a diagnostic for earnings management using changes.pdf

A Diagnostic for Earnings Management Using Changes

in Asset Turnover and Prot Margin*

IVO PH. JANSEN, Rutgers UniversityCamden

SUNDARESH RAMNATH, University of Miami

TERI LOMBARDI YOHN, Indiana University

1. Introduction

Identifying earnings management is important for nancial statement users to assesscurrent economic performance, to predict future protability, and to determine rm value.However, it is often difcult and time-consuming to identify earnings management, espe-cially in generic settings where an obvious incentive to manage earnings is absent. Whileacademic research has used numerous proxies for (or diagnostics of) earnings manage-ment, most recent studies use accruals models to decompose total accruals into a normal,economics-driven component and an abnormal, earnings management component.1

McNichols (2000) points out, however, that there is limited theory about how accrualsshould behave in the absence of discretion, and Fields, Lys, and Vincent (2001) argue thatthe use of existing accruals models may lead to serious inference problems.

In DuPont analysis, a rms return on assets is decomposed into asset turnover (ATO,the ratio of sales to net operating assets) and prot margin (PM, the ratio of operatingincome to sales), and nancial statement analysis textbooks broadly advocate making thisdecomposition when investigating protability and changes in protability (see, e.g.,White, Sondhi, and Fried 2003; Palepu, Bernard, and Healy 2004; Penman 2007; Stickney,Brown, and Wahlen 2004; Lundholm and Sloan 2004). In this study, we propose a simplediagnostic of earnings management that relies on the widely held notion underlyingDuPont analysis that sales is a fundamental driver of a rms investment and income, andthat net operating assets on the balance sheet and net operating income on the incomestatement should vary directly with sales. In other words, changes in ATO or PM warrantfurther investigation in quality of earnings analyses. Moreover, we note that changes inATO and PM in opposite directions could signal earnings management. We base this obser-vation on the articulation of the income statement and balance sheet, which ensures thatearnings management affects operating income and net operating assets in the same direc-tion, and thus causes ATO and PM to move in opposite directions. For example, for agiven level of sales, if a rm manages earnings upward by understating bad debt expense,both net income relative to sales and the net realizable value of accounts receivable relative

* Accepted by K.R. Subramanyam. We thank Patricia Faireld for her contributions to the paper, as well as

Bill Baber, Walt Blacconiere, Bill Brown, Dave Burgstahler, Prem Jain, Chris Jones, Bin Ke, Jim Ohlson,

Scott Richardson, D. Shores, and seminar participants at George Washington University, Georgetown Uni-

versity, Michigan State University, University of Washington, Morgan State University, University of Min-

nesota, Rutgers UniversityCamden, Suffolk University, Loyola Marymount University, University of New

Hampshire, Villanova University, the Financial Economics and Accounting Conference, and the University

of Utah Winter Accounting Conference. We also thank Glass Lewis & Co. for the restatement data. Teri

Yohn acknowledges the generous support of the PricewaterhouseCoopers Fellowship.

1. See, for example, Healy 1985; DeAngelo 1986; Jones 1991; Dechow, Richardson, and Tuna 2003; and

Kothari, Leone, and Wasley 2005.

Contemporary Accounting Research Vol. 29 No. 1 (Spring 2012) pp. 221251 CAAAdoi:10.1111/j.1911-3846.2011.01093.x

to sales will be overstated. The increase in net income relative to sales will lead to anincrease in PM, while the increase in net accounts receivable relative to sales will lead to adecrease in ATO. In general, upward earnings management causes PM to increase andATO to decrease, while downward earnings management causes PM to decrease and ATOto increase.

We rely on this observation to argue that contemporaneous, directionally oppositechanges in a rms ATO and PM can serve as a signal of potential earnings management.Specically, we propose and investigate the usefulness of contemporaneous increases inPM and decreases in ATO as a diagnostic for upward earnings management, and ofcontemporaneous decreases in PM and increases in ATO as a diagnostic for downwardearnings management.

The above relations between earnings management and ATO PM hold when the rela-tion between net operating assets and sales is stable and when earnings are managedthrough expenses. When earnings are managed through sales, the relations will hold if theprot margin on the managed sales is greater than the prot margin on unmanaged salesand if the asset turnover of the managed sales is less than the asset turnover of unmanagedsales. The ATO PM diagnostic further assumes that a company has not changed its strat-egy and has constant growth rates in investment. A Type I error could occur if a companychanges its strategy or experiences unexpected growth. The ATO PM diagnostic alsoassumes that a company does not manage earnings through cash ows. A Type II errorcould occur if a rm manages earnings upward by delaying, for example, advertising orresearch and development expenditures.

Because earnings management is not directly observable, we cannot perform directtests to validate the ATO PM earnings management diagnostic. Instead, we rely on priorresearch which documents situations and outcomes indicative of earnings managementand show that the ATO PM diagnostic is associated with these earnings managementscenarios. We also suggest that directionally opposite changes in ATO and PM can be auseful complement to abnormal accruals in detecting earnings management in academicresearch. Therefore, in all tests, we compare the relative and incremental informationcontent of the ATO PM diagnostic to performance-adjusted abnormal accruals, a widelyaccepted proxy for earnings management. (For recent studies that use abnormal accrualssee, e.g., Cohen, Dey, and Lys 2008; Gong, Louis, and Sun 2008; Zhao and Chen2008.)2

Relying on prior research which suggests that rms manage earnings upward to meetor beat analyst forecasts (e.g., Burgstahler and Eames 2006; Matsumoto 2002), we rstexamine the association between the ATO PM diagnostic and rms propensity to meetor beat earnings expectations. We nd that the ATO PM diagnostic provides informa-tion about the likelihood of a rm meeting or beating expectations, even after control-ling for performance-adjusted abnormal accruals. Moreover, we nd that the ATO PMmeasure has signicantly greater discriminating ability than performance-adjusted abnor-mal accruals in identifying rms that meet or beat earnings expectations.

Second, we argue that when rms beat or miss earnings expectations by a wide mar-gin, they are more likely to manage earnings downward (i.e., smooth earnings or take abath). We nd that the ATO PM diagnostic provides information about the likelihoodof a rm experiencing an extreme earnings surprise. Once again the ATO PM diagnostic

2. We estimate performance-adjusted abnormal accruals using the abnormal accruals model from Dechow et

al. 2003 (259, model 3), augmented with protability as an independent variable to control for perfor-

mance. All results reported in the paper are qualitatively similar if we omit the performance adjustment,

or if we estimate more basic versions of the abnormal accruals model.

222 Contemporary Accounting Research

CAR Vol. 29 No. 1 (Spring 2012)

has signicantly greater discriminating ability than performance-adjusted abnormal accru-als in identifying rms that report extreme earnings surprises.

Third, a growing body of research uses subsequent earnings restatements as an indica-tor of earnings management (Richardson, Tuna, and Wu 2002; Kedia 2003). We nd thatthe ATO PM diagnostic provides information about the likelihood of a rm subsequentlyrestating earnings. In comparison, performance-adjusted abnormal accruals are unrelatedto future earnings restatements.

Finally, earnings management temporarily inates or deates earnings articially andshould therefore lead to a reversal in future protability (Penman 2007: 633). In addition,failure of the stock market to see through the earnings management will lead to predictablefuture returns (e.g., Xie 2001). We nd that both the ATO PM diagnostic and performance-adjusted abnormal accruals are useful, and incrementally informative, for identifying futureearnings reversals and future abnormal returns.

Based on these ndings, we conclude that contemporaneous, directionally oppositechanges in ATO and PM are informative about earnings management, even after control-ling for performance-adjusted abnormal accruals, a widely accepted earnings managementproxy. In addition, the ATO PM diagnostic has signicantly greater discriminating abilitythan performance-adjusted abnormal accruals in identifying rms that meet or beat expec-tations, report extreme earnings surprises, and subsequently restate earnings.

As a diagnostic of earnings management, the ATO PM measure has several appeal-ing features. First, the measure relies on fundamental relations in the accounting model,as opposed to estimated relations typically used in abnormal accruals models. Second,ATO and PM are primary ratios in nancial statement analysis that are likely to beinvestigated by many users of nancial statements, even when they are not explicitlyconsidering earnings management. In addition, unlike abnormal accruals measures whichrequire nancial statement data from a substantial time series or even an entire industry,the ATO PM diagnostic can be computed for any rm using very few years of rm-leveldata. In practical terms, our ndings suggest that nancial statement users will benetfrom investigating the possibility that a rm has managed its earnings upward (down-ward) when there is a contemporaneous increase (decrease) in the rms PM anddecrease (increase) in its ATO. In short, we believe that the ATO PM diagnostic can beused in academic and investment research as a (complementary) diagnostic of earningsmanagement.

The paper proceeds as follows. The next section provides the background and motiva-tion for our earnings management diagnostic. In section 3, we discuss our sample, variablemeasurement, and descriptive statistics. In section 4, we describe our analyses and reportour ndings. In section 5, we summarize and conclude the paper.

2. Background and motivation

Prior literature

Healy and Wahlen (1999) argue that earnings management occurs when managers usejudgment in nancial reporting to alter nancial reports to mislead stakeholders about theunderlying economic performance of the company. Dechow and Skinner (2000) argue thatearnings management could arise from accounting choices that are fraudulent or fromchoices that are aggressive, but acceptable, uses of accounting discretion. Identifying bothforms of earnings management is important to investors for assessing rm value; however,it is often difcult to do so, in part because many discretionary earnings components arenot separately observable. This task is further complicated when earnings managementoccurs in the absence of an obvious incentive to manage earnings, such as preceding anequity offering or a leveraged buyout. Thus, it is important for nancial statement users

A Diagnostic for Earnings Management 223


and academic researchers to have diagnostics for earnings management that are informa-tive even when no obvious incentive to manage income exists.

