Sei sulla pagina 1di 24

Discretionary Revenues as a Measure of Earnings Management

Author(s): Stephen R. Stubben


Source: The Accounting Review, Vol. 85, No. 2 (MARCH 2010), pp. 695-717
Published by: American Accounting Association
Stable URL: http://www.jstor.org/stable/20744146
Accessed: 23-10-2017 04:58 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/20744146?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

American Accounting Association is collaborating with JSTOR to digitize, preserve and


extend access to The Accounting Review

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
THE ACCOUNTING REVIEW American Accounting Association
Vol. 85, No. 2 DOI: 10.2308/accr.2010.85.2.695
pp. 695-717

Discretionary Revenues as a Measure of


Earnings Management
Stephen R. Stubben
The University of North Carolina at Chapel Hill

ABSTRACT: This study examines the ability of revenue and accrual models to detect
simulated and actual earnings management. The results indicate that revenue models
are less biased, better specified, and more powerful than commonly used accrual mod
els. Using a simulation procedure, I find that revenue models are more likely than ac
crual models to detect a combination of revenue and expense manipulation. Using a
sample of firms subject to SEC enforcement actions for a mix of revenue- and expense
related misstatements, I find that, although revenue models detect manipulation, ac
crual models do not. These findings provide support for using measures of discretion
ary revenues to study earnings management.

Keywords: revenues; earnings management; discretionary accruals.

Data Availability: Dafa are available from public sources.

I. INTRODUCTION
This study examines the ability of revenue and accrual models to detect simulated
and actual earnings management. Despite repeated criticisms of accrual models over
the past 15 years (e.g., Dechow et al. 1995; Bernard and Skinner 1996; Guay et al.
1996; McNichols 2000; Thomas and Zhang 2000; Kothari et al. 2005), many studies have
addressed and continue to address earnings management using these models, presumably
because few viable alternatives exist.1 Accrual models have been criticized for providing
biased and noisy estimates of discretion, which calls into question the conclusions from
studies that use them (Bernard and Skinner 1996). The objective of this study is to evaluate
a different measure of earnings management?discretionary revenues?that permits more
reliable and conclusive inferences than existing models.

1 For example, between 2005 and 2008, The Accounting Review, Journal of Accounting and Economics, and
Journal of Accounting Research published at least 40 articles that use a measure of discretionary accruals.

This study is based on my dissertation titled "The Use of Discretionary Revenues to Meet Earnings and Revenue
Targets." I thank my dissertation committee of Mary Barth (co-chair), Bill Beaver (co-chair), and Maureen
McNichols, as well as Steven Kachelmeier (the senior editor), Mark Trombley (the editor), two anonymous re
viewers, Yonca Ertimur, Fabrizio Ferri, Wayne Landsman, and workshop participants at Stanford University, Uni
versity of California, Los Angeles, The University of Chicago, Massachusetts Institute of Technology, Harvard
Business School, The University of North Carolina at Chapel Hill, the 2005 Accounting Research Symposium at
Brigham Young University, and the 2006 EAA Doctoral Colloquium for invaluable comments and suggestions.
Editor's note: Accepted by Mark Trombley.
Submitted: June 2008
Accepted: July 2009
Published Online: March 2010

695

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
696 Stubben

The most common approaches to estimating earnings management (i.e., the extent
management exercises discretion over reported earnings) use aggregate accruals. However,
several studies have suggested focusing on one component of earnings, which has the
potential to provide more precise estimates of discretion (McNichols and Wilson 1988;
Bernard and Skinner 1996; Healy and Wahlen 1999; McNichols 2000). Modeling a single
earnings component permits incorporating key factors unique to that component, thereby
reducing measurement error. In addition, focusing on earnings components provides insights
into how earnings are managed. Revenues is an ideal component to examine as it is the
largest earnings component for most firms and is subject to discretion. Dechow and Schrand
(2004, 42) find that over 70 percent of SEC Accounting and Auditing Enforcement Releases
involve misstated revenues. Furthermore, revenues are the most common type of financial
restatement (Turner et al. 2001).
I model a common form of revenue manipulation?premature revenue recognition?
and its effect on the relation between revenues and accounts receivable. I define prematurely
recognized revenues as sales recognized before cash is collected using an aggressive or
incorrect application of Generally Accepted Accounting Principles. My revenue model is
similar to existing accrual models (Jones 1991; Dechow et al. 1995), but with three key
differences.
First, I model the receivables accrual, rather than aggregate accruals, as a function of
the change in revenues. Of the major accrual components, receivables have the strongest
empirical and most direct conceptual relation to revenues. Explaining other working capital
accruals by revenues alone contributes to the noise and bias that plague accrual models.
Because I model receivables instead of aggregate accruals, the model is one of revenues
rather than earnings.
Second, I model the receivables accrual as a function of the change in reported reve
nues, rather than the change in cash revenues (Dechow et al. 1995). Although this choice
systematically understates estimates of discretion in revenues, it is less likely to overstate
estimates of discretion for firms whose revenues are less likely to be realized in cash by
year-end (e.g., growth firms). The modified Jones model (Dechow et al. 1995) treats un
collected credit sales as discretionary.
Third, I model the change in annual receivables as a linear function of two components
of the change in annual revenues: (1) change in revenues of the first three quarters, and (2)
change in fourth-quarter revenues. Because revenues in the early part of the year are more
likely to be collected in cash by the end of the year, these have different implications for
year-end receivables than a change in fourth-quarter revenues. Revenue and accrual models
that fail to consider fourth-quarter revenues separately overstate (understate) discretion
when fourth-quarter revenues are relatively high (low).
I also estimate a version of the model that allows the relation between receivables and
annual revenues to vary depending on firms' credit policies. Models that fail to incorporate
differences across firms in the relation between revenues and receivables (or accruals) will
misstate estimates of discretion.
Discretionary revenues is the difference between the actual change in receivables and
the predicted change in receivables based on the model. Abnormally high or low receivables
indicate revenue management. To benchmark against existing models, I compare the ability
of the revenue model and commonly used accrual models (Jones 1991; Dechow et al. 1995;
Dechow and Dichev 2002; Kothari et al. 2005) to detect combinations of revenue and
expense management.
Findings indicate that measures of discretionary revenues do in fact produce estimates
with substantially less bias and measurement error than those of accrual models. Using

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 697

simulated manipulation (Kothari et al. 2005), I find that the revenue model produces esti
mates of discretion that are well specified for growth firms. Each of the accrual models
tested exhibits substantial misspecification. The results for growth firms indicate a bias of
over 2 percent of total assets for each of the accrual models.
The results indicate further that discretionary revenues can detect not only revenue
management, but also earnings management (via revenues), whereas accrual models do
not.2 Using a sample of firms targeted by SEC enforcement actions, I find that the revenue
model is able to detect earnings manipulation, but accrual models are not.3 The simulation
analysis reveals that the revenue model is more likely than accrual models to detect earnings
management for equal amounts of revenue and expense manipulation. These findings sug
gest that the revenue model is less likely than accrual models to falsely indicate earnings
management, and more likely than accrual models to detect earnings management when it
does occur.
Finally, the results have implications for studies that use accrual models. First, the Jones
model (Jones 1991) exhibits better specification than the modified Jones model (Dechow
et al. 1995), which suggests including reported revenues, rather than cash revenues, in
accrual models.4 Second, the Dechow-Dichev model (Dechow and Dichev 2002; McNichols
2002), which was originally developed to estimate earnings quality, exhibits greater mis
specification than other accrual models when used to estimate discretionary accruals. I am
unaware of any other study that evaluates the ability of the Dechow-Dichev model to
estimate discretionary accruals. Last, the innovations I incorporate into the revenue model,
including separating fourth-quarter revenues and allowing the relation between revenues
and accruals to vary across firms, could improve the performance of accrual models.

II. MOTIVATION AND RELATED RESEARCH


Earnings Management
Earnings management has received significant attention in the popular press and aca
demic accounting literature (McNichols 2000). Recently, various accounting scandals (e.g.,
Enron, WorldCom, and Parmalat), and subsequent regulation (e.g., Sarbanes-Oxley) have
fueled this attention.
As Beneish (1999, 24) writes, "The extent to which earnings are manipulated has long
been of interest to analysts, regulators, researchers, and other investment professionals."
For analysts and investors, understanding the extent to which managers exercise discretion
in earnings is important in assessing the quality of earnings. Understanding which firms
manage earnings and how they do it is useful for standard-setters and regulators. Frequent
earnings management warrants new standards or additional disclosures, and evidence on
the specific accruals managed allows standard-setters to narrow their focus in this regard.
Understanding the motives for earnings management allows the SEC to narrow its focus in
enforcing standards. In each case, a reliable measure of earnings management and the
specific accruals used is required.

