Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Earnings Management to Sustain Consecutive Earnings Increase and Market Response
JENNIFER KAO, The University of Alberta
KYUNGA NA, The University of Alberta
April 19, 2013
Acknowledgement: This paper is based on Kyunga Na’s thesis completed at the University of Alberta. We are grateful to her examination committee members, Professors Christopher Frei, Yonghua Ji, Bin Ke (external examiner), Tom Schneider and Yao Tian for their helpful comments and guidance. We would also like to thank participants and discussants, Professors Sunwha Choi (Lancaster University), Pascale Lapointe-Antunes (Brock University) and Wendy Wilson (Southern Methodist University) at the 2010 Korean Accounting Association Conference, 2011 Canadian Academic Accounting Association Conference and 2011 American Accounting Association Conference for their valuable suggestions on earlier versions of the paper. All the remaining errors are the authors’ responsibilities.
Abstract
In this paper, we examine the pattern of accrual and real activity management undertaken by firms with a
string of at least five-year consecutive earnings increases. Results indicate that both types of earnings
management intensify as firms move towards the end of an earnings string. While the extent of accrual
management declines sharply in the year when the string finally ends, firms continue to maintain a
moderately high level of real activity management. We also study market reaction to earnings strings
achieved through earnings management and find that firms with a high level of accrual management
during an earnings string receive a relatively lower earnings response coefficient, but the market is unable
to distinguish firms with a high vs. a low level of real activity management. These results are invariant to
sample choices (i.e., multiple earnings strings of varying lengths, one earnings string only or five-year
earnings strings only) and earnings management measures (i.e., performance-matched, performance-
unadjusted or performance-adjusted). Our study extends the literature by showing that earnings growth
may be artificially sustained through accrual as well as real activity management. Moreover, these two
methods have different market consequences and exhibit different patterns at the break of an earnings
string. Findings that the investors have a greater ability to understand the implication of accrual
management than real activity management add to the growing evidence questioning the existence of
accrual anomaly.
JEL classification: M41; C12; C33; M42;
Keywords: Accrual Management; Real Activity Management; Earnings String; Earnings Response
Coefficient
2
1. Introduction
Many firms view attaining a string of consecutive earnings increases (labelled an earnings string) as an
important goal because the capital market rewards consistent earnings growth and reacts negatively to the
break to this pattern (Barth, Elliott and Finn 1999). Sustaining earnings growth is also important to firm
managers who are remunerated based on performance (Ke 2004). While firms with competitive
advantages in the product market can outperform their competitors (Porter 1985), maintaining a pattern of
earnings increases is difficult due to the cyclical nature of underlying economic conditions. Thus, firm
managers are often motivated to take actions to ensure the continuity of an earnings string for as long as
possible. One method whereby growth in earnings can be artificially sustained is through earnings
management. Myers, Myers and Skinner (2007) interpret a larger than expected number of firms with
earnings momentum as prima facie evidence of earnings management. Crucial to this argument is the
assumption that insiders have superior information about earnings, compared to outside investors.
Information asymmetry allows insiders to predict or time the break to an earnings string by selling their
shares three to nine quarters in advance of actual earnings reports announcing the reversal to an upward
earnings trend, even though institutional investors normally do not start trading until one or two quarters
before the break (Ke and Petroni 2004; Ke, Huddart and Petroni 2003). A break to an earnings string may
eventually occur however, if it cannot be sustained with the help of earnings management. Previous
studies have shown that firms tend to manage earnings through accruals in order to extend an earnings
string (Yong 2009; Baik, Farber, Johnson and Yi 2008). However, these studies do not consider other
earnings management methods, such as real activity management, or the pattern of earnings management
when the string finally breaks. They also do not speak to market reaction to earnings strings achieved
through earnings management.
The purpose of this paper is to fill the gap in the literature by studying not only the pattern of, but
also the market response to, accrual and real activity management undertaken by firms with a string of
consecutive earnings increases for at least five years (labelled ES firms). Specifically, we address the
3
following three research questions: First, do ES firms undertake more aggressive earnings management
towards the end of an earnings string, compared to the early part of an earnings string? Second, is the year
immediately following the end of an earnings string (i.e., the break year) characterized by a different
pattern of earnings management from that within an earnings string? Third, can the capital market
appropriately price earnings management used by ES firms to sustain an earnings string?
To test the pattern of earnings management, we partition an earnings string into four sub-periods,
i.e., Early-ES (the first two years of an earnings string), Mid-ES (from the third year of an earnings string
to three years before the break), −2Break (two years before the break) and −1Break (one year before the
break), and use performance-matched discretionary accruals (DACCPM) and performance-matched
abnormal discretionary expenses multiplied by −1 (RDISXPM) to proxy for accrual and real activity
management, respectively. We employ a research design that treats firm-year observations from Early-ES
as the reference group and regress DACCPM and RDISXPM separately on four indicator variables,
representing firm-year observations that fall in the remaining three sub-periods and during the break year
(labelled MidES, −2Break, −1Break and Break). For a sample of 1,138 earnings strings from 1,043
firms between 1989 and 2010, we find that after controlling for the potential effects of covariates the
coefficient estimates on MidES, −2Break and −1Break are positive and significant in the DACCPM and
RDISXPM regressions. Moreover, both the coefficient estimate and the associated t-statistics increase in
strength as we move from MidES to −2Break and then to −1Break. These results suggest that ES firms
start to manage their discretionary accruals and real activities in the middle of an earnings string and
intensify such efforts towards the end. By comparison, the coefficient on the Break variable is negative
and significant in the DACCPM regression, but statistically insignificant in the RDISXPM regression.
To study the market response to earnings management by ES firms, we partition our sample into
the high and the low earnings management groups, defined as ES firms whose values of DACCPM or
RDISXPM are above and below the median value of respective sample distribution, and then regress the
market-adjusted return (Return) on the interaction between changes in earnings and an indicator variable,
4
representing ES firms with above-the-median values of DACCPM or RDISXPM (labelled DACCPM_H or
RDISXPM_H). For a reduced sample of 949 firms with complete return data, we find that, after controlling
for factors known to affect Return, the coefficient estimate on the interaction term involving DACCPM_H
is negative and significant and that involving RDISXPM_H is statistically insignificant. While the market
penalizes ES firms with a high level of accrual management by assigning a relatively low earnings
response coefficient (ERC), it is nonetheless unable to distinguish ES firms with a high vs. a low level of
real activity management.
All the above results continue to hold when we replicate the analysis separately on subset of ES
firms that have a single earnings string or firms whose earnings strings last exactly five years. They also
remain invariant to the use of alternative proxies for earnings management, i.e., performance-unadjusted
and performance-adjusted earnings management measures. Our study extends the literature by providing
evidence on the progression in the intensity of earnings management within an earnings string and by
showing that earnings growth may be artificially sustained through not just accrual, but also real activity
management. Moreover, we demonstrate that these two earnings management methods have different
market consequences and exhibit different patterns at the break of an earnings string. Findings that the
investors have a greater ability to understand and price the implication of accrual management, compared
to real activity management add to the growing evidence questioning the existence of accrual anomaly.
They also enhance one`s understanding of real activity management, which has attracted considerable
interest from academic researchers in recent years.
The remainder of this paper is organized as follows. Section 2 presents a review of related
literature and the development of hypotheses. Section 3 summarizes data collection procedure and our
sample. Section 4 presents research design and results from the earnings management analysis, followed
by similar discussions for the market response analysis in Section 5. Section 6 concludes this study.
5
2. Literature Review and Hypothesis Development
2.1 Earnings Management as a Means to Sustain an Earnings String
Burgstahler and Dichev (1997) find that small positive earnings changes occur more frequently than small
negative earnings changes. Based on simulations, Myers et al. (2007) also report that the number of firms
with a string of non-decreasing earnings for at least 20 quarters is much larger than expected. These
findings point to the likely presence of earnings management to avoid earnings declines.
The notion that firms have the ability to sustain an appearance of continued earnings growth
through earnings management until it is no longer viable is addressed more formally in two recent studies.
Yong (2009) shows that firms with earnings growth for at least three consecutive years tend to use large
discretionary accruals in the last two years of earnings strings. Unlike us, Yong (2009) does not speak to
the question of whether earnings management in the latter part of an earnings string is different from that
in the early part or if there is any change to the pattern of earnings management when an earnings string
comes to an end, as the final year of his sample period is allowed to be part of an earnings string.1 Baik,
Farber, Johnson and Yi (2008) examine the role of earnings management within an earnings string using
quarterly data and find that discretionary accruals increase significantly in four quarters before the break
during which the growth in fundamentals starts to decline. A direct comparison of our study with Baik et
al. (2008) is difficult however due to differences in research design. In particular, their design is chosen to
further examine the role of accounting fundamentals in maintaining an earnings string.2
According to a survey of 400 CFOs by Graham, Harvey and Rajgopal (2005), 80 percent of CFOs
favor reducing discretionary expenses as a means to achieving an earnings target, a finding that has
generated much interest from academic researchers in recent years. Roychowdhury (2006) for example
reports that firms manage discretionary expenses and other real activities in order to avoid losses. Gunny
1 For example, earnings strings that include the final year of his sample period, i.e., 2005, may have ended in 2005 or continued beyond 2005. 2 The authors acknowledged that some important accounting fundamentals suppressed from their model (e.g., effective tax rates, the number of employees, auditor quality and corporate governance) could not be measured using quarterly data.
6
(2010) also documents empirical support for real activity manipulations by firms whose earnings are
either just above zero or just meet or beat the last year’s level.
Zang (2012) turns her attention to the trade-offs between accrual and real activity management
and finds that managers view these two methods as substitutes based on their relative costs. Consistent
with Zang (2012), Cohen, Dey and Lys (2008) show that accrual management increased steadily from
1987 until the passage of the Sarbanes-Oxley Act (SOX) in 2002. While the trend was reversed after SOX,
the converse is true for real activity management. Cohen et al. (2008) attribute these patterns to rising
costs of using accrual management since the enactment of SOX. Finally, Cohen and Zarowin (2010)
document a high level of accrual as well as real activity management around 1,511 seasoned equity
offerings made between 1987 and 2006. If continuous earning growth is indeed sustained by aggressive
earnings management, then Cohen and Zarowin’s (2010) findings imply that firms are more likely to
manage earnings upwards using both earnings management tools when an earnings string is near the end,
as doing so would allow insiders to postpone the bad news about upcoming break to the earnings string
and sell their shares at a higher price. The above discussion leads to the first two hypotheses for the study:
Hypothesis 1A. Ceteris paribus, ES firms are expected to report larger discretionary accruals near
the end of an earnings string, compared to the early part of an earnings string.
Hypothesis 1B. Ceteris paribus, ES firms are expected to report smaller abnormal discretionary
expenses near the end of an earnings string, compared to the early part of an earnings string.
Earnings strings normally do not last for an indefinite period of time due to the unpredictability of
macroeconomic and firm-specific circumstances. Sustaining the appearance of continued earnings growth
through aggressive accrual management can be particularly difficult if earnings become too low, as
accruals are typically reversed in the following period. In this case, firms may revert back to “normal”
reporting without attempting to manage earnings. Some firms may even opt for overly conservative
accounting policies by taking a big-bath and manage earnings downward in order to increase accounting
reserve for future period. Either accrual reversal or a big-bath strategy is expected to result in a
7
significantly reduction in discretionary accruals during the year when an earnings string is finally broken,
compared to years leading to the break.3,4
The prediction for real activity management in the break year is less clear-cut. While reduction in
such activities through increases to the level of discretionary expenses can benefit firms in the long run,
they tend to lower income and cash flows in the short run. Since firms are already faced with a declining
fortune that halted the earnings string in the first place, they may be reluctant to increase spending on
discretionary expense during the break year. On the other hand, one may argue that with break to an
earnings string all but certain firms may choose to fix up the basic level of operations neglected due to an
unusually low level of spending on discretionary expenses near the end of an earnings string. On balance,
we expect to see little difference in the level of real activity management in the break year vs. that in the
early part of an earnings string. Thus, we have the next two hypotheses for the study:
Hypothesis 2A. Ceteris paribus, ES firms are expected to report smaller discretionary accruals
during the break year, compared to the early part of an earnings string.