In quality of earnings analyses, one is generally concerned when growth in net oper-ating assets exceeds growth in sales. Such a scenario could suggest that a company isinappropriately recording costs on the balance sheet instead of the income statement.The most popular proxy for earnings management abnormal accruals, estimated usingsome version of the Jones 1991 model builds on this idea and adjusts total accruals(i.e., growth in working capital less depreciation expense) for the amount of accrualsexplained by changes in sales and property, plant, and equipment. There are severalvariations of this model. For example, the modied Jones model (see Dechow, Sloan,and Sweeney 1995) subtracts growth in accounts receivable from growth in sales whencalculating nondiscretionary accruals, to avoid the assumption, implicit in the Jones 1991model, that earnings are not managed through sales. Dechow et al. (2003) furtherenhance the model by including an estimation of the relation between the change inreceivables and the change in sales to avoid the assumption implicit in the modiedJones model that the entire change in accounts receivable stems from revenue-based earn-ings management and by including prior total accruals. Kothari et al. (2005) suggestthat to isolate abnormal accruals researchers should control for rm performance in theestimation model as well. Regardless of the specic model, the model parameters are gen-erally estimated using annual, cross-sectional regressions within two-digit SIC codes, orusing time-series regressions by rm. The estimates are then used to calculate nondiscre-tionary accruals as the predicted value of total accruals, and the difference between totalaccruals and nondiscretionary accruals is deemed discretionary and is used as a proxy formanaged earnings.

Bernard and Skinner (1996: 31617) argue that there are likely to be important omit-ted variables in explaining working capital accruals, and that any nonlinearity in the rela-tion between growth in working capital and the explanatory variables will createmeasurement error in estimating discretionary accruals. In addition, when the model isestimated cross-sectionally in an industry, it is assumed that all rms in that industry havethe same strategy and relation between accruals and the explanatory variables. Alterna-tively, when the model is estimated by rm, one needs a sufcient estimation period to cal-culate the parameter estimates. Bernard and Skinner (1996) argue that the cross-sectionalestimates are very imprecise, even when estimated within two-digit industry codes, andthat the time-series estimates are even less precise. The abnormal accruals models, there-fore, make signicant assumptions that may or may not hold. Moreover, the choice ofexplanatory variables to capture the drivers of accruals is ad hoc, and the models lacktheoretical support.

Barton and Simko (2002) develop a measure of past earnings management that ismore grounded in the accounting model. They argue that rms that have aggressivelycapitalized expenditures will have high net operating assets relative to sales and, there-fore, rms with bloated balance sheets are more likely to have managed earnings upwardin the past. However, their metric assumes that all bloat in the balance sheet is due toearnings management and does not consider other plausible reasons why some rmsmay have a higher ratio of net operating assets to sales than others, such as differencesin strategy or protability. In addition, while the metric is potentially useful for identify-ing past earnings management, it is less useful for identifying earnings management inthe current period.

In short, existing proxies for earnings management build on the intuition that growthin net operating assets should be accompanied by growth in sales. As discussed above,however, these proxies have several limitations. In this study, we propose a new diagnosticfor earnings management that builds on similar intuition, but exploits the accounting



model to gain insight into when changes in the ratio of sales to net operating assets thatis, changes in the asset turnover ratio are likely due to earnings management.3

Hypotheses

It is well accepted that sales is the fundamental driver of net operating income on theincome statement and net operating assets on the balance sheet. Indeed, most nancialstatement analysis textbooks advocate forecasting income statement and balance sheet lineitems based on sales forecasts. For example, Penman (2007: 559) provides a frameworkfor forecasting and states that sales forecasting is the starting point. This intuition,which also underlies the widely used DuPont analysis, is that there should be a stable rela-tion between sales and both operating income on the income statement and net operatingassets on the balance sheet. These relations are captured by the ATO and PM ratios:ATO = Sales Net operating assets;PM = Operating income Sales.

We argue that ATO and PM should remain relatively constant in a stable operatingenvironment and that, in quality of earnings analysis, changes in ATO and or PM warrantfurther investigation. We also argue that one should be particularly concerned when ATOand PM change in opposite directions as this could signal earnings management. This isbased on the fact that the articulation of the income statement and balance sheet forcesearnings management to affect operating income and net operating assets in the samedirection. This is apparent from the denition of net operating assets:

Net operating assetst Net operating assetst1 DWorking capitaltDepreciation expenset DLong-term net operating assetst

Net operating assetst1 Operating incomet Cash from operationst DLong-term net operating assetst

Thus, assuming that earnings are not managed through cash ows (e.g., real earningsmanagement), any upward management of operating income will also overstate netoperating assets. Because operating income is the numerator of PM and net operatingassets is the denominator of ATO, upward earnings management increases PM anddecreases ATO, while downward earnings management decreases PM and increases ATO.We therefore propose that directionally opposite changes in the PM and ATO ratios canbe used as a diagnostic for earnings management. We state our rst set of hypotheses asfollows:4

Hypothesis 1a. Contemporaneous increases in PM and decreases in ATO signal upwardearnings management.

Hypothesis 1b. Contemporaneous decreases in PM and increases in ATO signal down-ward earnings management.

3. McNichols (2000) suggests three methods to identify earnings management: (i) using aggregate accrual

models such as the Jones 1991 model, (ii) examining the behavior of specic accruals, and (iii) examining

the distribution of earnings after management. Our diagnostic is in the spirit of the aggregate accrual

models, in that the aim is to identify general earnings management. We do not examine specic accruals

because earnings management is likely to occur in accounts that may not be separately reported in the

nancial statement and footnotes. We use the distribution of realized earnings to test whether our diagnos-

tic is informative about earnings management.

4. All hypotheses are stated in the alternative form.



The ATO PM earnings management diagnostic is in the same spirit as abnormal accrualsmodels, in that it also attempts to identify discretionary growth on the balance sheet; how-ever, the ATO PM diagnostic exploits the accounting model to produce additional insightsabout earnings management over those obtained from abnormal accruals models. Forexample, consider a rm that invests in current operating assets in anticipation of futuresales growth. In this case, abnormal accruals would likely be positive, even in the absenceof upward earnings management, because additional investments in working capital (i.e.,additional accruals) would not necessarily be accompanied by current sales growth. TheATO PM diagnostic, on the other hand, would not suggest upward earnings managementfor this scenario, because even though ATO would decrease, investments in operatingassets in anticipation of future sales growth will not affect current PM. We therefore pro-pose that directionally opposite changes in ATO and PM can be a useful complement toabnormal accruals in detecting earnings management in academic research. This leads toour second hypothesis:

Hypothesis 2. The ATO PM diagnostic provides incremental and greater relativeinformation content over abnormal accruals in identifying earnings management.

More insights into the ATO PM earnings management diagnosticThe argument made above, which underlies the ATO PM diagnostic, is that under fairlygeneral conditions upward earnings management increases PM and decreases ATO anddownward earnings management decreases PM and increases ATO. These relations holdfor (noncash) expense management when there is no change in business strategy and whenthere is neutral accounting or aggressive conservative accounting with constant growth innet operating assets.5 The relations do not hold for all revenue management cases, how-ever, because the numerators and denominators of both ATO and PM would change. Thediagnostic will correctly classify (upward) earnings management if the revenue manage-ment causes PM to increase and ATO to decrease. In other words, the diagnostic willcorrectly signal earnings management if: (i) the prot margin on the managed revenues isgreater than the prot margin on unmanaged revenue and (ii) the asset turnover of themanaged revenue is less than the unmanaged asset turnover. This rst condition is verylikely because prot margin on normal revenue will be reduced by both product and per-iod costs, whereas managed revenue is likely accompanied by only additional productcosts, not additional period costs.

The second condition is less clear-cut. If the company does not accrue any expensesrelated to the managed component of sales, the ATO of managed sales will be equal toone because the numerator (sales) will equal the denominator (receivables). In our sam-ple, 82 percent of the rm-year observations have ATO greater than one, which makesit likely that, in the absence of related expenses being accrued, the second condition willalso generally be satised. However, sales management will most likely be accompaniedby some accrued expenses which will decrease net operating assets (e.g., decreases ininventory, increases in payables) resulting in the ATO of managed sales being greaterthan one. In these instances, the second condition may or may not be satised. Expensesaccrued on managed sales are not observable, and therefore we are unable to provideempirical estimates of how frequently rms in our sample meet or do not meet the sec-ond condition. In short, the ATO PM diagnostic will identify most expense-based earn-ings management, but can only capture sales management in certain situations, which is

5. See Penman 2007 (593603) and Rajan, Reichelstein, and Soliman 2007 for discussions of the interaction

between growth and accounting methods and their effects on nancial ratios.



a limitation of the diagnostic. As a result, the diagnostic is more prone to Type II errors(failure to identify earnings management) than Type I errors (falsely agging earningsmanagement).

The ATO PM diagnostic assumes that a company has not changed its strategy andhas constant growth rates in investment. A Type I error could occur if a company changesits strategy or experiences unexpected growth. For example, if a rm changes its strategyfrom low-margin high-turnover to high-margin low-turnover, the ratios would most likelymove in opposite directions even in the absence of earnings management. Under this sce-nario, however, it is also likely that a rms abnormal accruals would be signicantly non-zero, because the change in strategy would probably be accompanied by signicantaccruals (i.e., a signicant change in working capital) that would be characterized byaccruals models as abnormal. In addition, it seems more reasonable to assume that a rmwill continue the same strategy over time than to assume that all rms in the same indus-try have the same strategy, which is the implicit assumption in cross-sectional estimates ofthe Jones model.6

The ATO PM diagnostic also assumes that a company does not manage earningsthrough cash ows. A Type II error could occur if a rm manages earnings upward bydelaying advertising or research and development expenditures. In this case, PM wouldincrease because of higher operating income, but ATO would be unaffected because of theabsence of an accrual on the balance sheet. Of course, given the latter, accruals modelswould also likely fail to detect such earnings management.

