Deecing Earnings Managemen Using Cross-Secional Abnormal Accruals Models * K. V. Peasnell, P. F. Pope and S. Young Lancaser Universiy Draf: December 1999 * We are graeful for helpful commens and suggesions provided by David Mes, Ashni Singh, Andrew Sark, paricipans a he 1999 Annual Meeing of he American Accouning Associaion, paricipans a he 1999 Financial Reporing and Business Communicaion Conference, and wo anonymous reviewers. Financial suppor was provided by he Research Board of he Insiue of Charered Accounans in England and Wales, The Leverhulme Trus, and he Economic and Social Research Council. Correspondence o: Seven Young, I.C.R.A., The Managemen School, Lancaser Universiy, Lancaser, LA1 4YX, U.K., Tel: (+44)-1524-594242, Fax: (+44)-1524-594334, E-mail: s.young@lancaser.ac.uk
Deecing Earnings Managemen Using Cross-secional Abnormal Accruals Models Absrac This paper examines specificaion and power issues in relaion o cross-secional models used o esimae abnormal accruals. In addiion o esing he sandard-jones (Jones, 1991) and modified-jones (Dechow e al., 1995) models, we also develop and es a new specificaion, labeled he margin model. Empirical ess sugges ha all hree models are well specified when applied o a random sample of firm-years. However, he margin model appears o generae relaively beer specified esimaes of abnormal accruals when cash flow performance is exreme. Analysis of he models abiliy o deec arificially induced earnings managemen indicaes ha all hree procedures are capable of generaing relaively powerful ess for economically plausible levels of accruals managemen (e.g., less han en percen of lagged oal asses). Regarding heir relaive performance, he sandard-jones and modified-jones models are found o be more powerful for revenue and bad deb manipulaions. In conras, he margin model appears o be more powerful a deecing non-bad deb expense manipulaions. These resuls sugges ha differen models may be required in differen circumsances. 2
Deecing Earnings Managemen Using Cross-Secional Abnormal Accruals Models 1. Inroducion An imporan issue in financial reporing is he exen o which managers manipulae repored earnings. Following Healy (1985), accruals-based measures are now widely employed in ess of he earnings managemen hypohesis. A significan obsacle associaed wih he implemenaion of his approach, however, is he need o accuraely separae repored accruals ino heir managed (discreionary) and unmanaged (non-discreionary) componens. 1 The mos frequenly used echniques for achieving his separaion are he sandard-jones model (Jones, 1991) and he modified-jones model (Dechow e al., 1995). However, recen evidence repored by Guay e al. (1996), Dechow e al. (1995) and Kang and Shivaramakrishnan (1995) for he US suggess ha boh models esimae discreionary accruals wih considerable imprecision. These findings have led some commenaors o quesion he reliabiliy of he emerging body of empirical evidence on he uses made by managers of accouning accruals o manipulae repored numbers. Furher research aimed a developing more effecive procedures for esimaing he discreionary componen of accruals, as well as providing addiional evidence on he performance of exising models, is herefore imporan for he developmen of he earnings managemen lieraure. This paper evaluaes he performance of hree alernaive cross-secional procedures for esimaing he managed componen of working capial accruals. In addiion o examining he performance of he sandard-jones and modified-jones 1 Following convenional pracice, we use he erms managed accruals, discreionary accruals, and abnormal accruals inerchangeably hroughou he remainder of he paper. Similarly, we also use he erms unmanaged accruals, non-discreionary accruals and normal accruals inerchangeably. 1
models, we also develop and es a new specificaion, labeled he margin model. 2 Following Dechow e al. (1995), model performance is evaluaed in erms of specificaion (i.e., he probabiliy of a Type I error) and power (i.e., he probabiliy of a Type II error). Using a sample of UK non-financial companies, resuls sugges ha all hree procedures are well specified when applied o a random sample of firmyears. Addiional ess sugges ha he margin model is relaively beer specified when cash flow performance is exreme. We assess he power of each model o deec accrual managemen using simulaion procedures ha allow for hree differen forms of earnings managemen: revenue manipulaion, expense manipulaion (excluding bad debs) and bad deb manipulaion. Findings sugges ha all hree procedures are capable of generaing reasonably powerful ess for economically plausible levels of all hree forms of accruals managemen (i.e., accruals less han en percen of lagged oal asses). However, imporan differences in he relaive power of he models are eviden. In paricular, he sandard-jones and modified-jones models are shown o be subsanially more powerful for deecing revenue and bad deb manipulaions. In conras, he margin model is significanly beer a deecing non-bad deb expense manipulaions. The paper conribues o he lieraure in hree ways. Firs, in conras o he sudies by Guay e al. (1996), Dechow e al. (1995) and Kang and Shivaramakrishnan (1995) ha examine ime-series versions of he models, we evaluae cross-secional specificaions. An evaluaion of he performance of cross-secional models is warraned as such procedures are now widely employed in he earnings managemen 2 Kang and Shivaramakrishnan (1995) develop a model ha uses an insrumenal variables echnique o isolae he discreionary componen of oal acrruals. Based on simulaion resuls, hey presen evidence ha heir model is more powerful han he sandard-jones model a deecing abnormal accruals. We do no, however, evaluae his model in he presen paper for wo reasons. Firs, oher han in he original Kang and Shivaramakrishnan (1995) paper, we are no aware of any oher sudy ha has used his 2
lieraure. Second, we develop an alernaive procedure for esimaing managed accruals which is shown o ouperform he sandard- and modified-jones models under cerain condiions. Third, he resuls offer some pracical guidance for deecing earnings managemen aciviy. Specifically, our findings sugges eiher ha he choice of accrual model should depend on he prediced form of earnings managemen aciviy (e.g., revenue-based or expense-based), or ha he use of several models in combinaion may afford he bes opporuniy of deecing accrual managemen, he specific form of which is unpredicable. The remainder of he paper is organized as follows. Secion 2 presens a brief review of he sandard-jones and modified-jones models and develops our alernaive model for esimaing abnormal accrual aciviy. Secion 3 provides deails of he research design and sample selecion procedure. Secion 4 presens our empirical findings. Secion 5 conains a summary and conclusions. 2. Modelling abnormal accruals 2.1 The sandard-jones and modified-jones procedures The sandard-jones model (Jones, 1991) uses a wo-sage approach o pariion oal accruals ino heir managed and unmanaged componens. 3 In he firs sage, oal accruals (TA) are regressed on he change in sales ( REV) and he gross level of propery, plan, and equipmen (PPE) for each sample firm, using he longes available ime-series of daa immediaely prior o he even year. In he second sage, he esimaed parameers from his regression are combined wih TA, REV and PPE daa from he even year o deermine he abnormal componen of oal accruals. The procedure o esimae accrual manipulaion. Secondly, he original ime-series formulaion of he procedure is no amenable o he cross-secional esimaion echniques ha are he focus of his paper. 3 In he discreionary accruals lieraure, oal accruals are ypically defined as he change in non-cash working capial accouns, minus depreciaion and amorisaion. 3