Hypothesis 2B. Ceteris paribus, ES firms are expected to report a similar level of abnormal
discretionary expenses during the break year, compared to the early part of an earnings string.
2.2 Market Response to Earnings Management by Firms with Consecutive Earnings Increases
Barth et al. (1999) report that firms with patterns of increasing earnings receive higher price-earnings (PE)
multiples than other firms, but the incremental PE multiples are reduced significantly in the year when an
earnings string is broken. Ghosh, Gu and Jain (2005) extend Barth et al. (1999) by classifying firms based
on whether consecutive earnings growth is supported by concurrent revenue growth over the same time
period. Results indicate that firms with revenue momentum have higher earnings quality and larger ERC,
compared to those without revenue momentum. Lev, Ryan and Wu (2008) examine market response to
3 It is beyond the scope of this study to distinguish between these two competing arguments. 4 Prolonged earnings management may also lead to restatements when an earnings string is finally broken. Richardson, Tuna and Wu (2002) show that restatement firms tend to have longer strings of consecutive earnings growth than non-restatement firms.
8
the announcements of accounting restatements and find that the market reacts more negatively if
restatements eliminate or shorten the history of consecutive earnings increases. None of these studies
however speak directly to the question of how the market reacts to earnings strings that are maintained
through earnings management, as we do in this paper.5
The effect of accrual management on ERC for firms reporting consecutive earnings growth is
broadly speaking related to a large volume of literature on the informativeness of accounting accruals.
Traditionally, investors are viewed as being fixated on earnings, failing to fully incorporate accruals into
stock price on a timely basis (Xie 2001; Sloan 1996). As a result, a trading strategy that buys stocks of
firms with low accruals and sells stocks of firms with high accruals can yield significant abnormal returns.
Several recent studies however have shown that accrual anomaly either disappears after 2001 (Green,
Hand and Soliman 2011) or exists only for a certain type of firms.6 Moreover, the market is largely
capable of differentiating accrual quality. Francis, LaFond, Olsson and Schipper (2005) for example find
that firms with poor accrual quality tend to have relatively higher costs of debt and equity capital. Balsam,
Bartov and Marquardt (2002) report that investors react negatively to firms suspected to have used
aggressive discretionary accruals to meet consensus analyst forecasts around their 10-Q filings and that,
unlike average investors, sophisticated institutional investors can respond to firms’ accrual management
even before 10-Q filing date. DeFond and Park (2001) document relatively lower ERCs for firms whose
abnormal accruals inflate earnings surprises in either direction. Baber, Chen and Kang (2006) find that
stock return is inversely related to the extent of discretionary accruals if, in addition to earnings, both
balance sheet and cash flow information are also disclosed. In a similar vein, Levi (2008) shows that
voluntary disclosure of accrual information in earnings releases helps investors better assess earnings
quality which in turn mitigates accrual anomaly. Finally, Louis, Robinson and Sbaraglia (2008) present 5 Under the assumption that revenues are less susceptible to manipulations than costs, one may interpret Ghosh et al.’s (2005) findings as implying that ERC is smaller when an earnings string is sustained by aggressive accrual management. 6 Examples include firms with low disclosure quality (Drake, Myers and Myers 2009), those that are profitable (Dopuch, Seethamraju and Xu 2010) and small firms in high sentiment period during which individual investors overestimate the expected returns of stocks (Ali and Gurun 2009). Kraft, Leone and Wasley (2006) also argue that accrual anomaly is likely to occur due to outliers, rather than the investor’s mispricing of accrual information.
9
evidence that the market correctly prices accruals for firms that disclose accrual information at earnings
announcements, but accrual anomaly remains if such information is not disclosed.
Collectively, the above literature suggests that the capital market can identify accrual
management and impound it into stock prices when investors either have reasons to believe that such
activity may have taken place or have access to timely disclosure of accrual information. Since earnings
momentum over a period of time is expected to receive much closer attention from the market than the
case when firms merely meet or beat an earnings target for a single year, we expect firms that sustain the
momentum through aggressive accrual management to be rewarded with smaller ERCs than firms that do
not resort to accrual management. While empirical evidence supports the notion that the market is capable
of correctly pricing accrual management, it may not be readily extended to real activity management
because these two earnings management tools have their own costs. In particular, accruals are closely
scrutinized by external auditors, whereas real activity management represents deviations from optimal
business operations. Suboptimal operating decisions lower firm value in the long run, but are of little
concern to external auditors so long as manipulations are properly accounted for in the financial
statements. Without a formal vetting process, real activity management is expected to be far more
difficult for the capital market to detect than accrual management. Thus, to the extent that the market
penalizes ES firms that manage reported earnings to sustain earnings strings we expect the reduction in
market rewards to be more pronounced for firms that rely on accrual management, compared to those that
use real activity management. The above discussion leads to the final two hypotheses for the study:
Hypothesis 3A. Ceteris paribus, the market is expected to react more negatively to ES firms that
undertake a high level of accrual management, compared to ES firms that undertake a low
level of accrual management.
Hypothesis 3B. Ceteris paribus, the market is expected to react similarly to ES firms that
undertake a high level of real activity management, compared to ES firms that undertake a
low level of real activity management.
10
3. Data and Sample
The initial sample consists of 85,576 firm-year observations over a 22-year (1989-2010) sample period7
and it is obtained by applying the following filters to the COMPUSTAT Fundamental Annual database
between 1987 and 2011: First, firms must not belong to the financial and regulated industries. Second,
firms must not have missing data required for the earnings management analysis. Third, each two-digit
SIC industry-year group must have at least 20 observations to calculate both earnings management
measures, after deleting earnings changes in the extreme 1 percent of distribution.
Following the convention of Barth et al. (1999), we work with firms that report at least five-year
consecutive increases in split-adjusted annual earnings per share. Among the initial sample, 1,417
earnings strings (or 8,932 firm-year observations) fit this definition. We then eliminate 279 strings (or
1,836 firm-year observations)8 that do not have an identifiable break year to yield a total of 1,138
earnings strings (or 7,096 firm-year observations). Combining these observations with another 1,138
observations from the break year yields the final sample of 8,234 firm-year observations that we use for
testing the predictions of Hypotheses 1A, 1B, 2A and 2B (labelled ES Sample).
For the market response analysis, we collect additional data, discussed in Section 5.1, from the
COMPUSTAT Fundamental Annual Database and the CRSP Monthly Stock Database for each firm in
the ES Sample. Deleting firms with missing data and eliminating the top and the bottom 1 percent of
return and price distributions yield 6,575 firm-year observations from 949 distinct firms available for
calculating model variables used to test the predictions of Hypotheses 3A and 3B (labelled ERC Sample).
Panel A of Table 1 summarizes the above sample filter rules. As is evident in Panel B of Table 1,
the ES Sample consists of 1,138 earnings strings from 1,043 distinct firms. Most of the earnings strings
7 The two endpoints for the sample period are chosen because firms must have one- to two-year lag data as well as one-year lead data to calculate model variables and earnings strings. The starting point of 1989 in particular allows us to calculate total accruals using the conceptually superior income statement approach (Hribar and Collins 2002), which requires data from Cash Flow Statement available since 1987. 8 Among them, 111 came from firms whose earnings strings continued beyond 2010; 93 (3) from firms that were merged or acquired (bankrupt or liquidated); 66 from firms that no longer filed with SEC for other reasons; 3 firms became private companies; and finally 3 were the second string of firms whose first earnings string had previously been excluded due to missing break year.
11
represent the first earnings string for a particular firm and only a few are the second string (i.e., 1,043 vs.
95). It would appear that firms rarely succeed in putting together another earnings string following a
break to the first string. In total, 948 distinct firms have only one earnings string during the entire sample
period and 95 firms have two strings. Panel C presents the frequency distribution by the length of
earnings strings. Almost half (46.13 percent) of the ES Sample has strings that last exactly five years and
most of the 1,138 earnings strings (96.83 percent) last 10 years or less. The ES Sample is distributed over
35 two-digit SIC industries, where six firms (or 39 firm-year observations) come from the Mining &
Construction industries and 172 firms (or 1,428 firm-year observations) are drawn from the Retail
industry (see Panel D). This pattern of industry distribution is largely comparable to that for the overall
COMPUSTAT population. The break of an earnings string occurred more frequently during the dot.com
bubble and the global financial crisis (i.e., 94 in 2001; 91, 149 and 109 in 2007–2009; see Panel E). While
not reported to conserve space, the ERC Sample exhibits similar patterns to those depicted in Panels B-E.
[Insert Table 1 about Here]
4. Analysis of Earnings Management Used to Sustain an Earnings String
4.1 Research Design
To test the prediction of Hypotheses 1A and 2A (1B and 2B), we estimate Equation 1 (Equation 2) below
by pooling across the ES Sample of 1,043 distinct firms, or equivalently 8,234 firm-year observations:
퐷퐴퐶퐶 = 훽 + 훽 푀푖푑퐸푆 + 훽 − 2퐵푟푒푎푘 + 훽 − 1퐵푟푒푎푘 + 훽 퐵푟푒푎푘 + 훾 푆푖푧푒 + 훾 퐵푇푀 + 훾 퐿푒푣푒푟푎푔푒 + 훾 퐶퐹푂 + 훾 퐿표푠푠 + 훾 푁푒푤퐼푠푠푢푒 + 훾 퐿푖푡푖푔푎푡푖표푛 + 훾 퐵푖푔푁 + 훾 퐼푛푑푢푠푡푟푦퐷푢푚푚푦 + 훾 푌푒푎푟퐷푢푚푚푦 + 휀 (1)
푅퐷퐼푆푋 = 훽 + 훽 푀푖푑퐸푆 + 훽 − 2퐵푟푒푎푘 + 훽 − 1퐵푟푒푎푘 + 훽 퐵푟푒푎푘 + 훾 푆푖푧푒 + 훾 퐵푇푀 + 훾 퐿푒푣푒푟푎푔푒 + 훾 푁푒푤퐼푠푠푢푒
+ 훾 퐼푛푑푢푠푡푟푦퐷푢푚푚푦 + 훾 푌푒푎푟퐷푢푚푚푦 + 휀 (2)
where subscripts i, t and j represent sample firm i in year t and industry j. All the continuous control
variables in Equations 1 and 2 are winsorized at the top and the bottom 1 percent to mitigate the effects of
12
outliers. We use robust standard errors to correct for potential problems associated with heteroskedasticity
and firm clustering when reporting t-values (Petersen 2009). Definitions and measurements for all the
variables in Equations 1 and 2 are summarized in Appendix 1.
Dependent Variables in Equations 1 and 2
The dependent variable in Equation 1 is used to proxy for the extent of accrual management. To mitigate
potential measurement errors discussed in Kothari, Leone and Wasley (2005), we work with
performance-matched discretionary accruals in the main analysis, calculated using a two-step procedure:
First, we estimate the following forward-looking modified-Jones model cross-sectionally for each two-
digit SIC industry-year group over the ES Sample (Dechow, Richardson and Tuna 2003):
= 훼 + 훼 ( )∆ ∆ + 훼 + 훼 + 훼 ∆ + 휀 . (3)
where TAC is total accruals, defined as income before extraordinary items minus cash flows from
operating activities; TA denotes total assets; ΔS is changes in sales; ΔREC is changes in account
receivables from trade; k is the estimated slope coefficient from regressing ΔREC/TA on ΔS/TA for each
two-digit SIC industry-year grouping and is restricted to between zero and one; PPE denotes gross
property, plant and equipment; S denotes sales. Residuals from Equation 3 represent discretionary
accruals for sample firm i in year t (labelled DACCit). Second, we match every ES firm with a non-ES
firm that has the closest return on assets within the same industry-year group.9 The performance-matched
discretionary accruals for ES firm i are its discretionary accruals minus the matched non-ES firm’s
discretionary accruals (labelled DACCPMit).