3. Sample, variable denitions, and descriptive statistics

Sample

We obtain nancial statement data from the 2006 COMPUSTAT Annual Industrial, FullCoverage, and Research tapes. Because funds from operations (COMPUSTAT data item#110) which which we need to compute cash from operating activities in years before1988 is not available prior to 1971, and because we need year-ahead data for severalvariables, our sample spans the years 1971 through 2005 (we use the COMPUSTAT yearconvention). The nancial statement variables we use in our study are available for118,679 rm-year observations. We eliminate observations in which net operating assetsare negative in year t)1 or year t, because ATO is undened for negative net operatingassets (5,578 observations). We also eliminate all nancial rms from our analyses (SIC60006999) because it is difcult to distinguish between operating and nancial activitiesfor these rms (2,931 of the remaining observations). After applying the above screens,our primary sample consists of 110,170 rm-year observations.7 Missing stock returnsaround the earnings announcement date of the rst scal quarter of the next year reducethe sample size to 67,075 observations for our abnormal returns tests. When we intersectthe primary sample with nonmissing analyst forecast data from I B E S, the sample sizedecreases to 46,522 observations. For the restatement analysis, the sample size decreases to

6. Consider, for example, Walmart and Macys, which are in the same four digit SIC code but have very dif-

ferent strategies. It seems more realistic to assume that each rm continues its strategy over time than to

assume that the rms have similar strategies (i.e., forcing the abnormal accruals model parameters to be

equal across all rms in the industry).

7. As a robustness test, we also ran the analyses after eliminating rms that were involved in any divestitures

and or mergers or acquisitions, because these transactions can cause the articulation between balancesheet changes and the income statement to break down. Specically, we eliminated rm-year observations

in which a rm discontinued operations or was involved in a merger or acquisition (COMPUSTAT annual

footnote code #1) in year t)1, year t, or year t + 1. We also deleted rm-years with increases in goodwillin year t)1, year t, or year t + 1. The results are qualitatively similar to those reported in the tables.



22,160 because we only have restatement data available for the 20002005 period. Weobtain our base sample of 2,319 restatements from Glass Lewis.8

Variable measurement

We provide a detailed description of the denition and measurement of all variables inTable 1. We dene the change in PM (DPMt) and the change in ATO (DATOt) as follows:

DPMt operating incomet=salest operating incomet1=salest1; andDATOt salest=net operating assetst salest1=net operating assetst1:

We argue that a contemporaneous increase in PM and decrease in ATO signals poten-tial upward earnings management, and that a contemporaneous decrease in PM andincrease in ATO signals potential downward earnings management. We dene two corre-sponding indicator variables that represent our diagnostic for earnings management:EM_UP for upward earnings management and EM_DN for downward earnings manage-ment. They are dened as follows:

EM UPt one if DPMt>0;DATOt

TABLE 1

Variable denitions

DPMt = change in PM = (operating incomet salest (COMPUSTAT data item #12);where operating incomet = salest (#12) (cost of goods sold (#41) + selling,

general and administrative expenses (#189) + depreciation and amortization

expense (#14))t;

DATOt = change in ATO = (salest (COMPUSTAT data item #12) net operating assetst) (salest)1 net operating assetst)1); where net operating assetst = net assetst(#216) net nancial assetst; and net nancial assetst = cash and short term

investments (#1) interest-bearing liabilitiest (#34 + #9).

EM_UPt = 1 if DPMt > 0, DATOt < 0, and EM_DNt)1 1, and 0 otherwise.EM_DNt = 1 if DPMt < 0, and DATOt > 0, and EM_UPt)1 1, and 0 otherwise.PABNACt = performance adjusted abnormal accruals = the tted residual from the follow-

ing model:

TACt TAt)1 = a1(1 TAt)1) + a2((1 + k)(DREVt )DRECt) TAt)1)+ a3(PPEt TAt)1) + a4(TACt)1 TAt)1) + a5(RNOAt TAt)1)

+ et,

where:

TACt = income before extraordinary itemst (#18) cash from opera-

tions (CFO)t;

TAt)1 = total assetst)1 (#6),

DREVt = changes in salest (#12),DRECt = change in receivablest (#2);PPEt = gross property, plant, and equipmentt (#7);

RNOAt = return on net operating assets;

CFOt = net cash ow from operating activities (#308) in 1988 and

thereafter;

CFOt = funds from operationst (#110) change in current assetst (#4)

+ change in cash and short-term investmentst (#1) + change

in current liabilitiest (#5) change in short term debtt (#34)

prior to 1988;

where all variables above are deated by total assetst)1 (#6); and k = the

slope 1 coefcient from the following model:

DRECt = b0 + kDREVt + u.PAB_UP = 1 if PABNAC is positive, and 0 otherwise;

PAB_DN = 1 if PABNAC is negative, and 0 otherwise.

SURP = actual earnings less the median analyst forecast of earnings closest to the earn-

ings announcement date. MBE = 1 if SURP is between 0 and 2 cents, and 0

otherwise;

EES = 1 if SURP is in the top or bottom decile of SURP, and 0 otherwise;

RESTATE = the difference (in $ millions) between restated and the originally reported oper-

ating income for year t;

DN_RESTATEt = 1 if the rm subsequently restates its operating earnings down for year t, and 0

otherwise;

UP_RESTATEt = 1 if the rm subsequently restates its operating earnings up for year t, and 0

otherwise;

NO_RESTATEt = 1 if the rm does not subsequently restated its operating earnings for year t,

and 0 otherwise;

(The table is continued on the next page.)



our diagnostic is intended to ag all potential earnings management, we do not use perfor-mance-matched abnormal accruals in our analyses, but instead use the alternativeapproach suggested in Kothari et al. 2005, in which rm performance is included in theaccruals estimation.

Accordingly, we add the rms return on net operating assets (RNOA) for the currentperiod to the enhanced version of the Jones 1991 model suggested in Dechow et al. 2003(259, model 3) and estimate abnormal accruals.11 Specically, we estimate performance-adjusted abnormal accruals (PABNACt) as the tted residual from the following model:

12

TACtTAt1

a1 1TAt1

a2 1 kDREVt DRECt1

TAt1

a3 PPEt

TAt1

a4 TACt1TAt1

a5 RNOAt

TAt1

et 1

where:

TACt = income before extraordinary itemst cash from operations (CFO)t;TAt)1 = total assetst)1,;DREVt = changes in salest;DRECt = change in receivablest;PPEt = gross property, plant and equipmentt;RNOAt = return on net operating assets;CFOt = net cash ow from operating activities (for rm-years from 1988);

= funds from operationst change in current assetst + change in cash andshort-term investmentst + change in current liabilitiest change in shortterm debtt (for rm-years prior to 1988); and

k = the slope coefcient from the following model:DRECt = b0 + kDREVt + u.

TABLE 1 (Continued)

ABRETt+1 = the size-adjusted cumulative abnormal return over the three-day period centered on

next years rst scal quarter earnings announcement;

IRNOAt = average of RNOAt for the 2-digit SIC code to which the rm belongs (excluding the

rm);

MTBt = the market-to-book ratio dened as the ratio of the rms market value of equity

(#25 * #199) divided by its book value (#60) at the end of the scal year;

MVEt = the rms market value of equity (#25 * #199);

RNOAt = return on net operating assets = operating incomet average net operating assetst;where average net operating assetst = (net operating assetst + net operating

assetst)1) 2;DRNOAt = change in return on net operating assets = RNOAt RNOAt)1;NOAt = net operating assetst salest; andDNOAt = change in net operating assets = (net operating assetst net operating assets

t)1) net operating assetst)1.

11. Dechow et al. (2003) also modify the Jones model by including forward-looking sales growth. We do not

use this variation because our purpose is to identify earnings management using current, not future, nan-

cial statement information. In addition, Dechow and Dichev (2002) suggest an alternative model that cap-

tures earnings quality over an extended time period. We do not examine this measure because our focus is

on short-horizon earnings management behavior, not long-run earnings quality.

12. We suppress rm subscripts.



Following DeFond and Jiambalvo 1994, Subramanyam 1996, and Xie 2001, weestimate the model in cross-section, for each two-digit SIC code and year combination.13

While PABNAC is a continuous measure in our primary analyses, we also examine anindicator variable, PAB_UP (PAB_DN), which equals one if PABNAC is positive (nega-tive), and zero otherwise, as indicators of upward (downward) earnings management.

We examine ve variables that are intended to capture the consequences of earningsmanagement, or earnings management outcomes, to determine if the ATO PM diagnos-tic is predictive of these outcomes. Specically, we examine if the diagnostic is useful inidentifying rms that meet or just beat analyst forecasts (MBE), report extreme earningssurprises (EES), subsequently restate earnings upward (UP_RESTATE) or downward(DN_RESTATE), experience a reversal in year-ahead protability (DRNOAt+1) or pro-duce predictable year-ahead abnormal returns (ABRETt+1). We dene the variables asfollows:MBE = one if SURP is greater than or equal to zero and less than $0.02, and

zero otherwise;whereSURP = actual earnings per share less the consensus median analyst forecast

of earnings per share closest to the earnings announcement date;14

EES = one if SURP is in the top or bottom decile of SURP, and zero other-wise;

DN_RESTATEt = one if the rm subsequently restates its operating earnings down foryear t, and zero otherwise;

UP_RESTATEt = one if the rm subsequently restates its operating earnings up for yeart, and zero otherwise;

DRNOAt = RNOAt+1 RNOAt; andABRETt+1 = the size-adjusted cumulative abnormal return over the three-day per-

iod centered on next years rst scal quarter earnings announcement.We dene control variables as follows:15

IRNOAt = average RNOAt for two-digit SIC code to which the rm belongs(excluding rm);

MVE = market value of equity;MTB = MVE book value of equity at the end of the scal year;RNOAt = operating incomet average net operating assetst;DRNOAt = RNOAt RNOAt)1;NOAt = net operating assetst salest;16 andDNOAt = (NOAt NOAt)1) NOAt)1.

