CFA Institute

The Detection of Earnings Manipulation

Author(s): Messod D. Beneish
Source: Financial Analysts Journal, Vol. 55, No. 5 (Sep. - Oct., 1999), pp. 24-36
Published by: CFA Institute
Stable URL: http://www.jstor.org/stable/4480190
Accessed: 13-03-2016 06:53 UTC

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

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

CFA Institute is collaborating with JSTOR to digitize, preserve and extend access to Financial Analysts Journal.

http://www.jstor.org

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earnings Manipulation

Messod D. Beneish

Presented are a profile of a sample of earnings manipulators, their

distinguishing characteristics, and a suggested model for detecting

manipulation. The model's variables are designed to capture either the

financial statement distortions that can result from manipulation or

preconditions that might prompt companies to engage in such activity. The

results suggest a systematic relationship between the probability of

manipulation and some financial statement variables. This evidence is

consistent with the usefulness of accounting data in detecting manipulation

and assessing the reliability of reported earnings. The model identifies

approximately halfof the companies involved in earnings manipulation prior

to public discovery. Because companies that are discovered manipulating

earnings see their stocks plummet in value, the model can be a useful

screening device for investment professionals. The screening results,

however, require determination of whether the distortions in the financial

statement numbers result from earnings manipulation or have another

structural root.

T he extent to which earnings are manipu- the SEC's accounting enforcement actions or were

lated has long been of interest to analysts, identified as manipulators by the news media.

regulators, researchers, and other invest-

To identify companies subject to accounting

ment professionals. The U.S. SEC's recent

enforcement actions by the SEC, I used Accounting

commitment to vigorous investigation of earnings

and Auditing Enforcement Releases (AAERs) #132

manipulation (see Levitt 1998) has sparked renewed

through #502, issued from 1987 to 1993. After elim-

interest in the area, but the academic and profes-

ination of inappropriate companies from the 363

sional literature contains little discussion of the

AAERs examined (#372 to #379 were not assigned

detection of earnings manipulation.

by the SEC), this search yielded a final sample of

This article presents a model to distinguish

103 AAERs relating to 49 companies that violated

manipulated from nonmanipulated reporting.1

GAAP.2

Earnings manipulation is defined as an instance in

An extensive media search on LEXIS-NEXIS

which a company's managers violate generally

for January 1987 to April 1993 identified 80 compa-

accepted accounting principles (GAAP) to favorably

nies mentioned in articles discussing earnings

represent the company's financial performance. To

manipulation.3 After elimination of eight compa-

develop the model, I used financial statement data

nies that were also identified by the AAER search

to construct variables that would capture the effects

and of other inappropriate companies from this

of manipulation and preconditions that might

search, the final sample from the new search was

prompt companies to engage in such activity.

40 companies.4

For the remaining 40, I required ex post evi-

Sample

dence of manipulation. That is, I required that the

The sample consisted of 74 companies that manip-

companies had to restate earnings to comply with

ulated earnings and all Compustat companies

GAAP at the request of auditors or following the

matched by two-digit SIC numbers for which data

conclusion of an internal investigation. This

were available for the 1982-92 period. The 74 earn-

requirement made this sample consistent with the

ings manipulators were either companies subject to

sample from the AAER search, in the sense that a

restatement (in addition to a permanent injunction

Messod D. Beneish is an associae professor at the Kelley

from future violations of security laws) is usually

School of Business, Indiana University.

the outcome of a successful SEC investigation.

24 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earnings Manipulation

turing (SIC 30-39) and personal and business ser-

This criterion was imposed to exclude compa-

vices (SIC 70-79), which together represented 45

nies that managed earnings within GAAP and to

percent of the sample. Of the 63 four-digit SIC codes

reduce the likelihood that companies would be con-

represented, only 4 companies were in software (SIC

sidered manipulators in the sample on the basis of

7372), 3 were in computers (SIC 3571), and 3 were in

media articles based on self-serving rumors by

audiovisual retail stores (SIC 5731).

interested parties. For example, articles by Hector

Table 1 contains a comparison of manipula-

(1989) and Khalaf (1992) discussed changes in "use-

tors' financial characteristics with those of

ful lives" at General Motors Corporation, unusual

industry-matched controls in the fiscal year prior

charges at General Electric Company, and short

to the year that contained the public disclosure of

sellers' interest in Advanta Corporation. None of

the earnings manipulation. In the prior year,

the companies was subsequently required to

manipulators were smaller (when size was mea-

reverse the effects of its accounting decisions; thus,

sured either in terms of total assets or in terms of

these companies were excluded from the manipula-

sales), less profitable, and more leveraged than

tor sample. Similarly, some companies-for exam-

control companies. Manipulators also experienced

ple, Battle Mountain Gold Company and

higher growth. The median sales growth of manip-

Blockbuster Entertainment-voluntarily changed

ulators (34.4 percent) is significantly larger than

their accounting choices or estimates as a result of

that of controls (9.4 percent), which raises the ques-

pressure from the investment community. Because

tion: Is growth exogenous to earnings manipula-

their accounting choices were initially within GAAP

tion or a result of manipulation? I found that in the

and they did not restate earnings, they also were not

year prior to the fiscal year of manipulation, manip-

considered manipulators. These eliminations

ulators also had significantly higher growth than

reduced the news media sample by 15 companies.

nonmanipulators (medians were 29.4 percent ver-

The 25 manipulators identified by the news

sus 10.6 percent), which suggests that growth orig-

media search had size, leverage, liquidity, profit-

inates exogenously. This profile of manipulators as

ability, and growth characteristics that were similar

companies with high-growth prospects could

to those of the 49 SEC manipulators, which suggests

explain why I found manipulators to have, on aver-

that manipulators found by the different searches

age, lower total assets than control companies but

were not drawn from different populations.

similar market values of equity.