The dependent variable in Equation 2 proxies for the extent of real activity management. The
performance-matched abnormal discretionary expenses are calculated by first estimating the following
model adapted from Roychowdhury (2006) cross-sectionally for each two-digit SIC industry-year:
= 훼 + 훼 + 훼 + 휀 . (4)
9 Return on assets is defined as net income deflated by opening total assets. In the event of ties among multiple non-ES firms, the firm with the closest firm size (i.e., natural logarithm of market value of equity) is selected as the match.
13
where DIS denotes discretionary expenses; TA and S are as defined in Equation 3.10 Residuals from
Equation 4 represent abnormal discretionary expenses for firm i in year t (labelled DISXit). We then match
every ES firm with a non-ES firm that has the closest return on assets within the same industry-year group.
The performance-matched abnormal discretionary expenses (labelled RDISXPMit) are given by the
difference between ES firm i’s abnormal discretionary expenses and the matched non-ES firm’s abnormal
discretionary expenses, multiplied by −1, so that a large RDISXPMit implies a greater extent of real
activity management which raises accounting earnings.
Model variables in Equations 3 and 4 are winsorized at the top and the bottom 1 percent of their
respective distributions. Appendix 2 (3) reports the mean values of estimated coefficients in Equation 3 (4)
and the associated t-statistics calculated using the mean value of standard errors across industry-years and
the mean value of adjusted R-square. All the estimated coefficients have the same signs as those in
Dechow et al. (2003) and Roychowdhury (2006) and are significant at the conventional levels.
Test and Control Variables in Equations 1 and 2
Equations 1 and 2 includes four test variables: MidESit, set equal to one if a firm-year observation falls in
the third year of an earnings string to three years before the break and zero otherwise; −2Breakit,
(−1Breakit) set equal to one if an observation comes from two years (one year) before the break and zero
otherwise; and Breakit, set equal to one if an observation falls in the break year and zero otherwise.
Observations in the first two years of an earnings string serve as the reference group in both equations.
We do not offer any prediction on the first test variable, MidESit. A positive and significant coefficient
estimate on each of the next two test variables, −2Breakit and −1Breakit, is consistent with the
predictions of Hypotheses 1A and 1B, whereas the prediction of Hypothesis 2A (Hypothesis 2B) implies
a significantly negative (insignificant) coefficient on the last test variable, Breakit. In the ensuing
discussion and all the tables, we report one-tailed p-value if there is a prediction and two-tailed p-value
otherwise.
10 Unlike Roychowdhury (2006), we do not consider abnormal cash flows from operations and abnormal production costs because the income effect of the former is mixed and the latter is not available for non-manufacturing firms.
14
Equation 1 also includes eight control variables found to affect a firm’s incentives for accrual
management in prior literature:11 firm size (Size it-1), book-to-market ratio (BTMit-1), debt-to-asset ratio
(Leverageit-1), cash flows from operations (CFOit), prior year loss (Lossit-1), new equity issues (NewIssueit),
litigation risks (Litigationit) and audit quality (BigNit). Equation 2 on the other hand includes four control
variables known to affect a firm’s abnormal discretionary expenses:12 firm size (Sizeit-1), book-to-market
ratio (BTMit-1), debt-to-asset ratio (Leverageit-1) and new equity issues (NewIssueit). Since our sample
spans over 22 years (1989–2010) and covers a large number of two-digit SIC industry groups, we also
include IndustryDummy and YearDummy variables to control for potential industry and year effects in
both equations. To ease exposition, subscripts i and t are suppressed throughout subsequent discussions.
4.2 Descriptive Statistics
Panel A of Table 2 presents the distribution of model variables in Equations 1 and 2 over the entire ES
Sample of 8,234 firm-year observations. The mean values of DACCPM and RDISXPM are close to zero (i.e.,
–0.0143 and –0.0017), with DACCPM ranging from –1.8450 to 1.4115 and RDISXPM from –3.7334 to
3.9109. On average, ES firms have positive cash flows (CFO = 0.1090), larger market value than book
value (BTM = 0.4434) and less debts than assets (Leverage = 0.1835). Only a few ES firms reported loss in
the previous year (Loss = 0.1647) and less than 40 percent of ES firms belong to highly litigious industries
(Litigation = 0.3594). The vast majority of ES firms issue equity capital in the current period (NewIssue =
0.8528) and retain a Big-N auditor (BigN = 0.8549).
[Insert Table 2 about Here]
Panel A (B) of Table 3 reports the correlation matrix among pairs of model variables, other than
Industry and Year dummies, in Equation 1 (2). The first earnings management measure, DACCPM, is
11 See Cohen and Zarowin (2010); Ball and Shivakumar (2008); Lim and Tang (2008); Zang (2012); Barton and Simko (2002); Erickson and Wang (1999); Francis, Maydew and Sparks (1999); Healy and Wahlen (1999); Teoh, Welch and Wong (1998); Dechow, Sloan and Sweeny (1996; 1995); Dechow (1994); Francis, Philbrick and Schipper (1994). 12 See Cohen and Zarowin (2010) and Roychowdhury (2006).
15
negatively associated with Size, CFO and BigN, but positively associated Leverage and Loss. The pair-wise
Pearson correlation coefficients, appearing above the diagonal, are –0.1143, –0.3333, –0.0593, 0.0278 and
0.1145, respectively, all significant at the 5 percent level. The second earnings management measure,
RDISXPM, on the other hand is significantly correlated with BTM, Leverage and NewIssue, with the Pearson
correlation coefficients of 0.0520, 0.0495 and –0.0520, respectively. Spearman correlation coefficients,
appearing below the diagonal in Panels A–B, exhibit qualitatively similar patterns. While many control
variables are highly correlated, un-tabulated results indicate that the largest variance inflation factor and the
largest condition index in the DACCPM (RDISXPM) regression are 1.5665 and 2.2044 (1.1870 and 1.5498),
respectively. Thus, multicollinearity is unlikely to be a major concern.
[Insert Table 3 about Here]
4.3 Main Results Based on the Earnings Management Analysis
Rows 1–5, Panel A of Table 4 summarize the mean and median values of DACCPM and RDISXPM,
calculated over 2,276, 2,544, 1,138, 1,138 and 1,138 firm-year observations from the following five sub-
periods: Early-ES, Mid-ES, – 2Break, –1Break and Break. For both measures of earnings management,
these values increase as we move forward along the earnings string before reaching their respective peak
in −1Break. At the break of an earnings string, both the mean and the median values of DACCPM decline
sharply from the level in −1Break, whereas they go down moderately to about the same level as Early-
ES for RDISXPM.13 These time trends are depicted in Figure 1A. Univariate tests indicate that the mean
values of DACCPM and RDISXPM in –1Break exceed those in Early-ES by 0.0181 and 0.0302, respectively,
significant at the 1 percent level (see Row 8). On average, DACCPM from Break is significantly less than
that from Early-ES by –0.0264, at the 1 percent level, but there is no difference between these two sub-
13 Take the mean values of DACCPM for example, they are −0.0120, −0.0187, −0.0051, 0.0061 and −0.0384 in Early-ES, Mid-ES, – 2Break, –1Break and Break, respectively. The corresponding figures for RDISXPM are −0.0148, −0.0008, 0.0069, 0.0154 and −0.0035, respectively.
16
periods for RDISXP (see Row 9). Comparisons based on median values are qualitatively similar. These
results lend preliminary support for the predictions of Hypotheses 1A, 1B, 2A and 2B.14
Panel B (C) presents the DACCPM (RDISXPM) regression results from estimating Equation 1 (2)
over the entire ES Sample of 8,234 firm-year observations. After controlling for the potential effects of
covariates, we find that the coefficient estimates on −2Break and −1Break in the DACCPM regression
are all positive and significant at the 1 percent level (one-tailed), whereas that on the Break variable is
significantly negative at the 1 percent level based on a one-tailed test (−2Break = 0.0274; t-stat. = 4.48;
−1Break = 0.0386; t-stat. = 5.65; Break = –0.0230; t-stat. = –3.05; see Panel B). While no prediction is
offered on MidES, it is also significantly positive at the 5 percent level (two-tailed). Compared to the
early part of an earnings string, firms would appear to undertake a progressively higher level of accrual
management as we move towards the end of an earnings string but significantly less in the break year, as
predicted in Hypotheses 1A and 2A. Turning next to the RDISXPM regression, we find that every pre-
break test variable once again has a positive and significant coefficient estimate and moreover both the
coefficient estimate and the associated t-statistics increase in strength from one pre-break sub-period to
the next (MidES = 0.0264, t-stat. = 1.97; −2Break = 0.0368, t-stat. = 2.25; −1Break = 0.0462, t-stat. =
2.65; see Panel C). However, unlike the DACCPM regression, the coefficient estimate on the Break
variable is now statistically insignificant (two-tailed). While firms resort to an increasingly higher level of
real activity management from the third year to the end of an earnings string, they modestly reduce such
activities in the break year, lending support for the predictions of Hypotheses 1B and 2B.
[Insert Table 4 and Figure 1A about Here]
4.4 Robustness Checks
14 As further univariate support, we also compare the actual vs. expected numbers of earnings strings whose peak level of earnings management occurs in each of the five sub-periods. Un-tabulated results indicate a higher (lower) than expected number of earnings strings reaches the peak level of accrual management in –1Break and –2Break (Break). For real activity management, we also find a higher than expected number of earnings strings reaching their peak level in –1Break and –2Break, but no difference in actual vs. expected number of peaks during the break year.
17
Recall from Panel B of Table 1 that 95 of the 1,043 ES firms in the ES Sample have two earnings strings
during the 22-year (1989–2010) sample period. On one hand, prior history of earnings strings may subject
these ES firms to closer scrutiny from the auditors and regulators, reducing their incentives for using
earnings management to sustain the second earnings string. On the other hand, one may argue that ES
firms with multiple strings likely face heavy pressure from investors to ensure that the second string does
not break, as did the previous one. To provide a “cleaner” analysis of our earnings management
hypotheses, we now replicate Equations 1 and 2 on a reduced sample of 948 firms with a single earnings
string, or equivalently 6,867 firm-year observations (labelled Subsample 1).15 As is evident in Panel A of
Table 5, all the main regression results continue to hold. In particular, the coefficient estimates on
−2Break and −1Break in both sets of regressions are significantly positive at the 5 percent level or better
(one-tailed), whereas that on the Break variable is significantly negative in the DACCPM regression only.16
In selecting the ES Sample, we have imposed a minimum length of five years on earnings strings,
but not the upper bound. Just over half of the ES Sample, or 613 out of 1,138 earnings strings, last from
six to 19 years, whereas the remaining 525 earnings strings have a length of exactly five years (see Panel
C of Table 1). This implies significant variations in the length of Mid-ES, ranging from one year for a
five-year earnings string to 15 years for a 19-year earnings string. To address the concern that our results
may be sensitive to cross-sectional variations in the length of Mid-ES, we impose a further requirement
that all earnings strings have the same length of five years, thus forcing Mid-ES to consist of exactly one
year for all firms. This requirement reduces the sample to 505 firms, or equivalently 3,150 firm-year
observations (labelled Subsample 2).17 Re-estimating Equations 1 and 2, we find that our main DACCPM
15 Subsample 1 is smaller than the original ES sample by 1,367 observations, of which 704 relate to the first earnings string and 663 to the second string. 16 The coefficient estimates (t-stat.) on MidES, −2Break, −1Break and Break in the DACCPM regression are 0.0122 (2.07), 0.0278 (4.02), 0.0404 (5.16) and –0.0202 (–2.37), respectively. The corresponding figures for the RDISXPM regression are 0.0273 (1.80), 0.0410 (2.16), 0.0479 (2.41) and 0.0235 (1.14), respectively. 17 Subsample 2 consists of 2,625 firm-year observations within an earnings string (525 strings x 5 years) and 525 observations from the break year for a total of 3,150. Among the 525 earnings strings, 20 represent the second earnings string for a particular firm. Thus, the number of distinct firms is 505 (525 – 20 = 505).