2 Consistent with this idea, Stubben (2006) finds evidence of earnings management to meet several earnings
targets using estimates of discretionary revenues, despite the estimates of discretionary accruals being too noisy
to produce significant results.
3 Richardson et al. (2006) find a positive correlation between noncurrent operating accruals and SEC enforcement
actions. However, I limit my focus to the more commonly used accrual models.
4 Relatedly, Kothari et al. (2005) note the misspecification of the modified Jones model for firms with extreme
growth and caution its use unless the researcher is confident that credit sales represent accrual manipulation.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
698 Stubben

Findings of Prior Literature


Accrual Models
Studies of earnings management around firm-specific events commonly use models of
aggregate accruals.5 A variety of accrual models have been used in recent years. Most
accrual models are some variation of the cross-sectional Jones model (Jones 1991; DeFond
and Jiambalvo 1994). In these models, nondiscretionary, or normal, accruals are usually
estimated as a linear function of change in revenues and gross property, plant, and equip
ment. The models are usually estimated by industry and year, and the residual is the dis
cretionary accruals estimate. The modified Jones model (Dechow et al. 1995) uses change
in cash revenues rather than change in total revenues because some credit revenues may
be discretionary. Other models add additional conditioning variables, such as cash flows
(Dechow and Dichev 2002; McNichols 2002).
Despite the prevalence of these models, they have been criticized for providing
biased and noisy estimates of discretionary accruals (Dechow et al. 1995; Kang and
Sivaramakrishnan 1995; Bernard and Skinner 1996; Guay et al. 1996; Thomas and Zhang
2000). Dechow et al. (1995) find that discretionary accrual models generate tests of low
power for earnings management of economically plausible magnitudes and misspecified
tests of earnings management for firms with extreme financial performance, which is fre
quently correlated with the earnings management incentive under investigation (e.g., IPOs
and SEOs). The results in Guay et al. (1996) are consistent with the modified Jones model
estimating discretionary accruals with considerable imprecision and/or misspecification.
Thomas and Zhang (2000) assert that the commonly used models "provide little ability to
predict accruals."
Inferences relating to earnings management depend on the researcher's ability to ac
curately estimate discretionary accruals. As a consequence of accrual models' poor per
formance, studies using accrual models have often produced mixed results. Burgstahler and
Eames (2006) find that firms with small positive forecast errors have higher unexpected
accruals using the Jones (1991) model, in contrast to the absence of such an association in
Phillips et al. (2003), who employ the modified Jones model. Matsumoto (2002) finds
evidence of firms using discretionary accruals to meet earnings forecasts in one specifica
tion, but not in another. Dechow et al. (2003) find similar magnitudes of discretionary
accruals for small loss and small profit firms and conclude that if firms overstate earnings
to report profits, then their accrual model is not powerful enough to detect it.

Specific Accruals
A limitation of aggregate discretionary accrual measures is that they do not provide
information as to which components of earnings firms manage. These models do not dis
tinguish discretionary increases in earnings through revenues or components of expenses.
McNichols (2000) suggests that future progress in the earnings management literature is
more likely to come from the examination of specific accruals (see also Bernard and Skinner
1996; Healy and Wahlen 1999; Beneish 2001).6

5 Other approaches to identify earnings management include distributional tests (e.g., Burgstahler and Dichev
1997), accounting procedure changes (Healy 1985; Healy and Palepu 1990; Sweeney 1994), and components
of discretionary cash flows (Dechow and Sloan 1991).
6 Examples of specific accruals that have been modeled are the allowance for bad debt (McNichols and Wilson
1988), loan loss provisions in the banking industry (Beaver et al. 1989; Moyer 1990; Scholes et al. 1990;
Whalen 1994; Beatty et al. 1995; Collins et al. 1995; Beaver and Engel 1996), loss reserves of property and
casualty insurers (Petroni 1992; Beaver and McNichols 1998; Nelson 2000; Petroni et al. 2000), and tax expense
(Phillips et al. 2003; Dhaliwal et al. 2004).

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 699

The advantage of using specific accruals, such as loan loss reserves, is that they are
material and a likely object of judgment and discretion. However, many accruals are in
dustry specific (e.g., banking and insurance industries) and the analysis cannot be applied
to firms outside the industry. In contrast, McNichols and Wilson (1988) examine discretion
in the allowance for bad debts. This accrual is common across industries. However, bad
debt expense is often only a small portion of reported earnings, so only a relatively small
portion of the total amount of a firm's discretion is likely to be captured. An ideal specific
accrual for study is one that is (1) common across industries, (2) subject to discretion, and
(3) represents a large portion of the earnings discretion available to firms. Based on these
criteria, revenues is a natural candidate.

Revenues as a Means of Earnings Management


Three studies that examine revenue management are Plummer and Mest (2001),
Marquardt and Wiedman (2004), and Caylor (2009). Plummer and Mest (2001) study the
discretion in earnings components using distributional tests similar to those of Burgstahler
and Dichev (1997), finding evidence suggesting firms manage earnings upward to meet
earnings forecasts by overstating revenues and understating some operating expenses. They
do not attempt to estimate discretionary revenues.
Marquardt and Wiedman (2004) estimate the unexpected portions of several accrual
components, including receivables, to determine which components of earnings firms ma
nipulate in certain settings. They find evidence that firms with small earnings increases
understate special items but do not overstate revenues. They also find evidence that firms
use discretion in revenues to increase (decrease) earnings prior to equity issuances (man
agement buyouts). Caylor (2009) uses discretionary revenues to test their use for avoiding
reporting negative earnings surprises and finds evidence that managers use discretion in
revenues that affects both accounts receivable and deferred revenues to report positive
earnings surprises. I seek to validate a measure of discretionary revenues that can be used
to test hypotheses in studies such as these.7

III. RESEARCH DESIGN


Discretionary revenues take a number of forms. Some involve the manipulation of real
activities (e.g., sales discounts, relaxed credit requirements, channel stuffing, and bill and
hold sales), and others do not (e.g., revenues recognized using an aggressive or incorrect
application of Generally Accepted Accounting Principles (GAAP), fictitious revenues, and
revenue deferrals). I model premature revenue recognition and its effect on the relation
between revenues and accounts receivable. Premature revenue recognition includes channel
stuffing and bill and hold sales, if customers do not pay cash for the inventory, and revenues
recognized using an aggressive or incorrect application of GAAP.8

7 The model used in Caylor (2009) is similar to my revenue model but does not split fourth-quarter revenues and
includes the change in cash flows from the current to following year. Untabulated results reveal that the model
used by Caylor (2009) produces statistically indistinguishable results compared to my revenue model in
the simulation analyses but fails to detect revenue manipulation by SEC firms. In addition, my model has the
advantage of not relying on future data; it can be applied in real time.
8 By focusing on premature revenue recognition, I will not capture discretionary cash revenues. These might
include not deferring enough revenue, or taking real actions that increase cash collected from revenues. In fact,
to the extent these cash-based discretionary revenues occur, the model will understate the amount of accrual
based discretionary revenues. Caylor (2009) provides a model of deferred revenues. I focus on receivables in
order to obtain a measure that can be used broadly across many firms; most firms do not report a significant
amount of deferred revenues.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
700 Stubben

I focus on premature revenue recognition because evidence suggests that it is the most
common form of revenue management.9 For example, Feroz et al. (1991) find that more
than half of SEC enforcement actions issued between 1982 and 1989 involved overstate
ments of accounts receivable resulting from premature revenue recognition. Other forms of
revenue manipulation, such as sales discounts, could be profit-maximizing business deci
sions and not merely attempts manage revenues to meet a performance benchmark.

Model of Discretionary Revenues


Reported (i.e., managed) revenues (R) is the sum of nondiscretionary revenues (RUM)
and discretionary revenues (bRM):10

R = RUM + ?*M

A fraction, c, of nondiscretionary revenues remains uncollected at year-end, and I assume


there are no cash collections of discretionary revenues. Thus, accounts receivable (AR)
equals the sum of uncollected nondiscretionary revenues (c x RUM) and discretionary rev
enues (8 ):11

ARit = cX R + 8f.

Discretionary revenues increase accounts receivable and revenues by the same amount;
discretionary receivables equals discretionary revenues.
Because nondiscretionary revenues are not observable, I rearrange terms and express
ending receivables in terms of reported revenues. I then take first differences to arrive at
the following expression for the receivables accrual:12

LARit = c X ARit + (1 - c) X A8f.