inuiion underlying he choice of PPE and REV as explanaory variables in he firssage regression is ha hey help o conrol for unmanaged accruals associaed wih he depreciaion charge and changes in economic aciviy, respecively. 4 As Dechow e al. (1995) discuss, a weakness of he sandard-jones specificaion lies in is inabiliy o capure he impac of sales-based manipulaion, since changes in sales are assumed o give rise o non-discreionary accruals. In an aemp o capure sales-based manipulaions, Dechow e al. (1995) proposed a modificaion o he sandard-jones model. The modified-jones model is idenical o he sandard-jones model, wih he excepion ha he change in debors ( REC) is subraced from REV a he second sage. In effec, he model herefore implicily assumes ha all changes in credi sales in he even period resul from earnings managemen. 5 In ess comparing he power of he modified-jones model wih ha of he sandard-jones model, Dechow e al. (1995) find ha he former procedure is indeed significanly beer a deecing sales-based earnings managemen. However, in absolue erms, boh models are found o generae ess of low power for earnings managemen of economically plausible magniudes (e.g., accruals of one o five percen of oal asses). In addiion, boh models are shown o be poorly specified when financial performance is exreme, wih each generaing a significan proporion of Type I errors when applied o firms wih exreme cash flows. These resuls cas doub on he effeciveness of boh he sandard-jones and modified-jones models a 4 The coefficien on PPE is prediced o be negaive since propery, plan, and equipmen are linked o he income-decreasing depreciaion accrual. In conras, he prediced sign for he REV coefficien is more ambiguous since a given change in sales can cause income-increasing changes in some working capial accouns (e.g., rade debors) and income-decreasing changes in ohers (e.g., rade crediors). The ambiguous predicions for he esimaed coefficien on REV highlighs he somewha ad hoc naure of he model. For her sample of 23 firms facing impor relief invesigaions by he Unied Saes Inernaional Trade Commission, Jones repors an insignifican posiive mean coefficien on REV (0.035) and a significan negaive mean coefficien on PPE (-0.033). 5 As Beneish (1998) discusses, o he exen ha REV REC can be re-wrien as CR REV -1, where CR equals cash received in period, he modified-jones model lacks economic inuiion. 4
isolaing accruals manipulaion. Consisen wih his scepical view, Guay e al. (1996) evaluae he sandard-jones and modified-jones models using a marke-based procedure and conclude ha neiher procedure generaes a reliable measure of accrual managemen. In an aemp o refine and improve he original formulaion of he sandard- Jones model, Teoh e al. (1998) and DeFond and Jiambalvo (1994) focus exclusively on he working capial componen of oal accruals. As Beneish (1998) and Young (1999, p. 842) discuss, his formulaion is poenially more appealing since coninuous (i.e., year-on-year) earnings managemen via he depreciaion accrual is likely o have limied poenial. In addiion, Young (1999) repors ha Jones-syle models based on a measure of oal accruals (i.e., inclusive of he depreciaion charge) induce subsanial measuremen error in he resuling esimae of managed accruals. Finally, as a pracical maer, he original ime-series formulaion of he sandard-jones and modified-jones models has proven resricive when implemening he procedures empirically because of he need for a sufficienly long ime-series of daa o allow effecive esimaion of he firs-sage regression parameers. This requiremen raises several concerns. Firs, issues of survivorship bias naurally arise. 6 Secondly, he assumpion ha he coefficien esimaes on REV and PPE remain saionary over ime may no be appropriae. Finally, he self-reversing propery of accruals may inroduce specificaion problems in he form of serially-correlaed residuals. In an effor o overcome hese problems, recen sudies have begun o use cross-secional versions of he models (e.g., Becker e al., 1998; Subramanyam, 1996; DeFond and Jiambalvo, 1994). 7 Under his approach, he firs sage regression is 6 For example, Jones (1991) requires ha her sample firms have a leas 10 ime-series observaions. 7 One should no inerpre he curren preference for cross-secional models in he lieraure as evidence of heir improved abiliy o deec earnings managemen. I is imporan o recognise ha he replacemen of ime-series models by heir cross-secional counerpars inroduces new problems. For 5
esimaed separaely for each indusry-year combinaion, afer which he resuling indusry- and ime-specific parameer esimaes are combined wih firm-specific daa o generae esimaed discreionary accruals. 8 Cross-secional versions of he sandard-jones and modified-jones models now dominae he earnings managemen lieraure. However, lile research has been conduced o dae ha evaluaes he effeciveness of hese models a deecing earnings managemen. 9 2.2 The margin model In his secion, we derive an alernaive procedure for esimaing abnormal accruals. Similar o he sandard-jones and modified-jones procedures, we esimae abnormal accruals using a wo-sage procedure, where he firs sage involves regressing accouning accruals on a vecor of explanaory variables designed o capure unmanaged accruals. In conras o he wo Jones models, however, he explanaory variables included in our firs-sage regression are derived from a formal model linking sales, accruals and earnings. 10 Following he discussion in he preceding secion, we exclude depreciaion from our measure of accruals on he grounds ha his iem represens an unlikely or unsuiable vehicle for sysemaic earnings managemen. The saring poin for our model is herefore working capial accruals. We hen seek o model normal changes in he following hree key working capial accrual componens: socks ( STOCK), debors ne of bad deb allowance example, cross-secional models are less likely o capure he effecs of (a) mean reversion in accruals, (b) dynamic accrual managemen sraegies, and (c) indusry-wide earnings managemen. Wih valid argumens for and agains boh esimaion approaches, he relaive power of cross-secional and imeseries models o deec earnings managemen remains an open empirical quesion ha is beyond he scope of he presen paper. 8 In he even ha he sample firm is included in he firs-sage regression model, is esimae of discreionary accruals is equal o he corresponding regression residual and consequenly he wo-sage esimaion procedure collapses o a single sage. 9 An excepion is he sudy by Jeer and Shivakumar (1999) which evaluaes cross-secional models using quarerly and annual daa for a sample of US firms. 10 Our model is consisen wih ha developed by Dechow e al. (1998). 6
