The Balance Sheet as an Earnings Management Constraint

Author(s): Jan Barton and Paul J. Simko

Source: The Accounting Review, Vol. 77, Supplement: Quality of Earnings Conference (2002),
pp. 1-27
Published by: American Accounting Association
Stable URL: http://www.jstor.org/stable/3203322
Accessed: 02/12/2010 20:42

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=aaasoc.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

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 [email protected].

American Accounting Association is collaborating with JSTOR to digitize, preserve and extend access to The
Accounting Review.

http://www.jstor.org
THEACCOUNTING REVIEW
Vol. 77
2002
Supplement
pp. 1-27

The Balance Sheet as an Earnings

Constraint
Management
Jan Barton
Emory University
Paul J. Simko
University of Virginia
ABSTRACT: The balancesheet accumulatesthe effects of previousaccount-
ing choices, so the level of net assets partlyreflectsthe extent of previous
earningsmanagement.Wepredictthat managers'abilityto optimistically bias
earningsdecreases with the extentto which the balance sheet overstatesnet
assets relativeto a neutralapplicationof GAARTo test this prediction,we
examinethe likelihoodof reportingvariousearningssurprisesfor 3,649 firms
during1993-1999. Consistentwithour prediction,we findthatthe likelihood
of reportinglargerpositiveor smallernegativeearningssurprisesdecreases
withourproxyfor overstatednet asset values.

between incomestatementand balancesheet; earn-

Keywords:articulation
ingsmanagementconstraints; measurementandrecognitionguid-
ance; earningssurprises;analysts'forecasts.
Dataare availablefromsources identifiedin the paper.
Data Availability:

I. INTRODUCTION
W x e examine the extent to which GenerallyAcceptedAccountingPrinciples
(GAAP)andtheirimplementation guidelinesconstrainmanagersfromoptimis-
ticallybiasingearnings.Becausethe balancesheet accumulatesthe effectsof
previous accountingchoices, the level of net assets partly reflects the extent of previous
earnings management.We predict that managers'ability to optimisticallybias earnings
decreases with the extent to which net assets are already overstated on the balance
sheet.' To test our prediction, we focus on a particularlystrong earnings management

We use the term "overstated" to describe net assets measured and recorded at values exceeding those that would
have been reported under a neutral application of GAAP, without implying a violation of GAAP.

We appreciate the helpful comments of Katherine Schipper (the editor), an anonymous referee, Linda Bamber,
George Benston, Walt Blacconiere, Paul Irvine, Connie Kertz, Steven Schmitt, Greg Waymire, and seminar partic-
ipants at The University of Alabama, University of Arkansas, Emory University, Georgia State University, Indiana
University, University of Virginia, The Accounting Review Conference on Quality of Earnings, and The University
of Texas at Austin's 2002 BMAS Conference. We also thank Ron Harris and Xiao Lan Wang for research assistance,
the Goizueta Business School for generous funding, and I/B/E/S International Inc. for access to data on analysts'
earnings forecasts.

1
2 The AccountingReview, 2002 Supplement

incentive-meeting or beating analysts' earnings expectations. Meeting or beating expec-

tations is not always the result of optimistic bias in financial reporting, so our empirical
analyses control for other reasons behind an earnings surprise. Our analyses suggest that
managers are less likely to report a predetermined earnings surprise by optimistically bi-
asing earnings when their firms' net assets are overstated.
The articulation between the income statement and the balance sheet ensures that ac-
crualsreflectedin earningsalso are reflectedin net assets. Therefore,an optimisticbias in
earnings implies net assets measured and recorded temporarily at values exceeding those
based on a neutral application of GAAP. Managers' generous assumptions about recognition
and measurement in one period reduce their ability to make equally generous assumptions
in later periods, if managers want to stay within the guidance provided by accounting
regulators and professional groups.2 Therefore, managers' ability to optimistically bias earn-
ings decreases with the extent to which net assets are already overstated.
We use the beginning balance of net operating assets relative to sales as a proxy for
managers' previous biased reporting choices. This proxy is consistent with overstated net
assets being less efficient at generating a given level of sales, all else equal. If this proxy
is valid, then firms with larger levels of net operating assets relative to sales will have
reported larger cumulative levels of income-increasing accruals in the past. In our sample
of 3,649 nonfinancial, nonregulated firms during 1993-1999, we find that firms with larger
levels of net operating assets (relative to sales) reported larger cumulative levels of abnormal
accruals in the previous 20 quarters, consistent with prior income-increasing earnings man-
agement leading to overstated net assets.
We then examine the association between our sample firms' quarterly earnings surprises
and the beginning balance of net operating assets relative to sales. We focus on earnings
surprises because "[p]erhaps the single most important cause [of earnings management] is
the pressure imposed on management to meet analysts' earnings projections" (Johnson
1999). Investors have become especially unforgiving of firms that fail to meet earnings
expectations(Skinnerand Sloan 2001).3 Previous researchdocumentsan increasinginci-
dence of reported earnings that just meet or slightly beat analysts' earnings forecasts (Brown
2001; Burgstahler and Eames 2001; Matsumoto 2002), an outcome that is apparently
achieved in part through earnings management (Payne and Robb 2000; Burgstahler and
Eames 2001; Matsumoto 2002). If missing expectations is costly and managing earnings
toward expectations is common, then why do some managers miss expectations even by a
small amount? In fact, why do they meet expectations rather than beat them altogether?
We suggest some managers do so because, all else equal, they have limited discretion to
repeatedly bias earnings upward. Specifically, we predict that managers' ability to report
larger positive or smaller negative earnings surprises decreases with the extent to which
net assets are already overstated on the balance sheet.
We model the level of earnings surprise as a function of managers' previous recognition
and measurement decisions, as reflected in the beginning balance of net operating assets
relative to sales. Our model controls for other constraints on earnings management, man-
agerial incentives to meet or beat analysts' forecasts, and firm performance and size. We
estimate the model using a generalized ordered logit regression on 35,950 quarterly earnings
surprises for our sample. Consistent with our prediction, we find that firms with larger

2
Departing from GAAP may result in substantial costs to managers and their firms (Dechow et al. 1996).
For example, Compaq Computer, Hewlett-Packard, Merck, Sears, Starbucks, and UAL all missed their consensus
analyst forecasts by 1? per share at least once during 1993-1999, and suffered an average size-adjusted cumulative
abnormal return of about -12 percent over the three days surrounding their earnings announcements.
Bartonand Simko-Earnings ManagementConstraint 3

beginning balances of net operating assets relative to sales are less likely to report a pre-
determined earnings surprise. These findings are robust to using alternative econometric
techniques, separating net operating assets into their current and long-term components,
controlling for industry and fiscal-quarter effects, using alternative definitions of earning
surprises, and excluding loss firms or firms reporting special items.
Our paper extends research on earnings management in three ways. First, while we do
not measure directly the amount of financial reporting discretion actually available to man-
agers, our approach provides a basis for doing so in future research. Second, prior research
examines how managers trade off financial reporting discretion available across several
balance sheet accounts when meeting various objectives (e.g., Beatty et al. 1995; Hunt et
al. 1996; Gaver and Paterson 1999). We extend this literature by providing evidence that
the balance sheet accumulates the effects of prior optimism in financial reporting, reducing
managers' ability to optimistically bias future earnings. Therefore, consistent with Hunt et
al. (1996), our findings suggest that managers likely trade off available financial reporting
discretion over time. Last, prior research provides evidence that managing earnings to meet
expectations is widespread (Degeorge et al. 1999; Payne and Robb 2000; Abarbanell and
Lehavy 2001la; Burgstahler and Eames 2001; Matsumoto 2002), but assumes that the ability
to manage earnings is random or even constant across firms. Our results suggest that the
ability to report a predetermined earnings surprise varies across firms as a function of
previous optimism in financial reporting.
Section II develops our main hypothesis. Section III describes our research design.
Specifically, this section explains the sample selection criteria, reports the frequencies and
market reactions associated with various levels of earnings surprise, describes the charac-
teristics of sample firms, discusses our proxy for overstated net asset values, and develops
our regression model. Section IV describes our main empirical results and sensitivity tests.
Finally, Section V offers some concluding remarks.

