Predicting Material Accounting Misstatements*

Patricia M. Dechow
The Haas School of Business
University of California, Berkeley
Berkeley, CA 94705
[email protected]

Weili Ge
Michael G. Foster School of Business
University of Washington
Mackenzie Hall, Box 353200
Seattle, WA 98195
[email protected]

Chad R. Larson
Washington University in St. Louis
Olin Business School
St. Louis, MO 63130
[email protected]

Richard G. Sloan
The Haas School of Business
University of California, Berkeley
Berkeley, CA 94705
[email protected]

First version: January 8, 2007

This version: April 21, 2010

Contemporary Accounting Research, forthcoming

*
We appreciate the comments of the workshop participants at the University of Michigan, the UBCOW Conference
at the University of Washington, New York University 2007 Summer Camp, University of California, Irvine and
University of Colorado at Boulder, Columbia University, University of Oregon, the Penn State 2008 Conference,
University of California, Davis 2008 Conference, American Accounting Association meetings 2007, FARS 2008
meetings, the University of NSW Ball and Brown Conference in Sydney 2008 and the 2009 George Mason
University Conference on Corporate Governance and Fraud Prevention. We thank Ray Ball, Sid Balachandran,
Sandra Chamberlain, Ilia Dichev, Bjorn Jorgensen, Bill Kinney, Carol Marquardt, Mort Pincus, and Charles Shi for
their comments and Seungmin Chee for research assistance. We would like to thank the Research Advisory Board
established by Deloitte & Touche USA LLP, Ernst & Young LLP, KPMG LLP and PricewaterhouseCoopers LLP
for the funding for this project. However, the views expressed in this article and its content are those of the authors
alone and not those of Deloitte & Touche USA LLP, Ernst & Young LLP, KPMG LLP or PricewaterhouseCoopers
LLP. Special thanks go to Roslyn Hooten for administering the funding relationship. This paper is dedicated to the
memory of our colleague, friend and research team member Nader Hafzalla, who was a joy to all who knew him.
 Predicting Material Accounting Misstatements

Abstract
We examine 2,190 SEC Accounting and Auditing Enforcement Releases (AAERs) issued
between 1982 and 2005. We obtain a comprehensive sample of firms that are alleged to have
misstated their financial statements. We examine the characteristics of misstating firms along
five dimensions: accrual quality, financial performance, non-financial measures, off-balance
sheet activities, and market-based measures. We compare misstating firms to themselves
during non-misstatement years and misstating firms to the broader population of all publicly
listed firms. We find that managers appear to be hiding diminishing performance during
misstatement years. We find that accruals are high and that misstating firms have a greater
proportion of assets with valuations that are more subject to managerial discretion. In addition,
the extent of leasing is increasing and there are abnormal reductions in the number of
employees. Misstating firms are raising more financing, have higher price-to-fundamental
ratios, and have strong prior stock price performance. We develop a model to predict
accounting misstatements. The output of this model is a scaled logistic probability that we
term the F-Score, where values greater than one suggest a greater likelihood of a misstatement.

Keywords: earnings quality; accounting misstatement; accrual quality; F-Score

JEL Classifications: G3, M4

1
1. INTRODUCTION

What causes managers to misstate their financial statements? How best can investors,

auditors, financial analysts and regulators detect misstatements? Addressing these questions is

of critical importance to the efficient functioning of capital markets. For an investor it can lead

to improved returns, for an auditor it can mean avoiding costly litigation, for an analyst it can

mean avoiding a damaged reputation, and for a regulator it can lead to enhanced investor

protection and fewer investment debacles. Our research has two objectives. First, we develop a

comprehensive database of financial misstatements. Our objective is to describe this database

and make it broadly available to other researchers to promote research on earnings

misstatements. Second, we analyze the financial characteristics of misstating firms and develop

a model to predict misstatements. The output of this analysis is a scaled probability (F-Score)

that can be used as a red flag or signal of the likelihood of earnings management or

“misstatement.”