13. Consistent with these studies, we require at least six observations for a given combination to be included

in the sample. We also performed the analyses using the Jones model and the modied Jones model to cal-

culate abnormal accruals. The results (untabulated) using these alternative abnormal accrual models are

similar to those reported in the tables.

14. We also dened MBE as analyst forecast errors greater than or equal to zero but less than $0.01. The

results are similar to those reported in the tables.

15. Note that we use ending net operating assets in computing ATO as well as for the change in net operating

assets so as to better capture earnings management that occurs at the end of the year. On the other hand,

we compute return on net operating assets using average net operating assets in the denominator because

the variable is included to control for the overall protability of the company for the year. Redening

RNOA using ending net operating assets in the denominator does not alter our conclusions.

16. We use the term NOA to describe this variable to be consistent with Barton and Simko 2002 but note that

since net operating assets is scaled by sales, the ratio reects the inverse of the ATO ratio. Excluding this

variable from the analyses does not alter our conclusions.



Descriptive statistics

Table 2 presents descriptive statistics, sample correlations, and univariate analyses. Allvariables are winsorized at the 1st and 99th percentiles to mitigate the inuence of outliers.The descriptive statistics in panel A suggest that the ATO PM diagnostic identies poten-tial upward earnings management in 14.8 percent of the sample observations and down-ward earnings management in 17.3 percent of the sample observations. EM_UP and

TABLE 2

Descriptive statistics

Panel A: Descriptive statisticsa

Variable Mean Std Dev 25% Median 75%

EM_UPt 0.148 0.355 0.000 0.000 0.000

EM_DNt 0.173 0.378 0.000 0.000 0.000

PABNACt )0.008 0.386 )0.071 )0.002 0.065PAB_UPt 0.495 0.499 0.000 0.000 1.000

DPMt 0.005 0.275 )0.027 0.000 0.025DATOt 0.052 1.369 )0.225 0.023 0.265IRNOAt )0.038 0.092 )0.080 )0.022 0.034MTBt 2.38 3.41 0.87 1.55 2.78

MVEt 969.28 3291.29 14.33 67.16 388.33

RNOAt 0.000 0.508 )0.030 0.095 0.194DRNOAt )0.008 0.316 )0.070 )0.002 0.056NOAt 0.795 1.083 0.333 0.510 0.796

DNOAt 0.207 0.658 )0.056 0.071 0.259MBEt 0.243 0.429 0.000 0.000 0.000

EESt 0.200 0.399 0.000 0.000 0.000

DN_RESTATEt 0.016 0.126 0.000 0.000 0.000

UP_RESTATEt 0.011 0.105 0.000 0.000 0.000

DRNOAt+1 )0.018 0.352 )0.070 )0.002 0.055ABRETt+1 0.006 0.093 )0.034 0.000 0.039

Panel B: Sample Pearson correlations

Variable DPMt DATOt IRNOAt MTBt MVEt RNOAt DRNOAt NOAt DNOAt PABNACt EM_UPt EM_DNt

DPMt 1.00 0.08 )0.01 0.08 0.00 0.01 0.48 )0.02 0.06 0.10 0.10 )0.15

DATOt 1.00 )0.01 )0.01 )0.01 )0.11 0.05 )0.07 )0.56 )0.08 )0.22 0.20

IRNOAt 1.00 )0.17 )0.12 0.30 )0.02 )0.14 )0.07 )0.05 0.01 0.00

MTBt 1.00 0.22 0.01 0.11 )0.03 0.15 0.03 0.09 )0.07

MVEt 1.00 0.10 0.02 0.01 )0.01 )0.03 0.03 )0.03RNOAt 1.00 0.24 )0.10 0.02 0.01

0.10 )0.08DRNOAt 1.00 0.03 0.14 0.07 0.16 )0.18NOAt 1.00 0.15 )0.01 )0.10

)0.05DNOAt 1.00 0.14 0.25 )0.20PABNACt 1.00 0.09 )0.09EM_UPt 1.00 )0.19EM_DNt 1.00




TABLE 2 (Continued)

Panel C: EM_UP, EM_DN, and PAB_UP by industry

Industry by rst digit of SIC code N EM_UP EM_DN PAB_UP

0 Agriculture and shing 338 14% 17% 44%

1 Extraction and construction 8,583 13% 16% 44%

2 Commodity production 19,812 15% 18% 46%

3 Manufacturing 42,815 14% 16% 50%

4 Utilities and transportation 5,949 15% 18% 49%

5 Wholesale and retail 15,696 17% 19% 50%

7 Business services and entertainment 12,128 16% 18% 55%

8 Health and other services 3,853 17% 19% 54%

9 Public administration 996 12% 19% 5%

Panel D: EM_UP, EM_DN, and PAB_UP by year

Year n EM_UP EM_DN PAB_UP

1971 1,929 12% 20% 50%

1972 1,894 12% 18% 50%

1973 1,908 14% 15% 50%

1974 1,991 12% 20% 50%

1975 2,215 9% 18% 48%

1976 2,913 12% 17% 49%

1977 2,767 15% 17% 51%

1978 2,697 17% 17% 49%

1979 2,589 13% 19% 50%

1980 2,510 11% 21% 51%

1981 2,548 13% 18% 47%

1982 2,673 11% 17% 44%

1983 2,788 16% 13% 45%

1984 2,953 15% 14% 52%

1985 3,020 14% 19% 47%

1986 3,031 16% 17% 48%

1987 2,665 15% 16% 45%

1988 3,003 12% 21% 50%

1989 3,283 14% 19% 45%

1990 3,318 14% 21% 46%

1991 3,404 13% 20% 46%

1992 3,507 14% 18% 46%

1993 3,596 17% 17% 49%

1994 3,791 18% 14% 49%

1995 3,935 19% 15% 50%

1996 4,096 19% 15% 52%

1997 4,453 21% 14% 49%

1998 4,390 19% 14% 50%

1999 4,137 18% 15% 58%




EM_DN thus identify approximately 32 percent of rms as having managed earnings,which is in contrast to PABNAC which, by construction, identies all rms as havingmanaged earnings either up or down. The distributions of the control variables are consis-tent with prior research (e.g., Faireld and Yohn 2001). The descriptive statistics on theearnings management outcome variables suggest that 24.3 percent of the observations meetor beat analyst expectations (MBE) while, by construction, 20 percent of the observationsare classied as extreme earnings surprises (EES). Consistent with restatements being anunusual event, only 1.6 (1.1) percent of the observations experiences downward (upward)restatements of reported earnings. The mean year-ahead change in return on net operatingassets is )1.8 percent and size-adjusted abnormal returns around the earnings announce-ment date of the rst scal quarter of the subsequent year are 0.60 percent.

Panel B of Table 2 reports Pearson correlation coefcients between the explanatoryvariables used in the study. Consistent with information overlap in both diagnostics, thereis a signicant positive correlation of 0.09 between EM_UP and PABNAC and a signi-cant negative correlation of )0.09 between EM_DN and PABNAC. EM_UP and EM_DNare signicantly correlated with most of the control variables. None of the correlationsexceed 30 percent, however, suggesting that EM_UP and EM_DN capture a substantialamount of unique information relative to the other variables. All reported correlations aresufciently low to mitigate concerns about multicollinearity when estimating multivariatemodels.

In panel C of Table 2, we report the percentage of observations with EM_UP,EM_DN and PAB_UP (the observations for which PABNAC is positive) by industry.17

We dene industries by the rst digit of the rms primary SIC code. We note that there issome variation in the frequency of EM_UP and EM_DN across industries. Although wedo not perform formal tests for differences across industries, we include industry xedeffects in all of our tests to control for industry clustering.

In panel D of Table 2, we report the percentage of observations with EM_UP,EM_DN, and PAB_UP by year. We note that the frequency of EM_DN observations isgreater than that of EM_UP in 24 of the 35 years. The years in which EM_DN is lessfrequent are concentrated in the period 1993 through 1999 when the EM_UP percentage is

TABLE 2 (Continued)

Year n EM_UP EM_DN PAB_UP

2000 4,109 14% 19% 58%

2001 4,018 9% 23% 46%

2002 3,832 14% 17% 60%

2003 3,715 15% 19% 53%

2004 3,573 16% 16% 52%

2005 2,932 15% 18% 47%

Notes:

a ABRETt+1 includes 67,075 observations; MBE EES includes 46,522 observations; RESTATEincludes 22,160 observations; all other variables include 110,170 observations.

indicates that the correlation coefcient estimate is not signicantly different from zero at the

10 percent level (two-tailed). All remaining correlation coefcient estimates are signicantly

different from zero at the 10 percent level (two-tailed).

17. We do not report PAB_DN because the abnormal accruals model classies all observations as either man-

aging earnings up or down, and therefore, PAB_DN = 1 ) PAB_UP.



markedly higher relative to other years. It is possible that rms were under greater pres-sure to manage earnings up during the bull market of the late 1990s. Alternatively, thesurge in EM_UP observations may be related to the rise of Internet and high-growth rmsduring this period. We also note that the EM_UP percentage drops markedly and theEM_DN percentage increases in 2000 and 2001, likely due to rms reporting conserva-tively subsequent to the accounting scandals that surfaced during this period. We also notethat the PAB_UP percentage is greater than 50 percent in only nine of the 35 years sug-gesting that, similar to the ATO PM diagnostic, PABNAC is more often negative thanpositive. However, in contrast to the ATO PM diagnostic, the PAB_UP percentage tendsto be greater than 50 percent in the later years of the sample.

4. Results

We present and discuss our results in two sections. First, we present contingency tables toprovide insight into the univariate relations between EM_UP, EM_DN, PAB_UP,PAB_DN, and the earnings management outcome variables. Next, we present results frommultivariate analyses that investigate the relation between the earnings management out-come variables and the ATO PM diagnostic.