II. HYPOTHESIS
The accruals basis of accounting requires managers to report the effects of economic
transactions based on expected cash realizations (SFAC No. 1, FASB 1978; SFAC No. 6,
FASB 1985). Although GAAP requires managers to make numerous judgments and as-
sumptions to report their firms' performance under accrual accounting, managers do not
have unlimited discretion in doing so. For example, accounting regulators like the Securities
and Exchange Commission and the Financial Accounting Standards Board, along with
professional groups like the American Institute of Certified Public Accountants, provide
detailed guidelines for implementing measurement, recognition, and disclosure rules. Au-
ditors and the courts help enforce these guidelines.
Because of the reversing nature of accrual accounting, managers' biased estimates and
judgments in one period reduce their ability to make similarly biased estimates and judg-
ments in subsequent periods (Hunt et al. 1996; Abarbanell and Lehavy 2001a). The artic-
ulation between the income statement and the balance sheet ensures that biased assumptions
reflected in earnings are also reflected in net asset values. Therefore, we predict that man-
agers' ability to optimistically bias earnings decreases with the extent to which net assets
are already overstated on the balance sheet, relative to what their values would have been
under a neutral implementation of GAAP.
Prior research shows that overstated assets and revenues are the most common reason
for restatements of financial statements (Palmrose et al. 2001), Securities and Exchange
Commission's accounting enforcement actions (Feroz et al. 1991; Dechow et al. 1996), and
lawsuits against auditors (St. Pierre and Anderson 1984; Lys and Watts 1994). Prior research
4 The AccountingReview, 2002 Supplement

also shows that managers may understate assets temporarily to regain the ability to opti-
misticallybias earningsin futureperiods (Moehrle2002).
Managershave various incentives to bias earnings (see Fields et al. 2001). However,
we focus on one particularlystrongincentive-meeting or beating analysts'earningsfore-
casts. Dechow and Skinner (2000) argue that the bull marketof the late 1990s and the
increaseduse of stock options for managerialcompensationhave increasedmanagers'in-
centives to manipulate earnings to preserve high stock valuations. Consistent with this
argument, Barth et al. (1999) provide evidence that firms with patterns of increasing earn-
ings have higher price-to-earningsratios, and that these ratios decline significantlywhen
firmsbreakthe patternby reportingearningsdecreases.Moreover,Skinnerand Sloan (2001)
report significant negative returns for firms that fail to meet analysts' earnings forecasts.
Finally,evidence also suggests widespreadearningsmanagementto meet or beat analysts'
forecasts (e.g., Payne and Robb 2000; Burgstahler and Eames 2001; Matsumoto 2002).
Based on these findings, we predict that managers will use available financial reporting
discretion to report higher levels of earnings surprises, all else equal. Therefore, we test
the following hypothesis:

Ha: The likelihood of reporting larger positive or smaller negative earnings surprises
decreases with the extent to which net assets are already overstated on the balance
sheet.

III. RESEARCH DESIGN

Sample
Our initial sample contains all firm-quarterson both the Compustatand I/B/E/S
databaseswith complete data for our test variablesduring 1993-1999. We consider this
period because pressure to meet analysts' forecasts appears to be a recent phenomenon
(Levitt 1998; Brown 2001; Matsumoto 2002), and because Abarbanell and Lehavy (2001b)
suggest that differences between pre- and post-1993 I/B/E/S forecast errors may be due
to changes in the way I/B/E/S calculates actual earnings per share (EPS). We exclude
utilities and financial services firms (two-digit SIC codes 49 and 60-67) because they are
subject to regulatory requirements that would unnecessarily complicate our research design.
Our final sample consists of 35,950 quarterly observations for 3,649 firms.
We define an earningssurprise(SURPRISE)as the I/B/E/S actual EPS for quartert
less the consensus analyst EPS forecast for quartert, both roundedto the nearestpenny.
The consensus forecast is the average of analysts' most recent forecasts of EPS for quarter
t available on I/B/E/S prior to the earnings announcement for quarter t (Brown and Kim
1991). To keep our analyses tractable, we combine firms missing the consensus forecast
by 50 or more (i.e., SURPRISE < -50) into one category and firms beating the fore-
cast by 50 or more (i.e., SURPRISE - 50) into another. SURPRISE thus consists of 11
categories coded sequentially from -5 to 5.
Table 1 reportscharacteristicsof quarterlyearningssurprisesfor our sample.Here and
in our remaininganalyses,we winsorize all variablesat the upperand lower one percentile
of their distributions. Across levels of SURPRISE, Panel A of Table 1 reports the means
of actual EPS, consensus EPS forecast, and forecast error (i.e., actual EPS less consensus
EPS forecast) as a percentage of the consensus EPS forecast. Panel A shows these three
variables increasing in the level of SURPRISE. The panel also shows that firms missing
(beating) earnings expectations by 10 per share reported EPS averaging about 6.3 percent
lower (6.7 percent higher) than the consensus forecast. On the other hand, firms missing
(beating) earnings expectations by 50 or more reported EPS averaging about 31 percent
lower (20.4 percent higher) than the consensus forecast.
Bartonand Simko-Earnings ManagementConstraint 5

TABLE 1
Characteristics of Quarterly Earnings Surprises

Panel A: Mean of Actual EarningsPer Share (EPS), ConsensusAnalysts'Forecasts,and Earnings

Surpriseas a Percentageof Forecasts
Consensus SURPRISE/
SURPRISE n Actual EPS EPS Forecast ConsensusEPS Forecast
< -5? 5,054 2? 170 -31.0%
-4 871 16 20 -16.1
-3 1,196 17 20 -13.5
-2 1,707 20 22 -8.3
-1 2,811 23 24 -6.3
0 8,426 22 22 0.6
1 5,163 25 24 6.7
2 3,050 27 25 10.7
3 1,927 28 25 12.2
4 1,236 28 24 13.9
- 5 4,509 38 28 20.4
All firm-quarters 35,950 220 230 -0.4%

Panel B: AnnualFrequencyof Firm-QuartersReportinga Given EarningsSurprise

Fiscal Year
SURPRISE 1993 1994 1995 1996 1997 1998 1999 Total
' -5? 16.4% 13.8% 17.1% 14.0% 13.3% 13.3% 11.9% 14.1%
-4 3.7 2.7 2.8 2.3 1.8 2.4 1.9 2.4
-3 4.2 3.9 3.7 3.1 2.8 3.6 2.5 3.3
-2 6.2 5.3 4.8 5.4 4.6 4.2 3.4 4.8
-1 9.8 8.6 7.7 7.8 7.6 7.7 6.4 7.8
0 25.4 24.5 24.0 24.8 24.1 22.5 20.1 23.4
1 10.4 13.4 13.7 13.7 15.6 15.9 16.0 14.4
2 6.6 7.6 7.2 8.5 9.1 9.3 10.0 8.5
3 4.2 4.7 4.8 4.9 6.2 5.6 6.4 5.4
4 2.4 3.3 3.1 3.3 3.2 3.5 4.8 3.4
- 5 10.7 12.2 11.1 12.2 11.7 12.0 16.6 12.5
Total 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0%
n 3,806 4,371 4,743 5,305 5,757 6,175 5,793 35,950
(Continuedon nextpage)
6 The Accounting Review, 2002 Supplement

TABLE 1 (Continued)

Panel C: Mean Three-Day Cumulative Abnormal Return (CAR) around Quarterly Earnings
Announcement Datea
t-Statistic for
SURPRISE n CAR Ho: CAR = 0 CAR / SURPRISE

' -5? 4,877 -2.81% -25.02 *** -0.27%

-4 846 -1.84 -7.60*** -0.46
-3 1,162 -1.98 -9.24*** -0.66
-2 1,670 -1.41 -8.44*** -0.71
-1 2,740 -1.16 -8.95 *** -1.16
0 8,221 -0.01 -0.12 NA
1 4,983 1.04 10.49 *** 1.04
2 2,930 2.00 15.90 *** 1.00
3 1,857 2.17 13.47 *** 0.72
4 1,177 2.81 13.28 *** 0.70

- 5 4,278 3.18 28.81 *** 0.42

All firm-quarters 34,741 0.26% 6.51 *** 0.19%

*** denotessignificantat the 0.01 level, based on two-tailedtests.

a
We calculatethe cumulativeabnormalreturn(CAR) as the with-dividendreturnover the threedays surrounding
the earningsannouncementfor quartert, less the returnover the same period on a size-matchedportfolio.To
to decile portfoliosbased on market
calculatethe size-matchedportfolio'sreturn,we allocate all firm-quarters
value of commonequity at the beginningof quartert, and then we calculatea value-weightedreturnfor each
portfoliofor the three-dayperiodsurroundingeach earningsannouncement.
The sample consists of 35,950 firm-quarterspertainingto 3,649 firms included in both the Compustatand
I/B/E/S databaseswith completedata for our primarytests over 1993-1999, excludingutilities and financial
services firms (two-digitSIC codes 49 and 60-67). Panel C also excludes 1,209 firm-quarters with incomplete
datato calculateabnormalreturns.SURPRISEis the I/B/E/S actualEPS for quartert less the consensusforecast
for quartert, both roundedto the nearestpenny.The consensusforecastis the meanof analysts'most recentEPS
forecastsfor quartert availableon I/B/E/S priorto the earningsannouncementfor quartert. In PanelsA andC,
n denotesthe numberof firm-quarters in each categoryof SURPRISE;in PanelB, n denotesthe numberof firm-
quartersper fiscal year.