We compile our database through a detailed examination of firms that have been subject

to enforcement actions by the Securities and Exchange Commission (SEC) for allegedly

misstating their financial statements. Since 1982, the SEC has issued Accounting and Auditing

Enforcement Releases (AAERs) during or at the conclusion of an investigation against a

company, an auditor, or an officer for alleged accounting and/or auditing misconduct. These

releases provide varying degrees of detail on the nature of the misconduct, the individuals and

entities involved and the effect on the financial statements. We examine the 2,190 AAERs

1
released between 1982 and 2005. Our examination identifies 676 unique firms that have

misstated at least one of their quarterly or annual financial statements.1

Using AAERs as a source to investigate characteristics of firms that manipulate financial

statements has both advantages and disadvantages. The SEC has a limited budget, so it selects

firms for enforcement action where there is strong evidence of manipulation. Firms selected

often have already admitted a “mistake” by restating earnings or having large write-offs (e.g.,

Enron or Xerox); other firms have already been identified by the press or analysts as having

misstated earnings (see Miller 2006); in addition, insider “whistleblowers” often reveal problems

directly to the SEC. Therefore, one advantage of the AAER sample is that researchers can have

a high level of confidence that the SEC has identified manipulating firms (the Type I error rate is

low). However, one disadvantage is that many firms that manipulate earnings are likely to go

unidentified, and a second disadvantage is that there could be selection biases in cases pursued

by the SEC. For example, the SEC may be more likely to pursue cases where stock performance

declines rapidly after the manipulation is revealed, since the identifiable losses to investors are

greater. Selection biases may limit the generalizability of our results to other settings. It is

worth noting, however, that problems with selection bias exist for other samples of

“manipulators” identified by an external source (e.g., shareholder litigation firms, SOX internal

control violation firms, or restatement firms).2 Bias concerns also exist for discretionary accrual

measures (Dechow, Sloan, and Sweeney 1995). Thus selection bias is a general concern when

analyzing the determinants of earnings manipulation and is not unique to AAER firms.

1
Throughout the paper we use the terms earnings management, manipulation, and misstatement interchangeably.
Although fraud is often implied by the SEC’s allegations, we use the term misstatement since firms and managers
typically do not admit or deny guilt with respect to the SEC allegations.
2
Shareholder lawsuit firms are biased towards firms that have had large stock price declines, SOX internal violation
firms are biased towards younger firms with less developed accounting systems, and restatement firms are biased
towards firms that have made a mistake that is not necessarily intentional.

2
 In our tests we focus on variables that can be easily measured from the financial

statements since we want our analysis to be applicable in most settings facing investors,

regulators, or auditors. Our tests focus only on AAER firm-years that have overstated earnings.

We examine (i) accrual quality, (ii) financial performance, (iii) non-financial measures, (iv) off-

balance sheet activities, and (v) market-based measures for identifying misstatements.

We investigate several measures of accrual quality. We examine working capital

accruals and the broader measure of accruals that incorporates long-term net operating assets

(Richardson, Sloan, Soliman and Tuna 2005). We provide an analysis of two specific accruals:

changes in receivables and inventory. These accounts have direct links to revenue recognition

and cost of goods sold, both of which impact gross profit, a key performance metric. We

measure the percent of “soft” assets on the balance sheet (defined as the percent of assets that are

neither cash nor PPE). We predict that the more assets on the balance sheet that are subject to

changes in assumptions and forecasts, the greater the manager’s flexibility to manage short-term

earnings (e.g., Barton and Simko 2002; Richardson et al. 2005). We find that all measures of

accrual quality are unusually high in misstating years relative to the broad population of firms.

We also find that the percent of soft assets is high, which suggests that manipulating firms have

more ability to change and adjust assumptions to influence short-term earnings.

In time-series tests that focus only on misstating firms, we find that the reversal of

accruals is particularly important for detecting the misstatement. We find that in the years prior

to the manipulation, all accrual measures are unusually high and in fact are not significantly

different from those of manipulation years. There are two explanations for this finding. First,

managers are likely to utilize the flexibility within GAAP to report higher accruals and earnings

before resorting to the aggressive manipulation identified by the SEC. Therefore, growing

3
accruals in earlier years is consistent with “within GAAP” earnings management. Second, the

positive accruals in earlier years could reflect an over-investment problem. Managers in

misstating firms could be relaxing credit policies, building up inventory and fixed asset capacity

in anticipation of future growth. When that growth is not realized, managers then resort to the

manipulation identified by the SEC. The two explanations are not mutually exclusive, since a

manager who is optimistic and over-investing is also likely to be optimistic in terms of

assumptions and forecasts that relate to asset values and earnings.