Sample manipulators typically overstated

Sample versus Control Companies

earnings by recording fictitious, unearned, or uncer-

The final sample consisted of 74 companies that

tain revenues, recording fictitious inventory, or

manipulated earnings, which were matched to 2,332

improperly capitalizing costs. The context of earn-

Compustat nonmanipulators by two-digit SIC

ings manipulation was an annual report or a 10-K

industry and year for which the financial statement for about two-thirds of the sample and was a secu-

data used in the model were available.5 The distri- rity offering prospectus (initial, secondary, debt

bution of manipulators by two-digit SIC groups offering) for the remaining third. Sample manipula-

indicates a concentration of companies in manufac- tors were relatively young growth companies,

Table 1. Characteristics of Sample and Control Companies: Fiscal Year Prior to Public Disclosure,

1982-92 Data

Manipulators Nonmanipulators Wilcoxon Z Median X2

Characteristic Mean Median Mean Median p-Value p-Value

Size (millions)

Total assets $467.33 $43.20 $1,140.37 $ 95.84 0.003 0.004

Sales 469.87 53.56 1,295.22 122.54 0.001 0.007

Market value 323.72 74.90 813.35 64.59 0.884 0.701

Liquidity/leverage

Working capital to total assets 0.26 0.28 0.30 0.31 0.472 0.345

Current ratio 2.54 1.83 2.54 2.11 0.103 0.473

Total debt to total assets 0.58 0.58 0.51 0.52 0.027 0.098

Profitability/growth

Return on assets -1% 3% 3% 5% 0.063 0.078

Sales growth 58 34 13 9 0.000 0.001

Note: The Wilcoxon rank-sum and the median X2 tests were used to evaluate the null hypothesis that the size, liquidity, profitability,

and growth characteristics of manipulators and nonmanipulators indicate that the groups were drawn from the same population.

September/October 1999 25

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
Financial Analysts Journal

which have characteristics that make a company financial statement data. First, I considered signals

more likely to come under the scrutiny of regulators about future prospects that appear in the academic

(see Beneish forthcoming 1999).

and practitioner literature.7 The presumption was

that earnings manipulation is more likely when

companies' future prospects are poor. Second, I con-

The Model

sidered variables based on cash flows and accruals

I estimated a model for detecting earnings manip-

(Healy 1985; Jones 1991). Third, I considered vari-

ulation using sample manipulators and

ables drawn from positive theory research, which

industry-matched companies in the 1982-88 period

hypothesizes contract-based incentives for earnings

and evaluated the model's performance on a

management (Watts and Zimmerman 1986).

holdout sample in the 1989-92 period. The model

The result of the search for variables based on

for detection of earnings manipulation is

financial statement data was a model that includes

eight variables.8 The variables were measured from

Mi= 3'Xi+ ?if

data from the fiscal year of the first reporting vio-

where M is a dichotomous variable coded 1 for lation (e.g., the first year for which the company

manipulators and 0 otherwise, X is the matrix of was subsequently required to restate). I designated

explanatory variables, and ? is a vector of residuals. seven of the eight variables as indexes because they

are intended to capture distortions that could arise

Estimation. In this study, earnings manipula-

from manipulation by comparing financial state-

tors were oversampled relative to their true propor-

ment measures in the year of the first reporting

tion in the population. The econometric justification

violation with those in the year prior.9

is that such a state-based sample is likely to generate

The computation of each variable and its Com-

a larger number of manipulators than a random

pustat number are in the notes to Table 2.

sample would generate, which would make the

X Days' sales in receivables index. The DSRI is

identification of a model for classifying earnings

the ratio of days' sales in receivables in the first year

manipulation difficult. However, because estima-

in which earnings manipulation was uncovered

tion of a dichotomous-state model that ignores the

(year t) to the corresponding measure in year t - 1.

state-based sample procedures would yield asymp-

This variable gauges whether receivables and rev-

totically biased coefficient estimates, I used

enues are in or out of balance in two consecutive

weighted exogenous sample maximum likelihood

years. A large increase in days' sales in receivables

(WESML) probit as well as unweighted probit.6 The

could be the result of a change in credit policy to

estimation sample spanned the 1982-88 period and

spur sales in the face of increased competition, but

consisted of 50 manipulators and 1,708 controls.

disproportionate increases in receivables relative to

Using WESML required an estimate of the sales could also suggest revenue inflation. Thus, I

proportion of companies in the population that expected a large increase in the DSRI to be associ-

manipulate earnings. Assuming that the popula- ated with a higher likelihood that revenues and

tion from which the companies were sampled is the earnings are overstated.

population of Compustat companies, one estimate X Gross margin index. The GMI is the ratio of

of the proportion of manipulators equals 0.0069 the gross margin in year t -1 to the gross margin in

(50/7,231). Because I have no way of assessing the year t. When the GMI is greater than 1, gross mar-

validity of this assumption, I also evaluated the gins have deteriorated. Lev and Thiagarajan sug-

gested that deterioration of gross margin is a

sensitivity of the model to other specifications of

negative signal about a company's prospects. So, if

the prior probability of manipulation.

companies with poorer prospects are more likely to

Variables: Can Accounting Data Be Used to engage in earnings manipulation, I expected a pos-

Detect Earnings Manipulation? If financial state- itive relationship between GMI and the probability

ment manipulations encompass not only earnings of earnings manipulation.10

but also other signals that investors and analysts rely 0 Asset quality index. Asset quality in a given

on, then the discriminatory power of accounting year is the ratio of noncurrent assets other than

data is diminished, the results of this study are property, plant, and equipment (PP&E) to total

biased against rejection of a null hypothesis on the assets and measures the proportion of total assets

variables' coefficients, and the usefulness of for which future benefits are potentially less certain.

accounting information for detecting earnings The asset quality index (AQI) is the ratio of asset

manipulation is limited. In the absence of an eco- quality in year t to asset quality in year t - 1. The

nomic theory of manipulation, I relied on three AQI is an aggregate measure of the change in asset

sources for choosing explanatory variables based on realization risk, which was suggested by Siegel. If

26 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earnings Manipulation

Table 2. Distribution of Variables for Manipulators and Nonmanipulators in

the Estimation Sample

Manipulators Nonmanipulators

(N = 50) (N = 1,708) Wilcoxon Z Median Test

Characteristic Mean Median Mean Median p-Value p-Value

DSRI 1.465 1.281 1.031 0.996 0.000 0.000

GMI 1.193 1.036 1.014 1.001 0.006 0.007

AQI 1.254 1.000 1.039 1.000 0.096 0.246

SGI 1.607 1.411 1.134 1.106 0.000 0.000

DEPI 1.077 0.966 1.001 0.974 0.307 0.774

SGAI 1.041 0.960 1.054 1.010 0.271 0.389

LVGI 1.111 1.030 1.037 1.000 0.394 0.077

TATA 0.031 0.034 0.018 0.013 0.000 0.002

N= number of companies.