18
regression results on −2Break, −1Break and Break continue to hold18 and that the RDISXPM regression
results are weakened with none of above three test variables statistically significant at the 1 percent level
(see Panel B, Table 5).
Up till now, we have performance-matched both earnings management measures. While
matching each ES firm with a non-ES firm along the industry, year and firm performance dimensions has
the advantage of holding constant potential confounding factors that may also contribute to cross-
sectional variations in earnings management, this design assumes an equal number of ES and non-ES
firms, an assumption that is unlikely to be representative of the actual distribution of these two types of
firms in the market. To ensure that our findings remain robust, we replicate all the regression analyses
using performance-unadjusted and performance-adjusted accrual (real activity) management measures,
labelled DACC and DACCPA (RDISX and RDISXPA), respectively. We define DACC as residuals from
Equation 3 and follow Francis, LaFond, Olsson and Schipper (2005) to define DACCPA as the difference
between discretionary accruals of ES firm i and the median discretionary accruals of industry-return on
assets decile excluding firm i. RDISX and RDISXPA are defined analogously by reference to Equation 4,
where both residuals and the difference are once again multiplied by −1 to facilitate interpretations.
Results from estimating the DACC and RDISX regressions over the entire ES Sample of 8,234 firm-year
observations appear in Panel C of Table 5. The corresponding results from estimating the DACCPA and
RDISXPA regressions appear in Panel D. Consistent with the main results, the coefficient estimates on the
two key pre-break test variables, −2Break and −1Break, in all four regressions are positive and
significant at the 1 percent level (one-tailed). Moreover, the magnitude of coefficient estimate increases as
we move from MidES to −2Break before reaching their peak at −1Break.19 While the coefficient
estimates on the Break variable remain significantly positive at the 1 percent level (or insignificant) based
18 The coefficient estimates (t-stat.) on −2Break, −1Break and Break are 0.0256 (2.67), 0.0427 (4.08) and –0.0287 (–2.44), respectively. 19 Take the DACC regression for example, the coefficient estimates (t-stat.) on MidES, −2Break and −1Break are 0.0033 (1.07), 0.0219 (5.39) and 0.0445 (8.78), respectively. The corresponding figures in the RDISX regression are 0.0377 (4.12), 0.0442 (4.02) and 0.0604 (4.67), respectively.
19
on a two-tailed test in the RDISX (or RDISXPA) regression, they are nonetheless significantly negative at
the 1 percent level (one-tailed) in the DACC and DACCPA regressions.
[Insert Table 5 about Here]
4.5 Summary
Taken together, we find consistent support for the predictions of Hypotheses 1A, 1B, 2A and 2B.
Compared to the early part of an earnings string, ES firms undertake a significantly higher level of accrual
as well as real activity management, starting from the third year of the earnings string until the end. For
both measures, the intensity of earnings management peaks in the year right before the break. While ES
firms significantly reduce the extent of accrual management during the break year from the level in Early-
ES, they maintain a similar level of real activity management in the break year as that in the first two years
of the earnings string. These results extend to cases where we limit ES firms to include those with one
earnings string only or firms whose strings last exactly five years. They are also invariant to the choice of
proxies for accrual and real activity management (i.e., performance-matched vs. performance-unadjusted or
performance-adjusted).
5. Analysis of Market Response to Earnings Management Used to Sustain an Earnings String
5.1 Research Design
In this section, we test the market response to the intensity of ES firms’ earnings management activities
within an earnings string by estimating the following pooled regression model based on Barth et al.
(1999)20 over the ERC Sample of 6,575 firm-year observations for the 22-year (1989–2010) sample period:
푅푒푡푢푟푛 = 훽 + 훽 ∆퐸 + 훽 퐷퐴퐶퐶 _H (표푟 푅퐷퐼푆푋 _H) *∆퐸 + 훽 퐺푟표푤푡ℎ *∆퐸 + 훽 퐸푉퐴푅 *∆퐸 + 훽 퐷퐸 *∆퐸 + 훽 ∆퐵푉 + 훽 푌푒푎푟퐷푢푚푚푦 + 훽 푌푒푎푟퐷푢푚푚푦*∆퐸 + 휀 (5)
20 As in Barth et al. (1999), we also consider price-level and changes-in-price-level regressions. Results are qualitatively similar and hence they are not reported to conserve space.
20
where subscripts i and t, suppressed below, represent firm i in year t. Definitions and measurements for all
the variables in Equation 5 appear in Appendix 1.
Dependent Variable in Equation 5
The dependent variable in Equation 5, Return, is defined as a firm’s compound annual buy-and-hold
return over the 12-month period ending three months after the fiscal yearend minus the compound annual
buy-and-hold return of value-weighted market index over the same 12-month period.21
Test and Control Variables in Equation 5
To address the question of whether ERC varies with the level of earnings management within an earnings
string, we include an interaction term between DACCPM_H (or RDISXPM_H) and ΔE, where DACCPM_H
(or RDISXPM_H), represents ES firms with a high level of accrual (or real activity) management and it is
set equal to one if the value of DACCPM (or RDISXPM) is above the median value for the ERC Sample and
zero otherwise; ΔE is defined as changes in earnings scaled by stock price at the end of previous year. A
negative and significant coefficient estimate on the test variable, DACCPM_H *ΔE, is consistent with the
prediction of Hypothesis H3A that ERC is lower for ES firms with a high level of accrual management,
compared to those with a low level of accrual management. The prediction that ES firms undertaking
aggressive real activity management receive the same ERC as those with limited real activity
management (Hypothesis H3B) implies a statistically insignificant coefficient on RDISXPM_H *ΔE.
Following Barth et al. (1999), we also control for changes in the book value of equity per share
scaled by stock price at the end of previous year (ΔBV) and interact ΔE with three control variables, five-
year compounded annual growth rate of book value of equity (Growth), variance of the past six years’
percentage changes in earnings (EVAR) and debt-to-equity ratio (DE), in Equation 5. Finally, Year
dummies enter the equation both directly and through interaction with ΔE.
21 We calculate return over the 12-month period ending three months after the fiscal yearend and measure stock price at that time to ensure that accounting information is publicly available to investors (see Kraft, Leone and Wasley 2006; Alford, Jones and Zmijewski 1994). Results (untabulated) are qualitatively similar when we use the equal-weighted market index return or exchange-matched size decile return as the benchmark.
21
5.2 Descriptive Statistic
Panel B of Table 2 reports the distribution of model variables (other than Year dummies) for Equation 5
over the entire ERC Sample of 6,575 firm-year observations. The dependent variable, Return, is skewed to
the right, ranging from −0.9531 to 4.7440 with the mean value of 0.1329. The variables ΔE and ΔBV have
positive mean values of 0.0240 and 0.0252, respectively. On average, the ERC Sample tends to have high
growth rate of book value of equity (Growth = 0.1841), large earnings variability (EVAR = 7.1298), and
less debt than equity (DE = 0.4521).
5.3 Main Results Based on the Market Response Analysis
Rows 1–5, Panel A of Table 6 summarize the mean and median values of Return, calculated over 1,511,
2,140, 959, 987 and 978 firm-year observations from the following five sub-periods, i.e., Early-ES, Mid-
ES, – 2Break, –1Break and Break. Both mean and median values are significantly positive in the first
three sub-periods, whereas they are significantly negative in the remaining two sub-periods. Comparing
the time trend for mean values of Return, depicted in Figure 1B, with the corresponding mean values of
accrual and real activity management in Figure 1A, we find that the sharp decline in market premium
coincides with the time when the level of accrual and real activity management reaches its peak, i.e., the
year immediately preceding the break. Row 6 (7), Panel A of Table 6 presents the mean and median
values of Return in the high (low) DACCPM group, consisting of 3,287 (3,288) firm-year observations. We
find that on average the low DACCPM group has much higher Return than the high DACCPM group, i.e.,
0.1592 vs. 0.1065, with the difference of −0.0526 significant at the 1 percent level (see Row 8). By
comparison, the mean value of Return in the low RDISXPM group is marginally higher than that in the
high RDISXPM group at the 10 percent level (0.1434 vs. 0.1223, difference = −0.0211; see Rows 9–11).
We now turn to a more formal analysis by linking the level of earnings management within an
earnings string with the magnitude of market response over the same time period. Panel B1 (B2) of Table
6 presents the DACCPM (RDISXPM) regression results, estimated based on Equation 5 over the entire ERC
22
Sample of 6,575 firm-year observations. Of particular interest to us is the incremental ERC received by
ES firms in the high DACCPM group and the high RDISXPM group. After controlling for the potential
effects of covariates, the coefficient estimates (t-statistics) on the two interaction terms, DACCPM_H*ΔE
and RDISXPM_H*ΔE, are –0.5252 (–3.94) and 0.0391 (0.32), respectively. Consistent with the prediction
of Hypothesis 3A, ES firms with a high level of accrual management within an earnings string receive a
significantly lower ERC than those with a low level of accrual management based on a one-tailed test. By
comparison, ES firms with a high level of real activity management are rewarded with a statistically
similar ERC as those that undertake a low level of real management, as predicted in Hypothesis 3B.
[Insert Table 6 and Figure 1B about Here]
5.4 Robustness Checks
As before, we check for the robustness of our main results by re-estimating Equation 5 first on a subset of
the ERC Sample with one earnings string and then on firms whose earnings strings last exactly five years
(Tests 1-2).22 Finally, we replicate the analysis on the high vs. the low earnings management group,
defined separately by reference to the performance-unadjusted (i.e., DACC or RDISX) and the
performance-adjusted (i.e., DACCPA or RDISXPA) earnings management measures (Tests 3-4). Key
regression results are summarized at the lower half of Panels A–D in Table 5. Briefly, the coefficient
estimates (t-statistics) on the interaction terms involving the accrual management measure are –0.5384
(−3.95), –0.7210, (−4.29) –0.3774 (−2.50) and –0.5821 (−4.39) in Tests 1-4, respectively, all significant
at the 1 percent level (one-tailed). On the other hand, none of the interaction terms involving the real
activity management measure are significantly different from zero at the 5 percent level or better (two-
tailed). These results lend further support for the predictions of Hypotheses 3A and 3B.
5.5 Summary
22 These two subsamples consist of 854 and 449 firms, or 5,341 and 2,335 firm-year observations, respectively.
23
In short, evidence suggests that the capital market can identify ES firms with aggressive accrual
management activities by penalizing these firms with a substantially discounted market premium during
the earnings strings. However, it seems incapable of distinguishing ES firms with a high vs. a low level of
real activities management.
6. Conclusion
In this paper, we have examined the pattern of accrual and real activity management for firms reporting a
string of at least five-year consecutive earnings increases and studied the capital market response to ES
firms’ earnings management activities during an earnings string over a 22-year (1989–2010) period.