We examine various models of discretionary accruals developed in prior accounting

research including the cross-sectional modified Jones model (Dechow et al. 1995; DeFond and

Jiambalvo 1994), the performance matched discretionary accruals model (Kothari, Leone, and

Wasley 2005), and a “signed” version of the earnings quality metric developed by Dechow and

Dichev (2002). Our results indicate that the residuals from the modified Jones model and the

performance matched Jones model have less power to identify manipulation than unadjusted

accrual measures (i.e., working capital accruals and the broader measure of accruals) or the

“signed” Dechow and Dichev model. This suggests that conventional approaches of controlling

for industry and performance induce considerable estimation error into the estimation of

discretionary accruals.

We examine whether the manipulations occur to hide diminishing firm performance. We

find that returns on assets are generally declining; however, contrary to our initial expectations,

we find that cash sales are increasing during misstatement periods. We failed to anticipate the

cash sales result because we expected firms to boost sales by overstating credit sales. There are

two explanations for the unexpected cash sale result. First, misstating firms tend to be growing

their capital bases and increasing the scale of their business operations. The greater scale of

4
operations should lead to increases in both cash and credit sales. Second, an inspection of the

AAERs reveals that many firms misstate sales through transaction management - for example,

encouraging sales to customers with return provisions that violate the definition of a “sale,”

selling goods to related parties, or forcing goods onto customers at the end of the quarter.

We find that one non-financial measure, abnormal reductions in the number of

employees, is useful in detecting misstatements. This measure is new to the literature and is

measured as year-over-year change in employee headcount less year-over-year change in total

assets. This result can be interpreted in two ways. First, reductions in the number of employees

are likely to occur when there is declining demand for a firm’s product. In addition, cutting

employees directly improves short-run earnings performance by lowering wage expenses.

Second, if physical assets and employees are complements, then a decrease in employees relative

to total assets could signal overstated asset balances.

Our examination of off-balance sheet information focuses on the existence and use of

operating leases and the expected return assumption on plan assets for defined benefit pension

plans. Operating leases can be used to front-load earnings and reduce reported debt. We find

that the use of operating leases is unusually high during misstatement firm-years. In addition,

more firms begin leasing in manipulation years (relative to earlier years). We also find that

misstating firms have higher expected returns on their pension plan assets than other firms. The

effect of higher expected return assumptions is to reduce reported pension expense. The results

for leases and pensions are consistent with misstating firms exhausting ‘legal’ earnings

management options before resorting to more aggressive financial misstatements.

Our final set of variables relates to stock and debt market incentives. Dechow et al.

(1995) suggest that market incentives are an important reason for engaging in earnings

5
management. Teoh, Welch, and Wong (1998) and Rangan (1998) provide corroborating

evidence that accruals are unusually high at the time of equity issuances. However, the evidence

in Beneish (1999b) suggests that leverage and stock issuances do not motivate misstatements.

Therefore, revisiting this question using our more comprehensive data is warranted. We find that

the comparison group is critical for evaluating whether raising financing is a motivation for the

misstatement. Inconsistent with Beneish, we find that misstating firms are actively raising

financing in misstating years relative to the broad population of firms. However, consistent with

Beneish, we find no significant difference in the extent of financing when we compare earlier

years to manipulation years for the same AAER firm. These results can be reconciled by the fact

that we find misstating firms are actively raising financing before and during the manipulation

period. Thus, one interpretation of these findings is that managers of misstating firms are

concerned with obtaining financing and this motivates earnings management in earlier years, as

well as the more aggressive techniques identified by the SEC in misstating years. Also

consistent with Beneish, we do not find evidence that misstating firms tend to have higher

financial leverage than non-misstating firms.

We examine the growth expectations embedded in misstating firms’ stock market

valuations. We find that the price-earnings and market-to-book ratios are unusually high for

misstatement firms compared to other firms, suggesting that investors are optimistic about the

future growth opportunities of these firms. We also find that the misstating firms have unusually

strong stock return performance in the years prior to misstatement. This is consistent with

managers engaging in aggressive techniques in misstating years in the hopes of avoiding

disappointing investors and losing their high valuations (Skinner and Sloan 2000).