Notes: The definition and Compustat data item number (in brackets) for each variable are as follows

(LTD = long-term debt; t and t - 1 are defined in the text):

Receivablest[2] /Salest[12]

DSRI = Receivablest_I /Salest-I

(Salest_1[12] - Cost of goods soldt_ [41] )/Salest_1[12]

(Salest[12] - Cost of goods soldt[41])/Salest[12]

(1 - Current assetst[4] + PP&E[8])/Total assetst[6]

(1 - Current assetst_l + PP&Et_1) /Total assetst_

S[Salest12]

SGI = Sls~

Salest_l

Depreciationt_ [14 - 65] /(Depreciationt_I + PP&Et_1[8]

Depreciationt/(Depreciationt + PP&Et)

Sales, general, and administrative expenset[189] /Sales [12]

SGA =i 1 - Sales, general, and administrative expenset_1 /Salest1

(LTDt[9] + Current liabilitiest[5]) /Total assetst[6]

L VGI = - (LTDt_ + Current liabilitiest _)/Total assetst_i

ACurrent assetst[4] - ACasht[I] - ACurrent liabilitiest[5]

- ACurrent maturities of LTDt[44] - Alncome tax payablet[71]

-Depreciation and amortizationt[14]

TA TA = -Total assets [6]

The Wilcoxon and median tests compared the distribution of manipulator companies' characteristics

with the corresponding distribution for nonmanipulators. The reported p-values indicate the smallest

probability of incorrectly rejecting the null hypothesis of no difference.

the AQI is greater than 1, the company has poten- because their financial positions and capital needs

tially increased its involvement in cost deferral." put pressure on managers to achieve earnings tar-

An increase in asset realization risk indicates an gets (National Commission on Fraudulent Financial

increased propensity to capitalize, and thus defer, Reporting 1987; National Association of Certified

costs. Therefore, I expected to find a positive rela- Fraud Examiners 1993). In addition, concerns about

tionship between the AQI and the probability of controls and reporting tend to lag operations in

earnings manipulation.
periods of high growth (National Commission on

X Sales growth index. The SGI is the ratio of Fraudulent Financial Reporting; Loebbecke, Eining,

sales in year t to sales in year t - 1. Growth does not and Willingham 1989). If growth companies face

imply manipulation, but growth companies are large stock price losses at the first indication of a

viewed by professionals as more likely than other slowdown, they may have greater incentives than

companies to commit financial statement fraud, nongrowth companies to manipulate earnings.

September/October 1999 27

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
Financial Analysts Journal

About this point, Fridson (1993) commented: problem when the denominator is small. To deal

Almost invariably, companies try to dispel the with this problem, I winsorized the data at the 1

impression that their growth is decelerating,

percent and 99 percent percentiles for each vari-

since that perception can be so costly to them.

able. In addition, the denominator of the AQI vari-

(pp. 7-8)

able was sometimes zero because assets in the

I thus foresaw a positive relationship between the

reference year (period t - 1) consisted exclusively

SGI and the probability of earnings manipulation.

of current assets and PP&E. Because the AQI was

0 Depreciation index. The DEPI is the ratio of

not defined in such cases, I set its value to 1 (its

the rate of depreciation in year t - 1 to the corre-

neutral value) instead of treating the observation as

sponding rate in year t. The depreciation rate in a

missing. Similarly, I set the DEPI and SGAI to val-

given year is equal to Depreciation/(Depreciation

ues of 1 when elements of the computation (amor-

+ Net PP&E). A DEPI greater than 1 indicates that

tization of intangibles, Compustat #65, and SG&A,

the rate at which assets are being depreciated has

Compustat #189) were not available on the Com-

slowed-raising the possibility that the company

pustat tapes. I found that estimating the model with

has revised upward the estimates of assets' useful

and without those observations yielded similar

lives or adopted a new method that is income

results.

increasing.'2 I thus expected a positive relation-

Table 2 contains a comparison of the distribu-

ship between the DEPI and the probability of

tion of these variables for manipulators and non-

manipulation.
manipulators in the estimation sample. The results

X Sales, general, and administrative expenses

indicate that, on average, manipulators have signif-

index. The SGAI is the ratio of sales, general, and

icantly larger increases in days' sales in receivables,

administrative expenses to sales in year t relative to

greater deterioration of gross margins and asset

the corresponding measure in year t -1. The use of

quality, higher growth, and larger accruals than

this variable follows the recommendation of Lev

nonmanipulators.

and Thiagarajan that analysts interpret a dispro-

portionate increase in sales as a negative signal

Results

about a company's future prospects. I expected to

Panel A of Table 3 reports the results of the WESML

find a positive relationship between the SGAI and

probit and unweighted probit estimations of the

the probability of manipulation.

model. The likelihood ratio test indicates that for

a Leverage index. The LVGI is the ratio of total

both estimations, with X2 statistics (p-values) of 34.5

debt to total assets in year t relative to the corre-

(0.00) and 129.2 (0.00), the model has significant

sponding ratio in year t - 1. An LVGI greater than

power. The pseudo-R2s of 0.306 and 0.371 for,

1 indicates an increase in leverage. This variable

respectively, WESML and unweighted probit indi-

was included to capture incentives in debt cove-

cate that the model has descriptive validity.

nants for earnings manipulation. Assuming that

Because the coefficient estimates have similar

leverage follows a random walk, the LVGI implic-

magnitudes and significance across estimations, I

itly measures the leverage forecast error. I used the

focus on the results of the unweighted probit esti-

change in leverage in a company's capital structure

mation. The variable DSRI has a positive coefficient

on the basis of evidence in Beneish and Press (1993)

(0.920) and is significant at the 5 percent level with

that such changes are associated with the stock

an asymptotic t-statistic of 6.02. This result is consis-

market effect of technical default.

