value of the difference between the mean earnings estimate
and the actual earnings divided by the actual earnings. We calculate DISPER as the standard
deviation of earnings forecasts scaled by the absolute value of the mean earnings forecast and
times 100. Nobs is the number of non-missing observations. We compute all the variables after
we match the ExecuComp database with the TAQ or I/B/E/S database. The sample period is
1993-2001 for PIN and is 1992-2001 for RESCOV, FERROR, and DISPER. Panel B presents
the bivariate Pearson correlations among the four price informativeness variables in a balanced
sample containing 4,460 firm-year observations. The significance level of each correlation
coefficient is reported in parenthesis.
Panel A: Summary statistics of the price informativeness measures
Variables Mean Std Dev Median Min Max Nobs
PIN1 0.163 0.051 0.157 0 0.797 8,456
RESCOV 0.023 1.134 0.264 -3.280 2.574 17,204
FERROR 0.444 5.014 0.066 0 422.5 9,420
DISPER 12.116 66.754 3.175 0 2,550 9,161
Panel B: Bivariate correlations among the price informativeness measures
PIN RESCOV FERROR DISPER
PIN1 1.0000
RESCOV -0.0155 1.0000
(0.301)
FERROR 0.0184 2e-5 1.0000
(0.219) (0.999)
DISPER 0.0403 0.0322 0.0158 1.000
(0.007) (0.031) (0.293)
Note 1: The sample period for PIN is 1993-2001, since we calculate PIN using data from the TAQ
database which starts its data coverage from 1993.
37
40.
Table 3 – Median Regression of Stock Based Pay-Performance Sensitivity (PPS): Using PIN
and RESCOV Measures as Proxies for Price Informativeness, 1992-2001
We define stock-based pay-performance sensitivity (PPS) as the change in value of the CEO’s
stock and stock options holdings per $1,000 change in shareholder value. We specify the model
as: PPSi,t = γ0 + γ1*INFOVi, t-1 + γ2* Sizei, t-1 + γ3* annvoli, t + γ4*Tobinqi, t-1 + γ5*Tenurei, t +
industry dummies + year dummies+ εi, t . We use raw measures of the two variables, PIN and
RESCOV, and their cumulative distribution functions (CDF) as proxies for stock price
informativeness. Coefficient estimates on industry dummies and year dummies are not reported
for brevity. All monetary variables are in 1992 constant dollars. We report standard errors based
on 20 bootstrap replications in parentheses. The sample period for the regressions involving PIN
or CDF(PIN) is 1994-2001, since the PIN measure which we calculate from the TAQ database
covers the period 1993-2001 and we need one-year lag for such regressions.
*, **, *** - significant at the 10%, 5% and 1% levels, respectively.
Independent Variables (1) (2) (3) (4)
Intercept 15.858*** 17.626*** 25.591*** 26.348***
(1.213) (1.324) (0 .848) (0 .929)
PINi, t-1 23.999***
(2.721)
CDF(PINi, t-1) 4.070***
(0.569)
RESCOVi, t-1 0. 414***
(0. 093)
CDF(RESCOVi, t-1) 1.599***
(0.389)
SIZEi, t-1 -3.140*** -3.115*** -3.981*** -3.977***
(0.099) (0.114) (0.080) (0.086)
ANNVOLi, t 20.143*** 19.927*** 18.055*** 18.197***
(1.004) (1.096) (0.757) (0.808)
Tobin’s Qi, t-1 1.633*** 1.558*** 1.453*** 1.444***
(0.104) (0.113) (0.041) (0.043)
Tenurei, t 0.943*** 0.947*** 1.224*** 1.224***
(0.014) (0.016) (0.014) (0.015)
Sample Size 7007 7007 16207 16207
Pseudo R-square 0.104 0.104 0.111 0.111
38
Clipping is a handy way to collect and organize the most important slides from a presentation. You can keep your great finds in clipboards organized around topics.
Be the first to comment