Panel B of Table 1 shows that the frequency of firms beating earnings expectations
increased over the sample period, consistent with findings reported in Brown (2001) and
Matsumoto (2002). A Pearson x2 test (not tabulated) rejects the null of no association
between SURPRISE and fiscal year (X2(6odf = 583.73, p < 0.01).
Panel C of Table 1, which presents mean three-day cumulative abnormal returns (CAR)
around the earnings announcement date across levels of SURPRISE, confirms the well-
known positive association between CARs and earnings surprises.4 The mean CAR for

4 As in Skinnerand Sloan (2001), our CARs arebased on size-matchedportfolios.Specifically,we calculateeach

firm-quarter's CAR as the with-dividendreturnover the threedays surroundingthe earningsannouncement, less
the returnon a size-matchedportfolioover the same period.To calculatethe size-matchedportfolio'sreturn,we
to decile portfoliosbased on marketvalues of commonequity at the beginningof the
allocateall firm-quarters
quarter.We then calculatea value-weightedreturnon each portfolioaroundeach earningsannouncementdate.
We are unableto obtainreturnsfor 1,209 firm-quarters (3.4 percentof our sample).
Bartonand Simko-Earnings ManagementConstraint 7

firmsmeetingexpectationsis -0.01 percent,statisticallyinsignificantat conventionallevels.

The last column of the panel shows that the averageproportionateloss in equity value is
more severe for firmsmissing expectatios by 10 thanfor firmsmissing by a largeramount
(t = -5.39, p < 0.01, not tabulated).For example, an EPS surpriseof -10 is associated
with a mean CAR of -1.16 percent,whereas an EPS surpriseof -40 (four times larger)
is associated with a mean CAR of -1.84 percent, about -0.46 percent per penny of
surprise.
Table 2 reportsdescriptivestatistics for selected firm characteristics.On average,our
sample firms have a market capitalization of $2.3 billion, total assets of $1.8 billion, and
quarterlysales of $471 million. About 44.1 percentof the firms are tradedon NASDAQ,
32.5 percentof their assets are fixed (i.e., consisting of property,plant, and equipment,net
of accumulateddepreciation),and 61.7 percent of their total assets are financedthrough
debt. Although the average net income is $24.5 million per quarter,18.6 percent of the
firm-quarters reportlosses. The averagequarterlyoperatingcash flow is $102 million, about
four times largerthan averagenet income.
The last column of Table 2 reportsSpearmanrank correlationsbetween the selected
firm characteristicsand SURPRISE.Although significantat conventionallevels, most cor-
relationsare less than 0.20 in absoluteterms.The correlationssuggest that firmsreporting
higher (i.e., largerpositive or smallernegative)levels of SURPRISEare largerin termsof
marketcapitalization,total assets, sales, net income, and operatingcash flows. These firms
also are less levered, and less likely both to reportlosses and to be tradedon NASDAQ.

Measuring Overstatement in Net Asset Values

We use the beginningbalanceof net operatingassets relativeto sales (NOA) as a proxy
for bias in the implementationof GAAP measurementandrecognitionguidelines.We define
net operatingassets as shareholders'equity less cash and marketablesecurities,plus total
debt. Therefore, NOA consists mainly of accrual-basedmeasures of net assets used in
operations.This proxy is consistentwith the assumptionthat overstatednet assets are less
efficient at generatinga given level of sales, all else equal. However,reportedsales may
also reflect managers'biased implementationof GAAP. Unlike net asset amounts, sales
reflect reportingbias exerted only during the currentperiod. If sales also are overstated,
then using NOA will work against finding supportfor Ha.
To provide evidence that NOA capturesprevious optimism in financialreporting,we
test whether firms with larger NOA also reportedlarger cumulative levels of income-
increasingaccrualsin the past. To this end, we first estimate abnormalaccrualsfor firm i
in quarter t using the residual from the following regression, estimated by two-digit SIC
code and fiscal year using 1988-1999 data (see Jones 1991; Dechow et al. 1995; Han and
Wang 1998):

TOT_ACCVt/TAt_j = ((l1/TA,it) + 42[(AREVit - AREC1)/TAt_j]

+ )3(PPEit/TAit-) + 4Qlit + (5Q2it + (6Q3it
+ +7Q4it + E,it (1)

where TOT_ACC is total accruals (i.e., earnings before extraordinary items and discontin-
ued operations less operating cash flows); TA is total assets; AREV is the quarterly change
in revenues; AREC is the quarterly change in accounts receivable; PPE is gross property,
plant, and equipment; and Ql, Q2, Q3, and Q4 are fiscal-quarter indicators. The (AREV
- AREC) term controls for normal working capital accruals related to sales, the PPE term
 TABLE 2
Descriptive Statistics for Selected Firm Characteristics and Rank Correlations with Ea

Standard First
Firm Characteristic Mean Deviation Quartile Median
Market value of
commonequity ($) 2,311.9 6,267.0 161.7 459.0
Totalassets ($) 1,811.8 4,397.9 129.1 371.1
Sales ($) 470.9 1,028.1 37.5 111.9
Traded on NASDAQ (%) 44.1 49.7 0.0 0.0
Ratio of fixed assets to total assets (%) 32.5 22.4 14.9 26.5
Ratio of total debt to total assets (%) 61.7 111.9 5.2 38.2
Net income ($) 24.5 78.9 0.8 4.5

Reporting losses (%) 18.6 38.9 0.0 0.0

Operating cash flows ($) 102.0 322.5 0.6 12.7

*** denotessignificantat the 0.01 level, based on two-tailedtests.

Tabulatedamountsare basedon datafor quartert. Dollar ($) amountsare in millions. The sampleconsists of 35,950 firm-qu
both the Compustatand I/B/E/S databaseswith complete data for our primarytests over 1993-1999, excludingutilities an
codes 49 and 60-67). SURPRISEis the I/B/E/S actualEPS for quartert less the consensusforecastfor quartert, both rou
forecastis the meanof analysts'most recentEPS forecastsfor quartert availableon I/B/E/S priorto the earningsannouncem
< -50 into one categoryand SURPRISE2 50 into another.Fixed assets consist of property,plant,and equipment,net of acc
BartonandSimko-EarningsManagement
Constraint 9

controls for normal depreciation and related deferred tax accruals, the TA deflator controls
for scale, and the fiscal-quarterindicatorscontrol for seasonality.
Equation(1) scales abnormalaccrualsby lagged TA, so we unscalethemby multiplying
the regression residuals by their corresponding lagged TA before accumulating them back-
ward in time. To avoid inducing a spurious correlation between prior cumulative (unscaled)
abnormal accruals and NOA, we rescale the cumulative abnormal accruals by sales for
quartert - 1, the same deflatorin NOA.
Table3 reportsmean levels of (rescaled)abnormalaccrualsaccumulatedover the prior
4, 8, 12, 16, and 20 quarters, across quintiles of adjusted NOA. Consistent with the esti-
mation of Equation (1), we adjust NOA only in this table by subtracting the mean NOA
in the same two-digit SIC code and fiscal year. The table shows that abnormal accruals
accumulated over up to 20 previous quarters are larger (at the 0.10 significance level or
better) for firms in the upper quintile of adjusted NOA than for firms in the lower quintile.
These results suggest that prior optimistic accounting choices are associated with larger
levels of net operatingassets relativeto sales.

Model
We consider two issues in developing a model to test Ha. First, recall that we define
SURPRISEas actualEPS less the consensusEPS forecast.Because we measureSURPRISE
in pennies, treatingit as an ordinalvariable,we use an estimationmodel specifically de-
signed for ordinaldependentvariables.Second, we suspect that the effects of the indepen-
dent variables vary across various earnings surprise benchmarks; for example, managers
may have less accounting discretion or weaker incentives to beat expectationsby 50 or
more than to beat expectationsby just 10. Our model thus allows the parametersof the
independentvariablesto vary across predeterminedearningssurprisebenchmarks.We in-
corporate both of these design choices using the following generalized ordered logit model,
derived in the Appendix:

Pr(SURPRISE11? k)/Pr(SURPRISE,, < k) =

exp(Po,k + Pl,kNOAit + Pk'CONTROLSit) (2)

where the left-hand-side expression is the odds of reporting an earnings surprise of at least
k0, a predetermined benchmark; SURPRISE is the signed EPS surprise; NOA is our proxy
for overstated net assets already on the balance sheet; CONTROLS is a vector of control
variables;Ps are parametersallowed to vary with k; and i and t denote firm and quarter.
For m categories in SURPRISE, Equation (2) yields m - 1 uniquely identified equations
that can be estimated jointly through maximum likelihood techniques.5 Thus, because we
coded SURPRISE using 11 categories (i.e., -5, -4, -3,..., 3, 4, 5), our empirical imple-
mentation of Equation (2) will yield a set of parameter estimates for each of ten uniquely
identified equations.
Ha predicts a negative association between SURPRISE and NOA. Therefore, we expect
the coefficients on NOA to be negative in all ten equations. For example, consider k = 0,
a zero earnings surprise. For k = 0, a negative coefficient P1,0on NOA implies that higher
values of NOA are associated with lower odds of a zero or positive earnings surprise.
Equation (2) also allows us to estimate how changes in the independent variables affect the

5 Forfurtherdetailson
generalizedorderedlogit modelsandtheirestimation,see Clogg andShihadeh(1994, 146-
147) and Fu (1998).
 TABLE 3
Average Prior Cumulative Abnormal Accruals across Quintiles of Net Operating Assets