tent with disproportionate increases in receivables,

X Total accruals to total assets. Total accruals

raising the likelihood that a company has inflated

were calculated as the change in working capital

revenues. The variable GMI has a positive coeffi-

accounts other than cash less depreciation. Either

cient (0.528) that is more than 2 standard deviations

total accruals or a partition of total accruals was

from zero. This result is consistent with companies

used in prior work to assess the extent to which

that face poor prospects having greater incentives

managers make discretionary accounting choices

for earnings manipulation. The AQI also has a sig-

to alter earnings (see, for example, Healy or Jones).

nificant positive coefficient (0.404, t-statistic 3.20),

I used total accruals to total assets (TATA) to proxy

which is consistent with the likelihood of earnings

for the extent to which cash underlay reported

manipulation increasing when companies change

earnings, and I expected higher positive accruals

their accounting treatment of cost deferral. The SGI

(less cash) to be associated with a higher likelihood

has a positive coefficient that is more than 5 stan-

of earnings manipulation.
dard deviations from zero, which is consistent with

X Distribution of variables. The explanatory growth companies that are facing growth decelera-

variables in the model were primarily based on tion having incentives to manipulate earnings. The

year-to-year changes, which introduces a potential TATA variable has a significant positive coefficient,

28 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earnings Manipulation

-MO Ln O N 0 m a-, O'\ W

C) all "d a-, 0 o CL, m Un w

cq

z CD c CD CD

41 .4 4. C)

Co C;,, N !t CD C) "C

OC) 00 01 m O CIr) C) u

N 01 C C) \C C) C)

CU
4-1 4 6
CD CD CD 14.

C)

M N C-4 4-1

+ C-4 00 M C-4 m

4-1

;-o

4 4-1

4.1 4-

C) C) 01 C) 4- C)

CD C:, o 8 CD C)

CD "C N CL) Z m a)

m m 't m 00 (1)

0c) 4.

C7\

rz C)

4 C)

C) OC) -C

O CZ

4-1 Ci

.-q 7

+ C) C) -61 4.1 14. 4.1

(A CZ C) -.0

C:) CD CD I t-, C7\ CD C) CA

z m N 00 CN C:) \0 C:) CD

N N C3" 0 q m \0 CD

C:) CD CD C:) C:)

4.1

CZ

u>
CD

CA 4' --4

Ct 4.,

T:s CA

M Ln 4-1 1-4

b r-4 r-..q N C) %4-.(

CZ -,, '-t 0

4-1 7-01

:S C)

-0 CIO -0

CZ cn O

CD lzt C7 CD C)

C) CD 01 C C:)

8 C,

(t,-, 4-1

4-1
C'4 C4 CZ

CD cn O" 0,., CZ N

+ 00 m CZ

In CD 4-1 C)

0 't ";

r-., CD Ln M Q)

CZ C)

ON C4 OIN O CD C4 CD 01 0

= N CD C3.,

6 C) -Q
0 --4

4.1 -+-, m

m oo 4 CA W

cu (3)

CL V 75

4.4 75

0 C;,\ 00

Lf) C4 C) m

(L cO 't C) Lr) C-4

E C) 4 - 0

C) M C:) C-4 m m M C)
- 4 C) CI4 r-- C) C)

4) C C:) \-O CD C) OC)

8 C o CD

CD

ci 0) a) -0 -4-

00 C) C4 N u

4-1 C)

r, cu u

OIN C:) Lr) C:)

Co O N Lr) N CD It 00 C

00 C) 0 C) O N C >, 75

CZ r-L, r! ri . . .
4-1 1-4
Lr) C) "t C) O C) 06 6 CY) C) 44
CL CA .,
C\ 00 00 C) C4

I IT r--

Lr

0 CZ

L-

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 Financial Analysts Journal

which is consistent with manipulators having less

to be a problem because none of the 36 Pearson

cash behind their accounting income.

correlation coefficients was greater than 0.25, I

The coefficients on the LVGI, DEPI, and SGAI

dropped up to four variables from the model to

variables are not significant. Possibly, these vari-

assess the stability of the coefficient estimates.

ables are associated with earnings management, not

Dropping the DEPI, LVGI, SGAI, and TATA vari-

manipulation. For example, a change from acceler-

ables one at a time and in combination yielded

ated to straight-line depreciation or a revision that

results similar to results for the remaining variables.

lengthens assets' useful lives would result in higher

Second, I assessed the sensitivity of the WESML

values of the DEPI but are an instance of earnings

estimation results to the specification of the prior

management, so the company would not be

probability of manipulation. In addition to the esti-

included in the manipulator sample. Similarly, for

mations based on prior probabilities of 0.0069 and

the LVGI variable, incentives to comply with debt

the 0.02844 implicit in unweighted probit (50 /1,758),

covenants might be insufficient to induce earnings

I estimated the model with four other prior proba-

manipulation because the costs of noncompliance

bilities of earnings manipulation; as explained in the

are small (Beneish and Press estimated these costs

notes to Table 3, these estimates yielded similar

at 1-2 percent of market value of equity). 13

results in terms of both the descriptive validity of the

Panel B of Table 3 shows the estimated proba-

model and the sign and significance of the coefficient

bilities of earnings manipulation for both the estima-

estimates.

tion and the holdout samples. For the estimation

Third, although the holdout sample was cho-

sample, the model estimated by using WESML pre-

sen to be independent from the estimation sample,

dicted a higher average (median) probability of

I assessed the sensitivity of the results to the choice

earnings manipulation for manipulators, 0.107

of estimation and holdout samples. To do so, I

(0.024), than for nonmanipulators, 0.006 (0.003).

generated 100 random samples of 50 manipulators

Similarly, the model estimated by using unweighted

and 1,500 controls with a basic random-number

probit predicted a higher average (median) proba-

generator in SAS and used them to estimate the

bility for manipulators, 0.237 (0.099), than for non-

model 100 times. Similarly, I obtained 100 random

manipulators, 0.022 (0.011). Wilcoxon and median

holdout samples by treating the complement of 24

tests rejected the null hypothesis that estimated

manipulators and 832 controls to each random esti-

probabilities for manipulators and nonmanipula-

mation sample as a holdout sample and repro-

tors were drawn from the same distribution.

duced the estimated probabilities tests. The results

Results for the holdout sample of 24 manipu-

are reported in Table 4, and the evidence suggests

lators and 624 controls were similar to the estima-

that the findings are not sensitive to the choice of

tion sample findings. The model predicted that

estimation/holdout samples.