Univariate analyses and ndings

Panel A in Table 3 shows the relation between EM_UP and PAB_UP, and betweenEM_DN and PAB_DN. In the rst contingency table, we report the percentage of obser-vations for which EM_UP is equal to zero or one and PAB_UP is equal to zero or one.When PAB_UP is equal to zero, EM_UP is equal to zero in 89.8 percent of observationsand equal to one in 10.2 percent of observations. When PAB_UP is equal to one, EM_UPis equal to zero in 80.6 percent of observations and equal to one in 19.4 percent of obser-vations. This suggests that the ATO PM diagnostic ags upward earnings managementalmost twice as often when PABNAC is positive than when it is negative (19.4 percent vs.10.2 percent). Similarly in the EM_DN and PAB_DN contingency table in panel A, theATO PM diagnostic ags downward earnings management almost twice as often whenPABNAC is negative than when it is positive (22.6 percent vs. 11.9 percent). A test of pro-portions shows that these ratios are highly signicant, which indicates that EM_UP agsupward earnings management much more often when PABNAC is positive than when it isnegative, and that EM_DN ags downward earnings management much more often whenPABNAC is negative than when it is positive.

Panel B in Table 3 shows the relation between EM_UP, PAB_UP, and rms that meetor beat expectations (MBE). We nd that PAB_UP correctly ags more of the meet orbeat observations than EM_UP (51.2 percent versus 21.3 percent); however, PAB_UP alsoincorrectly ags more of the non-MBE observations than EM_UP as having managedearnings (48.9 percent versus 14.7 percent). This is not surprising because there are morePAB_UP than EM_UP observations with a value of one. To make a more meaningfulcomparison between the two diagnostics, we rely on the following observation: if a diag-nostic is useful in identifying a particular event, we expect the diagnostic to ag the eventmore often when the event occurs (= 1) than when the event does not occur (= 0). Inother words, the ratio of proportions, the former proportion over the latter proportion,should be signicantly greater than one. The ratio of proportions for PAB_UP is 1.05(51.2 percent over 48.9 percent), which is signicantly greater than one (Z-statistic = 4.08;p-value < 0.001) based on a test of proportions (Hildebrand, Ott, and Gray 2005). Thecorresponding ratio for EM_UP is 1.45 (21.3 percent over 14.7 percent), which is also sig-nicantly greater than one (Z-statistic = 15.48; p < 0.001). Thus, both diagnostics areuseful for identifying MBE, but EM_UP appears to be more discriminating than PAB_UPbased on a comparison of the two ratios. It is also evident from the contingency tables



TABLE 3

Contingency tablesa

Panel A: The association between EM_UP and PAB_UP and between EM_DN and PAB_DN

EM_UP

N

EM_DN

N0 1 0 1

PAB_UP 0 89.8% 10.2% 55,351 PAB_DN 0 88.1% 11.9% 54,819

1 80.6% 19.4% 54,819 1 77.4% 22.6% 55,351

Ratio of proportions 1.90 Ratio of proportions 1.90

Z-statistic 43.16 Z-statistic 47.59

p-value

that PAB_UP is more prone to Type I errors (agging observations as having managedearnings when they have not), while EM_UP is more prone to Type II errors (not identify-ing rms that may have managed earnings). Indeed, all panels in this table conrm thehigher propensity for Type I errors for the PABNAC diagnostic, and for Type II errorsfor the ATO PM diagnostic.18

Panel C in Table 3 shows the relation between EM_DN, PAB_DN, and rms thathave extreme earnings surprises (EES). As discussed earlier, rms with EES are morelikely to manage earnings down; consequently, we expect PAB_DN and EM_DN to bemore frequently equal to one for EES rms than for other rms. The contingency tablesconrm our expectation: the ratio of proportions where EM_DN is equal to one whenEES is equal to one, compared to when EES is equal to zero, is 1.40 (Z-statistic = 13.12;p < 0.0001). This indicates that EM_DN ags downward earnings management muchmore frequently when a rm reports extreme earnings surprises than when it does not.Similarly, the corresponding ratio of proportions for PAB_DN equals 1.10 (Z-statis-tic = 8.76; p < 0.001), conrming the usefulness of PABNAC for identifying rms withextreme earnings surprises as well. The evidence is consistent with the ability of PAB_DNand EM_DN to identify possible downward earnings management.

TABLE 3 (Continued)

EM_DN

N

PAB_DN

N0 1 0 1

Ratio of

proportions

1.18 Ratio of

proportions

1.05

Z-statistic 1.31 Z-statistic 1.01

p-value

Finally, panels D and E in Table 3 provide similar details with respect to downward andupward earnings restatements. The proportion of observations classied as having managedearnings up is not signicantly higher for either EM_UP or PAB_UP when rms restateearnings downward as opposed to when they do not. In other words, neither EM_UP norPAB_UP appears to discriminate between rms that may have initially managed earningsup from rms that do not manage earnings. When rms subsequently restate earnings up,on the other hand, EM_DN is signicantly more likely to ag these observations as havinginitially managed earnings down (ratio = 1.58; Z-statistic = 3.73; p < 0.001). However,the proportion of rms that PAB_DN identies as having managed earnings down is not sig-nicantly higher when rms subsequently restate earnings up than when they do not.

Taken together, Table 3 provides preliminary evidence of the usefulness of theATO PM diagnostic for identifying earnings management associated with rms meetingor beating analyst forecasts, reporting extreme earnings surprises, and subsequently restat-ing earnings upward. In addition, it appears that the ATO PM diagnostic is at least asreliable in identifying earnings management as PABNAC. We next investigate the relativeand incremental information content of the two diagnostics in multivariate analyses, wherewe also control for other variables known to be related to the earnings management out-come variables.

Multivariate analyses and ndings

Meet or beat expectations

Prior research (e.g., Burgstahler and Eames 2006; Matsumoto 2002; Moehrle 2002; Chengand Wareld 2005) provides evidence that rms manage earnings upward to avoid missinganalyst forecasts of earnings. This suggests that rms that meet or just beat analyst fore-casts are more (less) likely to have managed their earnings upward (downward) than otherrms. We therefore predict that EM_UP (EM_DN) is positively (negatively) associatedwith the incidence of rms meeting or just beating analyst forecasts.19

Descriptive statistics for analyst forecast errors (SURP), EM_UP, EM_DN andPABNAC by MBE are reported in panel A of Table 4. We note that 11,286 rm-yearsmeet or just beat the analyst forecast. We nd a mean EM_UP of 21.34 percent for MBErms compared to a mean of only 14.69 percent for all other rms. This difference is sta-tistically signicant at the one percent level. Our results also show a mean EM_DN of13.61 percent for MBE rms compared to a mean of 17.52 percent for all other rms. Thisdifference is also signicant at the one percent level. These univariate results are consistentwith EM_UP and EM_DN identifying rms that meet or just beat analyst forecasts. PanelA also provides evidence that PABNAC is signicantly higher for MBE rm-years relativeto non-MBE rm-years.

In panel B of Table 4, we use multivariate analyses to evaluate the ability of EM_UPand EM_DN to identify rms that meet or just beat analyst forecasts. Specically, we runlogistic regressions with MBE as the dependent variable and EM_UP, EM_DN, andPABNAC as the primary explanatory variables.20 We also include industry RNOA(IRNOAt), market to book ratio (MTBt),

21 and market value of equity (MVEt) as control

19. Barton and Simko (2002) point out that the pressure to meet analyst forecasts appears to be a relatively

recent phenomenon and that I B E S changed the formulae to calculate actual earnings per share in1993. Given this, we also performed the analysis after excluding observations prior to 1993. The results

are similar to those reported in the tables.

20. It is possible that rm-years in which the rm missed the analyst forecast by a large amount reect signi-

cant economic changes, which could confound our analyses. Therefore, we repeated the tests including

only those rm-year observations with earnings surprises between )$0.02 and +$0.02. The results are sim-ilar to those reported in the tables.

21. We repeated all the multivariate tests after including the price to earnings ratio and sales growth as addi-

tional control variables. The results are similar to those reported.



TABLE4

AssociationbetweenEM_UP,EM_DN,andmeetingorjustbeatingexpectations(M

BE)

Panel

A:Characteristics

ofMBErm

srelativeto

other

rm

sa

MBE

Allother

rm

sTestofdifference

#observations

11,286

35,236

MeanSURP

0.0063

)0.1108

48.793***

MedianSURP

0.0051

)0.0100

71.79***

MeanEM_UP

0.2134

0.1469

15.48***

MeanEM_DN

0.1361

0.1752

)10.29***

MeanPABNAC

0.0120

0.0007

3.52***

MedianPABNAC

0.0021

)0.0018

4.08***

Panel

B:MultivariateanalysisoftheassociationbetweenEM_UP,EM_DN,andMBEb

MBEtc 0c 1IRNOAtc 2MTBtc 3MVEtc 4DATOtc 5DPM

tc 6RNOAtc 7DRNOAtc 8NOAtc 9DNOAtc 1

0PABNACtc 1

1EM

UPtc 1

2EM

DNtn t1

IRNOA

MTB

MVE

DATO

DPM

RNOA

DRNOA

NOA

DNOA

PABNAC

EM_UP

EM_DN

PseudoR2

1.3244***

0.0560***

0.0001***

)0.0365***

0.2583***

0.0588

6.03%

(29.42)

(258.66)

(106.29)

(16.81)

(28.74)

(2.25)

1.2586***

0.0526***

0.0001***

)0.0021

0.1592***

0.3354***

)0.1834***

6.57%

(25.60)

(225.86)

(98.95)

(0.05)

(10.57)

(127.39)

(31.15)

1.2717***

0.0525***

0.0001***

)0.0016

0.1560***

0.0266

0.3345***

)0.1825***

6.57%

(26.99)

(225.41)

(99.22)

(0.03)

(10.05)

(0.46)

(126.47)

(30.79)

0.5606**

0.0468***

0.0001***

0.0344***

0.2505***

0.67808***

)0.0044

)0.0295*

0.0754***

0.0073

0.2701***

)0.1330***

8.33%

(5.09)

(158.79)

(48.94)

(8.30)

(13.79)

(431.59)

(0.01)

(3.08)

(12.20)

(0.03)

(78.39)

(16.04)

Notes:

aTestofdifference

presentsat-statistic(Z-statistic)from

at-test(W

ilcoxonsigned

ranktest)ofdifference

inmeans(m

edians).bModelsare

estimatedusing

44,237rm

-yearobservations,controllingforxed

effectsbyyearandindustry

(two-digitSIC

code).v2statisticsare

reported

inparentheses.See

Table1

forvariabledenitions.