Accumulation Quintiles of NOA

Period
(Quarters) n Upper (1) (2) (3) (4) Lower (5)

[t - 1, t - 4] 26,146 0.32 -0.07 -0.12 -0.14 -0.06

[t - 1, t - 8] 21,104 0.19 -0.22 -0.27 -0.28 -0.22

[t - 1, t - 12] 17,478 0.08 -0.38 -0.41 -0.40 -0.34

[t - 1, t - 16] 14,969 -0.17 -0.50 -0.52 -0.50 -0.46

[t - 1, t- 20] 11,555 -0.43 -0.63 -0.59 -0.57 -0.58

*** and * denotesignificantat the 0.01 and0.10 levels, respectively,based on one-tailedtests. The t-statisticsfor differencesin
arebasedon unequalvariancesacrossquintiles.
The initial sampleconsistsof 35,950 firm-quarters pertainingto 3,649 firmsincludedin both the Compustatand I/B/E/S dat
tests over 1993-1999, excludingutilitiesandfinancialservicesfirms(two-digitSIC codes 49 and60-67). Of these firm-quart
abnormalaccrualsover a particularaccumulationperiod.NOA is net operatingassets (i.e., shareholders'equityless cash and m
beginningof quartert, scaledby sales for quartert - 1. Consistentwith the estimationof Equation(1) below, we adjustNOA
NOA in the same two-digitSIC code and fiscal year.We estimateabnormalaccrualsfor firmi in quartert using the residualfr
two-digitSIC code and fiscal yearusing 1988-1999 data (see Jones 1991; Dechow et al. 1995; Han and Wang 1998):
 TABLE 3 (Continued)

TOT_ACCit/A (lAit_ == (l/TA _1) + 2[(AREV, - ARECi,)/TAi_] + 13(PPEit/TAit ) + 4Qlit + 5Q2it

where:

TOT_ACC= total accruals(i.e., earningsbefore extraordinary

items and discontinuedoperationsless operatingcash
TA = total assets;
AREV= quarterlychange in revenues;
AREC= quarterlychangein accountsreceivable;
PPE = gross property,plant,and equipment;and
Q1, Q2, Q3, and Q4 = indicators.
fiscal-quarter

The abnormalaccrualsfrom Equation(1) are implicitlyscaled by lagged TA, so we unscale them by multiplyingthe regress
TA beforeaccumulatingthembackwardin time. To avoidinducinga spuriouscorrelationbetweenpriorcumulativeabnormalacc
abnormalaccrualsby sales for quartert - 1, the same deflatorin NOA. The tabulatedamountsare mean levels of (rescaled
prior4, 8, 12, 16, and 20 quarters,acrossquintilesof adjustedNOA.
12 The AccountingReview, 2002 Supplement

odds of a given earnings surprise. For instance, an increase of 8 in NOA changes the odds
of meeting or beating expectationsby 100[exp(8 x B1,O)- 1] percent. We derive the
expressionfor the percentagechange in odds in the Appendix.6
The vector CONTROLS includes variables capturingother constraintson earnings
management, managerial incentives to meet or slightly beat forecasts, and firm performance
and size. We discuss these in turn.
One constrainton earningsmanagementis the numberof sharesoutstanding.Managers
of firmswith more sharesoutstandingmay find it more difficultto manageearningstoward
expectations, because a penny short in EPS translates into more dollars of actual earnings
for a firm with more shares outstanding than for a firm with fewer shares outstanding. To
capture this effect, we include the weighted average number of common shares outstanding
during quarter t (SHARES), the denominator of basic EPS, and expect its coefficients to
be negative.7
Anotherconstrainton earningsmanagementis auditquality.Firmswith Big 5 auditors
have lower levels of abnormal accruals (Becker et al. 1998; Payne and Robb 2000). How-
ever, Libby and Kinney (2000) find that few Big 5 auditors expect a client to correct
otherwise quantitatively immaterial income-increasing misstatements if such corrections
would result in the client missing earnings forecasts. Therefore, we include an indicator
variablecoded 1 (0 otherwise)if the firmhas a Big 5 auditor(BIG5) in quartert, but make
no predictionaboutthe sign of its coefficients.
Managersare more likely to reportearningsthatmeet or beat expectationsif they have
strong incentives to do so-when their firms have a high price-to-book ratio (Skinner and
Sloan 2001), high litigation risk (Soffer et al. 2000; Matsumoto 2002), a large analyst
following (Johnson 1999), or a previous pattern of meeting or beating expectations.8 We
measure the price-to-book ratio (PB) as the market value of common equity divided by the
book value of shareholders'equity,both at the end of quartert. We measurelitigationrisk
(LTGN-RISK) as an indicatorvariablecoded 1 (0 otherwise)if the firm is in one of four
industries identified in prior research (e.g., Francis et al. 1994; Ali and Kallapur 2001) as
most susceptible to securities litigation-pharmaceuticals/biotechnology (SIC codes 2833-
2836, 8731-8734), computers (3570-3577, 7370-7374), electronics (3600-3674), and re-
tail (5200-5961). We measure analyst following (ANALYSTS) as the number of analysts
in the I/B/E/S consensusEPS forecastfor quartert. Last, we measurethe previousearnings
surprise pattern (PREV_MB) as an indicator variable coded 1 (0 otherwise) if, based on
I/B/E/S, the firm reported a nonnegative earnings surprise (i.e., SURPRISE - 0) in quarter
t - 1. We expect the coefficientson these variablesto be positive.
Managers are more likely to report earnings that miss expectations if the expectations
are imprecise (Payne and Robb 2000). We measure this imprecision by the coefficient of

6
The advantageof interpretingeach variable'seffect in termsof a percentagechangein oddsratherthana discrete
or marginalchangein probabilitiesis thatthe changein odds for thatvariabledoes not dependon the levels of
othervariablesin the model.
7
Alternatively,managersmay meet or beat expectationsby reducingthe denominatorof EPS throughshare
repurchases.We do not distinguishbetweenthese two reasonssince they both predictnegativecoefficientson
SHARES.
8
Includingthe firm'sprice-to-earningsratioinsteadof its price-to-bookratioyields similarresults.
Bartonand Simko-Earnings ManagementConstraint 13

variationin analysts'most recentforecastsfor quartert (CV_FORECAST),and expect the

coefficients on this variableto be negative.9
In addition to managing accruals, managers may report earnings that meet or beat
expectationsby "talkingdown" analysts'forecasts (Matsumoto2002), that is, by guiding
analysts'expectationsdown priorto announcingearnings.If this strategyis successful,then
we expect firms with downwardforecast revisions to be more likely to at least meet ex-
pectations.However,if analystsunderreactto the "bad"news (e.g., AbarbanellandBernard
1992; Easterwoodand Nutt 1999), then firms with downwardforecast revisions may still
miss expectations.We measuredownwardforecastrevisions(DOWN_REV)as an indicator
variablecoded 1 (0 otherwise)if at least one analystrevised his or her forecastdown after
the earnings announcementfor quartert - 1 but before the earnings announcementfor
quarter t (Bartov et al. 2002). We make no prediction for the sign of the coefficients on
this variable.
Previousresearch(e.g., Abarbanelland Lehavy 2001a; Skinnerand Sloan 2001) sug-
gests that the level of earningssurpriseis increasingin firm performance.We control for
performanceby including sales growth (SALES-GRW), defined as sales for quartert di-
vided by sales for quarter t - 4, less 1; return on equity (ROE), defined as net income for
quarter t divided by shareholders' equity at the end of quarter t; and the change in the
return on equity (AROE), defined as ROE in quarter t less ROE in quarter t - 4. We expect
the coefficients on these variables to be positive.
Finally, we control for firm size because analysts tend to issue less optimistic forecasts
for larger firms (Das et al. 1998). We measure firm size (MKT_CAP) as the natural log-
arithm of the market value of common equity at the end of quarter t, and expect the
coefficients on this variable to be positive.