manipulators are, on average, about 10 times more

Overall, the estimation results provide evidence

likely to manipulate earnings. The distributions of

of a systematic relationship between the likelihood

estimated probabilities for manipulators and non-

of manipulation and some financial statement data.

manipulators based on unweighted probit illus-

The results are robust to different estimates of the

trate these differences. For example, although not

prior probability of earnings manipulation, several

reported in the tables, nearly all the nonmanipula-

specifications of the model, and various transforma-

tors in the estimation sample (93.4 percent) had an

tions of the explanatory variables. The results are

estimated probability of manipulation of less than

also insensitive to the choice of estimation and hold-

0.05 whereas the manipulators had a 38.0 percent

out samples.
probability of manipulating. Similarly, in the hold-

out sample, 56.1 percent of the nonmanipulators

had an estimated probability of manipulation of

The Model as a Classification Tool

less than 0.01, compared with 20.8 percent of the

Because the model distinguishes manipulators

manipulators.

from nonmanipulators, I assessed its usefulness as

The evidence indicates that the probability of

a classification tool. This section discusses the prob-

manipulation increases with (1) unusual increases

ability cutoffs associated with different costs of

in receivables, (2) deteriorating gross margins, (3)

making classification errors. The model can make

decreasing asset quality, (4) sales growth, and (5)

two types of errors: It can classify a company as a

increasing accruals.

nonmanipulator when it manipulates (a Type I

error), or it can classify a company as a manipulator

Robustness

when it does not manipulate (a Type II error). The

I assessed the robustness of the results in three probability cutoffs that minimize the expected costs

ways. First, even though collinearity was not likely

of misclassification depend on costs associated with

30 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earnings Manipulation

uO0

u D>o O O OtN

OLr)

N
.u U > O n O n 6 t- O X 7 x

<~~~~~l c i \ 0 00 00 C7\ H c )

C)~~~~~~~~~~~~~~~~'

000C O~C z I

U a) > ? ? o C) ? N ?) N 4 ? r>

G o o o o. o o o. ?. ?. o 0 C CD

00~~~~~

0 O'0 0

E~~~~~~~~~~~~~~~

0 U 0 ~ ~ ~ ~ ~ ~ a 00" .

z 00 M N n 01- Ln O N

CD 00 O 0r--~ C) C4

Ct~~~~.

CD m C C ItN t m N a

0 2 -t -0a .e

X "t CD ? ? l

CZ 0 U m m m 0 Ln m r N

O ~~~~~~~~~~~~~~~X

r, X -1 o. CD M N NNt?
S u0 :E

Sepemer/ctbe 199 31 CZ C) a,

0 e 4 an-1 x-00 00 0C4 0 ?.

O ~ ~ ~ ~ C t:) 00 Lf L O O O O O' C-4 r

cn~~~ ~~~ C) C:, CD C00 C:) c tNHt $

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
Financial Analysts Journal

in the quarter containing the discovery of the

the relative cost of making an error of either type.

manipulation. Assuming that, on a similar basis, a

Although decision makers' objective functions

typical company's equity appreciates 1-2 percent a

are not observable, classification error costs are

quarter, 20-40 nonmanipulators would be needed

likely to differ by decision maker. For example,

in the investor's portfolio to offset a single manip-

investors are likely to have high Type I error costs

ulator in that quarter. Therefore, one possibility is

because the investment loss associated with the

that investors would view a Type I error as 20-40

discovery of the manipulation is dramatic, whereas

times as costly as a Type IL error.

their Type II error costs would be low because of

I computed the probability cutoffs that would

the availability of substitutes. A regulator's objec-

minimize the expected costs of misclassification;

tive function, however, requires balancing the pro-

the equation for computing the costs and the results

tection of the investing public against the costs of

are presented in Table 5. The results are similar

falsely accusing a company. Their relative costs

across estimation methods, so I focus on the

cannot be measured, but it is likely that their Type

unweighted probit estimation in Panel B.

II error costs are higher than those of investors.

In the estimation sample, at relative error costs

The costs of Type I and Type II errors are not

of 10:1, the model classified companies as manipu-

amenable to objective measurement, but I consid-

lators when the estimated probabilities exceeded

ered relative costs of Type I to Type II errors rang-

0.0685 (a score greater than -1.49); it misclassified

ing from 1:1 to 100:1. (For investors, however, the

42 percent of the manipulators and 7.6 percent of

relevant range is likely between 20:1 and 30:1.) To

the nonmanipulators. At relative error costs of 20:1

explain: According to Beneish (forthcoming 1999),

or 30:1, the model classified companies as manipu-

the typical manipulator loses approximately 40 lators when the estimated probabilities exceeded

0.0376 (a score greater than -1.78); it misclassified

percent of its market value on a risk-adjusted basis

Table 5. Probability Cutoffs that Minimize the Expected Costs of Misclassification

Estimation Sample Holdout Sample

Relative Costs Cost of Model Cost of Model

of Type I and Cutoff Errors Relative to Errors Relative to

Type II Errors Probability Type I Type II Naive Strategya TypeI TypeII Naive Strategy

A. WESML

1:1 1.0000 1.0000 0.0000 1.000 1.0000 0.0000 1.000

10:1 0.2905 0.9000 0.0004 0.991 0.9166 0.0048 0.986

20:1 0.0512 0.5600 0.0409 0.855 0.7500 0.0112 0.830

30:1 0.0512 0.5600 0.0409 0.757 0.7500 0.0112 0.804

40:1 0.0223 0.4600 0.0632 0.688 0.6667 0.0240 0.753

60:1 0.0092 0.2800 0.1329 0.597 0.5000 0.0689 0.665

100:1 0.0087 0.2600 0.1417 0.464 0.5000 0.0753 0.608

B. Unweighted probit

1:1 1.0000 1.0000 0.0000 1.000 1.0000 0.0000 1.000

10:1 0.0685 0.4200 0.0761 0.680 0.6250 0.0353 0.746

20:1 0.0376 0.2600 0.1382 0.496 0.5000 0.0721 0.623

30:1 0.0376 0.2600 0.1382 0.417 0.5000 0.0721 0.582

40:1 0.0294 0.2400 0.1747 0.433b 0.4583 0.0913 0.628b

60:1 0.0294 0.2400 0.1747 0.562b 0.4583 0.0913 0.896b

100:1 0.0294 0.2400 0.1747 0.819b 0.4583 0.0913 1.535b

Notes: The expected costs of misclassification, ECM, were computed as

ECM = P(M)PIC1 + [1 -P(M)]PITCII

where P(M) is the prior probability of encountering earnings manipulators (0.0069 for WESML and 0.02844 for unweighted probit), PI

and PI, are the conditional probabilities of, respectively, Type I and Type II errors, and CI and CII are the costs of Type I and Type II

errors. Cutoff probabilities were chosen for each level of relative costs to minimize the expected costs of misclassification as defined in

this equation.