***,**and

*indicatethatvalueissignicantlydifferentfrom

zero

atthe1,5,and10percentlevel,respectively(two-tailed).



variables because prior research (e.g., Koh, Matsumoto, and Rajgopal 2008; Barton andSimko 2002) shows that these variables are informative for identifying rms that meet orjust beat expectations.22 We further include the change in asset turnover (DATOt) and thechange in prot margin (DPMt) as control variables to test whether EM_UP and EM_DNprovide incremental information over these fundamental ratios on which the diagnostic isbased. Finally, we include xed effects dummy variables by year and two-digit SIC codeto control for cross-sectional and time-series correlation in the errors.

In the rst model in panel B, we include the control variables and PABNAC. We ndthat rms with higher IRNOA, higher MTB, and higher MVE are more likely to meet orjust beat analyst forecasts than other rms. These results are consistent with prior research(e.g., Barton and Simko 2002). The signicant positive coefcient on DPM and signicantnegative coefcient on DATO are consistent with the notion that rms that manageearnings upward experience increases in PM and decreases in ATO. We also note thatPABNAC is not signicant at the 10 percent level, suggesting that abnormal accruals donot provide incremental information in identifying rms that are more likely to meet orbeat analyst forecasts.

In the second model, we include the control variables and EM_UP and EM_DN. Wedocument a positive and signicant coefcient on EM_UP and a negative and signicantcoefcient on EM_DN (both at the 1 percent level). The results suggest that the ATO PMdiagnostic for earnings management is successful in identifying rm-years that meet or justbeat analyst forecasts. We also note that the pseudo R2 is higher for the EM_UP EM_DNmodel than for the PABNAC model. We formally compare the relative predictive powerof models 1 and 2 by comparing the area under the Receiver Operating Characteristic(ROC) curves for the two models (DeLong, DeLong, and Clarke-Pearson 1988). We nd(untabulated) that the area under the ROC curve for the PABNAC model is 0.6377 com-pared to 0.6415 for the ATO PM diagnostic model, and the difference is signicant at lessthan the one percent level (v2 = 15.08). Consistent with the univariate ndings in the con-tingency tables presented in Table 3, this evidence conrms that, compared to the PAB-NAC model, the multivariate ATO PM model better discriminates between rms thatmeet or just beat expectations versus rms that do not.

In the third model, we include PABNAC and EM_UP and EM_DN to assess whetherEM_UP and EM_DN provide incremental information over this widely used proxy forearnings management. We nd that both EM_UP and EM_DN are still highly signicantwith the expected signs, whereas the coefcient on PABNAC is not signicant at conven-tional levels. The odds ratio for EM_UP in this specication is 1.40, which indicatesthat, after controlling for the effect of other variables, the odds of meeting or just beatinganalyst forecasts are 40 percent higher for observations where EM_UP ags earnings man-agement than when it does not. In the nal model, we follow Barton and Simko 2002 andinclude the current return on net operating assets (RNOAt), current change in return onnet operating assets (DRNOAt), current net operating assets (NOAt), and the currentchange in net operating assets (DNOAt) as additional control variables, to determine ifEM_UP and EM_DN provide incremental information over these nancial statementratios. Again, we nd an insignicant coefcient on PABNAC and a signicant positive(negative) coefcient on EM_UP (EM_DN). The results suggest, therefore, that the

22. We note that prior research (e.g., Barton and Simko 2002) has included additional control variables, such

as the rms litigation risk, whether the rm previously met or beat analyst forecasts, and auditor type, in

the analysis of rms that meet or beat analyst forecasts. We do not include these variables as our interest

is not in developing a comprehensive model to predict rms that meet or beat analyst forecasts but rather

to determine the nancial statement variables that are useful in predicting whether a rm meets or beats

analyst forecasts (i.e., managed earnings).



ATO PM diagnostic is useful for identifying rms that meet or beat expectations whilePABNAC has no association with rms that meet or just beat expectations.23

Extreme earnings surprises

In the previous section, we based our analyses on the notion that rms manage earningsupward to meet or just beat analyst forecasts. In this section, we argue that rms that missor beat the analyst forecast by a large amount are more likely to have managed earningsdownward. That is, rms with surprisingly low earnings have an incentive to take a bath,while rms with surprisingly high earnings have an incentive to save earnings for thefuture (i.e., create cookie-jar reserves). Further, given that they have already missed (orbeaten) expectations by a wide margin, these rms are unlikely to manage earningsupward. We therefore predict that EM_DN (EM_UP) is positively (negatively) associatedwith the incidence of extreme earnings surprises (EES).

Results of this test are reported in Table 5. We rst sort all earnings surprises intodeciles and then label the observations in the top and bottom deciles as extreme earningssurprises (EES). Panel A provides descriptive statistics. The mean (median) absolute earn-ings surprise in the extreme deciles is $6.82 ($0.40).24 The mean (median) absolute earningssurprise for the remaining deciles is $0.05 ($0.02). The mean EM_UP is 11.56 percent forEES rms compared to a mean of 17.48 percent for other rms. On the other hand, themean EM_DN is 21.46 percent for EES rms compared to a mean of 15.35 percent for allother rms. The mean and median PABNAC are also signicantly lower for EES rms rel-ative to the other rms.

In panel B of Table 5, we estimate logistic regressions with the indicator variable EESas the dependent variable. As in the previous analysis, we include EM_UP, EM_DN, con-trol variables, PABNAC, and year and industry xed effects. We expect a positive (nega-tive) coefcient on EM_DN (EM_UP) because rms are more (less) likely to havemanaged earnings downward (upward) when they miss or beat analyst forecasts by a largeamount.

The rst model includes the control variables and PABNAC. The coefcient signs ofthe control variables are as expected, and have the opposite association with EES com-pared to MBE. Consistent with PABNAC being useful for identifying earnings manage-ment, we nd a signicantly negative coefcient on PABNAC. In the second model, wend a signicantly positive (negative) coefcient on EM_DN (EM_UP), suggesting thatthe ATO PM diagnostic provides information for identifying rms with EES, consistentwith our expectation. The pseudo R2 is 4.77 percent for the PABNAC model comparedto a pseudo R2 of 5.25 percent for the EM_UP EM_DN model, providing evidence thatthe ATO PM diagnostic provides more information for EES than PABNAC. Consistentwith this, we nd (untabulated) that the area under the ROC curve is larger for theATO PM model (0.6335) than for the PABNAC model (0.6305), and the difference is sig-nicant at less than the ve percent level (v2 = 5.75). Moreover, when we include allthree variables, EM_UP, EM_DN and PABNAC (model 3), we nd that the ATO PMdiagnostic is incrementally informative to PABNAC. The (untabulated) odds ratio of 1.30for EM_DN in this specication suggests that after controlling for the effect of othervariables, the odds of a rm having EES is 30 percent higher when EM_DN equals one.

23. We also estimated all regression models in this paper with the indicator PAB_UP in place of the continu-

ous variable PABNAC, as well as with indicator variables (for observations above and below the median)

instead of continuous variables for each of the explanatory variables. The (untabulated) results are similar

to those reported in the tables.

24. We do not scale the earnings surprise by price because it is difcult to argue that investors assess the sur-

prise relative to price. In addition, Cheong and Thomas (2009) show that the magnitude of earnings sur-

prises do not seem to vary with scale.



TABLE5

AssociationbetweenEM_UP,EM_DN,andextrem

eearningssurprises(EES)

Panel

A:Characteristics

ofEESrm

srelativeto

other

rm

sa

EES

Other

rm

sTestofdifference

#observations

9,272

37,250

Mean|SURP|b

6.8192

0.0478

3.55***

Median|SURP|

0.4000

0.0200

107.62***

MeanEM_UP

0.1156

0.1748

)15.34***

MeanEM_DN

0.2146

0.1535

13.12***

MeanPABNAC

)0.0124

0.0074

)5.51***

MedianPABNAC

)0.0099

0.0008

)8.75***

Panel

B:MultivariateanalysisoftheassociationbetweenEM_UP,EM_DN,andEESc

EEStc 0c 1IRNOAtc 2MTBtc 3MVEtc 4DATOtc 5DPM


0PABNACtc 1

1EM

UPtc 1

2EM

DNtn t1

IRNOA

MTB

MVE

DATO

DPM

RNOA

DRNOA

NOA

DNOA

PABNAC

EM_UP

EM_DN

PseudoR2

)1.6405***

)0.0735***

)0.0001***

0.1000***

)0.3344***

)0.1970***

4.77%

(40.78)

(221.50)

(79.19)

(103.56)

(37.82)

(21.24)

)1.5310***

)0.0687***)0.0001***

0.0682***

)0.2297***

)0.3060***

0.2696***

5.25%

(35.79)

(195.58)

(74.99)

(44.23)

(17.69)

(67.63)

(75.12)

)1.6053***

)0.0683***

)0.0001***

0.0652***

)0.2087***

)0.1654***

)0.3014***

0.2642***

5.30%

(39.01)

(193.36)

(75.70)

(40.18)

(14.56)

(14.83)

(65.56)

(71.97)

)0.5615**

)0.0633***

)0.0001***

)0.0121

)0.0604

)0.7126***

)0.0640

0.0911***

)0.1845***

)0.1387***

)0.2176***

0.2216***

8.22%

(4.52)

(171.90)

(31.49)

(0.94)

(1.26)

(770.04)

(1.83)

(41.99)

(50.90)

(9.90)

(32.86)

(48.63)

Notes:

aTestofdifference

presentsthet-statistic(Z-statistic)from

at-test(W

ilcoxonsigned


inmeans(m

edians).

b|SURP|istheabsolutevalueofSURP.

cModelsare

estimatedusing44,237rm


effectsbyyearandtwo-digitSIC

code.v2

statisticsare

reported

inparentheses.