The estimate of a firm's discretionary revenues is the residual from the following equation:

AARit = a + $ARit + ziv

Because reported revenues include discretionary revenues, the amount of revenue manage
ment estimated by the revenue model will be understated by a factor of 1 - c (see Jones
1991, footnote 31). The modified Jones model (Dechow et al. 1995) conditions on the

9 Because of a possible selection bias, the type of earnings manipulation that is most commonly detected is not
necessarily the most commonly employed. However, without a reliable measure of revenue manipulation (such
as those presented in this study), it is difficult to assess the extent of revenue versus expense manipulation in
the general population of firms.
10 If revenues are managed each year, 8^ could be interpreted as net revenue management (i.e., revenue manage
ment net of reversals from the prior period).
11 I assume that receivables are collected within one year. I address the impact of this assumption in untabulated
tests by adding additional lags of sales to the model. Qualitative results are unchanged.
12 The advantage of modeling receivables as opposed to accruals is that they are directly related to revenues?the
relation is not as clear with other current accruals (Kang and Sivaramakrishnan 1995). Accounts payable relates
to purchases, and inventory relates to forecasted revenues for the next period, not current revenues (Dechow et
al. 1998). Forecasted revenues equals current revenues if revenues follow a random walk, but this is not true
for growing firms. The relation between other accruals and change in revenues is unclear. Because revenues
alone do not explain payables, inventory, and other accruals, accrual models produce noisy estimates of discre
tionary accruals. The estimates are also biased, for example, for growth firms if the noise is correlated with
growth.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 701

change in cash revenues rather than total revenues, which avoids systematically understating
the amount of earnings management. However, this approach introduces another problem:
uncollected credit revenues are treated as discretionary. Firms with a higher than average
portion of nondiscretionary revenues that are credit will have higher estimates of discre
tionary accruals. I condition on total revenues because this understates estimated earnings
management, which biases against finding in favor of typical alternative hypotheses.
One limitation of accrual models is that by conditioning on annual revenues they treat
revenues from early in the year the same as revenues from later in the year. Current accruals
are generally resolved within one year. Thus, sales made late in the year are more likely
to remain on account at year-end. Therefore, the revenue model I estimate allows the
estimated portion of revenues that are uncollected at year-end to vary in the fourth quarter:

kARit = a + ^R1_JU + ?2AJW? + eir (1)

In Equation (1), Rl-3 is revenues in the first three quarters, and R4 is revenues in the
fourth quarter. Even though Equation (1) incorporates quarterly revenues, I estimate dis
cretion on an annual level.13 Any revenue management in early quarters that reverses by
year-end will not be captured.
Another limitation of accrual models is that cross-sectional estimation implicitly as
sumes that firms in the same industry have a common accrual-generating process. Dopuch
et al. (2005) show that the relation between accruals and revenue changes depends on firm
specific factors such as credit and inventory policies. In cross-sectional estimation, devia
tions from the industry average credit policy, for example, would lead to nonzero estimated
unexpected accruals. Therefore, to control for these deviations, I also estimate a version of
the revenue model that allows the coefficient on revenues to vary with firms' credit
policies.14
Petersen and Rajan (1997) review theories of trade credit and the determinants of
accounts receivable. I use the implementation of the Petersen and Rajan (1997) credit policy
model described in Callen et al. (2008). A firm's investment in receivables is a function of
its financial strength, operational performance relative to industry competitors, and stage in
the business cycle. Firm size, measured as the natural log of total assets (SIZE), is a proxy
for financial strength. Firm size and firm age (AGE), measured as the natural log of the
firm's age in years, are proxies for the firm's stage in the business cycle. Following Petersen
and Rajan (1997), I also include the square of firm age (AGESQ) to allow for a nonlinear
relation between age and credit policy. As proxies for the operational performance of the
firm relative to industry competitors, I use the industry-median-adjusted growth rate in
revenues (GRR-P if positive, GRR-N if negative) and the industry-median-adjusted gross
margin (GRM) and its square (GRMSQ).
I refer to the following model as the conditional revenue model, as it is a cross-sectional
conditional estimation of the revenue coefficient:

13 I focus on annual accruals because most studies using discretionary accruals do so at the annual level. Stubben
(2006) develops models of revenues and accruals at the quarterly level. Similar to the findings of this study, the
quarterly revenue model is more powerful and better specified than quarterly accrual models.
14 The relation between receivables and revenues is also affected by economic events such as factoring accounts
receivable (McNichols 2000). Regarding factoring, Sopranzetti (1998) searches a database of over 4,000 publicly
traded firms from 1972-1993 and finds only 269 reports of factored accounts receivable by 98 firms. Nearly
half of the firms (47) were from the Textile and Apparel industries, and no other industry had more than 7 firms.
To control for the potential effects of factored receivables, I repeat my tests after excluding firms from the
Textile and Apparel industries and find similar results.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
702 Stubben

AARit = a + ?Mit + ?2A/?? X SIZEit + ?3A/^ X AG?,


+ ?4ARit x AGEJ?* + ?5A/?? X G/?/?_PlY + ?6A/?lY X GRR_Nit

+ ?7ARit x G/?M,, + ?8A^ x GRM_SQit + e,r (2)

As a benchmark for the revenue models, I estimate discretionary accruals using four com
mon approaches: the Jones model (Equation (3) below; Jones 1991), the modified Jones
model (Equation (4) below; Dechow et al. 1995), the Dechow-Dichev model (Equation (5)
below; Dechow and Dichev 2002; McNichols 2002), and performance-matched estimates
from the modified Jones model (Kothari et al. 2005). The Appendix summarizes the dis
cretionary revenue and accrual estimates I use in this study.

ACit = a + ?,ARit + ?2PPEit + eit; (3)

ACit = a + ^(ARit - AARit) + ?2PPEit + e*; (4)

ACit = a + ?.AR, + bPPEu + ?3A?* X CFOit_x + ?4CFO, + ?5CFOu+1 + ?,.


(5)
Data and Descriptive Statistics
The sample period begins in 1988 because prior to that date cash flow from operations
disclosed under Statement of Financial Accounting Standards No. 95 (FASB 1987) is un
available. The sample period ends in 2003 but uses 2004 data for cash from operations. I
exclude firms in regulated industries (financial, insurance, and utilities) because their rev
enues and accruals likely differ from those of other firms.
I obtain financial data from Compustat. I calculate accruals as earnings before extraor
dinary items less cash flow from operations.15 I obtain cash flow from operations and the
change in receivables from the cash flow statement (Hribar and Collins 2002). Revenues
of the first three quarters is the difference between annual revenues and fourth-quarter
revenues. I deflate all revenue and accrual variables by average total assets. Annual earnings
growth is the annual change in income before extraordinary items, deflated by average total
assets. Revenue growth is the ratio of current to prior-year revenues. Gross margin per
centage equals revenues less cost of revenues, divided by revenues. Firm age is the number
of years since the firm's initial appearance in the Compustat annual file with non-missing
financial information. Industries are as defined in Barth et al. (2005). I winsorize at 1 percent
each model input variable by industry and year.
Following Kothari et al. (2005), I estimate nondiscretionary accruals with scaled and
unsealed intercepts (by assets), to control for scale differences among firms (Barth
and Kallapur 1996). I exclude firms suspected of manipulation (e.g., firms targeted by the
SEC) when estimating the model coefficients. I use the estimated coefficients to calculate
discretionary revenues of suspected firms.
Table 1 presents summary statistics. Panel A indicates that mean (median) accruals are
-8 (-6) percent of average assets. The mean and median change in receivables is 1 percent
of average assets. Panel A also indicates that the mean (median) change in revenues is 9
(7) percent of average assets. On average, the revenue change is approximately evenly

15 Because the statement in cash flows presents the change in net receivables, discretionary revenues include any
discretion in the allowance for doubtful accounts. McNichols and Wilson (1988) find evidence of earnings
management using the allowance for doubtful accounts for firms with unusually high or low earnings.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 703

TABLE 1
Sample Summary Statistics
(n = 70,580)

Panel A: Descriptive Statistics


Variable Mean Std. Dev. Ql Median Q3
AC -0.08 0.17 -0.12 -0.06 -0.01
AAR 0.01 0.07 -0.01 0.01 0.04
A/? 0.09 0.37 -0.04 0.07 0.22
AR1-3 0.07 0.30 -0.03 0.05 0.17
AR4 0.03 0.13 -0.02 0.02 0.07
PPE 0.30 0.24 0.11 0.24 0.44
CEO 0.03 0.19 -0.01 0.06 0.13