( DEBT), and crediors ( CREDIT). The remaining non-cash componens of working capial are no explicily modeled because of heir diverse naures. Ignoring for he momen such expenses as wages and ren, we define he following hree componens of changes in working capial: STOCK PUR COGS (1) DEBT REVC CRC BDE (2) CREDIT PUR CPS, (3) where PUR is purchases of maerials, COGS is cos of finished goods sold, REVC is revenue from credi sales, CRC is cash received from cusomers, BDE is he bad deb expense, and CPS is cash paid o suppliers. Noe ha while STOCK in (1) includes invenories of maerials, work in progress and finished goods, all inermediae ransfers beween hese invenory caegories involve cancelling enries ha can be ignored when he invenories are aggregaed. Nex, we wrie working capial accruals (WCA) as: WCA = ( STOCK + DEBT ) CREDIT + OTHER = ( REVC COGS BDE ) + ( CPS CRC ) + OTHER = sm REVC cm CRC + OTHER, (4) where sm equals he gross margin on recorded sales, cm equals he gross cash conribuion on cash collecions from cusomers, and OTHER includes all non-cash curren asses oher han socks and rade debors and all curren liabiliies oher han crediors. We assume ha OTHER is orhogonal o REVC and CRC in equaion (4). Equaion (4) is inended o capure he accrual recogniion process before conaminaion by earnings managemen. Working capial is expressed as he sum of wo conribuion margins: he gross margin on sales and is cash flow analogue, he 7
margin on cash received (he cash margin ). Under his approach, working capial accruals ha do no resul from sales and cash collecions in he period are classified as abnormal in naure and are considered he mos likely manifesaion of earnings managemen. Equaion (4) is implemened empirically using he following OLS regression: WCA = λ 0 + λ1rev + λ2cr + η, (5) where REV is oal sales (Daasream iem 104) and represens our proxy for REVC, CR is oal sales minus he change in rade debors (Daasream iems 104 370) and represens our proxy for CRC, λ 0, λ 1 and λ 2 are regression coefficiens and η is he regression residual. The λ 1 coefficien represens an esimae of he sales margin and is prediced o be posiive, while he λ 2 coefficien represens an esimae of he cash margin and is prediced o be negaive. We refer o his procedure as he margin model in he remainder of he paper. The primary difference beween he margin model and he sandard-jones and modified-jones models is ha he margin model disaggregaes he change in revenue erm ino wo componens a he parameer fiing sage, subsiuing cash receips in he curren period for revenues in he prior period. The primary advanage of his approach is is improved economic inuiion, which should, in urn, lead o a more precise esimae of normal accruals. The downside, relaive o he modified-jones procedure, is ha unmanaged accruals are compued using a variable (REV) ha may iself be conaminaed by earnings managemen. As such, we anicipae ha he margin model will be less powerful han he modified- Jones model a deecing revenue-based manipulaions. 11 8
3. Research design 3.1 Esimaion of accruals models Our empirical ess compare he specificaion and power of commonly used es saisics across measures of abnormal accruals generaed by he sandard-jones (s-j), he modified-jones (m-j) and he margin models. To ensure comparabiliy wih he margin model, boh he s-j and m-j models are formulaed using working capial accruals (e.g., Teoh e al., 1998; DeFond and Jiambalvo, 1994). 12 The firs sage regressions are esimaed cross-secionally for each indusry (Daasream level-6) and year combinaion using all firms wih available accrual daa on he Daasream Live and Dead socks files. All variables (wih he excepion of he inercep) are scaled by lagged oal asses o reduce heeroskedasiciy. 13 Indusry-year porfolios conaining fewer han en observaions are dropped from he analysis o ensure more efficien esimaion of he regression coefficiens. 3.2 Empirical es procedures Model specificaion We evaluae model specificaion by examining he exen o which each of he hree cross-secional models incorrecly rejecs he null hypohesis of no earnings managemen. This is achieved using he following simulaion procedure: 11 We es his predicion empirically in secion 4. 12 Working capial accruals are defined as he change in non-cash curren asses minus he change in curren liabiliies, excluding he curren porion of long-erm deb [Daasream iems (376 375) (389 387)]. 13 As such, he s-j and m-j models repored in his paper differ slighly from hose esimaed in exan sudies where he inercep is also scaled by oal asses and he resuling regression is esimaed wih he rue consan erm suppressed. We did no adop his approach in he curren paper for wo reasons. Firs, here is no heoreical reason for forcing he regression hrough he origin (e.g., we have no reason o believe ha oal accruals will be zero when, say, REV is zero). Secondly, regressions esimaed wih he consan suppressed preclude an analysis of he goodness of fi of he models because he associaed R-square values are unreliable. However, o ensure comparabiliy wih exan research, we repeaed all ess using his alernaive specificaion. In all cases he conclusions were idenical o hose based on he findings repored in ables 2-5. 9
(a) Esimae he firs sage regressions for each of he hree abnormal accrual models (s-j, m-j and margin) for each Daasream level-6 indusry group in year ; (b) Selec 25 firms a random from year and consruc an indicaor variable (PART) defined as one if he firm has been seleced and zero oherwise; (c) Randomise all observaions in year and compue abnormal accruals (AA) for each of he hree models using he coefficien esimaes obained in (a); (d) Esimae he following univariae regression for each measure of abnormal accruals AAi = α + βpart i + εi, (9) and es wheher he esimaed coefficien on PART is significanly differen from zero. Seps (a) (d) are hen repeaed 100 imes for each sample year. Since observaions a sage (b) are seleced a random, hey should no be characerised by any sysemaic earnings managemen aciviy. As such, a well specified model is no expeced o rejec he null hypohesis of ˆ β = 0 a raes ha significanly exceed he appropriae es saisic (e.g., five percen or one percen levels). Accordingly, we inerpre rejecion frequencies close o (significanly differen from) he specified es level as evidence ha he accrual model in quesion is well (poorly) specified. In addiion o esing for general model specificaion using a random sample of observaions, we also evaluae model specificaion when financial performance (measured using operaing cash flows) is exreme. Empirical ess exploi he fac ha esimaed abnormal accruals may be decomposed ino wo componens, rue earnings managemen and measuremen error. Since he rue earnings managemen componen is invarian o he specific modelling procedure employed, any observed difference in esimaed abnormal accruals across alernaive procedures mus be 10