See

Table1forvariabledenitions.

***,**and

*indicatethatvalueissignicantlydifferentfrom

zero

atthe1,5and10percentlevel,respectively(two-tailed).



In the nal model, when additional nancial statement ratios are included in the regres-sion, we continue to nd that EM_UP and EM_DN provide incremental informationcontent for EES.

Restatements

Dechow and Skinner (2000) note that earnings management can occur through accountingchoices that are acceptable within U.S. generally accepted accounting principles (GAAP)or through accounting choices that violate U.S. GAAP. The previous analyses on MBEand EES likely capture earnings management that falls within the acceptable use of discre-tion available in U.S. GAAP. In order to test whether EM_UP and EM_DN are informa-tive about earnings management that violates U.S. GAAP, we examine whether EM_UPand EM_DN are predictive of earnings restatements.

We argue that rms that subsequently restate their earnings downward (upward) aremore likely to have initially managed their earnings up (down). We therefore expectEM_UP to be more (less) frequent in a sample of rms that subsequently restate theirearnings downward (upward). Similarly, we expect EM_DN to be more (less) frequent ina sample of rms that restate their earnings upward (downward). We control for allpreviously identied variables including year and industry-xed effects and examinethe association between the ATO PM diagnostic and the probability of subsequentrestatements.

We obtain restatement announcements from the Glass Lewis database and then deter-mine the income effect of the restatements by comparing originally reported and restatednumbers in the COMPUSTAT Point-in-Time database. We sum the quarterly data overthe scal year for both the originally reported and the restated numbers and identifyrestatements as observations where the originally reported operating income25 differs fromthe restated income in the periods around the restatement ling.26 The nal sampleincludes 604 rm-year observations involving operating income restatements (355 down-ward and 249 upward) and 21,556 nonrestatement rm-year observations from 2000through 2005.

We report the restatement analysis results in Table 6. The descriptive statistics in panelA show that the mean (median) downward restatement is $90.52 million ($2.4 million),whereas the mean (median) upward restatement is $97.22 million ($1.44 million). We nd,consistent with expectations, that the mean EM_UP is signicantly higher (t-statistic of2.30) for rm-year observations that resulted in a downward restatement than in anupward restatement, and mean EM_DN is signicantly higher in the upward restatementsample than the downward restatement sample (t-statistic of )3.02). We do not nd a sim-ilar, signicant difference for PABNAC. We also nd that EM_UP (EM_DN) is signi-cantly lower (higher) for UP_RESTATE rms relative to rms that do not subsequentlyrestate earnings (t-statistics of )1.96 and 3.72, respectively). We document a similar meandifference for PABNAC. Finally, we do not nd signicant differences in EM_UP,EM_DN or PABNAC, in comparing DN_RESTATE rms relative to rms that do notrestate earnings (NO_RESTATE).

In panel B, we report results of logistic regression analyses with DN_RESTATE,which is set equal to one (zero) if the rm subsequently restates (does not restate) earnings

25. Consistent with the rest of the paper, we dene operating income as sales (cost of goods sold + selling,

general and administrative expenses + depreciation and amortization expense).

26. The Point-in-Time database provides nancial information as originally reported in the announcement

month as well as for each month thereafter. We identify the impact of restatements by capturing changes

in the nancial variables reported within four months surrounding the month of the restatement ling.

This methodology was used to ensure that we capture the effect of the restatement of interest rather than

the effect of other events that led to restatements.



TABLE6

AssociationbetweenEM_UP,EM_DN,andrestatementsofoperatingincome

Panel

A:Characteristics

ofDN_RESTATEandUP_RESTATErm

srelativeto

other

rm

sa

DN_RESTATE

UP_RESTATE

NO_RESTATE

Testofdifferences

DN_RESTATEvs.

UP_RESTATE

DN_RESTATEvs.

NO_RESTATE

UP_RESTATEvs.

NO_RESTATE

#ofobservations

355

249

21,556

MeanRESTATE($

millions)

)90.52

97.22

0.00

4.70***

3.84***

3.02***

MedianRESTATE($

millions)

)2.40

1.44

0.00

20.94***

)120.96***

112.83***

MeanEM_UP

0.1634

0.1004

0.1380

2.30**

1.28

)1.96*

MeanEM_DN

0.1859

0.2932

0.1852

)3.02***

0.03

3.72***

MeanPABNAC

0.0197

)0.0444

0.0095

1.64

0.41

)1.77*

MedianPABNAC

0.0119

0.0025

0.0076

1.20

0.43

)1.08

Panel

B:MultivariateanalysisoftheassociationbetweenEM_UP,EM_DNandDN_RESTATEb

DNRESTATEtc 0c 1IRNOAtc 2MTBtc 3MVEtc 4DATOtc 5DPM


0PABNACtc 1

1EM

UPt

c 1

2EM

DNtn t

IRNOA

MTB

MVE

DATO

DPM

RNOA

DRNOA

NOA

DNOA

PABNAC

EM_UP

EM_DN

PseudoR2

)0.0441

)0.0102

0.0000

0.0183

)0.4450**

0.0335

10.16%

(0.00)

(0.39)

(0.13)

(0.33)

(5.39)

(0.08)

)0.1740

)0.0115

0.0000

0.0325

)0.4909**

0.1783

)0.1098

10.22%

(0.01)

(0.49)

(0.11)

(0.97)

(6.34)

(1.26)

(0.55)

)0.1545

)0.0116

0.0000

0.0325

)0.4940**

0.0234

0.1770

)0.1088

10.22%

(0.01)

(0.50)

(0.11)

(0.98)

(6.38)

(0.04)

(1.24)

(0.54)

(Thetableiscontinued

onthenextpage.)



TABLE6(Continued)

IRNOA

MTB

MVE

DATO

DPM

RNOA

DRNOA

NOA

DNOA

PABNAC

EM_UP

EM_DN

PseudoR2

)0.2798

)0.0161

0.0000

0.0953**

)0.5233**

0.3742***

)0.2962

)0.0164

0.2031**

0.0274

0.1353

)0.0975

10.64%

(0.02)

(0.88)

(0.01)

(6.06)

(4.22)

(9.02)

(2.27)

(0.07)

(5.18)

(0.05)

(0.70)

(0.42)

Panel

C:MultivariateanalysisoftheassociationbetweenEM_UP,EM_DNandUP_RESTATEb

UPRESTATEtc 0c 1IRNOAtc 2MTBtc 3MVEtc 4DATOtc 5DPM


0PABNACtc 1

1EM

UPt

c 1

2EM

DNtn t

IRNOA

MTB

MVE

DATO

DPM

RNOA

DRNOA

NOA

DNOA

PABNAC

EM_UP

EM_DN

PseudoR2

3.5931

0.0076

)0.0000*

0.0175

)0.6389***

)0.2187

10.79%

(2.40)

(0.18)

(3.29)

(0.21)

(8.34)

(2.26)

4.0816*

0.0113

)0.0000*

)0.0225

)0.5109**

)0.2396

0.5335***

11.26%

(3.15)

(0.39)

(2.98)

(0.30)

(4.97)

(1.12)

(12.34)

3.7852

0.0119

)0.0000*

)0.0229

)0.4852**

)0.1942

)0.2291

0.5267***

11.32%

(2.66)

(0.44)

(3.05)

(0.58)

(4.44)

(1.76)

(1.03)

(12.01)

3.5755

0.0132

)0.0000**

)0.0374

)0.7891**

0.3082**

0.1628

)0.0617

)0.1230

)0.1873

)0.2300

0.5196***

11.63%

(2.34)

(0.50)

(4.07)

(0.49)

(6.23)

(4.04)

(0.42)

(0.63)

(0.72)

(1.60)

(1.00)

(11.46)

Notes:

aTestofdifference

presentsthet-statistic(Z-statistic)from

at-test(W

ilcoxonsigned


inmeans(m

edians).

bModelsare

estimatedusing22,160rm


effectsbyyearandindustry

(two-digitSIC

code).v2

statisticsare

reported

inparentheses.See

Table1forvariabledenitions.

***,**and

*indicatethatthevalueissignicantlydifferentfrom

zero

atthe1,5,and10percentlevel,respectively(two-tailed).



downward, as the dependent variable. As explained earlier, we assume that rms thatsubsequently restate earnings down must have initially managed earnings up for that year.The only control variable that is consistently signicant in all variations of the logisticregression is DPM which has a negative coefcient. This suggests that rms with decreas-ing prot margins are more likely to subsequently restate earnings downward, which isconsistent with poorly performing rms attempting to manage earnings upward. The rstmodel also shows that the coefcient on PABNAC is positive but not signicant. In thesecond model, we nd that EM_UP (EM_DN) is positively (negatively) related toDN_RESTATE, but the coefcients are not signicant at conventional levels. The resultssuggest, somewhat surprisingly but consistent with our ndings from the univariateanalysis, that neither PABNAC nor the ATO PM diagnostic is useful in identifying rmswith subsequent downward restatements. The lack of association between the earningsmanagement proxies and the binary restatement variable may in part be due to earningsrestatement being a noisy partition of rms that manage or do not manage earnings. Spe-cically, some of the NO_RESTATE observations may in fact have managed earnings up,but managed to go undetected during our sample period.