IV. EMPIRICAL ANALYSES

Descriptive Statistics and Correlations
The first five columns of Table 4 report descriptive statistics for the independent vari-
ables. The mean (median) level of NOA is 2.66 (1.97), suggesting that net operating assets
are about twice as large or larger as sales for most firm-quarters. Sample firms have on
average 61 million shares outstanding, and 97 percent of them have a Big 5 auditor. The
mean price-to-book ratio is 3.42; about 33 percent of the firms are in highly litigious
industries; the average number of analysts following a firm is 6.39; and firms met or beat
analysts' expectations in the previous quarter 68 percent of the time. The mean (median)
coefficient of variation in analysts' forecasts is 0.20 (0.06), indicating a high degree of
consensus among analysts in most firm-quarters. In 23 percent of the cases, at least one
analyst revised downward his or her forecast for current-quarterEPS. Average sales growth
is 21 percent;however,the mean ROE is 8 percent, about 2 percentlower than ROE for
the same quarterin the previous year. The mean MKT_CAP is 6.28, a marketvalue of
common equity of about $534 million.
The last column of Table 4 reports Spearman rank correlations between SURPRISE
and each independent variable. As predicted, the correlation between SURPRISE and NOA
is negative (r = -0.11, one-tailed p < 0.01). The correlations between SURPRISE and the

9 Highervalues of CV-FORECASTalso are consistentwith less accurateanalysts'forecasts.Analyststend to

issue less accurateand more optimisticallybiased forecastsfor firms with more volatile earnings(Das et al.
1998; Lim 2001). As a sensitivitycheck, we include a controlvariablefor earningsvolatility,measuredas the
coefficientof variationin I/B/E/S actualEPS over the previouseight quarters.The coefficientson this variable
are significantlynegativeat the 0.10 level or better;the coefficientson the othervariablesremainqualitatively
unchanged.
 TABLE 4
Descriptive Statistics for Independent Variables and Rank Correlations with Earnin

Standard First Third

Independent Variable Mean Deviation Quartile Median Quartile
NOA 2.66 2.63 1.26 1.97 2.99
SHARES 60.69 116.70 12.03 23.50 52.40
BIG5 0.97 0.18 1.00 1.00 1.00
PB 3.42 3.57 1.60 2.49 3.99
LTGN_RISK 0.33 0.47 0.00 0.00 1.00
ANALYSTS 6.39 4.71 3.00 5.00 8.00
PREV_MB 0.68 0.47 0.00 1.00 1.00
CV-FORECAST 0.20 0.43 0.03 0.06 0.17
DOWN-REV 0.23 0.42 0.00 0.00 0.00
SALES-GRW 0.21 0.42 0.02 0.12 0.30
ROE 0.08 0.27 0.04 0.12 0.19
AROE -0.02 0.30 -0.07 -0.01 0.04
MKT_CAP 6.28 1.60 5.09 6.13 7.28
 TABLE 4 (Continued)

*** and * denotesignificantat the 0.01 and 0.10 levels, respectively,based on one-tailedtests for signed predictionsand two-
The sample consists of 35,950 firm-quarters pertainingto 3,649 firms includedin both the Compustatand I/B/E/S databa
excludingutilitiesandfinancialservicesfirms(two-digitSIC codes 49 and 60-67). The variablesare definedas follows:
SURPRISE= I/B/E/S actualEPS for quartert less the consensusforecastfor quartert, both roundedto the nearestp
of analysts'most recentEPS forecastsfor quartert availableon I/B/E/S priorto the earningsannouncem
< -5? into one categoryand SURPRISE - 5? into another;
NOA = net operatingassets (i.e., shareholders'equity less cash and marketablesecurities,plus total debt) at the b
quartert - 1;
SHARES= weightedaveragenumberof commonsharesoutstandingduringquartert;
BIG5 = indicatorvariablecoded 1 if the firmhas a Big 5 auditorin quartert, 0 otherwise;
PB = marketvalue of commonsharesdividedby shareholders'equity,both at the end of quartert;
LTGN-RISK= indicatorvariablecoded 1 if the firm is in one of the following industries:pharmaceuticals/biotechno
computers(3570-3577, 7370-7374), electronics(3600-3674), or retail(5200-5961), 0 otherwise;
ANALYSTS= numberof analystsin the I/B/E/S consensusEPS forecastfor quartert;
PREV_MB = indicatorvariablecoded 1 if, basedon I/B/E/S, the firmreporteda nonnegativeearningssurprise(i.e., SUR
CV_FORECAST= coefficientof variationin analysts'most recentforecastsfor quartert;
DOWN_REV= indicatorvariablecoded 1 if at least one of the firm's analystsrevisedhis or her forecastdown priorto t
announcementdate for quartert - 1, 0 otherwise;
SALES_GRW= sales for quartert dividedby sales for quartert - 4, less 1;
ROE = net income for quartert dividedby shareholders'equity at the end of quartert;
AROE= ROEfor quartert less ROEfor quartert - 1; and
MKT_CAP = naturallogarithmof marketvalue of commonsharesat the end of quartert.
16 The AccountingReview, 2002 Supplement

remaining variables are significant at conventional levels. Untabulated analyses reveal that
rank correlations between most independent variables also are significant at conventional
levels.

Regression Results
Table 5 reportsestimationsresults for Equation(2). Recall that the model allows the
effects of the independent variables to vary across levels of predetermined earnings surprise
benchmarks, k. Because SURPRISE consists of 11 categories coded -5 through 5, we
obtain ten sets of uniquely identified parameter estimates, one for each k ranging from -4
through 5.10We estimate the model using the generalized ordered logit regression technique
described in Fu (1998). The test statistics for all coefficients are heteroscedasticity-
consistent and, because our sample includes multiple observations for most firms, they are
adjustedfor residualcorrelationamong observationsfor the same firm.
Panel A of Table 5 presentscomplete estimationresults when the dependentvariable
is the odds of at least meeting earningsexpectations,that is, when k = 0. For each inde-
pendent variable,the last column of Panel A presentsresults of Wald tests proposedby
Brant (1990) for the null hypothesisthat the coefficients on the independentvariableare
constant across levels of k. These tests reject the null at the 0.01 level for all variables
except BIG5. Therefore,we also reportsummaryresults for all values of k. Specifically,
PanelB shows the percentagechangesin odds for each variableacrossk.11For a continuous
variable, the change in odds is based on a standard-deviation increase in the variable; for
indicators,it is based on a change from 0 to 1 (see the Appendix).
Panel A of Table 5 shows that the coefficient on NOA is negative (p < 0.01). A
standard-deviation increase in the beginning balance of net operating assets relative to sales
decreases by 7.7 percent the odds of at least meeting the consensus forecast, all else equal.
That is, consistent with Ha, the likelihood of meeting or beating analysts' earnings forecasts
by optimistically biasing earnings decreases with the extent to which net assets are already
overstated on the balance sheet.
With respect to the control variables, the coefficient on SHARES is negative (p
< 0.01), suggesting that the odds of at least meeting expectations decrease with shares
outstanding. The insignificant coefficient on BIG5 suggests that our proxy for audit quality
is unrelated to the odds of at least meeting analysts' expectations. The coefficients on PB,
LTGN_RISK, ANALYSTS, and PREV_MB are positive and, except for LTGN_RISK,
significant at the 0.10 level or better. These results suggest that the odds of at least meeting
expectations are increasing in the firm's price-to-book ratio, its record of meeting or beating
forecasts in the previous quarter, and the size of its analyst following. The coefficients on
CV_FORECAST and DOWN_REV are negative (p < 0.01), suggesting that the odds of
at least meeting expectations decrease with the dispersion in analysts' forecasts and with
the presence of downward forecast revisions. Finally, the coefficients on SALES-GRW,
ROE, AROE, and MKT_CAP are positive (p < 0.01), suggesting that the odds of at least
meeting expectations increase with the firm's performance and size.
Panel B of Table 5 shows that the percentage changes in odds for NOA are negative
(p < 0.01) across all k. A standard-deviation increase in the beginning balance of net

10The odds of SURPRISE> -50 can be recoveredfromthe odds of SURPRISE> -40 throughtediousalgebra.
Interestedreadersmay contactus.
1 We
reportpercentagechangesin oddsbecausethey captureboththe sign andrelativeeffect of eachindependent
variable.Therefore,unlikecoefficients,percentagechangesin odds arecomparableacrossindependentvariables.
 TABLE 5
Regression Results for Generalized Ordered Logit Model

Model:
Pr(SURPRISEi, - k)/Pr(SURPRISEt, < k) = exp((,8k + P,kIVOAt, + 32kSHARES,,+ 3,kBIGi,, + 34,
+ 6,ANALYSTS,, + 37,kPREV_MBit+ P8,kCV_FOREC
+ 3,o0kSALES_GRWi + 11I,PkOEit+ 312,kgROEit
+ P13,kKT_-CAPit + it)

Panel A: Regressicon Results for k = 0?, i.e., Odds of Meeting or Beating vs. Missing Analysts' Forecasts Pr(SUR

Independent Predicted Ch
Variable Sign Coefficient z-statistic Od

Intercept 7 -0.805 -7.25***

NOA -0.031 -4.98***
SHARES -0.001 -5.60***
BIG5 0.013 0.15
PB + 0.026 4.86***
LTGN_RISK + 0.030 0.94
ANALYSTS + 0.008 1.62*
PREV_MB + 1.004 33.18***
CV_FORECAST -0.244 -7.92***
DOWN_REV -0.594 - 19.82**
SALES GRW + 0.606 13.63***
ROE + 0.228 2.87***
AROE + 0.156 2.40***
 TABLE 5 (Continued)

Independent Predicted Ch
Variable Sign Coefficient z-statistic Od
MKT_CAP + 0.156 10.17***
Likelihoodratio X2(130 df 8,105.612**
MaximumlikelihoodR2 0.20

Panel B: Percentage Change in Odds of Reporting SURPRISE - k, Given a Change in the Independent Variabl