aThe expected cost of misclassification for a naive strategy that classified all companies as nonmanipulators was 0.0069C1 for the

WESML model and 0.02844C1 for the unweighted probit.

bIn these computations, the naive strategy classified all companies as manipulators. The switch to this naive strategy minimized the

expected costs of misclassification because the ratio of relative costs was greater than the population proportion of manipulators. The

switch occurred at 40:1 for unweighted probit (> 1/0.02844).

32 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earnings Manipulation

26 percent of the manipulators and 13.8 percent of simply classifies all companies as nonmanipula-

the nonmanipulators. In the holdout sample, at rel- tors. Table 5 contains a comparison of the model's

ative error costs of 20:1 or 30:1, the model classified expected costs of misclassification with those of the

companies as manipulators when the estimated naive strategy.

probabilities exceeded 0.0376 (a score greater than

A naive strategy makes no Type II errors (PI, =

-1.78); it misclassified 50 percent of the manipula-

0), and the conditional probability of a Type I error

tors and 7.2 percent of the nonmanipulators.

(PI) is 1. Thus, the naive strategy's expected costs of

Figure 1 and Figure 2 help clarify the classifi-

misclassification would be ECM(naive) = P(M)CI,

cation performance of the unweighted probit

or 0.0069C, for the WESML comparison and

model. The figures contain the following informa-

0.02844C, for the unweighted probit comparison.

tion: (1) the probability cutoffs associated with each

For both the estimation and the holdout sam-

relative error cost assumption, (2) the percentage of

ples, the model had lower expected misclassifica-

correctly classified manipulators, and (3) the per-

tion costs than the naive strategy when the cost of

centage of incorrectly classified nonmanipulators.

a Type I error was greater than that of a Type II

For the estimation sample in Figure 1, the percent-

error. For example, in the estimation sample in

age of correctly classified manipulators ranges

Panel B, the ratios of the cost of the model's errors

from 58 percent to 76 percent whereas the percent-

to the cost of errors from a naive strategy are 0.496

age of incorrectly classified nonmanipulators

and 0.417 at relative error costs of, respectively, 20:1

ranges from 7.6 percent to 17.5 percent. For the

and 30:1. Similarly, for the holdout sample, the

holdout sample in Figure 2, the percentage of cor-

ratios of the cost of the model's errors to the cost of

rectly classified manipulators ranges from 37.5 per-

errors from a naive strategy are 0.623 and 0.582 at

cent to 56.1 percent whereas the percentage of

relative error costs of 20:1 and 30:1. The evidence

incorrectly classified nonmanipulators ranges from

suggests that the model is cost-effective in relation

3.5 percent to 9.1 percent.

to a naive strategy that treats all companies as

These results suggest that, although the model

nonmanipulators.

identifies potential manipulators, it does so with

large error rates in the range of error costs that are

Conclusion

likely to be of most relevance to investors. Thus,

Some accounting variables can be used to identify

because instances of discovered manipulations are

rare, a question is raised about whether the model companies that are manipulating their reported

earnings. I found that, because manipulation typi-

is more useful to investors than a naive strategy that

Figure 1. Classification Performance of the Unweighted Probit Model:

Estimation Sample

90 -

80 74-OOo 76.0%

70 -

60 - 58.00o

;u 50-

30 -

20 -17.500

10:1, 6.85% 20:1, 3.76% 40:1, 2.94%

Error Cost Ratio, Cutoff Probability

1 Maniipulators Correctly Identified D Nonmanipulators Incorrectly Identified

September/October 1999 33

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
Financial Analysts Journal

Figure 2. Classification Performance of the Unweighted Probit Model:

Holdout Sample

100

90

80 -

70 -

60 - 56.1%

50 ~~~~~~~~50.00o

40-

30-

20-

10:1, 6.85% 20:1, 3.76% 40:1, 2.94%

Error Cost Ratio, Cutoff Probability

I Manipulators Correctly Identified C] Nonmanipulators Incorrectly Identified

cally consists of an artificial inflation of revenues or to a strategy of treating all companies as nonma-

deflation of expenses, variables that take into nipulators, its large rate of classification errors

account simultaneous bloating in asset accounts makes further investigation of the screening

have predictive content. I also found that sales results important. The model's variables exploit

growth has discriminatory power: The primary distortions in financial statement data that might

characteristic of sample manipulators was that they or might not result from manipulation. For exam-

had high growth prior to periods during which ple, the distortions could be the result of a material

manipulation was in force. acquisition during the period examined, a material

The evidence presented here was based on a

shift in the company's value-maximizing strategy,

sample of companies whose manipulation of earn-

or a significant change in the company's economic

ings was publicly discovered. Such companies

environment.

probably represent the upper tail of the distribution

One limitation of the model was that it is esti-

of companies that seek to influence their reported

mated using financial information for publicly

earnings-successful and undiscovered manipula-

traded companies. Therefore, it cannot be reliably

tors undoubtedly exist-so the evidence should be

used to study privately held companies. Another

interpreted in that light.

limitation is that the earnings manipulation in the

Given this caution, evidence has been presented

sample involved earnings overstatement rather than

here of a systematic association between earnings

understatement; therefore, the model cannot be reli-

manipulation and financial statement data that is of

ably used to study companies operating in circum-

interest to accounting researchers and investment

stances that are conducive to decreasing earnings.

professionals. The evidence suggests that account-

ing data not only meet the test of providing useful

I benefited from the comments of Vic Bernard, Jack

information, but they also enable an assessment of

Ciesielski, Linda DeAngelo, Martin Fridson, Cam Har-

the reliability of the reporting. The explicit classifi-

vey, David Hsieh, Charles Lee, Eric Press, Bob Whaley,

cation model described here requires only two years

Mark Zmijewski, and workshop participants at Dluke,

of data (one annual report) to evaluate the likelihood

the University of Maryland, the University ofMichigaui,

of manipulation and can be inexpensively applied

and the University of Quebec at Montreal. I am indebted

by the SEC, auditors, and investors to screen a large

to David Hsiehfor his generous econometric advice aind]

number of companies and identify potential manip- the use of his estimation subroutines. I thank Julie Azoli-

ulators for further investigation. lay, Pablo Cisilino, and Melissa McFadden for expert

Although the model is cost-effective relative assistance.