aribuable o measuremen error. As such, a comparison of he disribuional properies of abnormal accruals generaed by alernaive models provides a means of evaluaing heir relaive specificaion. 14 In he conex of exreme cash flow performance, prior research has found ha unusually high (low) operaing cash flows cause esimaed discreionary accruals o be negaively (posiively) biased (Young, 1999; Dechow e al., 1995). Accordingly, average abnormal accruals esimaed using a model ha conrols for high (low) cash flows are expeced o be less negaive (less posiive) han abnormal accruals esimaed using a procedure ha affords lile effecive proecion agains exreme cash flow performance, all else equal. The es is implemened by forming decile porfolios, ranked wihin each sample year, based on operaing cash flow scaled by lagged oal asses. 15 Abnormal accruals for each decile porfolio are hen compued using he hree cross-secional models. Finally, mean and median abnormal accrual esimaes generaed by he hree models are compared wihin each cash flow decile. 16 Power o deec earnings managemen We assess he power of alernaive cross-secional accruals models by examining heir abiliy o deec earnings managemen aciviy when i is known o 14 While his approach allows direc comparisons of he relaive magniude of he measuremen error associaed wih alernaive accrual models, a limiaion of he procedure is ha affords no measure of he absolue level of misspecificaion. 15 Operaing cash flow is measured as operaing profi (Daasream iem 137) minus working capial accruals (Daasream iems 376 375 389 + 387), plus depreciaion and amorisaion (Daasream iems 402 + 562). 16 Using a procedure equivalen o ha employed o es for general model specificaion, Dechow e al. (1995) evaluae model specificaion when cash flows are exreme by examining observed rejecion frequencies for samples of firms seleced from he op and boom deciles of he operaing cash flow disribuion. We do no abulae resuls based on his approach in he curren paper because as Guay e al. (1996) discuss, he es is based on he assumpion of no earnings managemen. While his assumpion is expeced o hold for he random samples used o es for general model misspecificaion, i is unlikely o hold in he case of firms wih exreme cash flow performance, given he srong incenives for earnings managemen faced by hese firms. For compleeness, however, we did perform such ess. Consisen wih Dechow e al. (1995), rejecion frequencies for all hree models significanly exceeded he specified es level for boh he high and low cash flow pariions. However, he rejecion frequencies for he margin model were always lower han hose obained for he s-j and m-j models, 11
exis. This is achieved by adding a pre-deermined amoun of posiive accruals o he repored accruals of a randomly seleced se of firms and hen examining he abiliy of he models o deec his arificial earnings managemen. The procedure is similar o ha described above for esing general model specificaion wih he excepion ha a sage (b), arificial income-increasing accruals are added o he repored accruals for firms where PART equals one. 17 As before, seps (a) (d) are hen repeaed 100 imes for each sample year. However, since he observaions where PART equals one are now known o conain income-increasing earnings managemen aciviy, we would expec a powerful model o rejec he null hypohesis ha ˆ β = 0, in favour of he alernaive ha ˆ β > 0, a raes ha significanly exceed he specified es level. All else equal, he higher he rejecion frequencies associaed wih a paricular model, he more powerful ha model is deemed o be a deecing earnings managemen aciviy. In implemening his procedure, we experimen wih differen magniudes of income-increasing accruals ranging from zero percen o 100 percen of firms lagged oal asses in he randomly-seleced PART reamen firms. In addiion, we also consider hree differen ypes of accruals manipulaion: (a) Expense Manipulaion (excluding bad debs) e.g., delayed recogniion of an expense (Dechow e al 1995, p. 201). This approach is implemened by adding an assumed amoun of expense manipulaion o working capial accruals. 18 (b) Bad Deb Manipulaion e.g., wrie-back of an exising bad deb provision. This approach is implemened by adding he assumed amoun of bad deb manipulaion o boh working capial accruals and rade debors. consisen wih he view ha he margin model is relaively beer specified when cash flows are exreme. 17 As such, he parameer esimaes from he firs sage regressions in par (a) are no conaminaed by he arificially induced earnings managemen. 12
(c) Revenue Manipulaion e.g., premaure recogniion of a sale, assuming all coss are fixed (Dechow e al. 1995, p. 202). This approach is implemened by adding he assumed amoun of revenue manipulaion o working capial accruals, oal sales and rade debors. 3.3 Sample The saring poin for our sample is he populaion of firms on he Daasream Live and Dead socks files wih he necessary accouning daa for he compuaion of he s-j, m-j and margin models and which have year-ends beween 30 June 1990 and 31 May 1997. From his iniial sample, we exclude all financial firms because (a) heir financial reporing environmens differ from hose of indusrial firms and (b) hey have fundamenally differen accrual processes ha are no capured very well by our expecaions models of normal accrual aciviy. Firm-years wih missing Daasream level-6 indusry codes are also omied. Since he esimaion of our crosssecional accrual models requires a leas en firms per indusry-year combinaion, we also exclude all Daasream level-6 indusry groups wih fewer han en observaions in a given sample year. Finally, because he disribuion of working capial accruals is characerised by a small number of exreme observaions, we also exclude all cases for which scaled working capial accruals lie ouside he one and niney-nine perceniles. These sampling crieria resul in a final sample of 4,352 firm-year observaions, comprising 837 individual firms from 35 Daasream level-6 indusry groups. Of hese, 418 firms (49.9%) are represened in all seven years, while 572 firms (68.3%) are represened a leas five imes. Firms wih only a single year of daa 18 Unlike Dechow e al. (1995) and Kang and Shivaramakrishnan (1995), because our models are esimaed cross-secionally we do no have o make corresponding adjusmens for accrual reversals in 13