P'red. k
Independent
Variable Ssign -4? -3? -2? -1? O0 1A 2?
NOA - -6.2*** -6.8*** -7.0*** -7.9*** -7.7*** -7.9*** -7.5***
SHARES - -0.6 -1.1 -1.1 -6.2** -11.3*** -13.1*** - 14.2***

BIG5 ? -6.9 -8.7 -2.3 -6.3 1.3 9.8 2.3

PB + 24.3*** 21.8*** 17.7*** 15.0*** 9.7*** -3.2 -8.6'
LTGN_RISK + 4.1 3.8 4.2 3.8 3.0 -14.9- -20.4-
ANALYSTS + 8.1** 9.1*** 10.0*** 8.1*** 3.6* -4.0 -6.0.
PREV-MB + 186.8*** 186.1*** 186.7*** 188.5*** 172.9*** 80.6*** 54.4***
CV_FORECAST - -19.7*** -19.0*** -17.2*** - 14.2*** -9.9_g*** -4.1*** 1.3
DOWN-REV ? -40.3*** -42.0*** -45.9*** -47.3*** -44.8*** -33.3*** -31.7***
SALES_GRW + 27.8*** 27.7*** 28.9*** 28.0*** 28.9*** 21.3*** 16.4***
ROE + 23.9*** 21.4*** 16.3*** 11.9*** 6.4*** -3.2 -9.4.
AROE + 1.8 1.9 2.2 3.0* 4.7*** 12.3*** 15.1***
MKT_CAP + 15.5*** 14.8*** 16.3*** 22.1*** 28.4*** 22.0*** 23.7***
 TABLE 5 (Continued)

***,**, and * denote significantat the 0.01, 0.05, and 0.10 levels, respectively,based on one-tailedtest for signed predicti
statisticsare heteroscedasticity-consistent
and adjustedfor residualcorrelationamongobservationspertainingto the same firm
* denotes significant at the 0.10 level or better, based on two-tailed tests, but opposite to our predicted sign.
a The percentagechangein odds is the effect of a changein the independentvariableon the odds of reportingSURPRISE : k
predetermined earnings surprise benchmark. For continuous variables, the percentage change in odds is 100[exp(sj3j,k) - 1],
of variable j (reported in Table 4) and Pj,kis the estimated regression coefficient for variable j with respect to earnings surp
the percentage change in odds is 100[exp(pj,k) - 1].
b The Brant
X2 statistic tests the null hypothesis that odds for the independent variable are proportional, i.e., that the coefficients
across k (Brant 1990).
The sampleconsists of 35,950 firm-quarters
pertainingto 3,649 firmsincludedin both the Compustatand I/B/E/S database
over 1993-1999, excluding utilities and financial services firms (two-digit SIC codes 49 and 60-67).
The variables are defined in Table 4.
20 The AccountingReview, 2002 Supplement

operating assets relative to sales decreases the odds of reporting a larger positive or smaller
negative earnings surprise by 5.7 to 7.9 percent. These results further support Ha.
Panel B of Table 5 also shows that some of the independent variables have a consistent
pattern across all k. For example, the percentage changes in odds for BIG5 are not signif-
icant at conventional levels for any value of k, suggesting that our proxy for audit quality
is unrelated to earnings surprises. The percentage changes in odds for PREV_MB,
SALES_GRW, and MKT_CAP are positive (p < 0.01) across all values of k, suggesting
that the odds of reporting larger positive or smaller negative earnings surprises are increas-
ing in the likelihood of a nonnegative earnings surprise in the previous quarter,sales growth,
and firm size. The percentage changes in odds for DOWN_REV are negative (p < 0.01)
across all k, suggesting that the odds of reporting larger positive or smaller negative earnings
surprises decrease when analysts make downward forecast revisions.
Other independent variables have a consistent pattern across only part of the range of
values for k. For example, the percentage changes in odds for AROE are positive (p
< 0.10) when k : -1, suggesting that the odds of reporting larger positive surprises are
increasing in AROE, but the odds of reporting earnings surprises lower than -1? are un-
related to AROE. The percentage changes in odds for SHARES are negative (p < 0.05)
when k - -1, suggesting that more shares outstanding decrease the odds of reporting
larger positive surprises, but have no effect on the odds of missing expectations by more
than 10. The percentage changes in odds for LTGN_RISK are negative (p < 0.10) when
k - 1, suggesting that litigation risk decreases the odds of larger positive surprises, but has
no effect on the odds of missing expectations.
Finally, the pattern of some variables switches signs around positive earnings surprises
of 10 or 20. For example, the percentage changes in odds for PB, ANALYSTS, and ROE
are positive (p < 0.10) when k < 1 but negative (p < 0.10) when k > 1. These results
suggest that firms with high price-to-book ratios, large analyst following, and high returns
on equity are more likely to miss expectations by smaller amounts than by larger amounts,
and they are less likely to beat expectations by larger amounts than by smaller amounts.
In contrast, the percentage changes in odds for CV_FORECAST are negative (p < 0.01)
when k < 2 but positive (p < 0.10) when k > 2, suggesting that firms with imprecise
forecasts are more likely to miss expectations by larger than smaller amounts, and are more
likely to beat expectations by more than 20.
In sum, the results reported in Table 5 are consistent with H, showing a negative
association between earnings surprises and our proxy for overstated net asset values. We
interpret these results as evidence that the likelihood of reporting larger positive or smaller
negative earnings surprises decreases with the extent to which net assets are already over-
stated on the balance sheet. That is, prior optimism in financial reporting accumulates on
the balance sheet, and reduces managers' ability to optimistically bias earnings in future
periods, if managers want to stay within GAAP and its implementation guidance. The
results also indicate that the relationship between earnings surprises and the independent
variables is complex. The presence of a Big 5 auditor is consistently unrelated to earn-
ings surprises. However, none of the remaining variables measuring constraints on earnings
management, managerial incentives to meet or beat analysts' forecasts, firm performance,
and size has a constant effect across earnings surprise levels.

Sensitivity Analyses
Alternative Econometric Techniques
We use a generalized ordered logit regression of SURPRISE on NOA and CONTROLS
because SURPRISE is an ordinal dependent variable and because we allow the coefficients
on all independent variables to vary across levels of SURPRISE. Using ordinary least
Bartonand Simko-Earnings ManagementConstraint 21

squares (OLS) or "regular" ordered logit regressions constrains the coefficients on each
variable to be constant across levels of SURPRISE, yielding only one set of coefficient
estimates. Moreover, OLS regression treats SURPRISE as a continuous dependent variable.
To assess the robustnessof our resultsunderalternativeeconometricspecificationsthat
constrainthe coefficients on each variableto be constantacross levels of SURPRISE,we
regress SURPRISE on NOA and CONTROLS using OLS and "regular"orderedlogit
regressions. The coefficient on NOA is -0.053 (t = -8.54, p < 0.01) based on OLS
regression;it is -0.028 (z = -4.31, p < 0.01) based on "regular"orderedlogit regression.
Thus, the results bearing on our hypothesis are robust to these alternativespecifications.
However, OLS and "regular"ordred logit regressionslead to substantialchanges in the
coefficients on some of the other independentvariables.

Composition of Net Operating Assets

To capturethe possibility that differentcomponentsof
nes t operatingassets
aet re subject
to varying degrees of managerial manipulation (Beatty et al. 1995; Hunt et al. 1996; Teoh
et al. 1998; Gaver and Paterson 1999), we split NOA into three components:
1) Working capital (WC), defined as current assets less cash, marketable securities,
and current liabilities, plus short-term debt and the current portion of long-term
debt, all at the beginningof quartert and scaled by sales for quartert - 1.
2) Net fixed assets (FIXED-ASSETS), defined as property, plant, and equipment, net
of accumulateddepreciation,at the beginningof quartert and scaled by sales for
quarter t - 1.
3) Other long-term assets (OTHER_LTASSETS), defined as NOA less WC and
FIXED ASSETS.
We separateworkingcapitalfrom long-termnet operatingassets becausepriorresearch
suggests that current accruals are likely easier to manage than long-term accruals (Hunt et
al. 1996; Beneish 1998; Teoh et al. 1998). We separate fixed assets from other long-term
assets because managing earnings through depreciation is transparent (e.g., firms are re-
quired to disclose the effects of changes in depreciation policies) or costly if managers
must time capital investments to make depreciation-related earnings management less trans-
parent(Beneish 1998).
Table 6 reportsselected regressionresults using WC, FIXED_ASSETS,and OTHER_
LTASSETSin place of NOA in Equation(2). Consistentwith Ha and the results reported
in Table 5, the coefficients on the three components of net operating assets are negative
and significantat the 0.05 level or better across the various surprisebenchmarksk. Brant
(1990) X2tests reject at the 0.01 level the null hypothesis that the coefficients on each
variable are constant across k. Finally, the coefficients on WC are about 9 to 28 (3 to 5)
times larger in absolute terms than the coefficients on FIXED_ASSETS (OTHER_LTAS-
SETS), suggesting that, relative to the level of long-termnet assets, the level of working
capital has a strongereffect on the odds of reportinga predeterminedearningssurprise.
Additional Robustness Tests
The level of net operatingassets relativeto sales tends to vary acrossindustries(Nissim
and Penman 2001). To control for this effect, we reestimate Equation (2) including indi-
cators for two-digit SIC codes. The coefficients on NOA (not tabulated)remain negative
and significantat the 0.01 level.
For reportingquarterlyperformance,GAAP requiresmanagersto estimate annualop-
erating expenses and allocate these costs across quarters(e.g., APB 1973; FASB 1974).
This opportunity to misreport quarterly performance is exacerbated by the fact that financial
 TABLE 6
Selected Regression Results for Generalized Ordered Logit Model Using Components of N