Results indicate that ES firms increase the intensity of both accrual and real activity management as they
move towards the end of an earnings string. While ES firms sharply reduce the level of accrual
management when the string finally comes to an end, they only make a very modest adjustment to real
activity management in the break year. Further analysis indicates that the market significantly discounts
the premium awarded to ES firms that undertake a high level of accrual management during an earnings
string. However, there is no evidence to suggest that the market reacts differently to ES firms with a high
vs. a low level of real activity management. Both earnings management and market reaction results
continue to hold regardless of our sample choices (i.e., multiple earnings strings of varying lengths vs.
one earnings string only or five-year earnings strings only) or how we calculate earnings management
measures (i.e., performance-matched vs. performance-unadjusted or performance-adjusted).
As directions for future research, it would be interesting to see whether the presence of stock
options has any effect on the incentive by managers of ES firms to sustain an earnings string through
earnings management and if different patterns of earnings management surrounding the break to an
earnings string would emerge when the break is accompanied by CEO changes. It would also be useful to
study the characteristics of ES firms that resort to accrual vs. real activity management. Insight from such
an analysis could help the market get around the difficulties of identifying real activity management.
24
Appendix 1 Definitions & Measurements of Variables Variables Definitions & Measurements EM Regression Model DACCPM Performance-matched discretionary accruals, defined as the difference between ES firm i’s discretionary accruals and those of a non-ES firm that has the closest ROA (return on assets, defined as net income deflated by opening assets) within the same two-digit SIC industry-year group. RDISXPM Performance-matched abnormal discretionary expenses multiplied by −1, defined as −1 times the difference between ES firm i’s abnormal discretionary expenses and those of a non-ES firm that has the closest ROA (return on assets) within the same two-digit SIC industry-year group. MidES = 1, if a firm-year observation falls in the period from the third year of an earnings string to three years before the break of an earnings string; = 0 otherwise. −2Break = 1, if a firm-year observation falls in two years before the break of an earnings string; = 0 otherwise. −1Break = 1, if a firm-year observation falls in one year before the break of an earnings string; = 0 otherwise. Break = 1, if a firm-year observation falls in the break year of an earnings string; = 0 otherwise. Size Natural logarithm of market value of equity at the beginning of the fiscal year. BTM Book-to-market ratio at the beginning of the fiscal year. Leverage Total debts at the beginning of the fiscal year scaled by opening total assets. CFO Cash flow from operations scaled by opening total assets. Loss = 1, if a firm reports a net loss in the previous year; = 0 otherwise. NewIssue = 1, if a firm issues new equity in the current period; = 0 otherwise. Litigation = 1, if a firm's business belongs to the following industries; = 0 otherwise. Biotechnology (SIC 2833–2836, 8731–8734), Computer (SIC 3570–3577, 7370–7374), Electronics (SIC 3600–3674), and Retailing (SIC 5200–5961). BigN = 1, if a firm retains a Big-N auditor; = 0 otherwise. IndustryDummy 37 dummies for two-digit SIC industry group YearDummy 21 dummies for fiscal year
(The appendix is continued on the next page.)
25
Appendix 1 (Continued) Variables Definitions & Measurements Accrual Management Estimation Model (See Appendix 2) TAC Total accruals (income before extraordinary items minus cash flow from operations). TA Total assets. ΔS Change in sales. ΔREC Change in account receivables from trade. k Estimated slope coefficient from a regression of ΔREC/TA on ΔS/TA for each two-digit SIC industry-year group. PPE Property, Plant & Equipment (Gross). S Sales (net). DACC Discretionary accruals, residuals in the forward-looking modified-Jones model. Real Management Estimation Model (See Appendix 3) DIS Discretionary expenses; the sum of advertising, R&D (research & development), and SG&A
expenses (selling, general & administration expenses); at least one of these three expenses
is available, other remaining expenses are set to zero if they are missing. DISX Abnormal discretionary expenses; residuals in the normal discretionary expense regression
model based on Roychowdhury (2006).
Return Regression Model
Ret Market-adjusted return (a firm’s compound annual buy-and-hold return over the 12-month period ending three months after the fiscal yearend minus the compound annual return of value-weighted market index over the same 12-month period). ΔE Change in earnings per share scaled by stock price at the end of previous year, (EPSt − EPSt-1) / Pt-1. DACCPM_H = 1, if a firm’s DACCPM is above the median of DACCPM distribution; = 0 otherwise; RDISXPM_H = 1, if a firm’s RDISXPM is above the median of RDISXPM distribution; = 0otherwise;
(The appendix is continued on the next page.)
26
Appendix 1 (Continued) Variables Definitions & Measurements Growth Five-year compounded annual growth rate of book value of equity, (BVEt / BVEt-5)1/5. EVAR Variance of the past six years’ percentage change in earnings, where the percentage of change in earnings is (EPSt − EPSt-1) / abs EPSt-1. DE Debt-to-equity ratio, (current debt + long-term debt)/shareholder’s equity. ΔBV Change in book value of equity per share (BV) scaled by stock price at the end of previous year, (BVt − BVt-1) / Pt-1. YearDummy 21 dummies for fiscal year.
Robustness Checks DACC Discretionary accruals (i.e., residuals) estimated from the forward-looking modified-Jones model. RDISX Abnormal discretionary expenses multiplied by −1, defined as −1 times the residuals from the normal discretionary expense regression in Roychowdhury (2006). DACCPA Performance-adjusted discretionary accruals, defined as the difference between an ES firm i’s discretionary accruals and the median discretionary accruals of its industry-ROA decile, where the median calculation excludes firm i. RDISXPA Performance-adjusted abnormal discretionary expenses multiplied by −1, calculated as −1 times the difference between an ES firm i’s abnormal discretionary expenses and the median abnormal discretionary expenses of the industry-ROA decile, where the median calculation excludes firm i. DACC_H = 1, if a firm’s DACC is above the median of DACC distribution; = 0 otherwise. RDISX_H = 1, if a firm’s RDISX is above the median of RDISX distribution; = 0 otherwise. DACCPA_H = 1, if a firm’s DACCPA is above the median of DACCPA distribution; = 0 otherwise. RDISXPA_H = 1, if a firm’s RDISXPA is above the median of RDISXPA distribution; = 0 otherwise.
27
Appendix 2 Estimation of Discretionary Accruals To calculate discretionary accruals, we follow Dechow, Richardson and Tuna (2003) and estimate the following forward-looking modified-Jones model cross-sectionally for every two-digit SIC industry-year group with at least 20 observations over 1989–2010:
= 훼 + 훼 ( )∆ ∆ + 훼 + 훼 + 훼 ∆ + 휀 .
All variables are winsorized at the top and bottom 1 percent of their respective distributions following the convention of prior literature. The residuals represent discretionary accruals (denoted as DACC). A summary of the mean value of each estimated coefficient across 770 industry-years, along with t-statistics calculated using the standard error of the mean across 770 industry-years, is presented below. The adjusted R2 is the mean adjusted R2 over 770 industry-years.
Mean Value of Estimated Coefficient (t-statistics)
1푇퐴 −0.3574
(−5.82)*** (1 + 푘)∆푆 − ∆푅퐸퐶
푇퐴
0.0184 (5.25)***
푃푃퐸푇퐴
−0.0683 (−40.29)***
푇퐴퐶푇퐴
0.1957 (21.01)***
∆푆푆 0.0244
(5.61)***
Mean Adjusted R2 (in %) = 45.61%
Notes: *, **, and *** represent significance levels of 0.1, 0.05, and 0.01 (two-tailed), respectively. Please refer to Appendix 1 for variable definitions and measurements.
28
Appendix 3 Estimation of Abnormal Discretionary Expenses To calculate abnormal discretionary expenses, we follow Roychowdhury (2006) and estimate the following regression model cross-sectionally for every two-digit SIC industry-year group with at least 20 observations over 1989–2010:
= 훼 + 훼 + 훼 + 휀
All variables are winsorized at the top and bottom 1 percent of their respective distributions. The residuals represent abnormal discretionary expenses (denoted as DISX). A summary of the mean value of each estimated coefficient across 770 industry-years, along with t-statistics calculated using the standard error of the mean across 770 industry-years, is presented below. The adjusted R2 is the mean adjusted R2 over 770 industry-years.
Mean Value of Estimated Coefficient (t-statistics)
Intercept 0.1498 (22.56)***
1푇퐴 1.6268
(12.53)*** 푆
푇퐴 0.1365 (29.19)***
Mean Adjusted R2 (in %) = 38.39%
Notes: *, **, and *** represent significance levels of 0.1, 0.05, and 0.01 (two-tailed), respectively. Please refer to Appendix 1 for variable definitions and measurements.
29
Reference
Alford, A. W., J. J. Jones, and M. E. Zmijewski. 1994. Extensions and Violations of the Statutory SEC Form 10-K Filing Requirements. Journal of Accounting and Economics 17 (1–2): 229–254.
Ali, A. and U. G. Gurun. 2009. Investor Sentiment, Accruals Anomaly and Accruals Management.
Journal of Accounting, Auditing and Finance 24 (3): 415 – 431. Baber, W. R., S. Chen, and S. H. Kang. 2006. Stock Price Reaction to Evidence of Earnings Management:
Implications for Supplementary Financial Disclosure. Review of Accounting Studies 11 (1): 5–19. Baik, B., D. B. Farber, M. F. Johnson, and H. Yi. 2008. What Determines an Earning String? Working
paper, Seoul National University. Ball. R., and L. Shivakumar. 2008. Earnings Quality at Initial Public Offerings. Journal of Accounting
and Economics 45 (2–3): 324–349. Balsam, S., E. Bartov, and C. Marquardt. 2002. Accruals Management, Investor Sophistication, and
Equity Valuation: Evidence from 10-Q Filings. Journal of Accounting Research 40 (4): 987–1011. Barth, M. E., J. A. Elliott, and M. W. Finn. 1999. Market Rewards Associated with Patterns of Increasing
Earnings. Journal of Accounting Research 37 (2): 387–413. Barton, J., and P. J. Simko. 2002. The Balance Sheet as an Earnings Management Constraint. The
Accounting Review 77 (Quality of Earnings Conference): 1–27. Burgstahler, D. C., and I. Dichev. 1997. Earnings Management to Avoid Earnings Decreases and Losses.
Journal of Accounting and Economics 24 (1): 99–126. Cohen, D. A., A. Dey, and T. Z. Lys. 2008. Real and Accrual-Based Earnings Management in the Pre-
and Post-Sarbanes-Oxley Periods. The Accounting Review 83 (3): 757–787. Cohen, D. A., and P. Zarowin. 2010. Accrual-Based and Real Earnings Management Activities around
Seasoned Equity Offerings. Journal of Accounting & Economics 50 (1): 2–19. Dechow, P. M. 1994. Accounting Earnings and Cash Flows as Measures of Firm Performance. The Role
of Accounting Accruals. Journal of Accounting and Economics 18 (1): 3–42. Dechow, P. M., R. G. Sloan, and A. P. Sweeny. 1995. Detecting Earnings Management. The Accounting
Review 70 (2): 193–225. Dechow, P. M., R. G. Sloan, and A. P. Sweeny. 1996. Causes and Consequences of Earnings
Manipulation: An Analysis of Firms subject to Enforcement Actions by the SEC. Contemporary Accounting Research 13 (1): 1–36.
Dechow, P. M., S. A. Richardson, and I. Tuna. 2003. Why Are Earnings Kinky? An Examination of the
Earnings Management Explanation. Review of Accounting Studies 8 (2–3): 355–384. DeFond, M. L., and C. W. Park. 2001. The Reversal of Abnormal Accruals and the Market Valuation of
Earnings Surprises. The Accounting Review 76 (3): 375–404.
30
Dopuch, N., C. Seethamraju, and W. Xu. 2010. The Pricing of Accruals for Profit and Loss Firms. Review of Quantitative Finance & Accounting 34: 505–516.