number 78 (9.3%). Annual sample sizes are 601, 590, 607, 580, 624, 657 and 693 for 1990, 1991, 1992, 1993, 1994, 1995 and 1996, respecively. The larges indusry group in any given year is general engineering, wih he number of firms ranging from 51 in 1993 o 57 in 1996. Deails of he sampling procedure are repored in able 1. 4. Resuls 4.1 Descripive saisics Table 2 presens descripive saisics for he Jones and margin models. Panel A repors he coefficien esimaes and associaed fi saisics for he firs-sage regressions. Panel B presens addiional informaion for he indusry-year esimaion porfolios, pariioned according o he goodness-of-fi of he regressions esimaed in panel A. Since he s-j and m-j models are equivalen a he esimaion sage, able 2 repors only a single se of resuls for hese wo models in he columns headed Jones Models. Examinaion of he coefficien esimaes and associaed goodness-of-fi saisics presened in panel A provides preliminary evidence regarding he models abiliy o capure, and hence conrol for, normal working capial accrual aciviy. 19 For he Jones models, he coefficien on REV is posiive in all years, wih he excepion of 1994. However, he average magniudes of he βs are generally very low, ranging from a high of 0.068 in 1993 o a low -0.007 in 1994. The associaed -saisics highligh he lack of associaion beween working capial accruals (WCA) and he REV erm for our sample: median -saisics for he indusry-specific esimaes of β never aain significance in any of he seven esimaion periods. The lack of fuure accouning periods. This holds for all hree ypes of accrual manipulaion examined. 19 Alhough a relaively high R-squared saisic does no necessarily guaranee ha a paricular model is effecively pariioning accrual aciviy ino normal and abnormal componens, a model ha accouns for only a small proporion of he variaion in working capial accruals seems inconsisen wih (a) he raionale underlying he use of a regression-based modelling procedure and (b) managemen s limied flexibiliy o manipulae revenues and expenses under GAAP. 14
associaion beween WCA and REV is also refleced in he explanaory power of he models: he median adjused R-squared saisic ranges from a maximum of 15% in 1993 o a minimum of 2% in 1992. Moreover, he adjused R-squared saisics are negaive for more han 45% of indusries in four ou of he seven sample years. In oher words, he Jones models have zero explanaory power for a large number of indusry years. Even a heir bes, he cross-secional Jones models fail o accoun for over 85% of he inra-indusry variaion in WCA, on average. Panel A of able 2 also repors he corresponding saisics for he margin model. As prediced, he coefficien on REV (λ 1 ) is posiive in all seven sample years. Furher, he magniude of he coefficien lies wihin he feasible range for he gross profi margin on sales. The median coefficien on he CR erm (λ 2 ) is negaive in all esimaion periods, as prediced. The coefficien magniudes of λ 2 are similar o hose repored for λ 1. The median adjused R-squared saisic ranges from 41% (1993) o 7% (1994) and is significanly higher (p < 0.001) han ha repored for he Jones models in six of he seven sample years. Consisen wih his, he proporion of indusries wih essenially zero fi (i.e., zero or negaive adjused R-squared saisics) in a given sample year is also significanly lower han he equivalen figures for he Jones models. These resuls provide evidence ha he margin model may be somewha beer specified, in erms of he firs sage esimaion, han eiher of he wo Jones models. Furher insighs may be gained by using he coefficien esimaes repored in panel A of able 2 o compue he elasiciy of working capial accruals wih respec o 15
he drivers of normal accruals for each of he hree accruals models. 20 For he margin model, he ne effec of a one percen change in REV and CR on working capial accruals ranges from 152% in 1993 o 5% in 1994. In conras, he effec of a one percen change in REV on working capial accruals for he s-j model ranges from 31% in 1993 o 0.18% in 1994. The equivalen elasiciies for he m-j model are 26% in 1993 and 0.14% in 1994. The resuls indicae he higher responsiveness of working capial accruals o he drivers of normal accruals in he case of he margin model. This has boh posiive and negaive implicaions regarding he relaive performance of he margin model. On he posiive side, a high elasiciy is expeced o resul in improved esimaes of abnormal accruals in siuaions where he drivers of working capial accruals are unconaminaed by earnings managemen (e.g., expense manipulaion, excluding bad debs), due o increased precision in he esimae of normal accruals. On he negaive side, a high elasiciy means ha he model will generae poorer esimaes of abnormal accruals when eiher REV or CR are conaminaed by earnings managemen (e.g., revenue manipulaion), since a larger fracion of observed accrual aciviy will be incorrecly explained by he (conaminaed) drivers of normal accruals. As an example of his effec, consider he elasiciies for he s-j and m-j models: he higher elasiciy associaed wih he s-j model reflecs is endency o uninenionally aribue a fracion of revenue manipulaion o normal accrual aciviy. In an effor o shed ligh on he reasons underlying he low explanaory power of all models for cerain indusry-year combinaions, panel B of able 2 repors addiional informaion for he indusry-year esimaion porfolios, pariioned 20 An elasiciy measures he effec of a one percen change in an independen variable (X j ) on he dependen variable (Y). Elasiciies, measured a he poin of he means of he independen variables, 16
according o he adjused R-squared saisics obained from he firs sage regression models. For boh he Jones and he margin models, an F-es fails o rejec he null hypohesis ha he mean number of observaions in he esimaion porfolios differs across he hree goodness-of-fi pariions. These findings indicae ha sparse daa is unlikely o be he primary facor driving he low explanaory power for cerain indusry-year combinaions. Similarly, here is no evidence ha a difference in he variabiliy of WCA wihin he indusry-year porfolios underlies he low adjused R- square values. Finally, panel B also repors he primary indusry groups (i.e., hose represened a leas four imes) in each goodness-of-fi pariion. No obvious indusry paerns are apparen. As such, he facors ha deermine indusry-year combinaions where he firs-sage regressions explain a subsanial fracion of he variaion in WCA from hose where he firs-sage regressions accoun for lile or no variaion in WCA remain unidenified. 4.2 Model specificaion Resuls of he ess designed o assess he general specificaion of he s-j, m-j, and margin models are presened in Table 3. Type I errors from one-ailed ess are repored for boh he null hypohesis ha abnormal accruals are greaer han or equal o zero (alernaive hypohesis: income-decreasing earnings managemen) and he null hypohesis ha abnormal accruals are less han or equal o zero (alernaive hypohesis: income-increasing earnings managemen). Recall ha since he sampling procedure helps ensure ha firms where PART equals one are unlikely o be characerised by sysemaic earnings managemen aciviy, a well-specified model should no rejec he null hypohesis of no earnings managemen a raes ha are compued as X j E j = βˆ j. Y 17