Model:
Pr(SURPRISEi, - k)/Pr(SURPRISEi, < k) = exp(yo,k + y,kWCil + y2,FIXED_ASSETSi, + 3,kOTHER_
+ y3,kBIGSi,+ y6,PBi, + y7LTGN_RISKi, + y8ANALYSTS
+ y,okCV_FORECASTi, + y,,kDOWN_REV,i + y2,~ALES_
+ Y4,kROEit + y5,lMKT_CAPit + vi)

WC FIXED ASSETS
Predicted Change in Change in
k Sign Coefficient z-statistic Odds (%)a Coefficient z-statistic Odds (%)a Co

-40 -0.204 -6.28*** -12.0 -0.018 -2.22** -4.2 -

-3 -0.223 -7.39*** -13.0 -0.020 -2.67*** -4.8 -
-2 -0.226 -8.04*** -13.2 -0.021 -3.04*** -5.0 -
-1 -0.245 -9.26*** -14.2 -0.026 -4.04*** -6.2 -
0 -0.263 -10.56*** -15.1 -0.024 -3.81*** -5.8 -
1 -0.296 - 10.49*** -16.9 -0.021 -3.20*** -5.0 -
2 -0.370 -10.83*** -20.6 -0.020 -2.74*** -4.7 -
3 -0.420 -10.72*** -23.1 -0.020 -2.51*** -4.7 -
4 -0.451 -9.92*** -24.6 -0.016 -1.96** -3.9 -
5 -0.484 -10.07*** -26.1 -0.018 -2.09** -4.4 -
Brant X29 df)b 92.81*** 108.03**
 TABLE 6 (Continued)

*** and**denotesignificantat the 0.01 and 0.05 levels, respectively,based on one-tailedtest for signed predictionsand two
and adjustedfor residualcorrelationamongobservationspertainingto the same firm.
heteroscedasticity-consistent
a The percentage change in odds is the effect of a change in the independent variable on the odds of reporting SURPRISE 2 k
predetermined earnings surprise benchmark. For continuous variables, the percentage change in odds is 100[exp(sjPj,k) - 1],
of variable j (reported in Table 4) and 3j,kis the estimated regression coefficient for variable j with respect to earnings surp
the percentage change in odds is 100[exp(ij k) - 1].
b The Brant
X2statistic tests the null hypothesis that odds for the independent variable are proportional, i.e., that the coefficients
across k (Brant 1990).
The sampleconsists of 35,950 firm-quarterspertainingto 3,649 firmsincludedin both the Compustatand I/B/E/S database
over 1993-1999, excludingutilitiesand financialservices firms(two-digitSIC codes 49 and 60-67). The variablesare define
NOA = net operatingassets (i.e., shareholders'equity less cash and marketablesecurities,plus total debt) at t
for quarter t - 1;
WC = currentassets less cash, marketablesecurities,andcurrentliabilities,plus currentdebt, all at the beginn
t - 1;
FIXEDASSETS = property, plant, and equipment, net of accumulated depreciation, at the beginning of quarter t, scaled b
OTHER_LTASSETS = NOA less WC less FIXED_ASSETS.
All other variables are defined in Table 4.
24 The AccountingReview, 2002 Supplement

statementsfor the firstthree quartersare usually auditedretrospectivelyduringthe annual

externalaudit(Elliottand Shaw 1988; Ettredgeet al. 2000). Consequently,securitymarkets
tend to react more strongly to negative earnings surprisesin earlier quartersthan in the
fourthquarter(Mendenhalland Nichols 1988), suggestingthatmanagersmay have stronger
incentivesto meet or beat expectationsearlierin the fiscal year. To capturethis potential
effect, we reestimateEquation(2) including fiscal-quarterindicators.The coefficients on
NOA (not tabulated)remainnegativeand significantat the 0.01 level.
Because regulators and the media routinely mention the consensus analyst forecast as
an earningstarget,we calculateSURPRISEusing the consensusforecastas a proxy for the
market'seaings expectations.However,marketexpectationsmay be based on the most
recentforecast(Brownand Kim 1991; Brown 2001). We recalculateSURPRISEusing the
most recent I/B/E/S forecastpriorto the earningsannouncementand reestimateEquation
(2). Results (not tabulated)are similarto those reportedin Table 5.
Abarbanelland Lehavy(2001b) point out thatdatabaseproviderssuch as I/B/E/S tend
to adjustboth actualand expected earningsfor nonrecurringitems. Because SURPRISEis
based on I/B/E/S data and some of our independentvariables(such as NOA) are based
on Compustat data, we repeat the tests reported in Table 5 using (1) only firms for which
I/B/E/S actual EPS agrees with Compustat EPS, and (2) only firms without special items
on Compustat. The resulting coefficients on NOA (not tabulated) remain negative and sig-
nificant at the 0.01 level.
Finally, prior research suggests that earnings management behavior and bias in analysts'
forecasts differ between firms reporting losses and firms reporting profits (e.g., Degeorge
et al. 1999; Easterwood and Nutt 1999; Brown 2001). We repeat the tests in Table 5
excludingfirmsreportinglosses (1) in quartert, (2) in quartert - 1, and (3) in both quarters
t and t - 1. For each set of tests, the coefficients on NOA (not tabulated) remain negative
and significant at the 0.01 level.

V. CONCLUSION
Because the balance sheet accumulates the effects of previous accounting choices, the
level of net operating assets partly reflects the extent of previous earnings management. We
present evidence consistent with the hypothesis that managers' ability to optimistically bias
earnings decreases with the extent to which net asset values are already overstated on the
balance sheet.
Using 1993-1999 quarterly data for a sample of 3,649 nonfinancial, nonregulated firms,
we estimate a generalized ordered logit model relating the level of a firm's earnings surprise
to our proxy for the degree of overstatement in net asset values, other constraints on earn-
ings management, managerial incentives to meet or beat analysts' earnings forecasts, and
firm performance and size. We measure the degree of overstatement in net asset values
using the beginning balance of net operating assets relative to sales, a proxy that additional
analysis shows is associated with prior optimistic bias in financial reporting.
Our empirical analysis shows that the likelihood of reporting larger positive or smaller
negative quarterly earnings surprises decreases with the beginning balance of net operating
assets relative to sales, suggesting that managers' ability to optimistically bias earnings
decreases with the extent to which net asset values are already overstated on the balance
sheet. Sensitivity tests suggest that our findings are robust. Nevertheless, our evidence
should be interpreted with caution given the limitations of our study. In particular,our study
focuses on only one incentive to manage earnings-meeting or beating analysts' earnings
forecasts. However, managers have other incentives to manage earnings, some of which
Bartonand Simko-Earnings ManagementConstraint 25

may lead to pessimistically rather than optimistically biased earnings. Our study does not
control for these other incentives.We leave this for futureresearch.

APPENDIX
We derive Equation(2) by assuming that the cumulativeprobabilityof reportingan
EPS surpriseof less than k? is:

Pr(SURPRISE< k I x) = F(-xPk),

where x is a vector of independentvariables,Pk is a vector of parametersfor a predeter-

mined earningssurprisebenchmarkk, and F is the cumulativelogistic distribution:

F(-X3k) = exp(-XPk) / [1 + exp(-xlk)].

To ensure that the sum of cumulative probabilities across all k equals 1, we impose the
constraint-xpik -XP,k-1 for all k. The odds of reporting an earnings surprise of at least
k? instead of less than k? are:

lk(x) = Pr(SURPRISE - k | x) / Pr(SURPRISE< k I x)

= [1 - F(-xpk)] / F(-xpk) = exp(xpk),

which is a general version of Equation(2).

To determinethe effect of a change in x on the odds of reportingan earningssurprise
of at least k?, suppose that x changes from x = xl to x = x2. The odds then change from
fk(xl) to flk(x2) by the factor:

lk(x2) / fk(Xl) = exp(x2Pk) / exp(xPk) = exp([x2 - l]Jk)

That is, the odds change by 100[exp([x2 - xl]Pk) - 1] percent. If only one variable, say
Xjwith parameterpj, changes by 8, then the odds of reportingan earnings surpriseof at
least k? will change by 100[exp(6 x j,k) - 1] percent, all else equal.