Drake, M. S., J. N. Myers, and L. A. Myers. 2009. Disclosure Quality and the Mispricing of Accruals and
Cash Flow. Journal of Accounting, Auditing and Finance 24 (3): 357–284. Erickson, M., and S. Wang. 1999. Earnings Management by Acquiring Firms in Stock for Stock Mergers.
Journal of Accounting and Economics 27 (2): 149–176. Francis, J., D. Philbrick, and K. Schipper. 1994. Shareholder Litigation and Corporate Disclosures.
Journal of Accounting Research 32 (2): 137–164. Francis, J., E. Maydew, and H. Sparks. 1999. The Role of Big Six Auditors in the Credible Reporting of
Accruals. Auditing: A Journal of Practice and Theory 18 (2): 17–34. Francis, J., R. LaFond, P. Olsson, and K. Schipper. 2005. The Market Pricing of Accruals Quality.
Journal of Accounting and Economics 39 (2): 295–327. Ghosh, A., Z. Gu, and P. C. Jain. 2005. Sustained Earnings and Revenue Growth, Earnings Quality, and
Earnings Response Coefficients. Review of Accounting Studies 10 (1): 33–57. Graham, J. R., C. R. Harvey, and S. Rajgopal. 2005. The Economic Implications of Corporate Financial
Reporting. Journal of Accounting and Economics 40 (1–3): 3–73. Green, J., J.M. Hand, and M. Soliman. 2011. Going, Going, Gone? The Apparent Demise of the Accruals
Anomaly. Management Science 57 (5): 797–816. Gunny, K. A. 2010. The Relation between Earnings Management Using Real Activities Manipulation and
Future Performance: Evidence from Meeting Earnings Benchmarks. Contemporary Accounting Research 27 (3): 855–888.
Healy, P. M., and J. M. Wahlen. 1999. A Review of the Earnings Management Literature and its
Implications for Standard Setting. Accounting Horizons 13 (4): 365–383. Hribar, P., and D. W. Collins. 2002. Errors in Estimating Accruals: Implications for Empirical Research.
Journal of Accounting Research 40 (1): 105–134. Ke, B. 2004. Do Equity-Base Incentive Induce CEOs to Manage Earnings to Report String of
Consecutive Earnings Increases? Working paper, Pennsylvania State University, PA. Ke, B., and K. Petroni. 2004. How Informed Are Actively Trading Institutional Investors? Evidence from
Their Trading Behavior before a Break in a String of Consecutive Earnings Increases. Journal of Accounting Research 42 (5): 895–927.
Ke, B., S. Huddart, and K. Petroni. 2003. What Insiders Know about Future Earnings and How They Use
It: Evidence from Insider Trades. Journal of Accounting and Economics 35 (3): 315–346. Kothari, S. P., A. J. Leone, and C. E. Wasley. 2005. Performance Matched Discretionary Accrual
Measures. Journal of Accounting and Economics 39 (1): 163–197.
31
Kraft, A., A. J. Leone, and C. Wasley. 2006. An Analysis of the Theories and Explanations Offered for the Mispricing of Accruals and Accrual Components. Journal of Accounting Research 44 (2): 297–339.
Lev, B., S. G. Ryan, and M. Wu. 2008. Rewriting Earnings History. Review of Accounting Studies 13 (4):
419–451. Levi, S. 2008. Voluntary Disclosure of Accruals in Earnings Press Releases and the Pricing of Accruals.
Review of Accounting Studies 13 (1): 1–21. Lim, C. Y., and H. T. Tan. 2008. Non-audit Service Fees and Audit Quality: The Impact of Auditor
Specialization. Journal of Accounting Research 46 (1): 199–246. Louis, H., D. Robinson, and A. Sbaraglia. 2008. An Integrated Analysis of the Association between
Accrual Disclosure and the Abnormal Accrual Anomaly. Review of Accounting Studies 13 (1): 23–54.
Myers, J. N., L. A. Myers, and D. J. Skinner. 2007. Earnings Momentum and Earnings Management.
Journal of Accounting, Auditing and Finance 22 (2): 249–284. Petersen, M. A. 2009. Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches.
Review of Financial Studies 22 (1): 435–480. Porter, M. E. 1985. Competitive Advantage: Creating and Sustaining Superior Performance. New York:
The Free Press. Richardson, S., I. Tuna, and M. Wu. 2002. Predicting Earnings Management: The Case of Earnings
Restatements. Working paper, University of Pennsylvania. Roychowdhury, S. 2006. Earnings Management through Real Activities Manipulation. Journal of
Accounting and Economics 42 (3): 335–370. Sloan, R. G. 1996. Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future
Earnings. The Accounting Review, 71 (3): 289–315. Teoh, S. H., I. Welch, and T. J. Wong. 1998. Earnings Management and the Underperformance of
Seasoned Equity Offerings. Journal of Financial Economics 50 (1): 63–99. Xie, H. 2001. The Mispricing of Abnormal Accruals. The Accounting Review, 76 (3): 357–373. Yong, K. O. 2009. Earnings Breaks and Earnings Management. Working paper, Singapore Management
University. Zang, A. Y. 2012. Evidence on the Trade-off between Real Activities Manipulation and Accrual-Based
Earnings Management. The Accounting Review, 87 (2): 675–703.
32
TABLE 1 Sample and Data
Panel A: Sample Selection Procedure
Number of
Earnings Strings Number of
Observations Available observations for calculating earnings stringsa
and model variables in EM analysis (1989–2010) b 85,576
Total number of earnings strings (ES)/observations 1,417 8,932 Less: Earnings strings that continue beyond 2010c (111) (840) Less: Earnings strings without an identifiable break yeard (168) (996)
Number of earnings strings/observations with clearly identifiable break year (excluding the break year) 1,138 7,096
Number of observations in the break yeare 1,138
ES sample (1989–2010) for EM analysis 1,138 8,234
ERC sample (1989–2010) for Return analysise,f 6,575 Notes: a An earnings string (ES) is defined as a string of increases in annual EPS (split-adjusted) for at least five years. b The initial sample is obtained through the following filters from the COMPUSTAT Fundamental Annual Database
between 1987 and 2011: (1) the financial and regulated industries are deleted, (2) any missing data for calculating variables in earnings management analysis are excluded, (3) after trimming earnings changes in the extreme 1 percent of distribution, any two-digit SIC industry-year group with less than 20 observations are also deleted. Since firms must have two-year lag data as well as one-year lead data to calculate model variables and earnings strings in earnings management analysis, the final sample period do not include observations from 1987, 1988, and 2011.
c Earnings strings are deleted if they continue beyond the final year of the sample period (2010) since the break years of those strings are unknown.
d Based on DLRSN of COMPUSTAT, 93 firms were merged or acquired and 3 firms were bankrupt or liquidated. 3 firms became private companies while 66 companies no longer file with SEC for other reasons or without reasons. Also, if the first string of a firm is deleted due to missing break year, then its second string is also excluded (3 firms).
e The break year is included in the original ES sample to test the prediction of Hypotheses 2A and 2B. f Data for the ERC Sample are collected from the COMPUSTAT Fundamental Annual Database and the CRSP
Monthly Stock Database (raw returns and stock price data) for each firm in the ES sample. Before calculating model variables in the Return regression, the extreme 1% of return and stock price distributions are eliminated. Among 8,234 ES sample, only 6,575 observations (or 949 firms) have complete data for calculating model variables in the Return regression
(The table is continued on the next page.)
33
TABLE 1 (Continued)
Panel B: Composition of the ES Sample (1989–2010)
Number Percentage Number of 1st earnings string 1,043 91.65%
Number of 2nd earnings string 95 8.35%
Total number of earnings strings 1,138 100.00%
Number of distinct firms with one earnings string 948 90.89%
Number of distinct firms with two earnings strings 95 9.11%
Total number of distinct firms with an earnings string 1,043 100.00%
Panel C: Frequency Distribution of the ES Sample by the Length of Earnings Strings (1989–2010)
Length of Earnings Strings Frequency Percentage 5 525 46.13% 6 288 25.31% 7 131 11.51% 8 80 7.03% 9 55 4.83% 10 23 2.02% 11 12 1.05% 12 10 0.88% 13 5 0.44%
14 3 0.26% 15 2 0.18% 16 1 0.09% 17 0 0.00% 18 1 0.09% 19 2 0.18%
Total 1,138 100.00%
(The table is continued on the next page.)
34
TABLE 1 (Continued)
Panel D: Frequency Distribution of the ES Sample by Industry (1989–2010)
Industry ES Sample
Compustat Populationa
(two-digit SIC code) # of Firms
# of Observations
# of Observations
N %
N %
N %
Mining & Construction 6 0.58%
39 0.47%
1,569 1.83% (01, 10-12,14-19)
Oil & Gas (13,29) 47 4.51%
314 3.81%
5,725 6.69% Food products (20-21) 41 3.93% 344 4.18% 2,551 2.98% Texitiles (22-27) 70 6.71%
544 6.61%
5,251 6.14%
Chemical products (28) 100 9.59%
839 10.19%
9,417 11.00% Manufacturing (30-34) 63 6.04%
491 5.96%
5,286 6.18%
Computer equipment (35) 83 7.96%
706 8.57%
6,715 7.85%
Electronic equipment (36) 111 10.64%
843 10.24%
9,223 10.78% Transportation (37,39-43) 47 4.51% 384 4.66% 3,662 4.28% Scientific instruments (38) 100 9.59%
782 9.50%
7,153 8.36%
Retail (50-59) 172 16.49%
1,428 17.34%
10,512 12.28% Services (70-89) 196 18.79%
1,471 17.86%
17,408 20.34%
Other (90-99) 7 0.67%
49 0.60%
1,104 1.29%
Total 1,043 100.00%
8,234 100.00%
85,576 100.00% Note: a This population is based on Panel A of Table 1.
(The table is continued on the next page.)
35
TABLE 1 (Continued)
Panel E: Frequency Distribution of the ES Sample by the Timing of Break Year (1989–2010)
Year Frequency Percentage 1994 36 3.16% 1995 43 3.78% 1996 59 5.18% 1997 61 5.36% 1998 83 7.29% 1999 73 6.41% 2000 75 6.59% 2001 94 8.26% 2002 28 2.46% 2003 40 3.51% 2004 43 3.78% 2005 49 4.31% 2006 72 6.33% 2007 91 8.00% 2008 149 13.09% 2009 109 9.58% 2010 33 2.90% Total 1,138 100.00%
Note: The distributions in Panels B-E are qualitatively similar for the ERC Sample and they are not reported to conserve space.