Panel B: Pearson (above), Spearman (below) Corre


AC AAR AR AR1-3 AR4 PPE CFO
AC 0.35 0.21 0.18 0.19 -0.05 0.04
AAR 0.40 0.48 0.38 0.49 -0.04 -0.07
AR 0.26 0.50 0.94 0.67 0.02 0.09
AR1-.3 0.23 0.41 0.93 0.40 0.02 0.09
AR4 0.24 0.54 0.72 0.50 0.01 0.05
PPE -0.14 -0.04 0.01 0.01 0.01 0.23
CFO -0.25 -0.06 0.14 0.14 0.10 0.28

Variables are deflated by average total assets. All correlations in Panel B are sig
(p < 0.01).
Variable Definitions:
AC = annual current accruals = earnings before extraordinary items - cash from operations;
AR = end of fiscal year accounts receivable;
R = annual revenues;
Rl-3 = revenues of the first three quarters;
R4 = revenues of the fourth quarter;
PPE = end of fiscal year gross property, plant, and equipment;
CFO = cash from operations; and
A = annual change.

distributed across quarters. The median change in revenues of the first three quarters is 5
percent of average assets (approximately 2 percent per quarter), and the median change in
fourth-quarter revenues is 2 percent of average assets.
Panel B of Table 1 presents correlations. Because the Pearson and Spearman correla
tions are similar, I focus on the Pearson correlations. All correlations are significantly
different from zero.16 Change in receivables is positively correlated with accruals (0.35),
largely by construction because change in receivables is typically a large component of
current accruals. However, the correlation between annual revenue change and change in
receivables (0.48) is larger than the 0.21 correlation between annual revenue change and
accruals. Additionally, change in receivables is more highly correlated with change in
fourth-quarter revenues than with the change in revenues of the first three quarters (0.49

16 I use the term significance to denote statistical significance at less than the 0.05 level, based on a one-sided test
when I have signed predictions and a two-sided test otherwise.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
704 Stubben

versus 0.38) or even change in annual revenues (0.49 versus 0.48). Taken together, these
correlations suggest that estimates from models of receivables are less noisy than estimates
from accrual models, and that using quarterly data to disaggregate annual change in reve
nues might lead to better specified models. However, I base my inferences on the tests
presented in the next section.

IV. RESULTS
Estimation of the Models
Table 2 presents results from the estimation of the revenue and accrual models. Panel
A presents summary results of the revenue models. In revenue model (1), the coefficient
on change in fourth-quarter revenues (0.21) is over four times higher than that of the change
in revenues of the first three quarters (0.05), although both are significantly positive.17 In
contrast, when conditioning on annual revenue change, the average coefficient is 0.08 and
the adjusted R2 decreases from 0.28 to 0.21. The coefficient on revenues in model (2) varies
significantly with size (SIZE), age (AGE), industry-adjusted revenue growth (GRR-P and
GRR-N), and gross margin (GRM). The adjusted R2 increases from 0.21 to 0.28 over the
model that does not allow for variations in credit policy.
Panel B of Table 2 presents summary results of the accrual models. The coefficient on
revenue change in the Jones model (3) is 0.10, as compared to 0.06 for the modified Jones
model (4). In addition, the adjusted R2 of the modified Jones model is lower (0.09) than
that of the Jones model (0.12). In the Dechow-Dichev model (5), the coefficients on past
and future cash flows (0.27 and 0.14, respectively) and the coefficient on current cash flows
(-0.46) are each significantly related to accruals. The adjusted R2 increases to 0.30 when
adding cash flows to the Jones model.
Untabulated results reveal that, when estimating the Jones model after excluding re
ceivables from accruals, the resulting coefficient on revenue change is 0.01 and is signifi
cantly positive in only 66 out of 285 industry-year groups. This finding indicates that the
change in receivables drives much of the correlation between accruals and change in rev
enues. As expected, the relation between other accruals and revenue change is weaker than
that of the receivables accrual and revenue change, leading to noisier estimates of discretion
for accrual models.18

Evaluation of the Models


I evaluate estimates of discretion from the various models by assessing the models'
abilities to detect simulated revenue and expense manipulation and by relying on actual
earnings and revenue manipulation in a sample of firms known to have misstated their
financial results.

Detection of Simulated Revenue Manipulation


Similar to Dechow et al. (1995) and Kothari et al. (2005; hereafter KLW), I use sim
ulations to test the power and specification of discretionary accrual models in the presence
of extreme performance. By comparing estimates of discretionary revenues and expenses

17 Untabulated results reveal that, as expected, the higher fourth-quarter coefficient is attributable to the fourth
quarter being closer in time to the fiscal year-end rather than an inherent difference in fourth quarter accruals
due to factors such as audit requirements. I find similar coefficient magnitudes using alternative annual aggre
gations of quarterly sales and accruals, ending at the first, second, or third quarter (see Jacob and Jorgensen
2007).
18 Finding little or no relation between aggregate accruals other than receivables and change in sales does not
imply that there is no relation between change in sales and individual accrual components. However, it does
support modeling specific accruals, such as receivables, rather than aggregate accruals.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 705

TABLE 2
Estimation of Revenue and Accrual Models

Panel A: Estimation of Revenue Models

AARit = a + ^ARl_3it + ?2AR4it + zit (1)


AARit = a + ^ARit + ?2ARit X SIZEit + ?3ARit X AGEit + ?4ARit X AGE_SQit
+ ?5ARit x GRR-Pit + ?6ARit x GRR-N? + ?,ARit x GRMit
+ ?8A/?,Y x GRMSQit + (2)
(1) (2)
Mean FM Mean FM Mean FM
Variable Estimate t-stat Estimate t-stat Estimate t-stat
AR 0.08 33.03** 0.19 9.84**
AR1-3 0.05 22.86**
AR4 0.21 30.00**
AR X SIZE -0.01 -6.20**
AR X AGE -0.04 -2.80**
AR X AGE-SQ 0.01 2.07*
AR X GRR-P -0.02 -5.12**
AR X G/ffi_N 0.07 7.70**
A/? x G#M 0.04 3.83**
A# X GRM-SQ 0.06 1.38
Adj. R2 0.28 0.21 0.28
Panel B: Estimation of Accrual Models

ACit = a + ?xARit + ?2PPEit + zit (3)


ACit = a + ?x(ARit - AARit) + ?2PPEit + e? (4)
ACit = a + ?,ARit + ?2PPEit + ?3AflJ> X CFO^ + ?4CF0ft + ?5CFOit+l + Eit (5)
(3) (4) (5)
Mean FM Mean FM Mean FM
Variable Estimate t-stat Estimate t-stat Estimate t-stat
AR 0.10 22.29** 0.10 24.09**
AR - AAR 0.06 15.73**
PPE -0.07 -15.55** -0.07 -15.03** -0.07 -18.07**
CFOt_x 0.27 23.62**
CFOt -0.46 -27.75**
CFOt+l 0.14 13.98**
Adj. R2 0.12 0.09 0.30

*, ** Indicates the coefficient estimate is significantly different from zero at the 0.05 and 0.01 level, using a
two-sided test.
Table 2 summarizes the results of separate estimations of revenue and accrual models for each of 285 industry
years. An intercept scaled by average total assets and an unsealed intercept are not tabulated. FM t-stat is the
Fama and Macbeth (1973) t-statistic. Variables are deflated by average total assets.

(continued on next page)

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
706 Stubben

TABLE 2 (continued)
Variable Definitions:
AR = end of fiscal year accounts receivable;
AC ? annual current accruals = earnings before extraordinary items - cash from operations;
R = annual revenues;
Rl-3 = revenues of the first three quarters;
R4 = revenues of the fourth quarter;
PPE = end of fiscal year gross property, plant, and equipment;
CFO = cash from operations;
SIZE = natural log of total assets at end of fiscal year;
AGE = age of firm (years);
GRR-P = industry-median-adjusted revenue growth (= 0 if negative);
GRR-N = industry-median-adjusted revenue growth (= 0 if positive);
GRM = industry-median-adjusted gross margin at end of fiscal year;
-SQ = square of variable; and
A = annual change.

against a known quantity of manipulation, I am able to obtain evidence of the bias, spec
ification, and power of competing models. I measure the bias of each model as the differ
ence between the mean estimate of discretion and the amount of manipulation I induce. If
the model is unbiased, then the difference will equal zero. I evaluate the specification of the
models by computing how often tests reject the null hypothesis of no manipulation for
samples with no manipulation induced. Finally, I evaluate the power of the models by
computing how often tests detect induced manipulation.
I perform this simulation first using subsamples from the entire set of firms and then
on subsamples of firms known to produce biased estimates of discretion?i.e., subsamples
with high growth (McNichols 2000; Kothari et al. 2005). With the simulation on all firms,
I compare the revenue and accrual models' power in detecting combinations of revenue
and expense manipulation. With the simulation on growth firms, I compare the specification
of the revenue and accrual models in the presence of extreme performance.
I repeat the following steps 250 times (either drawing firms from the entire population
or from the set of firms in each industry-year's highest quartile of earnings growth):

(1) Draw a random subsample of 100 firm-year observations without replacement.