REFERENCES
Abarbanell,J. S., and V. L. Bernard.1992. Tests of analysts'overreaction/underreactionto earnings
informationas an explanationfor anomalousstock price behavior.The Journal of Finance 47
(July): 1181-1207.
, andR. Lehavy.2001a. Biased forecastsor biasedearnings?The role of earningsmanagement
in explainingapparentoptimismand inefficiencyin analysts'earningsforecasts.Workingpaper,
The Universityof NorthCarolinaat ChapelHill, NC.
, and . 2001b. Differencesin commercialdatabasereportedearnings:Implicationsfor
inferencesconcerninganalyst forecast rationality,the associationbetween prices and earnings,
and firm reportingdiscretion.Workingpaper,The Universityof North Carolinaat ChapelHill,
NC.
AccountingPrinciplesBoard (APB). 1973. InterimFinancialReporting.APB OpinionNo. 28. New
York,NY: APB.
Ali, A., and S. Kallapur.2001. Securities price consequencesto the Private Securities Litigation
ReformAct of 1995 and relatedevents. TheAccountingReview76 (July):431-460.
26 The AccountingReview, 2002 Supplement

Barth,M. E., J. A. Elliott,andM. W. Finn. 1999. Marketrewardsassociatedwithpatternsof increasing

earnings. Journal of Accounting Research 37 (Autumn): 387-413.
Bartov,E., D. Givoly, and C. Hayn. 2002. The rewardsto meeting or beatingearningsexpectations.
Journal of Accounting and Economics 33 (June): 173-204.
Beatty,A., S. L. Chamberlain,andJ. Magliolo. 1995. Managingfinancialreportsof commercialbanks:
The influence of taxes, regulatorycapital, and earnings.Journal of AccountingResearch 33
(Autumn):231-261.
Becker,C. L., M. L. DeFond,J. Jiambalvo,and K. R. Subramanyam.1998. The effect of auditquality
on earnings management. Contemporary Accounting Research 15 (Spring): 1-24.
Review
Beneish,M. D. 1998. Discussionof: Are accrualsduringinitialpublicofferingsopportunistic?
of Accounting Studies 3: 209-221.
Brant,R. 1990. Assessing proportionalityin the proportionalodds model for ordinallogistic regres-
sion. Biometrics46 (December):1171-1178.
Brown, L. D., and K-J. Kim. 1991. Timely aggregateanalystforecastsas betterproxies for market
earnings expectations. Journal of Accounting Research 29 (Autumn): 382-385.
. 2001. A temporalanalysisof earningssurprises:Profitsversuslosses. Journalof Accounting
Research39 (September):221-241.
Burgstahler,D. C., and M. Eames. 2001. Managementof earningsand analystforecaststo achieve
zero earningssurprise.Workingpaper,Universityof Washington,Seattle,WA.
Clogg, C. C., and E. S. Shihadeh. 1994. Statistical Models for Ordinal Variables. Thousand Oaks,
CA: Sage.
1998. Earningspredictabilityand bias in analysts'
Das, S., C. B. Levine, and K. Sivaramakrishnan.
earnings forecasts.The AccountingReview73 (April):277-294.
Dechow, P. M., R. G. Sloan, and A. P. Sweeney. 1995. Detecting earningsmanagement.The Ac-
counting Review 70 (April): 193-225.
, and . 1996. Causesand consequencesof earningsmanipulations:An analysis
of firms subject to enforcement actions by the SEC. ContemporaryAccountingResearch 13
(Spring): 1-35.
, and D. J. Skinner.2000. Earningsmanagement:Reconcilingthe views of accountingaca-
demics, practitioners,and regulators.AccountingHorizons 14 (June):235-250.
Degeorge, F, J. Patel, and R. Zeckhauser.1999. Earningsmanagementto exceed thresholds.The
Journal of Business 72 (January): 1-33.
Easterwood,J., and S. Nutt. 1999. Inefficiencyin analysts'earningsforecasts:Systematicmisreaction
or systematicoptimism?The Journalof Finance 54 (October):1777-1797.
Elliott, J. A., and W. H. Shaw. 1988. Write-offsas accountingproceduresto manageperceptions.
Journal of Accounting Research 26 (Supplement): 91-119.
Ettredge,M. L., D. T. Simon, D. B. Smith, and M. S. Stone. 2000. The effect of the externalac-
countant'sreview on the timing of adjustmentsto quarterlyearnings.Journal of Accounting
Research38 (Spring):195-207.
Feroz, E. H., K. Park, and V. S. Pastena. 1991. The financialand marketeffects of the SEC's ac-
countingand auditingenforcementreleases. Journal of AccountingResearch29 (Supplement):
107-142.
Fields, T. D., T. Z. Lys, and L. Vincent.2001. Empiricalresearchon accountingchoice. Journalof
Accounting and Economics 31 (September): 255-307.
Financial Accounting Standards Board (FASB). 1974. Reporting Accounting Changes in Interim Fi-
nancial Statements.Statementof FinancialAccountingStandardsNo. 3. Stamford,CT: FASB.
1978. Objectives of Financial Reporting by Business Enterprises. Statement of Financial
AccountingConceptsNo. 1. Stamford,CT: FASB.
. 1985. Elementsof FinancialStatements.Statementof FinancialAccountingConceptsNo. 6.
Stamford,CT: FASB.
Francis, J., D. Philbrick,and K. Schipper.1994. Shareholderlitigation and corporatedisclosures.
Journal of Accounting Research 32 (Autumn): 137-164.
Bartonand Simko-Earnings ManagementConstraint 27

Fu, V. K. 1998. sg88: Estimatinggeneralizedorderedlogit models. Stata TechnicalBulletin44: 27-

30.
Gaver, J. J., and J. S. Paterson. 1999. Managinginsurancecompany financialstatementsto meet
regulatoryand tax reportinggoals. Contemporary AccountingResearch16 (Summer):207-241.
Han,J. C. Y., and S-W.Wang.1998. Politicalcosts andearningsmanagementof oil companiesduring
the 1990 PersianGulf crisis. TheAccountingReview73 (January):103-117.
Hunt, A., S. E. Moyer, and T. Shevlin. 1996. Managinginteractingaccountingmeasuresto meet
multiple objectives:A study of LIFO firms. Journal of Accountingand Economics21 (June):
339-374.
Johnson,N. S. 1999. CurrentSEC developments-"Managed earnings"and "The year of the ac-
countant."Remarksdeliveredat the Utah StateBar Mid-YearConvention,St. George,UT. March
6. Availableat: http://www.sec.gov/news/speech/speecharchive/1999/spch264.htm.
Jones, J. J. 1991. Earningsmanagementduringimportrelief investigations.Journal of Accounting
Research29 (Autumn):193-228.
Levitt, A. 1998. The numbersgame. Remarksdeliveredat New YorkUniversityCenterfor Law and
Business, New York, NY. September 28. Available at: http://www.sec.gov/news/speech/
speecharchive/1998/spch220.txt.
Libby, R., and W. R. Kinney, Jr. 2000. Does mandatedaudit communicationreduce opportunistic
correctionsto manageearningsto forecasts?TheAccountingReview75 (October):383-404.
Lim, T. 2001. Rationalityand analysts'bias. The Journalof Finance 56 (February):369-385.
Lys, T., and R. L. Watts. 1994. Lawsuitsagainstauditors.Journalof AccountingResearch32 (Sup-
plement):65-93.
Matsumoto,D. A. 2002. Management'sincentivesto avoid negativeearningssurprises.TheAccount-
ing Review77 (July):483-514.
Mendenhall,R. R., and W. D. Nichols. 1988. Bad news and differentialmarketreactionsto an-
nouncementsof earlier-quarters earnings.Journalof AccountingResearch
versus fourth-quarter
26 (Supplement):63-86.
Moehrle, S. R. 2002. Do firms use restructuringchargereversalsto meet earningstargets?The Ac-
countingReview77 (April):397-413.
Nissim, D., and S. H. Penman.2001. Ratio analysis and equity valuation:Fromresearchto practice.
Reviewof AccountingStudies6 (March):109-154.
Palmrose,Z-V.,V. J. Richardson,andS. Scholz. 2001. Determinantsof marketreactionsto restatement
announcements.Workingpaper,Universityof SouthernCalifornia,Los Angeles, CA.
Payne,J. L., and S. G. Robb.2000. Earningsmanagement:The effect of ex anteearningsexpectations.
Journalof Accounting,Auditingand Finance 15 (Fall):371-392.
Skinner,D. J., and R. G. Sloan. 2001. Earningssurprises,growthexpectations,and stock returnsor
don't let an earningstorpedosink your portfolio.Workingpaper,Universityof Michigan,Ann
Arbor,MI.
Soffer, L. C., S. R. Thiagarajan,and B. R. Walther.2000. Earningspreannouncementstrategies.
Reviewof AccountingStudies5 (March):5-26.
St. Pierre,K., and J. A. Anderson.1984. An analysis of the factorsassociatedwith lawsuitsagainst
public accountants.TheAccountingReview59 (April):242-263.
Teoh, S. H., T. J. Wong, and G. R. Rao. 1998. Are accrualsduringinitial public offeringsopportun-
istic? Reviewof AccountingStudies3: 175-208.