36
TABLE 2 Distribution of Model Variables Panel A: The ES Sample (1989–2010; N = 8,234 firm-year observations)
Mean Std Dev Min 25% 50% 75% Max N
DACCPM −0.0143 0.1719 −1.8450 −0.0849 −0.0096 0.0587 1.4115 8,234 RDISXPM −0.0017 0.4248 −3.7334 −0.1902 −0.0031 0.1892 3.9109 8,234 Size 5.9940 2.3602 0.6672 4.2906 6.0037 7.6165 11.5930 8,234 BTM 0.4434 0.4062 −0.5754 0.2097 0.3489 0.5660 2.2390 8,234 Leverage 0.1835 0.1824 0.0000 0.0181 0.1491 0.2869 0.9290 8,234 CFO 0.1090 0.1839 −0.8412 0.0660 0.1289 0.1947 0.5111 8,234 Loss 0.1647 0.3709 0.0000 0.0000 0.0000 0.0000 1.0000 8,234 NewIssue 0.8528 0.3543 0.0000 1.0000 1.0000 1.0000 1.0000 8,234 Litigation 0.3594 0.4798 0.0000 0.0000 0.0000 1.0000 1.0000 8,234 BigN 0.8549 0.3523 0.0000 1.0000 1.0000 1.0000 1.0000 8,234
Panel B: The ERC Sample (1989–2010; N = 6,575 firm-year observations)
Mean Std Dev Min 25% 50% 75% Max N
Return 0.1329 0.4879 −0.9531 −0.1628 0.0558 0.3205 4.7440 6,575 ΔE 0.0240 0.1097 −0.1450 0.0016 0.0054 0.0149 0.9100 6,575 ΔBV 0.0252 0.0999 −0.3953 −0.0039 0.0394 0.0718 0.3057 6,575 Growth 0.1841 0.2089 −0.2550 0.0599 0.1450 0.2637 1.0359 6,575 EVAR 7.1298 20.5111 0.0000 0.0432 0.2915 2.5972 100.0000 6,575 DE 0.4521 0.6116 0.0000 0.0268 0.2630 0.6089 3.5800 6,575
Notes: In Panel A, all continuous variables, except for dependent variables in the earnings management analysis (DACCPM and RDISXPM), are winsorized at the top and the bottom 1 percent of their respective distributions. In Panel B, all continuous variables, except for the dependent variable in the market reaction analysis (Return) and EVAR, are winsorized at the top and the bottom 1 percent of the distribution. EVAR is winsorized to 100 (see Barth et al. 1999). Please refer to Appendix 1 for variable definitions and measurements.
37
TABLE 3 Correlation Matrix among Model Variables for the ES Sample (1989–2010; N = 8,234 firm-year observations)
Panel A: DACCPM Regression
DACCPM Size BTM Leverage CFO Loss NewIssue Litigation BigN
DACCPM 1.0000 −0.1143 0.0154 0.0278 −0.3333 0.1145 −0.0183 0.0077 −0.0593
<.0001 0.1615 0.0116 <.0001 <.0001 0.0963 0.4863 <.0001
Size −0.1114
1.0000 −0.3444 −0.0297 0.3238 −0.3832 0.1787 0.0119 0.3670
<.0001 <.0001 0.0070 <.0001 <.0001 <.0001 0.2813 <.0001
BTM 0.0458 −0.3303
1.0000 −0.0386 −0.0304 0.0670 −0.2447 −0.1128 −0.0349
<.0001 <.0001 0.0005 0.0057 <.0001 <.0001 <.0001 0.0016
Leverage 0.0345 0.0593 0.0980
1.0000 −0.1221 0.1112 −0.0898 −0.1394 −0.0067
0.0017 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 0.5439
CFO −0.3597 0.2997 −0.2388 −0.1879
1.0000 −0.5296 0.0099 −0.0211 0.1376
<.0001 <.0001 <.0001 <.0001 <.0001 0.3675 0.0552 <.0001
Loss 0.0971 −0.3775 −0.0235 0.0404 -0.4271
1.0000 −0.0771 0.1131 −0.1415
<.0001 <.0001 0.0328 0.0002 <.0001 <.0001 <.0001 <.0001
NewIssue −0.0166 0.1685 −0.1863 −0.0722 0.0533 -0.0771
1.0000 0.1154 0.0964
0.1320 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001
Litigation 0.0021 0.0092 −0.1712 −0.1807 0.0531 0.1131 0.1154
1.0000 −0.0033
0.8465 0.4037 <.0001 <.0001 <.0001 <.0001 <.0001 0.7662
BigN −0.0595 0.3561 −0.0052 0.0492 0.1175 −0.1415 0.0964 −0.0033 1.0000 <.0001 <.0001 0.6384 <.0001 <.0001 <.0001 <.0001 0.7662
(The table is continued on the next page.)
38
TABLE 3 (Continued)
Panel B: RDISXPM Regression
RDISXPM Size BTM Leverage NewIssue
RDISXPM 1.0000 0.0173 0.0520 0.0495 −0.0520 0.1176 <.0001 <.0001 <.0001
Size 0.0141
1.0000 −0.3444 −0.0297 0.1787
0.2019 <.0001 0.0070 <.0001
BTM 0.0896 −0.3303
1.0000 −0.0386 −0.2447
<.0001 <.0001 0.0005 <.0001
Leverage 0.0598 0.0593 0.0980
1.0000 −0.0898
<.0001 <.0001 <.0001 <.0001
NewIssue −0.0547 0.1685 −0.1863 −0.0722
1.0000 <.0001 <.0001 <.0001 <.0001 Notes: Pearson correlation coefficients are reported above the diagonal and Spearman rank correlation coefficients are reported below the diagonal. The corresponding p-values appear underneath the correlation coefficients. Please refer to Appendix 1 for variable definitions and measurements.
39
TABLE 4 Results from the Earnings Management Analysis (1989–2010; N = 8,234 firm-year observations)
Panel A: Mean and Median Values of Earnings Management Measures along an Earnings String
(1). DACCPM (2). RDISXPM N (1a). Mean (1b). Median (2a). Mean (2b). Median
(1) Early-ES −0.0120 −0.0129 −0.0148 −0.0221 2,276 (2) Mid-ES −0.0187 −0.0122 −0.0008 0.0040 2,544
(3) −2Break −0.0051 −0.0006 0.0069 0.0047 1,138 (4) −1Break 0.0061 0.0029 0.0154 0.0112 1,138
(5) Break −0.0384 −0.0222 −0.0035 −0.0122 1,138
Total N 8,234
Hypotheses 1A and 1B
(6) (2) − (1) −0.0067 0.0007 0.0140 0.0261 * (7) (3) − (1) 0.0069 0.0124 ** 0.0216 0.0268 **
(8) (4) − (1) 0.0181 *** 0.0159 *** 0.0302 ** 0.0333 ***
Hypotheses 2A and 2B
(9) (5) − (1) −0.0264 *** −0.0092 *** 0.0113 0.0100
(The table is continued on the next page.)
40
TABLE 4 (Continued)
Panel B: DACCPM Regression Results
퐷퐴퐶퐶 = 훽 + 훽 푀푖푑퐸푆 + 훽 − 2퐵푟푒푎푘 + 훽 − 1퐵푟푒푎푘 + 훽 퐵푟푒푎푘 + 훾 푆푖푧푒 +훾 퐵푇푀
+훾 퐿푒푣푒푟푎푔푒 + 훾 퐶퐹푂 + 훾 퐿표푠푠 + 훾 푁푒푤퐼푠푠푢푒 + 훾 퐿푖푡푖푔푎푡푖표푛 + 훾 퐵푖푔푁
+훾 퐼푛푑푢푠푡푟푦퐷푢푚푚푦 + 훾 푌푒푎푟퐷푢푚푚푦 +휀
Prediction Coefficient t-statistics p-value Test Variable
MidES ? 0.0113 2.20 0.0283 −2Break + 0.0274 4.48 <.0001 −1Break + 0.0386 5.65 <.0001
Break − −0.0230 −3.05 0.0012 Control Variable
Size −0.0016 −1.35 0.1758 BTM 0.0027 0.40 0.6920
Leverage −0.0168 −1.08 0.2825 CFO −0.3793 −12.66 <.0001 Loss −0.0428 −5.67 <.0001
NewIssue −0.0062 −1.19 0.2350 Litigation −0.0181 −2.77 0.0058
BigN −0.0065 −0.94 0.3452 YearDummy Yes
IndustryDummy Yes
Adj. R2 13.38% N 8,234 obs. (1,043 firms)
(The table is continued on the next page.)
41
TABLE 4 (Continued)
Panel C: RDISXPM Regression Results
푅퐷퐼푆푋 = 훽 + 훽 푀푖푑퐸푆 + 훽 − 2퐵푟푒푎푘 + 훽 − 1퐵푟푒푎푘 + 훽 퐵푟푒푎푘 +훾 푆푖푧푒 + 훾 퐵푇푀
+훾 퐿푒푣푒푟푎푔푒 + 훾 푁푒푤퐼푠푠푢푒 + 훾 퐼푛푑푢푠푡푟푦퐷푢푚푚푦 + 훾 푌푒푎푟퐷푢푚푚푦 +휀
Prediction Coefficient t-statistics p-value Test Variable
MidES ? 0.0264 1.97 0.0486 −2Break + 0.0368 2.25 0.0125 −1Break + 0.0462 2.65 0.0041
Break insignificant/+ 0.0213 1.19 0.2349 Control Variable
Size 0.0162 3.21 0.0014 BTM 0.0993 4.20 <.0001
Leverage 0.1760 3.70 0.0002 NewIssue −0.0674 −3.71 0.0002
YearDummy Yes IndustryDummy Yes
Adj. R2 2.47%
N 8,234 obs. (1,043 firms) Notes: In Panels A–C, DACCPM and RDISXPM represent performance-matched discretionary accruals and performance-matched abnormal discretionary expenses multiplied by −1, respectively. DACCPM (RDISXPM) is defined as the difference (−1 times the difference) between an ES sample firm i’s discretionary accruals (abnormal discretionary expenses) and the discretionary accruals (abnormal discretionary expenses) of a non-ES firm that has the closest return on asset within the same industry-year group. In Panel A, Early-ES, Mid-ES, −2Break, −1Break, and Break represent sub-periods, defined as the first two years of an earnings string, the period from the third year to three years before the break, two years before the break, one year before the break and the year immediately following the end of an earnings string, respectively. Mean and median comparisons are based on t-tests and Wilcoxon tests, respectively. *, **, and *** represent significance levels of 0.1, 0.05, and 0.01 (two-tailed), respectively. In Panels B–C, MidES, −2Break, −1Break, and Break are the test variables, representing the period from the third year of an earnings string to three years before the break, two years before the break, one year before the break and the break year of an earnings string, respectively. All t-values are reported using robust standard errors to correct heteroskedasticity problem and firm clustering effect. The p-values are one-tailed if there is a prediction, and two-tailed otherwise. Please refer to Appendix 1 for the definitions and measurements of all the control variables.
42
TABLE 5 Results based on Robustness Checks
Panel A. Subsample 1 (One Earnings String Only) Panel B. Subsample 2 (Five-Year Earnings String Only)
(1). EM = DACCPM (2). EM = RDISXPM (1). EM = DACCPM (2). EM = RDISXPM Test Variables Coeff. t-stat p-value Coeff. t-stat p-value Coeff. t-stat p-value Coeff. t-stat p-value
EM Regressions MidES 0.0122 2.07 0.0385 0.0273 1.80 0.0724 0.0142 1.50 0.1331 −0.0512 −2.51 0.0124
−2Break 0.0278 4.02 <.0001 0.0410 2.16 0.0153 0.0256 2.67 0.0040 −0.0087 −0.38 0.3509 −1Break 0.0404 5.16 <.0001 0.0479 2.41 0.0081 0.0427 4.08 <.0001 0.0136 0.57 0.2839
Break −0.0202 −2.37 0.0090 0.0235 1.14 0.2549 −0.0287 −2.44 0.0075 0.0029 0.11 0.9120 Adj. R2 12.99% 2.73% 13.13% 2.79%
N 6,867 obs. (948 firms) 6,867 obs. (948 firms) 3,150 obs. (505 firms) 3,150 obs. (505 firms) Return Regression DACCPM_H* ΔE −0.5384 −3.95 <.0001 n/a n/a n/a −0.7210 −4.29 <.0001 n/a n/a n/a RDISXPM_H* ΔE n/a n/a n/a 0.0326 0.26 0.7955 n/a n/a n/a 0.1747 1.19 0.2337
Adj. R2 9.35% 9.01% 9.60% 8.91% N 5,341 obs. (854 firms) 5,341 obs. (854 firms) 2,335 obs. (449 firms) 2,335 obs. (449 firms)
Notes: In Panel A, Subsample 1 consists of ES firms with only one earnings string (see Panel B of Table 1). In Panel B, Subsample 2 consists of ES firms with an earnings string that lasted exactly five years (see Panel C of Table 1). Columns 1 and 2 of each panel show the regression results based on performance-matched discretionary accruals (DACCPM) and performance-matched abnormal discretionary expenses multiplied by −1 (RDISXPM), respectively. DACCPM (RDISXPM) is defined as the difference (−1 times the difference) between an ES sample firm i’s discretionary accruals (abnormal discretionary expenses) and the discretionary accruals (abnormal discretionary expenses) of a non-ES firm that has the closest return on asset within the same industry-year group. In the EM regression, MidES, −2Break, −1Break, and Break are the test variables, denoting the period from the third year of an earnings string to three years before the break, two years before the break, one year before the break and the break year of an earnings string, respectively. In the Return regression, DACCPM_H (RDISXPM_H) is an indicator variable which is set equal to 1 if a firm’s DACCPM (RDISXPM) is above the median of DACCPM (RDISXPM) distribution and zero otherwise. ΔE represents change in earnings per share scaled by stock price at the end of previous year. Results on control variables are not reported to conserve space. All t-values are reported using robust standard errors to correct heteroskedasticity problem and firm clustering effect. The p-values are one-tailed if there is a prediction, and two-tailed otherwise.