6
 Our final tests aim at developing a prediction model that can synthesize the financial

statement variables that we examine and provide insights into which variables are relatively

more useful for detecting misstatements. The model is built in stages based on the ease of

obtaining the information, and compares the characteristics of misstating firm-years to other

public firms. Model 1 includes variables that are obtained from the primary financial statements.

These variables include accrual quality and firm performance. Model 2 adds off-balance sheet

and non-financial measures. Model 3 adds market-related variables. The output of these models

is a scaled logistic probability for each firm-year that we term the F-Score.

We show that while only 20 percent of the public firms have an F-Score greater than 1.4,

over fifty percent of misstating firms have F-Scores of 1.4 or higher. We also investigate the

time-series pattern of F-Scores for misstating firms. We show that average F-Scores for

misstating firms increase for up to three years prior to the misstatement, but decline rapidly to

more normal levels in the years following the misstatement. This is consistent with the F-Score

identifying within-GAAP earnings management as well as the more aggressive techniques

identified by the SEC. We discuss interpretation issues concerning Type I and Type II errors

related to the F-Score, and provide marginal analysis and sensitivity analysis showing that

variation in the F-Score is not driven by one specific variable. We also conduct several

robustness tests that confirm the stability of the variables selected for our models, our coefficient

estimates and the predictive ability of the F-Score over time.

The remainder of the paper is organized as follows. Section 2 reviews previous research

on this topic. Section 3 describes database construction and research design. Section 4 presents

our analysis of misstatement firms and develops our misstatement-prediction model. Section 5

concludes.

7
2. PREVIOUS LITERATURE

Understanding the types of firms that will misstate financial statements is an extensive

area of research. We briefly discuss some of the key findings but do not attempt to document all

literature examining characteristics of AAER firms. Dechow, Ge, and Schrand (2009) provide a

comprehensive review of this literature.

Early work by Feroz, Park, and Pastena (1991) examines 224 AAERs issued between

April 1982 and April 1989 covering 188 firms, of which 58 have stock price information. Feroz

et al. document that receivables and inventory are commonly misstated. Two pioneering papers

analyzing misstating firms are Beneish (1997 and 1999a). Beneish (1997) analyzes 363 AAERs

covering 49 firms and a further 15 firms whose accounting was questioned by the news media

between 1987 and 1993. The 64 firms are classified as manipulators. He creates a separate

sample of firms using the modified Jones model to select firms with high accruals that he terms

“aggressive accruers”. His objective is to distinguish the manipulators from the aggressive

accruers. Beneish (1997) finds that accruals, day’s sales in receivables and prior performance

are important for explaining the differences between the two groups. Beneish (1999a) matches

the sample of manipulators to 2,332 Compustat non-manipulators by two-digit SIC industry and

year for which the financial statement data used in the model were available. For seven of the

eight financial statement ratios that he analyzes, he calculates an index, with higher index values

indicating a higher likelihood of an earnings overstatement. Beneish shows that the day’s sales

in receivables index, gross margin index, asset quality index, sales growth index, and accruals

(measured as the change in non-cash working capital plus depreciation) are important. He

provides a probit model and analyzes the probability cutoffs that minimize the expected costs of

misstatements.

8
 Our research builds on and is complementary to Beneish (1997, 1999a). We take a

different perspective from Beneish that leads us to make a number of different choices.

However, such differences should not be viewed as a critique of his approach; rather, they stem

from our objectives. One of our objectives is to develop a measure that can be directly calculated

from the financial statements. Therefore, we do not use indexes for any of our variables. A

second objective is to enable researchers and practitioners to calculate an F-Score for a random

firm and to easily assess the probability of misstatement. Therefore, we do not match AAER

firms to a control group by industry or size. Matching by industry and size provides information

on whether a variable is significantly different relative to a control firm. However, it is more

difficult when matching to determine Type I and Type II error rates that users will face in an

unconditional setting. Models could be developed for individual industries and size categories.

We choose not to do this since it would add greatly to the complexity of our analysis and the

presentation of our results. A third objective is to evaluate the usefulness of financial statement

information beyond that contained in the primary financial statements; therefore we include other

information disclosed in the 10-K either in Item 1 (discussion of the business), Item 5 (stock

price information), or the footnotes.