34 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
 The Detection of Earniings Manipulation

Notes

HUMAN ACHIEVEMENT

1. Beneish (1997) contains a model for detecting earnings By way of introduction, afew pertinentfacts. We are a gobalfinancial services powerhouse. An innovator who was

manipulation that differs from this study's model in three ear!y to recognize the energy that could be released lygetting one's corporate head around clients' needs, expectations

and desires. We've never strayedfrom the law Charles Merrill laid down in the '20's - "The interests of our customers
ways: That model was estimated with 64 sample companies

must comefirst. " Wejust happen to believe people can achieve almost anything fthtyput their mindto it. Optimistic?
(versus 74 companies in the present study), the control

Absolutely. But a necessay mindset ifyou are out to make dreams into realiiy...and the impossible happen.
companies were Compustat companies with the largest

unexpected accruals (versus Compustat companies in the

same industry in the present study), and the set of explan-

Merre d
atory variables in the present study provides a more parsi-

monious model than the previous model.

2. I eliminated 80 AAERs relating to financial institutions, 15

relating to auditing actions against independent Certified

Public Accountants, 9 relating to 10-Q violations that were

resolved in annual filings, and 156 relating to companies for

Interestingly enough, this unique belief system has made us immensely successful. With total

which no financial statement data were available on either

client assets of around $1.5 trillion, we are the undisputed leader in planning-based financial

Compustat, S&P Corporate Text, or 10-K microfiche.

advice and management for individuals worldwide. Research tells us that affluent investors regard

Merrill Lynch as the leading financial services firm and, financially, the strongest. We're seeking
3. The search encompassed the following specific

the following professionals in our Walnut Creek; San Ramon; Pleasanton; San Francisco;

LEXIS-NEXIS databases: Barron's, Business Week, Buisiness

Berkeley; Oakland; Stockton; Sacramento; Modesto; Roseville; Santa Rosa; Chico; Auburn;

Wire, Corporate Cash Flozv, Disclosuire Online, Forbes, Fortune,

Redding and Eureka, CA; Reno and Carson City, NV offices.

Itnstitutional Investor, Investor's Business Daily, Money, the

Financial Consultants

Courier Journal, the Nezw York Tinies, the Wall Street Journial,

The assumption is, you're committed to sales and in search of high-income potential. You want

the Washington Post, and the Reuter Business Report. I used

to do the right thing for clients-and, yourself. Our scope, visibility, financial strength, global

the following keywords: "earnings management," "earn- reach, inclination to lead, technical savvy, market insight, depth of talent and focus on achieve-

ment enable us to deliver on a high-earnings, fast-track promise.

ings manipulation," "cooking the books," "financial state-

ments" or "reports" (with adjectives such as "deceptive," Indeed, if you're the high caliber professional we're looking for-aggressive, confident, dynamic.

familiar with all types of business negotiation-Merrill Lynch offers you the opportunity to reach
"false," "fraudulent," "misleading," "illusive," "inappro-

your fullest potential selling our financial services to a broad range of clients in the retail market.

priate,""misstated," and "spurious"), and "inflated" or

Send us your resume. Sell us on your abiities. Together, we're destined for success. If it looks like

"overstated" (with "profits," "earnings," or "income").

a good match for both of us, you'll hear from us promptly. It may just be the best investment you

4. I eliminated 10 companies for which no financial statement

make. Please forward your resume, specifying location of interest, to; Merrill Lynch, ATTN: Leigh

data were available on Compustat, S&P Corporate Text, or

Budnitz, IoI California, 14th Floor, San Francisco, CA 941ii; Faxa (415) 274-6080; e-mail

[email protected]. VisitourWebsite atwww.ml.com. Merrill Lynch is an Equal Opportunity Employer.

10-K microfiche; 5 financial institutions; and 17 companies

mentioned in articles with no discussion of an accounting

or disclosure problem. For example, in an article on the

manipulation of earnings at Chambers Development, Flynn

and Zellner (1992) discussed other companies in the waste

management industry, such as Sanifill and Waste Manage-

W Merrill Lynch

ment, without referring to any accounting measurement or

disclosure problems.

5. I treated companies in the same industry for which my

cash-flow-based variables (the cash flow adequacy ratio

searches did not identify an instance of manipulation as

and the cash flow coverage of debt service), (4) signals of

nonmanipulators. Because successful manipulators would

earnings quality documented by Lex and Thiagarajan,

not be identified by the searches, the control sample of 2,332

namely, changes in the receivable provision, changes in

could contain manipulators. The effect would be to bias the

capital expenditures, changes in the effective tax rate,

tests against discriminating between manipulators and

changes in employee productivity, and a variable indicat-

nonmanipulators.

ing whether the firm used LIFO to value its inventory, and

6. Probit analysis is a form of regression analysis appropriate

(5) an index of days' sales in inventory similar to the index

for cases in which the dependent variable is dichotomous

of days' sales in receivables. None of these variables

(e.g., the firm either is or is not a manipulator). Probit

improved the model's performance and are not reported

coefficients are estimated by maximizing a likelihood func-

here.

tion and are indicative of how a particular variable affects

9. The variables were not measured contemporaneously with

the probability of a company being a manipulator. WESML

manipulation discovery because, in line with Feroz, Park,

probit weighs the likelihood function according to the pro-

and Pastena (1991), manipulation becomes public, on aver-

portion of earnings manipulators in the sample and in the

age, 19 months after the end of the fiscal year of the first

population to account for state-based sampling. For a dis-

reporting violation.

cussion of the implications of using state-based samples, see

10. Manipulation of inventories and other production costs

Hsieh, Mansky, and McFadden (1985). Prior research used

could lead to increasing gross margins, which suggests that

weighted probit to predict audit qualifications (Dopuch,

either increased or decreased gross margins can increase the

Holthausen, and Leftwich 1987) and bankruptcy (Zmijew-

likelihood of manipulation. Kellogg and Kellogg stated,

ski 1984).