significanly exceed he es level (e.g., 5% or 1%). Consisen wih he findings repored by Dechow e al. (1995), he frequency of Type I errors in able 3 corresponds o he specified es levels for all hree models: in all cases a binomial es fails o rejec he null hypohesis ha he observed rejecion frequencies equal he specified es levels. These findings sugges ha all hree cross-secional accrual models appear well-specified when applied o a random sample of firm-years. Resuls presened in able 4 assess he relaive specificaion of abnormal accrual models when cash flow performance is exreme. Resuls for he low (high) CFO pariion reveal ha all hree models generae posiive (negaive) abnormal accrual esimaes of around six o seven percen of lagged oal asses. This is consisen wih all hree models inducing posiive (negaive) measuremen error when cash flows are unusually low (high). 21 Examinaion of he relaive magniudes reveals ha he margin model produces less exreme esimaes of abnormal accruals han hose generaed by he wo Jones models. For example, in he low CFO porfolio, mean (median) abnormal accruals for he margin model are 6.0% (5.6%) of lagged oal asses, compared wih around 7.5% for he s-j and m-j models. These differences are significan a he 0.001 level. Similarly, in he high CFO porfolio, mean abnormal accruals for he margin model are 5.6% of lagged oal asses, compared wih 6.5% for he s-j and m-j models (difference significan a he 0.001 level). To he exen ha low operaing cash flows induce bias in esimaed abnormal accruals and vice versa, hese resuls are consisen wih he margin model being somewha beer specified han eiher he s-j model or he m-j model when cash flow performance is exreme. Noe, however, ha due o he naure of our empirical ess we are unable o 21 An alernaive explanaion for he posiive (negaive) abnormal accruals observed in he low (high) cash flow porfolio is ha hese firms are sysemaically managing heir earnings upwards (downwards) as a means of smoohing repored income. Indeed, i is he inabiliy o disinguish beween 18
make inferences regarding he absolue level of measuremen error. In paricular, we make no claim ha he margin model is compleely free of measuremen error. Furher analysis of able 4 indicaes several oher findings worhy of commen. Firs, he margin model s apparen superior specificaion does no appear o be confined o he wo exreme CFO porfolios. For example, relaive o esimaes generaed by he s-j and m-j models, margin model abnormal accruals are less posiive in CFO decile 2 and less negaive in CFO deciles 7-9. Secondly, abnormal accruals generaed by he s-j model are significanly lower han hose generaed by he m-j model when cash flows are unusually low (CFO deciles 1 and 2), suggesing he s-j model s relaive superioriy when applied o firms wih low operaing cash flows. Finally, esimaed abnormal accruals produced by he hree models are indisinguishable from each oher and close o zero in he mid cash flow porfolios (CFO deciles 5 and 6), suggesing ha all models are reasonably well-specified when applied o firms characerised by normal operaing performance. 4.3 Power o deec earnings managemen This secion repors he resuls of simulaions using arificially induced income-increasing earnings managemen. Figure 1 provides informaion regarding he relaive power of alernaive models a deecing earnings managemen. The graphs plo he frequency wih which he null hypohesis of no earnings managemen is rejeced (verical axis) agains he magniude of he induced earnings managemen (horizonal axis). 22 Separae graphs are presened for each assumed source of earnings manipulaion, using wo es levels (five percen and one percen). Rejecion raes for measuremen error and earnings managemen ha makes i impossible o draw inferences abou absolue model specificaion in hese porfolios. 22 While we experimen wih differen magniudes of income-increasing accruals managemen ranging from zero percen o 100 percen of firms lagged oal asses, he graphs in figure 1 only plo earnings 19
boh es levels are compued using a one-ailed es. The graphs in panel A provide he power funcions for expense manipulaion (excluding bad debs), evaluaed a he five percen and one percen levels. The equivalen resuls for bad deb manipulaion and revenue manipulaion are presened in panels B and C, respecively. A sriking feaure of he resuls presened in figure 1 is ha all hree models appear capable of generaing relaively powerful ess for economically plausible levels of accruals managemen. Specifically, all hree models generae rejecion frequencies for he null hypohesis of no income-increasing earnings managemen a (or close o) 100% for arificially induced accrual manipulaions of around six percen of lagged oal asses or greaer. These resuls hold for rejecion frequencies evaluaed a eiher he five percen or one percen es levels. Even for relaively low accruals manipulaions of around wo percen of lagged oal asses, he rejecion frequencies can be as high as 40%, depending on he paricular accrual model used and he specific ype of earnings managemen considered. These resuls conras wih he much lower rejecion frequencies repored by Dechow e al. (1995) for he ime-series versions of he Jones and modified-jones models. For example, Dechow e al. (1995) repor rejecion raes of less han 30% for earnings managemen equal o five percen of oal asses, wih he raes only approaching 100% when arificially induced earnings managemen exceeds 50% of oal asses. Regarding he models relaive power o deec earnings managemen, a sriking feaure of he power funcions presened in figure 1 is he apparen similariy in performance of he s-j and m-j models, paricularly in relaion o revenue manipulaion (panel C). Our resuls herefore fail o confirm he findings repored by Dechow e al. (1995) ha he m-j model is significanly more powerful a deecing managemen levels up o en percen of lagged oal asses as in mos cases, he deecion raes reached 100% a or well before his level. 20
revenue-based manipulaions. We aribue his resul o he small coefficien esimaes repored for he Jones models in able 2, which ac o dampen-down he effec of revenue manipulaions ha conaminae he REV erm. In oher words, when he esimaed coefficien on REV is as low as ha repored in able 2, i makes lile difference wheher or no REV is adjused for he change in debors associaed wih revenue manipulaion. While all hree models evaluaed in his sudy appear capable of deecing relaively low magniudes of accruals manipulaion, resuls presened in figure 1 sugges some imporan differences wih respec o he performance of he Jones models (s-j and m-j) and he margin model. For example, he findings repored in panel A for expense manipulaion (excluding bad debs) indicae he margin model s relaive superioriy for low levels of earnings managemen (i.e., accruals less han six percen of lagged oal asses). A Z-es reveals ha he rejecion frequencies obained for he margin model are significanly higher (p < 0.05) han hose obained for he s-j and m-j models for magniudes of arificially-induced earnings managemen beween 1.5% and 6.5% of lagged oal asses. As discussed above, his is consisen wih he margin model generaing a more precise esimae of normal accruals in siuaions where he drivers of normal accrual aciviy are no conaminaed by earnings managemen. In conras, he margin model is significanly worse han he s-j and m-j models a deecing boh bad deb manipulaion (panel B) and revenue manipulaion (panel C). This is because he smaller coefficien esimaes obained in he firs sage fi of he Jones models serve o dampen-down he effec of earnings managemen ha conaminaes he drivers of normal accruals. In conras, he larger coefficien esimaes repored in able 2 for he margin model are associaed wih a reducion in he model s power o deec earnings managemen because a greaer fracion of he 21