(2) Simulate revenue manipulation by adding 1 percent of average total assets to the
change in revenues, the change in fourth-quarter revenues, and the receivables
accrual, and 1 percent times the gross margin percentage to current accruals of
these 100 firm-years; or simulate expense manipulation by adding 1 percent to
current accruals.
(3) Estimate the models using all observations except the 100 subsample firm-years.
(4) Use each model's coefficient estimates to calculate estimates of discretion for the
100 subsample firm-years.
(5) Calculate the mean estimate of discretion from each model and test whether the
mean is significantly greater than zero.

Statistics from the 250 subsamples form the basis of the tests. I report the mean and
standard error of the 250 estimates of discretion, as well as the percent of the 250 times
that the model rejects the null hypothesis of no manipulation. A rejection rate of 5 percent
is expected when manipulation is not introduced, and based on the 95 percent confidence
interval, an actual rejection rate below 2 percent or above 8 percent indicates that the test
is misspecified. When manipulation is introduced, however, the rejection rate should be 100
percent.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 707

My procedure differs from that of KLW in three ways. First, I simulate combinations
of revenue and expense manipulation to evaluate the models under different forms of earn
ings management. Second, I calculate accruals using items from the statement of cash flows
rather than the balance sheet. Hribar and Collins (2002) find that the error in the balance
sheet approach of estimating accruals is correlated with firms' economic characteristics. As
KLW note, this error not only reduces the models' power to detect earnings management,
but also has the potential to generate incorrect inferences about earnings manage
ment. Finally, I winsorize variables before, rather than after, estimating the models. This
ensures that each model's mean estimate of discretion is zero.
Table 3, Panel A, presents descriptive statistics from the simulation on subsamples
drawn from the entire sample of firms. The table presents estimates of discretionary reve
nues and accruals based on four combinations of induced manipulation: no manipulation,
revenue manipulation of 1 percent of assets, expense manipulation of 1 percent of assets,
and both revenue and expense manipulation of 1 percent of assets.

TABLE 3
Bias, Specification, and Power of Revenue and Accrual Models Using Simulated Revenue and
Expense Manipulation on All Firms

Panel A: Mean Bias and Standard Error of Discretionary Revenue and Accrual Estimates
% Manip: (Rev, Exp)_ (0%,0%) (0%,0%)
Model Mean Std. Dev.
Revenue (1) -0.05 0.55
Conditional Revenue (2) -0.06 0.64
Jones (3) 0.05 1.63
Modified Jones (4) 0.03 1.63
Dechow-Dichev (5) 0.21 1.66
Performance-Matched Modified Jones (6) 0.19 1.86
Panel B: Rejection Rates (Ha: Discretion > 0) for Combinations of
Expense Manipulation
% Manip: (Rev, Exp)_ (0%,0%) (1%,0%)
Model Rate Rate Rate Rate
Revenue (1) 0.8%** 23.6%** 0.8%**
Conditional Revenue (2) 1.2%** 24.0%** 1.2%** 24.0%**
Jones (3) 5.6% 6.8% 13.6%** 11.6%**
Modified Jones (4) 5.2% 7.2% 13.2%** 14.0%**
Dechow-Dichev (5) 7.6% 8.8%** 18.8%** 19.2%**
Performance-Matched Modified Jones (6) 4.4% 6.0% 9.2%*
*, ** Indicates the rejection rate is significantly different from 5 percent at the 0.05 and 0.01 levels
Table 3 presents statistics on discretionary revenues and discretionary accruals from 250 random
firms. "% Manip" is the percent of revenue or expense manipulation induced in each of 250 ran
of 100 firm-years?either 0 percent or 1 percent of average assets.
Panel A presents the mean ("Mean") and standard deviation ("Std. Dev.") of the 250 sample m
discretionary accrual or revenue estimates, expressed as a percent of average assets.
Panel B presents rejection rates ("Rate"), which is the percent of the 250 sample means that are
greater than zero (a = 0.05).
The revenue models (Revenue and Conditional Revenue) and accrual models (Jones, Modified Jon
Dechow-Dichev, and Performance-Matched Modified Jones) are described in the Appendix.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
708 Stubben

Because mean discretionary revenues and accruals in the entire sample are zero by
construction, the mean bias of each of the models should approximate zero. Table 3, Panel
A, confirms this expectation. Table 3, Panel A, also presents standard errors. The standard
errors from the revenue models are about one-third those of the accrual models. A model
that produces estimates with lower standard errors is more likely to detect manipulation
when it occurs.
The first column of Table 3, Panel B, reports results on the specification of the models
under the null hypothesis of no discretion. None of the models over-rejects the null hy
pothesis; however, the revenue models under-reject. Rejection rates for the revenue and
conditional revenue models are 0.8 and 1.2 percent. Rejection rates for the Jones, modified
Jones, Dechow-Dichev (DD), and performance-matched modified Jones (PM) models are
5.6, 5.2, 7.6, and 4.4 percent, respectively.
The next three columns of Table 3, Panel B present evidence on the power of the
models. The revenue and conditional revenue models detect revenue manipulation of 1
percent in 23.6 and 24.0 percent of samples. The rejection rates for the Jones, modified
Jones, DD, and PM models are a substantially lower 6.8, 7.2, 8.8, and 6.0 percent, respec
tively. By construction, the revenue models do not detect expense manipulation. The Jones,
modified Jones, DD, and PM models detect expense manipulation of 1 percent in 13.6,
13.2, 18.8, and 9.2 percent of samples, respectively. The revenue and conditional revenue
models detect a combination of revenue and expense manipulation in 23.6 and 24.0 percent
of samples. The Jones, modified Jones, DD, and PM models detect the same combina
tion of manipulation in only 11.6, 14.0, 19.2, and 11.2 percent of samples, respectively.
The low rejection rates of the PM model relative to the modified Jones model indicate that
performance matching reduces the power of accrual models.
Table 4 presents results from the simulation on firms in the highest quartile of earnings
growth. Panel A of Table 4 reveals that each of the six models produces a positive estimate
of discretion for growth firms with zero induced manipulation, which indicates a positive
bias for growth firms. However, the bias is smaller for the revenue models than for the
accrual models. The revenue and conditional revenue model estimates are 0.41 and 0.40
percent of assets, respectively; accrual model estimates are 2.39, 2.75, 3.71, and 2.07 per
cent of assets for the Jones, modified Jones, DD, and PM models, respectively. The larger
estimates for the accrual models are consistent with accruals other than receivables
not being explained by the change in revenues alone, and the factors omitted from the
models being correlated with growth. For example, it is likely that growth firms invest in
inventory beyond what would be predicted by the change in current revenues alone.
Results in Table 4, Panel A, indicate that the modified Jones model is more biased than
the Jones model. This finding is consistent with growth firms having more uncollected
credit sales, which are treated as discretionary in the modified Jones model. The DD model
exhibits the greatest bias of all models tested. Although performance-matching reduces the
bias of the modified Jones model, the bias remains more than five times the amount from
either of the revenue models. Performance matching by income as suggested by KLW does
not fully correct for growth-related model bias.
Table 4, Panel A, also presents standard errors across models. The standard errors from
the revenue models are less than half of those of the accrual models. Of the accrual mod
els, the PM model produces estimates with the largest standard errors. Thus, the reduction
in bias accomplished by performance matching comes at a cost in efficiency.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 709

TABLE 4
Bias, Specification, and Power of Revenue and Accrual Models Using Simulated Revenue and
Expense Manipulation on Growth Firms

Panel A: Mean Bias and Standard Error of Discretionary Revenue and Accrual Estimates
% Manip: (Rev, Exp) (0%,0%) (0%,0%)
Model Mean Std. Dev.
Revenue (1) 0.41 0.64
Conditional Revenue (2) 0.40 0.70
Jones (3) 2.39 1.99
Modified Jones (4) 2.75 2.07
Dechow-Dichev (5) 3.71 1.56
Performance-Matched Modified Jones (6) 2.07 2.41