(The table is continued on the next page.)
43
TABLE 5 (Continued) Panel C. Performance-Unadjusted EM Measures Panel D. Performance-Adjusted EM Measures
Test Variables (1). EM = DACC (1). EM = RDISX (1). EM = DACCPA (1). EM = RDISXPA Coeff. t-stat p-value Coeff. t-stat p-value Coeff. t-stat p-value Coeff. t-stat p-value
EM Regressions MidES 0.0033 1.07 0.2871 0.0377 4.12 <.0001 0.0027 0.92 0.3580 0.0304 3.39 0.0007
−2Break 0.0219 5.39 <.0001 0.0442 4.02 <.0001 0.0196 5.28 <.0001 0.0354 3.17 0.0008 −1Break 0.0445 8.78 <.0001 0.0604 4.67 <.0001 0.0388 8.28 <.0001 0.0490 3.80 0.0001
Break −0.0296 −5.32 <.0001 0.0571 4.27 <.0001 −0.0264 −5.21 <.0001 0.0208 1.60 0.1090 Adj. R2 8.78% 10.76% 20.71'% 6.77%
N 8,234 obs. (1,043 firms) 8,234 obs. (1,043 firms) 8,234 obs. (1,043 firms) 8,234 obs. (1,043 firms) Return Regression
DACC_H*ΔE −0.3774 −2.50 0.0062 n/a n/a n/a n/a n/a n/a n/a n/a n/a RDISX_H*ΔE n/a n/a n/a 0.1948 1.25 0.2133 n/a n/a n/a n/a n/a n/a
DACCPA_H*ΔE n/a n/a n/a n/a n/a n/a −0.5821 −4.39 <.0001 n/a n/a n/a RDISXPA_H*ΔE n/a n/a n/a n/a n/a n/a n/a n/a n/a 0.2388 1.57 0.1173
Adj. R2 8.43% 8.31% 8.62% 8.33% N 6,575 obs. (949 firms) 6,575 obs. (949 firms) 6,575 obs. (949 firms) 6,575 obs. (949 firms)
Notes: In Panel C, DACC (RDISX) represents performance-unadjusted discretionary accruals (abnormal discretionary expenses multiplied by −1), defined as the residuals (−1 times the residuals) from Equation 3 (Equation 4). In Panel D, DACCPA (RDISXPA) denotes performance-adjusted discretionary accruals (abnormal discretionary expenses multiplied by −1), calculated as the difference (−1 times the difference) between an ES sample firm i’s discretionary accruals (abnormal discretionary expenses) and the median discretionary accruals (abnormal discretionary expenses) for its industry-ROA decile excluding firm i. In the EM regression, MidES, −2Break, −1Break, and Break are the test variables, denoting the period from the third year of an earnings string to three years before the break, two years before the break, one year before the break and the break year of an earnings string, respectively. In the Return regression, DACC_H (DACC PA_H) is an indicator variable, set equal to 1 if a firm’s DACC (DACC PA) is above the median of DACC (DACC PA) distribution and zero otherwise. RDISX_H (RDISX PA_H) is defined analogously. ΔE represents change in earnings per share scaled by stock price at the end of previous year. Results on control variables are not reported to conserve space. All t-values are reported using robust standard errors to correct heteroskedasticity problem and firm clustering effect. The p-values are one-tailed if there is a prediction, and two-tailed otherwise.
44
TABLE 6 Results from the Return Analysis (1989–2010; N = 6,575 firm-year observations)
Panel A: Mean and Median Values of Return Along the five Sub-periods of an Earnings String
Return N
Mean Median (1) Early-ES 0.2419 *** 0.1451 *** 1,511 (2) Mid-ES 0.2030 *** 0.1111 *** 2,140 (3) −2Break 0.1606 *** 0.0857 *** 959 (4) −1Break −0.0331 *** −0.0784 *** 987 (5) Break −0.0488 *** −0.0689 *** 978
Total N 6,575 Hypothesis 3A
(6) High DACCPM Group 0.1065 0.0305 3,287 (7) Low DACCPM Group 0.1592 0.0817 3,288 (8) (6) − (7) −0.0526 *** −0.0513 ***
Hypothesis 3B (9) High RDISXPM Group 0.1223 0.0508 3,287 (10) Low RDISXPM Group 0.1434 0.0593 3,288 (11) (9) − (10) −0.0211 * −0.0085
Notes: Return is the market-adjusted return, defined as the difference between ES firm i’s compound annual buy-and-hold return over the 12-month period ending three months after the fiscal yearend and the compound annual return of value-weighted market index return over the same period. Rows 1-5 show the mean and median values of ES firms’ Return along the following five sub-periods of an earnings string: Early-ES (the first two years of an earnings string), Mid-ES (the period from the third year of an earnings string to three years before the break), −2Break (two years before the break), −1Break (one year before the break) and Break (the break year of an earnings string). *, **, and *** represent significance levels of 0.1, 0.05, and 0.01 (two-tailed), respectively, using t-tests for means and sign tests for medians. Rows 6-7 show the mean and median values of Return in the high and the low DACCPM groups. A firm is assigned to the High (Low) DACCPM group if its DACCPM is above (below) the median of DACCPM distribution. DACCPM is performance-matched discretionary accruals, calculated as the difference between an ES firm i’s discretionary accruals and the discretionary accruals of a non-ES firm that has the closest ROA in the same industry-year group. Row 8 shows the comparisons of mean and median values of Return across the high vs. the low DACCPM group. Rows 9-10 show the mean and median values of Return in the high and the low RDISXPM groups. A firm is assigned to the High (Low) RDISXPM group if its RDISXPM is above (below) the median of RDISXPM distribution. RDISXPM is performance-matched abnormal discretionary expenses multiplied by −1 and is calculated as −1 times the difference between an ES firm i’s abnormal discretionary expenses and the abnormal discretionary expenses of a non-ES firm that has the closest ROA in the same industry-year group. Row 11 shows the comparisons of mean and median values of Return across the high vs. the low RDISXPM group. Mean comparisons are based on t-tests and median comparisons are based on Wilcoxon tests. *, **, and *** represent significance levels of 0.1, 0.05, and 0.01 (two-tailed), respectively.
(The table is continued on the next page.)
45
TABLE 6 (Continued)
Panel B: Return Regression
Model: 푅푒푡푢푟푛 = 훽 + 훽 ∆퐸 + 훽 퐴푀 _H *∆퐸 (or 훽 푅푀 _H *∆퐸 )
+ 훽 퐺푟표푤푡ℎ *∆퐸 + 훽 퐸푉퐴푅 *∆퐸 + 훽 퐷퐸 *∆퐸 + 훽 ∆퐵푉
+훽 푌푒푎푟퐷푢푚푚푦 + 훽 푌푒푎푟퐷푢푚푚푦*∆퐸 +휀
Panel B1. EM = DACCPM Panel B2. EM = RDISXPM
Prediction Coefficient t-statistics p-value Coefficient t-statistics p-value ΔE + 1.3525 2.25 0.0122 1.0729 1.91 0.0284
Test Variable
DACCPM_H*ΔE − −0.5252 −3.94 <.0001 n/a n/a n/a
RDISXPM_H*ΔE insignificant n/a n/a n/a 0.0391 0.32 0.7507 Control Variable
Growth*ΔE 0.5101 1.91 0.0570 0.6120 2.22 0.0263 EVAR*ΔE 0.0003 0.10 0.9190 −0.0008 −0.33 0.7391
DE*ΔE 0.0058 0.07 0.9472 0.0504 0.54 0.5903
ΔBV 0.4602 6.72 <.0001 0.4534 6.62 <.0001 YearDummy Yes Yes
YearDummy*ΔE Yes Yes
Adj. R2 8.56% 8.27%
N 6,575 obs. (949 firms) 6,575 obs. (949 firms) Notes: Return is market-adjusted return, defined as the difference between ES firm i’s compound annual buy-and-hold return over the 12-month period ending three months after the fiscal yearend and the compound annual return of value-weighted market index return over the same period. ΔE represents change in earnings per share scaled by stock price at the end of previous year. Panel B1 shows the regression results based on performance-matched discretionary accruals (DACCPM), where DACCPM is calculated as the difference between an ES firm i’s discretionary accruals and the discretionary accruals of a non-ES firm that has the closest ROA in the same industry-year group. The variable, DACCPM_H, is an indicator variable which is set equal to 1 if a firm’s DACCPM is above the median of DACCPM distribution and zero otherwise. Panel B2 shows the regression results based on performance-matched abnormal discretionary expenses multiplied by −1 (RDISXPM), where RDISXPM is calculated as −1 times the difference between an ES firm i’s abnormal discretionary expenses and the abnormal discretionary expenses of a non-ES firm that has the closest ROA in the same industry-year group. The variable, RDISXPM_H, is an indicator variable which is set equal to 1 if a firm’s RDISXPM is above the median of RDISXPM distribution and zero otherwise. In Panels B1-B2, All t-values are reported using robust standard errors to correct heteroskedasticity problem and firm clustering effect. The p-values are one-tailed if there is a prediction, and two-tailed otherwise. Please refer to Appendix 1 for the definitions and measurements of all the control variables.
46
FIGURE 1 Pattern of Earnings Management and Return (1989–2010)
Notes: The sample period covers from 1989 to 2010. In Figure 1A the ES Sample consists of 1,043 firms (or 8,234 firm-year observations), whereas in Figure 1B the ERC Sample includes 949 firms (or 6,575 firm-year observations). In both figures, the X axis represents the following five sub-periods of an earnings string: Early-ES (the first two years of an earnings string), Mid-ES (the period from the third year of an earnings string to three years before the break), −2Break (two years before the break), −1Break (one year before the break) and Break (the year when an earnings string is broken). In Figure 1A, the Y axis represents mean values of accrual and real activity management, as proxied by performance-matched discretionary accruals (DACCPM) and performance-matched abnormal discretionary expenses multiplied by −1 (RDISXPM). In Figure 1B, the Y axis represents mean values of market-adjusted return (Return). Patterns based on median values are similar and hence are not reported to conserve space.
-0.04
-0.03
-0.02
-0.01
1E-17
0.01
0.02
Early-ES Mid-ES -2Break -1Break Break
DAC
CPM
Figure 1A. Pattern of Earnings Management
DACCPM RDISXPM
-0.1
0
0.1
0.2
0.3
Early-ES Mid-ES -2Break -1Break Break
Ret
urn
Figure 1B. Pattern of Return