arificially induced accruals manipulaion is uninenionally aribued o normal accruals aciviy. 5. Summary and conclusions This paper examines specificaion and power issues relaing o he measuremen of abnormal accruals using cross-secional esimaion procedures. While Guay e al. (1996), Dechow e al. (1995) and Kang and Shivaramakrishnan (1995) have evaluaed he performance of accrual models esimaed in ime series, examinaion of he lieraure indicaes ha cross-secional models are now more widely used in conemporary earnings managemen sudies. Empirical evidence regarding boh he relaive and absolue performance of cross-secional accrual models is herefore called for. In addiion o evaluaing he performance of crosssecional versions of he models proposed by Jones (1991) and Dechow e al. (1995), we also develop and es an alernaive procedure, labelled he margin model. The margin model differs from exising procedures in ha he drivers of normal accruals are derived from a formal model linking sales, accruals and earnings. Resuls indicae ha all hree cross-secional models appear well specified when applied o a random sample of firm-years. However, addiional ess indicae ha he margin model generaes relaively beer specified esimaes of abnormal accruals when cash flow performance is exreme. Specifically, average abnormal accrual esimaes produced by he margin model are significanly lower (higher) han hose generaed by eiher he sandard-jones or modified-jones models when operaing cash flows are unusually high (low). Furher analysis designed o assess he models abiliy o deec known cases of accruals managemen indicaes ha all hree procedures appear capable of generaing high power ess for earnings managemen upwards of five percen of lagged oal asses: for earnings managemen exceeding 22
five percen of lagged oal asses, all hree models yield deecion raes close o 100%. In conras, Dechow e al. (1995) repor rejecion frequencies of beween 20% and 30% for heir ime-series models for arificially induced earnings managemen equal o five percen of oal asses. To he exen comparisons across differen sudies conduced in differen insiuional regimes are meaningful, hese resuls provide some suppor for Subramanyam s (1996) findings ha his cross-secional accrual model may be more powerful han is ime-series counerpar. Noe, however, ha he Dechow e al. (1995) models are based on a measure of oal operaing accruals (i.e., including depreciaion) while he models examined in his paper are defined in erms of working capial accruals. This raises he possibiliy ha he differences in rejecion frequencies beween he wo sudies may be due o differences in he way accruals are defined, raher han o he use of alernaive esimaion procedures. As such, an alernaive inerpreaion of he high rejecion raes documened for he models in his sudy is ha a pure working capial accruals measure may be more powerful han an operaing accruals measure when earnings are managed via working capial accouns. Fuure research aimed a formally comparing he power of (a) alernaive esimaion procedures and (b) alernaive accrual definiions would herefore represen an ineresing conribuion o he lieraure. Regarding he relaive power of he hree models, findings sugges ha he sandard-jones and modified-jones models are subsanially more powerful a deecing suble revenue and bad deb manipulaions (i.e., less han 10% of lagged oal asses in magniude). Thus, despie heir ad hoc naure, hese models sill appear o represen relaively powerful soluions o he problem of deecing cerain ypes of accrual managemen. The price for his improved deecion power, however, is greaer misspecificaion when cash flow performance is exreme. Conrary o Dechow e al. 23
(1995), we fail o documen any significan difference in he relaive power of he Jones and modified-jones models o deec revenue manipulaions. We aribue his resul o he relaively small coefficien esimaes obained from he firs sage regressions of working capial accruals on he change in revenue. In conras o he findings for revenue and bad deb manipulaions, he margin model ouperforms he sandard- and modified-jones models a deecing non-bad deb expense manipulaions ranging from 1.5% o 6.5% of lagged oal asses. We noe, however, ha he improved power of he margin model over ha of he s-j and m-j models in relaion o non-bad deb manipulaions is more han offse by is relaive inabiliy o deec revenue and bad deb manipulaions. Finally, in erms of providing pracical guidance on he quesion of which abnormal accruals model o use, our resuls imply ha his decision is likely o be coningen on he form he earnings managemen is expeced o ake. If earnings managemen via revenue or bad deb accouns is anicipaed, hen he sandard-jones and modified-jones models appear o offer he greaes chance of deecing i. Alernaively, if expense manipulaion (exclusive of bad debs) is anicipaed, hen he margin model may be more appropriae. In he absence of any srong priors regarding he specific ype of manipulaion used, our resuls sugges ha using all hree models in combinaion may afford he greaes chance of deecing earnings managemen. Irrespecive of he specific approach used, however, accurae deecion of small-scale accruals managemen (i.e., less han five percen of oal asses in magniude) remains fraugh wih difficuly. 24
References Becker, C., DeFond, M. L., Jiambalvo, J. and Subramanyam, K. R. (1998). The effec of audi qualiy on earnings managemen. Conemporary Accouning Research, 15: 1-24. Beneish, M. D. (1998). Discussion of Are accruals during iniial public offerings opporunisic?. Review of Accouning Sudies, 3: 209-221. Dechow, P., Kohari, S. P. and Was, R. (1998). The relaion beween earnings and cash flows. Journal of Accouning and Economics, 25: 133-168. Dechow, P., Sloan, R. and Sweeney, A. (1995). Deecing earnings managemen. The Accouning Review, 70: 193-225. DeFond, M. L. and Jiambalvo, J. (1994). Deb covenan violaion and he manipulaion of accruals. Journal of Accouning and Economics, 17: 145-176. Guay, W.R., Kohari, S. P. and Was, R. (1996). A marke-based evaluaion of discreionary accrual models. Journal of Accouning Research, 34: 83-105. Healy, P. (1985). The effec of bonus schemes on accouning decisions. Journal of Accouning and Economics, 7: 85-107. Jeer, D.C. and Shivakumar, L. (1999). Cross-secional esimaion of abnormal accrual models using quarerly and annual daa: effeciveness in deecing evenspecific earnings managemen. Accouning and Business Research, 29: 299-319. Jones, J. (1991). Earnings managemen during impor relief invesigaions. Journal of Accouning Research, 29: 193-228. Kang, S.-H. and Shivaramakrishnan, K. (1995). Issues in esing earnings managemen and an insrumenal variable approach. Journal of Accouning Research, 33(2): 353-367. Subramanyam, K. R. (1996). The pricing of discreionary accruals. Journal of Accouning and Economics, 22: 249-281. Teoh, S. H., Wong, T. J. and Rao, G. R. (1998). Are accruals during iniial public offerings opporunisic?. Review of Accouning Sudies, 3: 175-208. Young, S. (1999). Sysemaic measuremen error in he esimaion of discreionary accruals: an evaluaion of alernaive modelling procedures. Journal of Business, Finance and Accouning, 26: 833-862. 25
Table 1 Sample Selecion Crieria. The Sample Period Comprises all Non-Financial Firms on he Daasream Live and Dead Socks Files wih Year-Ends Beween 1 June 1990 and 31 May 1997 (Inclusive) Crieria Firm-Years Firm-years wih non-missing accrual daa: nonfinancials only 4,600 Less: Exreme accrual values (rimmed a 1% and (101) 99% levels) Indusry-year combinaions wih fewer han 10 observaions (81) Missing Daasream level-6 indusry codes (66) Final sample 4,352 26