Panel B: Rejection Rates (Ha: Discretion > 0) for Combinations of Induced Revenue and
Expense Manipulation
_% Manip: (Rev, Exp) (0%,0%) (1%,0%) (0%,1%) (!%,!%)
Model Rate Rate Rate Rate
Revenue (1) 7.3% 44.4%** 7.3% 44.4%**
Conditional Revenue (2) 7.2% 41.2%** 7.2% 41.2%**
Jones (3) 35.6%** 31.2%** 52.4%** 46.8%**
Modified Jones (4) 43.6%** 38.4%** 58.0%** 52.0%**
Dechow-Dichev (5) 68.8%** 59.6%** 84.4%** 76.4%**
Performance-Matched Modified Jones (6) 20.0%** 18.4%** 37.2%** 34.4%**

*, ** Indicates the rejection rate is significantly different from 5 percent at the 0.05 and 0.01 levels, respectively.
Table 4 presents statistics on discretionary revenues and discretionary accruals from 250 random samples of 100
growth firms. Growth firms are those in the highest quartile of change in earnings before extraordinary items
deflated by average total assets, where quartiles are determined by industry and year. "% Manip" is the percent
of revenue or expense manipulation induced in each of 250 random samples of 100 firm-years?either 0 percent
or 1 percent of average assets.
Panel A presents the mean ("Mean") and standard deviation ("Std. Dev.") of the 250 sample mean
discretionary accrual or revenue estimates, expressed as a percent of average assets.
Panel B presents rejection rates ("Rate"), which is the percent of the 250 sample means that are significantly
greater than zero (a = 0.05).
The revenue models (Revenue and Conditional Revenue) and accrual models (Jones, Modified Jones,
Dechow-Dichev, and Performance-Matched Modified Jones) are described in the Appendix.

The first column of Table 4, Panel B, reports results on the specification of the models
under the null hypothesis of no discretion.19 Only the revenue models produce well-specified
tests of manipulation for growth firms. Rejection rates for the revenue and conditional
revenue models are 7.3 and 7.2 percent, respectively. Each of the four accrual models over
rejects the null hypothesis of no manipulation. Rejection rates for the Jones, modified Jones,
DD, and PM models are 35.6, 43.6, 68.8, and 20.0 percent, respectively.

19 This analysis assumes zero discretionary revenues and accruals on average for growth firms. This does not,
however, assume no manipulation. Because models of discretion are estimated in cross-section, estimated ma
nipulation is relative to the industry-year average. Therefore, the assumption of this analysis is that growth firms,
on average, do not manipulate more than other firms in the same industry and year. To the extent this is not
true, I overstate the bias and misspecification of the models.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
710 Stubben

The next three columns of Panel B, Table 4 present evidence on the power of the
models. The revenue and conditional revenue models detect revenue manipulation of 1
percent in 44.4 and 41.2 percent of samples. The rejection rates for the Jones, modified
Jones, and DD models are 31.2, 38.4, and 59.6 percent, respectively. However, these rejec
tion rates largely reflect the general tendency of biased models to over-reject even when
no manipulation is induced. The PM model, which exhibits the least misspecification of
the accrual models, detects revenue manipulation in only 18.4 percent of samples. By
construction, the revenue models do not detect expense manipulation. The PM model de
tects expense manipulation of 1 percent in 37.2 percent of samples. The revenue and con
ditional revenue models detect a combination of revenue and expense manipulation in 44.4
and 41.2 percent of samples. The PM model detects the same combination of manipulation
in only 34.4 percent of samples.
In summary, simulation analysis on all firms indicates that the revenue models are more
likely than accrual models to detect an equal combination of revenue and expense manip
ulation. The simulation analysis on growth firms indicates that, although performance
matching improves the specification of the accrual models, only the revenue models are
well-specified with or without performance matching. Again, the revenue models are more
likely than the PM model to detect an equal combination of revenue and expense
manipulation.

Detection of Actual Revenue Manipulation


The second procedure I use to evaluate revenue and accrual models assesses the ability
of these models to detect revenue and expense manipulation in a sample of firms that are
known to have misstated their financial results. The known manipulators are a sample of
250 firm-years that were investigated by the SEC for accounting irregularities between 1995
and 2003.20
I divide sample firms into two groups: those that manipulated revenues and those that
manipulated expenses only. For each sample firm, I group observations into four time
periods: the manipulation period, the year before the manipulation, the year after the ma
nipulation, and all other years. I assume that sample firms overstate revenues and understate
expenses during the manipulation period.21 If the models are powerful, then mean discre
tionary revenue and accrual estimates should be significantly positive during the manipu
lation period. Assuming no manipulation took place the year prior to the manipulation
period, if the models are correctly specified, then mean discretionary revenue and accrual
estimates should not differ from zero.22
When studying SEC enforcement actions, one concern is selection bias. For this reason,
I present results after adjusting discretionary revenues and accruals of the SEC sample firms
by those of control firms chosen from the same industry and year. In the first case, I choose
a control firm for each sample firm from the same industry and year with the closest return
on assets, which is analogous to the performance matching approach suggested by KLW.

20 The sample was obtained by searching the EDGAR database at http://www.sec.gov. The initial search led to
enforcement actions for 265 firms (after excluding firms with enforcement actions on separate occasions) with
250 firm-years remaining after imposing data requirements. A similar attrition rate is found by Dechow et al.
(1996).
21 Consistent with an overstatement on average, Dechow et al. (1996) find positive discretionary accruals during
the manipulation period by firms subject to SEC enforcement actions.
22 This test of specification does not provide conclusive evidence, because it is possible that the sample firms used
discretionary accruals or revenues before the manipulation period. However, I assume that if this discretion was
sufficiently large, the SEC would have included the prior year in the enforcement action.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 711

In the second case, I choose a control firm from the same industry and year with the closest
revenue growth in the prior year. Prior research documents that firms targeted by SEC
enforcement actions tend to be firms with high revenue growth (Beneish 1999).
Table 5, Panel A, displays the distribution of sample firms through event time. Reve
nues were manipulated over 173 firm-years and expenses were manipulated over 77 firm
years, consistent with revenue manipulation being the most common form of earnings
management.
Panel B provides results based on the entire sample of both revenue and expense
manipulators. When choosing control firms based on profitability, evidence indicates that
the Jones, modified Jones, and DD models are misspecified for the sample of SEC firms.
Discretionary accruals before the manipulation period are estimated to be higher than con
trol firms by 6.30, 7.05, and 5.49 percent of assets, respectively. If these discretionary
accruals represented actual manipulation, however, then it is likely that they would have
been included in the alleged manipulation period. The evidence indicates the performance
matched revenue models are well specified.
The revenue and conditional revenue models detect discretion significantly higher than
that of control firms during the event period (1.80 and 1.63 percent of assets when matching
on profitability; 1.16 and 1.15 percent of assets when matching on growth). Each of the
performance-matched accrual models fails to detect discretion during the event period?
discretionary accruals are not significantly higher than those of control firms.
Panel C of Table 5 provides results based on the entire sample of firms that manipulated
revenues alone or revenues and expenses. Each of the performance-matched accrual models
is misspecified. Before the manipulation period, the Jones, modified Jones, and DD mod
els produce discretionary accruals that are significantly higher than those of control firms
by 9.07, 10.18, and 7.92 percent of assets when matching on profitability and by 6.86, 6.96,
and 5.47 percent of assets when matching on growth. The revenue models are well specified.
Only the revenue models detect manipulation during the event period. Discretionary
revenues from the revenue and conditional revenue models are significantly higher than
those of control firms by 2.19 and 1.82 percent of assets when matching on profitability
and 1.56 and 1.43 percent of assets when matching on growth. The performance-matched
accrual models fail to detect manipulation.
Panel D of Table 5 provides results based on the sample of firms that manipulated
expenses. As expected, the performance-matched revenue models do not detect discretion
in the sample of firms that did not manipulate revenues. However, neither do the accrual
models. Discretionary accruals during the manipulation period are not significantly different
than those of control firms. Because the sample studied in Table 5, Panel D is restricted to
expense manipulators, the accrual models should be more powerful than the revenue mod
els. However, neither the accrual nor revenue models detect manipulation by this subset of
SEC firms.