"Barring unusual circumstances, the percentage of gross

7. Specifically, I used constructs that analysts consider to be

profit to sales should remain the same from year to year"

indicators of future performance. See, for example, O'Glove

(pp. 10-16), but what "the same" means is difficult to deter-

(1987); Kellogg and Kellogg (1991); Siegel (1991); Fridson

mine. I considered a variable relating gross margin changes

(1993); Lev and Thiagarajan (1993).

to inventory changes, but it did not enhance the specifica-

8. To examine whether the model could be improved, I also

tion of the model.

considered, but did not include in the model, five other

11. Part of the increase might be attributable to acquisitions

types of variables: (1) variables to isolate the income effect

involving goodwill, but manipulators undertake few acqui-

of nonrecurring items (the ratio of unusual items to pretax

sitions and those they do undertake are primarily

income [Compustat #17/#170] and the ratio of net nonop-

stock-for-stock exchanges that are accounted for by using

erating income to income before extraordinary items [Com-

pooling of interests. Nevertheless, I also calculated the ratio

pustat #61/#181), (2) variables to capture the rate and

of noncurrent assets other than PP&E and goodwill to total

changes in the rate of intangible amortization and variables

assets and found results similar to those reported here.

that identify the funding status of pension funds, (3)

September/October 1999 35

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions
Financial Analysts Journal

12. To allow for the possibility that companies manipulate 13. I also considered alternative definitions of leverage-total

earnings by using lower depreciation rates than compara- debt to market value of equity, total debt to book value of

ble companies in their industry, I used the depreciation rate

equity, and long-term debt to total assets-as well as using

instead of changes in the depreciation rate to estimate the

level leverage variables instead of changes in variables.

model. This variable did not enhance the specification of the

None of the alternatives attained significance.

model and did not alter the magnitude or the significance

of the coefficients on the other variables.

References

Beneish, M.D., and E. Press. 1993. "Costs of Technical Violation Kellogg, I., and L.B. Kellogg. 1991. Fraud, Window Dressing, aInd

of Accounting-Based Debt Covenants." Accounting Review, vol.

Negligence in Financial Statements. New York: McGraw-Hill.

68, no. 2 (April):233-257.

Khalaf, R. 1992. "Fuzzy Accounting." Forbes (June 22):96.

Beneish, M.D. 1997. "Detecting GAAP Violation: Implications

Lev, B., and S.R. Thiagarajan. 1993. "Fundamental Information

for Assessing Earnings Management among Firms with

Analysis." Journal of Accounting Research, vol. 31, no. 2

Extreme Financial Performance." Journal ofAccountingand Public

(Autumn):190-215.

Policy, vol. 16, no. 3 (Fall):271-309.

Levitt, A. 1998. "The Numbers Game." Securities and Exchange

. Forthcoming 1999. "Incentives and Penalties Related to

Commission, Remarks by Chair Arthur Levitt at New York

Earnings Overstatements That Violate GAAP." Accounting

University Center for Law and Business (September 28).

Reviezv, vol. 74, no. 4 (October).

Available online at www.sec.gov/news /speeches /spch220.txt

Dopuch, N., R.W. Holthausen, and R.W. Leftwich. 1987.

Loebbecke, J.K., M.M. Eining, and J.J. Willingham. 1989.

"Predicting Audit Qualifications with Financial and Market

Variables." Accounting Reviezv, vol. 62, no. 3 (July):431-454. "Auditor's Experience with Material Irregularities: Frequency,

Nature, and Detectability." Auditing, vol. 9, no. 1 (Fall):1-28.

Feroz, E.H., K. Park, and V.S. Pastena. 1991. "The Financial and

Market Effects of the SEC's Accounting and Auditing

Maddala, G.S. 1983. Limited-Dependent and Qualitative Variables

Enforcement Releases." Journal of Accounting Research, vol. 29

in Econometrics. Cambridge, U.K.: Cambridge University Press.

(Supplement):107-148.

National Association of Certified Fraud Examiners. 1993.

Flynn, J., and W. Zellner. 1992. "Buying Trash in Big Holes-On

Cooking the Books: What Every Accoluntanit Should Know about

the Balance Sheet." Business Week (May 11):88-89.

Fraud. New York: NASBA.

Fridson, M.S. 1993. Financial Statement Analysis: A Practitioner's

National Commission on Fraudulent Financial Reporting. 1987.

Guide. New York: John Wiley & Sons.

Report of National Commission on Fraudulent Financial Reporting.

Healy, P.M. 1985. "The Effect of Bonus Schemes on Accounting

New York: AICPA.

Decisions." Journal of Accounting and Economics, vol. 7, no. 1-3

O'Glove, T.L. 1987. Quality of Earnings. New York: The Free

(April):85-107.

Press.

Hector, G. 1989. "Cute Tricks on the Bottom Line." Fortune (April

Siegel, J.G. 1991. How to Analyze Businesses, Financial Statements,

24):193-200.

and the Quality of Earnings. 2nd ed. Englewood Cliffs, NJ:

Hsieh, D.A., C.F. Mansky, and D. McFadden. 1985. "Estimation

Prentice-Hall.

of Response Probabilities from Augmented Retrospective

Watts, R.L., and J.L. Zimmerman. 1986. Positive Accounting

Observations." Journal of the American Statistical Association, vol.

Theory. Englewood Cliffs, NJ: Prentice-Hall.

80 (September):651-662.

Jones, J.J. 1991. "Earnings Management during Import Relief Zmijewski, M.E. 1984. "Methodological Issues Related to the

Investigations." Journal of Accoutnting Research, vol. 29, no. 2 Estimation of Financial Distress Prediction Models." Journal of

(Autumn):193-228. Accounting Research, vol. 22 (Supplement): 59-82.

36 ?Association for Investment Management and Research

This content downloaded from 130.82.1.40 on Sun, 13 Mar 2016 06:53:27 UTC
All use subject to JSTOR Terms and Conditions