Table 2 Descripive Saisics for he Jones (s-j and m-j) Models and he Margin Model, Esimaed Cross-Secionally for Each Indusry (Daasream Level-6) and Year Combinaion a Panel A: Descripive saisics for he firs-sage regression models Jones Models: WCA i = α + β REV i + e i Median for indusry groups b Margin Model: WCA i = λ 0 + λ 1 REV i +λ 2 CR i + η i Median for indusry groups b Year α α β β Adj-R 2 % 0 c λ 0 λ0 λ 1 λ1 λ 2 λ2 Adj-R 2 % 0 c 1990-0.002-0.179 0.046 0.458 0.064 29.03 0.008 0.138 0.545 2.464-0.552-2.514 0.311 22.58 1991-0.016-0.822 0.011 0.215 0.019 46.67 0.020 0.484 0.335 1.313-0.349-1.587 0.106 20.00 1992-0.007-0.434 0.024 0.567-0.017 54.84-0.001-0.015 0.464 1.855-0.492-1.855 0.173 9.68 1993-0.012-0.571 0.068 1.495 0.149 34.48-0.004-0.174 0.472 2.245-0.444-2.144 0.413 13.79 1994 0.004 0.220-0.007-0.193 0.075 14.94 0.007 0.266 0.261 1.045-0.284-1.005 0.075 32.26 1995 0.008-0.385 0.020 0.136 0.029 48.39 0.015 0.300 0.301 1.207-0.379-1.380 0.156 29.03 1996-0.002-0.107 0.029 0.475-0.014 53.13 0.006 0.193 0.391 1.638-0.432-1.828 0.192 25.00 Panel B: Indusry-year porfolios, pariioned according o he goodness-of-fi of he firs-sage regressions Jones Models Mean of indusry-year Very Poor Fi: Poor Fi: Good Fi: Very Poor Fi: porfolios (R 2 0) (0 < R 2 < 5%) (R 2 > 20%) F d (R 2 0) Margin Model Poor Fi: (0 < R 2 < 5%) Good Fi: (R 2 > 20%) F d Number of firms 18.716 20.130 19.807 0.490 18.277 18.762 21.275 1.728 Sand. dev. of WCA 0.075 0.074 0.077 0.231 0.075 0.067 0.077 1.654 Recurring indusries: e Clohing & foowear NA Breweries, pubs & resaurans House building NA Clohing & foowear Vehicle Elecrical Elecrical disribuors Oher building maerials Engin. vehicle componens equipmen Oil exploraion & producion equipmen Breweries, pubs & resaurans Paper, packaging & prining 27
Table 2 (Coninued) Aerospace Broadcasing Chain Sores House building Leisure faciliies Muli-sores Business suppor Publishers Oher consrucion Broadcasing Medical producs Engineering Fabricaors Elecronics Ind. componens Diversified indusrials a The Jones models and he margin model are esimaed separaely for each indusry-year combinaion. The minimum number of observaions in an esimaion porfolio is en. Since he s-j and m-j models are equivalen a he esimaion sage, we repor only a single se of resuls for hese wo models in he columns headed Jones Models. Working capial accruals (WCA) are defined as non-cash curren asses minus curren liabiliies (excluding he curren porion of long-erm deb). REV is oal sales revenue, CR is cash received, measured as oal sales minus he change in rade debors. b The number of indusry groups in 1990 (1991, 1992, 1993, 1994, 1995, 1996) is 31 (30, 31, 29, 31, 31, 32). c The percenage of indusry groups where he adjused-r 2 is negaive. d The F-saisic from a one-way ANOVA esing he difference in means across he hree goodness-of-fi caegories. e Daasream level-6 indusry groups where he adjused-r 2 mees he goodness-on-fi crieria in a leas four years during he sample period. 28
Table 3 Comparison of Type I Error Raes for Abnormal Accruals Esimaed Using he Sandard- Jones (s-j), he Modified-Jones (m-j) and Margin. Percenage of 700 Randomly Seleced Firm-Years for which he Null Hypohesis of No Earnings Managemen is Rejeced Using a One-Tailed Tes. a Alernaive Hypohesis Income-increasing accruals Income-decreasing accrual Null Hypohesis Earnings Managemen 0 Earnings Managemen 0 Tes Level: 5% 1% 5% 1% s-j model -es 5.4% 1.7% 5.9% 1.3% m-j model -es 6.3% 1.4% 5.4% 1.3% margin model -es 5.1% 0.9% 5.0% 1.3% a Simulaions for each abnormal accrual model are performed using 100 random samples of 25 firms in each of he seven sample years (1990-1996), resuling in a oal of 700 simulaions per model. * Significanly differen from he specified es level a he 5 percen level using a wo-ailed binomial es. ** Significanly differen from he specified es level a he 1 percen level using a wo-ailed binomial es. 29
Table 4 Descripive Saisics for Abnormal Accruals, Pariioned by Cash Flow from Operaions Scaled by Lagged Toal Asses (CFO). Abnormal Accruals are Esimaed Using he Sandard-Jones (s-j), he Modified-Jones (m-j) and Margin Models a Abnormal accrual model p-value for difference beween: b CFO Decile s-j m-j margin s-j vs. m-j s-j vs. margin m-j vs. margin Low Mean 0.0736 0.0762 0.0603 0.001 0.001 0.001 Median 0.0693 0.0705 0.0561 0.001 0.001 0.001 S. dev. 0.0883 0.0908 0.0783 2 Mean 0.0369 0.0381 0.0299 0.055 0.001 0.001 Median 0.0351 0.0352 0.0281 0.002 0.001 0.001 S. dev. 0.0552 0.0590 0.0486 3 Mean 0.0172 0.0174 0.0154 0.672 0.263 0.269 Median 0.0136 0.0143 0.0129 0.853 0.501 0.510 S. dev. 0.0499 0.0513 0.0474 4 Mean 0.0088 0.0102 0.0083 0.001 0.748 0.260 Median 0.0070 0.0074 0.0061 0.022 0.379 0.209 S. dev. 0.0515 0.0533 0.0457 5 Mean 0.0022 0.0028 0.0026 0.045 0.793 0.880 Median 0.0014 0.0016 0.0017 0.119 0.469 0.702 S. dev. 0.0475 0.0490 0.0451 6 Mean -0.0042-0.0038-0.0021 0.256 0.179 0.328 Median -0.0074-0.0068-0.0032 0.284 0.321 0.443 S. dev. 0.0503 0.0517 0.0465 7 Mean -0.0144-0.0141-0.0106 0.242 0.002 0.008 Median -0.0101-0.0092-0.0075 0.091 0.002 0.010 S. dev. 0.0486 0.0490 0.0465 8 Mean -0.0189-0.0178-0.0146 0.090 0.011 0.099 Median -0.0137-0.0134-0.0116 0.203 0.001 0.002 S. dev. 0.0573 0.0609 0.0527 9 Mean -0.0352-0.0349-0.0313 0.415 0.013 0.036 Median -0.0318-0.0324-0.0271 0.140 0.022 0.100 S. dev. 0.0642 0.0648 0.0593 High Mean -0.0651-0.0658-0.0563 0.273 0.001 0.001 Median -0.0519-0.0565-0.0495 0.965 0.001 0.001 S. dev. 0.0830 0.0831 0.0783 a Cash flow from operaions is compued as operaing income minus he change in non-cash working capial, plus depreciaion and amorisaion of inangible fixed asses, all scaled by lagged oal asses. The CFO deciles are formed by ranking all observaions wihin heir sample year. b The differences in means (medians) are evaluaed using a paired -es (Wilcoxon es). P-values relae o wo-ailed ess. 30
Figure 1 Simulaion Resuls of Power Funcions for Tess of Earnings Managemen Using Abnormal Accruals. Simulaions are Performed Using Arificially Induced Amouns of Earnings Managemen Ranging from Zero Percen o 10 Percen of Lagged Toal Asses Using One-ailed Tes Levels of Five Percen and One Percen. The Number of Simulaions is Equal o 700. Abnormal Accruals are Esimaed Using he Sandard-Jones (Solid Line), he Modified-Jones (Dashed Line) and Margin (Doed Line) Models Null Hypohesis Rejeced (%) Tes Level = 5% Tes Level = 1% Panel A: Expence Manipulaion 100 100 90 90 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0 0 2 4 6 8 10 0 2 4 6 8 10 Induced Earnings Managemen (%) Induced Earnings Managemen (%) Panel B: Bad Deb Manipulaion Null Hypohesis Rejeced (%) 100 90 80 70 60 50 40 30 20 10 0 0 2 4 6 8 10 Induced Earnings Managemen (%) 100 90 80 70 60 50 40 30 20 10 0 0 2 4 6 8 10 Induced Earnings Managemen (%) Null Hypohesis Rejeced (%) 100 90 80 70 60 50 40 30 20 10 0 Panel C: Revenue Manipulaion 0 2 4 6 8 10 Induced Earnings Managemen (%) 100 90 80 70 60 50 40 30 20 10 0 0 2 4 6 8 10 Induced Earnings Managemen (%) 31