V. CONCLUSION
This study provides evidence on the reliability of discretionary revenues and various
measures of discretionary accruals by assessing their ability to detect both simulated and
actual manipulation. The results indicate that the revenue model is less biased and better
specified than accrual models, such that estimates from revenue models could be useful as
a measure of revenue management or as a proxy for earnings management.
Although revenue models do not detect discretion in expenses, findings suggest that
accrual models have difficulty detecting discretion in expenses as well. The state-of-the-art

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
712 Stubben

TABLE 5
Detection of Revenue and Expense Manipulation by Firms Subject to
SEC Enforcement Actions

Panel A: Number of Sample Firm Observations through Event Time


No Revenue Revenue
Event Year Manipulation Manipulation
-1 25 54
0 77 173
1 28 69
Total 130 296

Panel B: Mean Discretionary Revenues and Accruals by Event Year?All Firms


Event Year -1 Event Year 0 Event Year +1
Mean t-stat Mean t-stat Mean t-stat
Adjusted for Control Firms
(matched on profitability)
Revenue (1) 0.86 0.90 1.80 2.99** -1.14 -1.10
Conditional Revenue (2) 1.71 1.72 1.63 2.93** -1.10 -1.02
Jones (3) 6.30 2.57* 0.89 0.70 -4.98 -2.32*
Modified Jones (4) 7.05 2.89** 1.39 1.08 -5.07 -2.35*
Dechow-Dichev (5) 5.49 2.40* 0.92 0.68 -3.85 -1.52
Adjusted for Control Firms
(matched on growth)
Revenue (1) 0.36 0.35 1.16 2.02* -2.17 -2.36*
Conditional Revenue (2) 1.24 1.07 1.15 1.97* -2.10 -2.01*
Jones (3) 3.95 1.57 -0.20 -0.11 -8.43 -3.50**
Modified Jones (4) 4.16 1.66 -0.43 -0.25 -8.94 -3.72**
Dechow-Dichev (5) 2.61 0.95 -0.92 -0.59 -3.81 -1.57
Panel C: Mean Discretionary Revenues and Accruals by Event Year?Revenue Manipulators
Event Year -1 Event Year 0 Event Year +1
Mean t-stat Mean t-stat Mean t-stat

Adjusted for Control Firms


(matched on profitability)
Revenue (1) 1.32 0.98 2.19 2.75** -2.04 -1.51
Conditional Revenue (2) 1.86 1.39 1.82 2.51** -1.95 -1.39
Jones (3) 9.07 2.85** 0.77 0.48 -5.37 -1.92
Modified Jones (4) 10.18 3.21** 1.55 0.95 -5.50 -1.95
Dechow-Dichev (5) 7.92 2.64* -0.02 -0.01 -3.68 -1.12
Adjusted for Control Firms
(matched on growth)
Revenue (1) 0.06 0.05 1.56 2.14* -1.67 -1.58
Conditional Revenue (2) 0.11 0.08 1.43 1.89* -1.77 -1.45
Jones (3) 6.86 2.30* -1.11 -0.54 -9.05 -3.02**
Modified Jones (4) 6.96 2.32* -1.33 -0.64 -9.53 -3.19**
Dechow-Dichev (5) 5.47 1.78 -0.26 -0.13 -5.62 -1.78
(continued on next page)

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 713

TABLE 5 (continued)

Panel D: Mean Discretionary Revenues and Accruals by Event Year?Expense Manipulators


Event Year -1 Event Year 0 Event Year +1
Mean t-stat Mean t-stat Mean t-stat
Adjusted for Control Firms
(matched on profitability)
Revenue (1) 0.52 0.47 0.91 1.16 1.02 0.81
Conditional Revenue (2) 1.95 1.46 1.14 1.41 0.95 0.70
Jones (3) -0.66 -0.21 1.17 0.59 -4.02 -1.41
Modified Jones (4) -0.80 -0.26 1.06 0.53 -4.00 -1.44
Dechow-Dichev (5) -1.76 -0.72 2.73 1.59 -4.43 -1.24
Adjusted for Control Firms
(matched on growth)
Revenue (1) 1.05 0.59 0.26 0.29 -3.41 -1.86
Conditional Revenue (2) 3.62 1.56 0.53 0.59 -2.95 -1.47
Jones (3) -2.43 -0.54 1.96 0.64 -6.88 -1.75
Modified Jones (4) -2.01 -0.46 1.69 0.54 -7.46 -1.90
Dechow-Dichev (5) -4.83 -0.89 -2.23 -0.99 0.61 0.20
*, ** Indicates a significant difference from zero at the 0.05 and 0.01 levels, respectively, using a one-sided te
in event year 0 and a two-sided test otherwise.
Table 5 presents mean discretionary revenue and accrual estimates before, during, and after SEC enforcement
actions.
"Revenue Manipulation" indicates whether the alleged violations involved a revenue misstatement. Event years
include -1, the year preceding the first year of the manipulation, 0, one or more years during the manipulation,
and 1, the first year after the manipulation.
Panels B through D present mean ("Mean") estimates of discretion from models described in the Appendix,
expressed as a percent of average assets. Mean estimates presented under "Adjusted for Control Firms (matched
on profitability)" ("Adjusted for Control Firms (matched on growth)") use discretionary revenue and accrual
estimates after adjusting for a control firm from the same industry and year with the closest return on assets
(revenue growth). The t-statistics represent a paired t-test of discretionary revenues or accruals of sample firms
versus control firms.

performance-matched modified Jones model (Kothari et al. 2005) detects simulated expense
manipulation in only 9.2 percent of samples of firms, and it fails to detect expense manip
ulation by firms subject to expense-related SEC enforcement actions. Still, the success of
the revenue model at detecting earnings management depends on the relative frequency
of revenue versus expense manipulation. For equal amounts of simulated revenue and ex
pense manipulation across the entire sample, the revenue model outperforms each of the
accrual models. The revenue model also detects earnings management by firms subject to
SEC enforcement actions, but the performance-matched accrual models do not. Overall, the
revenue model outperforms accrual models both in detecting and failing to detect earnings
management, as appropriate. Thus, revisiting research settings with the revenue model could
shed light on whether significant results were driven by misspecification of accrual models
or whether insignificant results were driven by the accrual models' lack of power.
Measures of discretionary revenues can also be useful by providing evidence on how
firms manage earnings or for studying revenue management. On the whole, relatively little
research has been conducted in the area of discretion in revenues. Although revenues is a
logical first step in examining individual components of earnings, future studies could model
discretion in various expense components of earnings.
Finally, this study has implications for studies that use accrual models. First, the Jones
model (Jones 1991) exhibits better specification than the modified Jones model (Dechow

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
714 Stubben

et al. 1995), which suggests including reported revenues, rather than cash revenues, in
accrual models. Second, the Dechow-Dichev model (Dechow and Dichev 2002; McNichols
2002), which was originally developed to estimate earnings quality, exhibits greater mis
specification than other accrual models when used to estimate discretionary accruals. Last,
separating fourth-quarter revenues and allowing the relation between revenues and accruals
to vary across firms could be applied to accrual models to improve their performance.

APPENDIX
SUMMARY OF DISCRETIONARY REVENUE AND ACCRUAL ESTIMATES
Revenue Model

Discretion = AARit - d - ?xARl-3it - ?2AR4ir

Conditional Revenue Model


Discretion = AARit - a - ^ARit - ?2ARit X SIZEit - ?3ARit X AGEit - $4ARit
X AGE_SQit - ?5ARit X GRR-PU - ?6ARit X GRRJ*U - ?7A/?,
X GRMit - $sARit X GRMSQir

Accrual Models
Jones Model (Jones 1991)
Discretion = ACit - a - ?xARit - ?2PPEir

Modified Jones Model (Dechow et ah 1995)


Discretion = ACit - a - ?i(ARit - AARit) - ?2PPEir

Dechow-Dichev (DD) Model (Dechow and Dichev 2002; McNichols 2002)


Discretion = ACit - a - ?xARit - ?2PPEit - $3CFOit-x - &4CFOit - ?5CFOit+l.