In panel C, we report results of logistic regression analyses with UP_RESTATE as thedependent variable. With respect to the control variables, MVE has a signicantly negativecoefcient, suggesting that smaller rms are more likely to subsequently restate earningsupward. In addition, the coefcient on DPM is signicantly negative, which is consistentwith poorly performing rms managing earnings down (i.e., big bath). The negative rela-tion between DPM and both DN_RESTATE and UP_RESTATE suggests that rms withincreasing protability are less likely to subsequently restate earnings upward or down-ward. The rst model further shows that the coefcient on PABNAC is negative but notsignicant at conventional levels. In the second model, we nd that the coefcient onEM_UP is also negative but not signicant at conventional levels. However, EM_DN ispositively related to the probability that rms subsequently restate upward and highly sig-nicant (p < 0.001). The pseudo R2 for the ATO PM model (11.26 percent) is higher thanthat of the PABNAC model (10.79 percent) suggesting that the ATO PM diagnostic hasgreater explanatory power for rms that subsequently restate earnings upward. A formaltest (untabulated) of the area under the ROC curve for the two logistic models conrmsthat the ATO PM model better discriminates UP_RESTATE rms from rms that do notsubsequently restate earnings upwards (v2 = 3.93; p < 0.05). In the third model, wherewe include PABNAC, EM_UP and EM_DN, the coefcient on EM_DN continues to besignicantly positive. The (untabulated) odds ratio of 1.69 for EM_DN in this specicationsuggests that after controlling for other variables, the probability of upward restatement isabout 69 percent higher for rms that are classied as EM_DN than when they are not.Overall, both EM_UP and EM_DN are associated with restatements in the expected direc-tion, although the effect is much stronger for subsequent upward restatements. We nd nosignicant association between earnings restatements and PABNAC.

Future performance

Penman (2007: 634) states that if a rm manipulates earnings upward (downward), futureprotability should fall (rise) as the income contribution from earnings management in thecurrent period reverses. Consistent with this, Dechow et al. (2003) use future reversals inprotability and stock returns as measures of earnings management. Relying on theseobservations, we predict that rms with EM_UP (EM_DN) will report lower (higher)year-ahead changes in performance than other rms. To test our prediction, we use regres-sion analysis with the year-ahead change in return on net operating assets (DRNOAt+1)and year-ahead abnormal returns (ABRETt+1) as dependent variables. We expect a nega-tive coefcient on EM_UP and a positive coefcient on EM_DN. We include the following



control variables that have been identied in previous research (e.g., Faireld and Yohn2001) to predict year-ahead performance: RNOA, DRNOA, NOA, DNOA, DATO, andDPM.27

We examine the information content of EM_UP and EM_DN for explaining year-ahead DRNOA in panel A of Table 7. In the rst model, we include the control variablesas well as PABNAC. The results for the control variables are consistent with mean rever-sion in RNOA (Freeman, Ohlson, and Penman 1982; Faireld, Sweeney, and Yohn 1996),positive serial correlation in DRNOA (Faireld and Yohn 2001), rms with bloated bal-ance sheets (NOA) being more likely to have engaged in upward earnings management inprior years (Barton and Simko 2002), and DATO being informative about future prot-ability changes (Faireld and Yohn 2001). We nd a signicant negative coefcient onPABNAC, which suggests that PABNAC identies earnings components that will likelyreverse in the next year, and is consistent with reversal in abnormal accruals documentedin prior research (Dechow et al. 2003).

In the second model, we include the control variables and EM_UP and EM_DN(without PABNAC). Consistent with our predictions, we document a signicant negativecoefcient on EM_UP and a signicant positive coefcient on EM_DN. Our ndings sug-gest, therefore, that the ATO PM diagnostic is successful in identifying protability rever-sals. We note, however, that the adjusted R2 for the PABNAC model is 4.08 percent whilethe adjusted R2 for the EM_UP EM_DN model is 4.05 percent, indicating that PABNAChas greater relative information content than the ATO PM diagnostic for explaining year-ahead DRNOA.28 Finally, in the third model, we include EM_UP, EM_DN and PABNAC,and nd that the coefcients on all variables remain signicant. These results suggest thatthe ATO PM diagnostic has incremental information content relative to PABNAC foridentifying future earnings reversals.

In panel B, we estimate the same three models as in panel A, except that we use as thedependent variable the size-adjusted abnormal return for the three-day window surround-ing the subsequent rst quarter earnings announcement (ABRETt+1). With respect to thecontrol variables, we nd that the coefcient estimates on RNOA, NOA, DNOA andDATO are signicantly different from zero, suggesting that the market does not correctlyprice the information in these variables. We nd a signicant negative coefcient onPABNAC, which is consistent with prior research (Xie 2001) and with the notion that thestock market overprices abnormal accruals. Consistent with our predictions, we also nd asignicant negative coefcient on EM_UP and a signicant positive coefcient onEM_DN. These results suggest that EM_UP and EM_DN predict year-ahead abnormalreturns. We note that the PABNAC model yields an adjusted R2 of 0.44 percent while theEM_UP EM_DN model yields an adjusted R2 of 0.39 percent, indicating that PABNAChas greater relative information content than the ATO PM diagnostic for explainingyear-ahead returns.29 Finally, in the third model, we nd that the ATO PM diagnostic hasincremental information content relative to PABNAC for identifying future abnormalreturns. Overall, the ATO PM diagnostic identies future protability reversals as well as

27. We also ran the analyses excluding NOA from the model, since NOA is dened as net operating assets

scaled by sales and is, therefore, the inverse of ATO. The conclusions regarding the relative and incremen-

tal information content of PABNAC and EM_UP and EM_DN are unchanged when NOA is excluded

from the model.

28. A Vuong 1989 test (untabulated) conrms the greater explanatory power of the PABNAC model over the

EM_UP EM_DN model. The Vuong statistic for comparison of these two models is 16.79, which is signif-icant at the 1 percent level.

29. A Vuong 1989 test (untabulated) conrms the greater explanatory power of the PABNAC model over the

EM_UP/EM_DN model. The Vuong statistic for a comparison of these two models is 13.41, which is sig-

nicant at the 1 percent level.



future abnormal stock returns, two commonly documented consequences of earningsmanagement.

Additional analyses

Controlling for the effect of contemporaneous sales on net operating assets and operatingincome is crucial in the computation of the ATO PM diagnostic we propose. To ensurethat contemporaneous sales is not merely acting as a deator, we repeat all the analysesreported in the paper after computing ATO and PM with two alternative variables lagged sales and market value of equity in place of contemporaneous sales. We ndthat the ATO PM diagnostic derived with these alternative deators is not associated withthe earnings management scenarios we examine in the paper (results not tabulated).

We also note that changes in ATO and PM could be driven by a change in rm per-formance. However, if changes in ATO and PM are driven by rm performance, then they

TABLE 7

Association between EM_UP and EM_DN and future protability and returnsa

Panel A: Multivariate analysis of the association between EM_UP, EM_DN and year

DRNOAt1 c0 c1RNOAt c2DRNOAt c3NOAt c4DNOAt c5DATOt c6DPMt c7PABNACt c8EM UPt c9EM DNt nt1

RNOA DRNOA NOA DNOA DATO DPM PABNAC EM_UP EM_DN Adj.R2

)0.1381*** 0.0132** )0.0034*** )0.0031 0.0039*** )0.0001 )0.0359*** 4.08%()61.07) (3.37) ()3.03) ()1.56) (4.21) ()0.33) ()8.95))0.1374*** 0.0156*** )0.0034*** )0.0025 0.0035*** )0.0018 )0.0167*** 0.0086*** 4.05%()60.66) (3.94) ()2.97) ()1.29) (3.78) ()0.40) ()5.40) (2.97))0.1782*** 0.0160*** )0.0034*** )0.0014 0.0030*** 0.0000 )0.0346*** )0.0159*** 0.0076*** 4.11%()60.59) (4.06) ()3.05) ()0.74) (3.23) (0.09) ()8.63) ()5.13) (2.61)

Panel B: Multivariate analysis of the association between EM_UP, EM_DN and year-ahead returns

ABRETt1 c0 c1RNOAt c2DRNOAt c3NOAt c4DNOAt c5DATOt c6DPMt c7PABNACt c8EM UPt c9EM DNt nt1

RNOA DRNOA NOA DNOA DATO DPM PABNAC EM_UP EM_DN Adj.R2

0.0021*** )0.0025* )0.0016*** )0.0021*** 0.0021*** )0.0025 )0.0096*** 0.44%(2.61) ()1.79) ()3.46) ()3.06) (6.05) ()1.42) ()7.00)0.0022*** )0.0019 )0.0015*** )0.0020*** 0.0020*** )0.0027 )0.0022** 0.0034*** 0.39%(2.80) ()1.39) ()3.33) ()2.94) (5.72) ()1.48) ()2.10) (3.40)0.0023*** )0.0019 )0.0016*** )0.0018*** 0.0019***)0.0019 )0.0093*** )0.0020** 0.0032*** 0.46%(2.87) ()1.34) ()3.42) ()2.58) (5.34) ()1.07) ()6.79) ()1.99) (3.14)

Notes:

a DRNOAt+1 (ABRETt+1) models are estimated using pooled regressions, on 110,170 (67,075)rm-year observations, controlling for xed effects by year and industry (two-digit SIC

code). t-statistics are reported in parentheses. See Table 1 for variable denitions ***, **

and * indicate coefcients signicantly different from zero at the 1, 5, and 10 percent level,

respectively (two-tailed).



are likely to move in the same direction. For example, if a rm has an increase in sales,ATO will likely go up because sales is in its numerator, and PM will likely go up because,in the presence of period costs, an additional dollar of sales increases the ratio of operat-ing income to sales. We nd, indeed, that rms with a signicant increase in sales (denedas rms in the top quintile of sales growth), experience an increase in ATO 6

a diagnostic for earnings management using changes.pdf

Documents

earnings management

university of washington

suffolk university

university of newhampshire

villanova university

diagnostics of earnings

net operating income

george washington university