Performance-Matched (PM) Modified Jones Model (Kothari et ah 2005)


Discretion = ACit - a - ?x(ARit - AARit) - ?2PPEit, less the same measure for the
firm from the same industry and year with the closest return on assets.

where:

AR = end of fiscal year accounts receivable;


AC = annual current accruals = earnings before extraordinary items - cash from
operations;
R = annual revenues;
Rl-3 = revenues of the first three quarters;
R4 = revenues of the fourth quarter;
PPE = end of fiscal year gross property, plant, and equipment;
CFO = cash from operations;
SIZE = natural log of total assets at end of fiscal year;
AGE = age of firm (years);
GRR-P = industry-median-adjusted revenue growth (= 0 if negative);
GRR-N = industry-median-adjusted revenue growth (= 0 if positive);
GRM = industry-median-adjusted gross margin at end of fiscal year;
-SQ = square of variable; and
A = annual change.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 715

REFERENCES
Barth, M., and S. Kallapur. 1996. The effects of cross-sectional scale differences on regression results
in empirical accounting research. Contemporary Accounting Research 13: 527-567.
-, W. Beaver, J. Hand, and W. Landsman. 2005. Accruals, accounting-based valuation models,
and the prediction of equity values. Journal of Accounting, Auditing & Finance 20: 311-345.
Beatty, A., S. Chamberlain, and J. Magliolo. 1995. Managing financial reports of commercial banks:
The influence of taxes, regulatory capital and earnings. Journal of Accounting Research 33:
195-212.
Beaver, W., C. Eger, S. Ryan, and M. Wolfson. 1989. Financial reporting, supplemental disclosures,
and bank share prices. Journal of Accounting Research 27: 157-178.
-, and E. Engel. 1996. Discretionary behavior with respect to allowances for loan losses and
the behavior of security prices. Journal of Accounting and Economics 22: 177-206.
-, and M. McNichols. 1998. The characteristics and valuation of loss reserves of property
casualty insurers. Review of Accounting Studies 3: 73-95.
Beneish, M. 1999. The detection of earnings manipulation. Financial Analysts Journal (Sep/Oct):
24-36.
-. 2001. Earnings management: A perspective. Managerial Finance 27: 3-17.
Bernard, V., and D. Skinner. 1996. What motivates managers' choice of discretionary accruals? Journal
of Accounting and Economics 22: 313-325.
Burgstahler, D., and I. Dichev. 1997. Earnings management to avoid earnings decreases and losses.
Journal of Accounting and Economics 24: 99-126.
-, and M. Eames. 2006. Management of earnings and analysts' forecasts to achieve zero and
small positive earnings surprises. Journal of Business, Finance & Accounting 33: 633-652.
Callen, J., S. Robb, and D. Segal. 2008. Revenue manipulation and restatements by loss firms. Au
diting: A Journal of Practice & Theory 27: 1-29.
Caylor, R. 2009. Strategic revenue recognition to achieve earnings benchmarks. Journal of Accounting
and Public Policy.
Collins, J., D. Shackelford, and J. Wahlen. 1995. Bank differences in the coordination of regulatory
capital, earnings and taxes. Journal of Accounting Research 33: 263-291.
Dechow, R, and R. Sloan. 1991. Executive incentives and the horizon problem: An empirical inves
tigation. Journal of Accounting and Economics 14: 51-89.
-,-, and A. Sweeney. 1995. Detecting earnings management. The Accounting Review 70:
193-225.
-,-, and-. 1996. Causes and consequences of earnings manipulation: An analysis
of firms subject to enforcement actions by the SEC. Contemporary Accounting Research 13:
1-36.
-, S. R Kothari, and R. Watts. 1998. The relation between earnings and cash flow. Journal of
Accounting and Economics 25: 131-168.
-, and I. Dichev. 2002. The quality of accruals and earnings: The role of accrual estimation
errors. The Accounting Review 11 (Supplement): 35-59.
-, S. Richardson, and I. Tuna. 2003. Why are earnings kinky? An examination of the earnings
management explanation. Review of Accounting Studies 8: 355-384.
-, and C. Schrand. 2004. Earnings Quality. Charlottesville, VA: The Research Foundation of
CFA Institute.
DeFond, M., and J. Jiambalvo. 1994. Debt covenant violation and manipulation of accruals. Journal
of Accounting and Economics 17: 145-176.
Dhaliwal, D., C. Gleason, and L. Mills. 2004. Last change earnings management: Using the tax
expense to achieve earnings targets. Contemporary Accounting Research 21: 431-459.
Dopuch, N., R. Mashruwala, C. Seethamraju, and T. Zach. 2005. Accrual determinants, sales changes
and their impact on empirical accrual models. Working paper, Washington University in St.
Louis.
Fama, E., and J. Macbeth. 1973. Risk, return and equilibrium: empirical tests. The Journal of Political
Economy 81: 607-636.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
716 Stubben

Feroz, E., K. Park, and V. Pastena. 1991. The financial and market effects of the SEC's Accounting
and Auditing Enforcement Releases. Journal of Accounting Research 29 (Supplement): 107
142.
Financial Accounting Standards Board. 1987. Statement of Cash Flows. Statement of Financial Ac
counting Standards No. 95. Norwalk, CT: FASB.
Guay, W., S. P. Kothari, and R. Watts. 1996. A market-based evaluation of discretionary accrual
models. Journal of Accounting Research 34 (Supplement): 83-105.
Healy, P. 1985. The effect of bonus schemes on accounting decisions. Journal of Accounting and
Economics 1: 85-107.
-, and K. Palepu. 1990. Effectiveness of accounting-based dividend covenants. Journal of Ac
counting and Economics 12: 97-124.
-, and J. Wahlen. 1999. A review of the earnings management literature and its implications
for standard-setting. Accounting Horizons 13: 365-383.
Hribar, P., and D. Collins. 2002. Errors in estimating accruals: Implications for empirical research.
Journal of Accounting Research 40: 105-135.
Jacob, J., and B. Jorgensen. 2007. Earnings management and accounting income aggregation. Journal
of Accounting and Economics 43: 369-390.
Jones, J. 1991. Earnings management during import relief investigations. Journal of Accounting Re
search 29: 193-228.
Kang, S., and K. Sivaramakrishnan. 1995. Issues in testing earnings management and an instrumental
variable approach. Journal of Accounting Research 33: 353-367.
Kothari, S. P., A. Leone, and C. Wasley. 2005. Performance matched discretionary accrual measures.
Journal of Accounting and Economics 39: 163-197.
Marquardt, C, and C. Wiedman. 2004. How are earnings managed? An examination of specific
accruals. Contemporary Accounting Research 21: 461-491.
Matsumoto, D. 2002. Management's incentives to avoid earnings surprises. The Accounting Review
77: 483-514.
McNichols, M., and P. Wilson. 1988. Evidence of earnings management from the provision for bad
debts. Journal of Accounting Research 26 (Supplement): 1-31.
-. 2000. Research design issues in earnings management studies. Journal of Accounting and
Public Policy 19: 313-345.
-. 2002. Discussion of the quality of accruals and earnings: The role of accrual estimation
errors. The Accounting Review 11 (Supplement): 61-69.
Moyer, S. 1990. Capital adequacy ratio regulations and accounting choices in commercial banks.
Journal of Accounting and Economics 13: 123-154.
Nelson, K. 2000. Rate regulation, competition and loss reserve discounting by property-casualty in
surers. The Accounting Review 15: 115-138.
Petersen, M., and R. Rajan. 1997. Trade credit: Theories and evidence. Review of Financial Studies
10: 661-691.
Petroni, K. 1992. Optimistic reporting in the property-casualty insurance industry. Journal of Ac
counting and Economics 15: 485-508.
-, S. Ryan, and J. Wahlen. 2000. Discretionary and non-discretionary revisions of loss reserves
by property-casualty insurers: Differential implications for future profitability, risk and market
value. Review of Accounting Studies 5: 95-125.
Phillips, J., M. Pincus, and S. Rego. 2003. Earnings management: New evidence based on deferred
tax expense. The Accounting Review 78: 491-521.
Plummer, E., and D. Mest. 2001. Evidence on the management of earnings components. Journal of
Accounting, Auditing & Finance 16: 301-323.
Richardson, S., R. Sloan, M. Soliman, and I. Tuna. 2006. The implications of accounting distortions
and growth for accruals and profitability. The Accounting Review 81: 713-743.
Scholes, M., G. Wilson, and M. Wolfson. 1990. Tax planning, regulatory capital planning, and finan
cial reporting strategy for commercial banks. Review of Financial Studies 3: 625-650.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms
Discretionary Revenues as a Measure of Earnings Management 717

Sopranzetti, B. 1998. The economics of factoring accounts receivable. Journal of Economics and
Business 50: 339-359.
Stubben, S. 2006. The use of discretionary revenues to meet earnings and revenue targets. Doctoral
dissertation, Stanford University.
Sweeney, A. 1994. Debt-covenant violations and managers' accounting responses. Journal of Ac
counting and Economics 17: 281-308.
Thomas, J., and X. J. Zhang. 2000. Identifying unexpected accruals: a comparison of current ap
proaches. Journal of Accounting and Public Policy 19: 347-376.
Turner, L., J. R. Dietrich, K. Anderson, and A. Bailey. 2001. Accounting restatements. Working paper,
United States Securities and Exchange Commission, The Ohio State University, Georgetown
University, and University of Illinois at Urbana-Champaign.
Whalen, J. 1994. The nature of information in commercial bank loan loss disclosures. The Accounting
Review 69: 455-478.

The Accounting Review March 2010


American Accounting Association

This content downloaded from 111.223.252.27 on Mon, 23 Oct 2017 04:58:28 UTC
All use subject to http://about.jstor.org/terms

Potrebbero piacerti anche