Concurrent research provides additional insights into variables that are useful for

detecting misstatements. Ettredge, Sun, Lee, and Anandarajan (2006) examine 169 AAER firms

matched by firm size, industry and whether the firm reported a loss. They find that deferred

taxes can be useful for predicting misstatements, along with auditor change, market-to-book,

revenue growth and whether the firm is an OTC firm. Brazel, Jones, and Zimbelman (2009)

examine whether several non-financial measures (e.g., number of patents, employees, and

products) can be used to predict misstatement in 50 AAER firms. They find that growth rates

9
between financial and non-financial variables are significantly different for AAER firms. Bayley

and Taylor (2007) study 129 AAER firms and a matched sample based on industry, firm size and

time period. They find that total accruals are better than various measures of unexpected

accruals in identifying material accounting misstatements. In addition, they find that various

financial statement ratio indices are incrementally useful. They conclude that future earnings

management research should move away from further refinements of discretionary accrual

models and instead consider supplementing accruals with other financial statement ratios. We

agree with Bayley and Taylor and view our work as moving in the direction that they

recommend.

There has also been work using AAER firms to examine the role of corporate governance

and incentive compensation in encouraging earnings manipulation (see, for example, Dechow,

Sloan and Sweeney (1996), Beasley (1996), Farber (2005), Skousen and Wright (2006) and, for a

summary, Dechow et al. (2009)). We chose not to investigate the role of governance variables

and compensation because these variables are available for only limited samples or must be hand

collected. Therefore, adding these variables would have limited our analysis to a smaller sample

with various biases in terms of data availability. However, a useful avenue for future research is

to analyze the role of governance, compensation, insider trading, short-selling, incentives to meet

and beat analyst forecasts, etc., and to determine the relative importance of these variables over

financial statement information in detecting overstatements of earnings.

3. DATA AND SAMPLE FORMATION

3.1 Data Sample

The objective of our data collection efforts is to construct a comprehensive sample of

material and economically significant accounting misstatements involving both GAAP violations

10
and the allegation that the misstatement was made with the intent of misleading investors. Thus

we focus our data collection on the SEC’s series of published Accounting and Auditing

Enforcement Releases.3

The SEC takes enforcement actions against firms, managers, auditors and other parties

involved in violations of SEC and federal rules. At the completion of a significant investigation

involving accounting and auditing issues, the SEC issues an Accounting and Auditing

Enforcement Release (AAER). The SEC identifies firms for review through anonymous tips and

news reports. Another source is the voluntary restatement of the financial results by the firm

itself, since restatements are viewed as a red flag by the SEC. The SEC also states that it reviews

about one-third of public companies’ financial statements each year and checks for compliance

with GAAP. If SEC officials believe that reported numbers are inconsistent with GAAP, then

the SEC can initiate informal inquiries and solicit additional information. If the SEC is satisfied

after such informal inquiries, then it will drop the case. However, if the SEC believes that one or

more parties violated securities laws, then the SEC can take further steps, including enforcement

actions requiring the firm to change its accounting methods, restate financial statements, and pay

damages.

There are a number of conceivable alternative sources for identifying accounting

misstatements. They are discussed briefly below, along with our reasons for not pursuing these

alternatives.

3
The Accounting and Auditing Enforcement Release series began on May 17, 1982 with the SEC’s issuance of
AAER No. 1. The SEC states in the first AAER that the series would include “future…enforcement actions
involving accountants” and “enable interested persons to easily distinguish enforcement releases involving
accountants from other Commission releases” (AAER No 1). Although the AAERs often directly involve
accountants, the AAER series also includes enforcement actions against non-accountant employees that result from
accounting misstatements and manipulations.

11
1. The GAO Financial Statement Restatement Database. This database consists of

approximately 2309 restatements between January 1997 and September 2005. This database was

constructed through a Lexis-Nexis text search of press releases and other media coverage based

on variations of the word ‘restate.’ There is some overlap between the AAER firms and the

GAO restatement firms since a) the SEC often requires firms to restate their financials as part of

a settlement, and b) restatements often trigger SEC investigations. The GAO database covers a

relatively small time period but consists of a relatively large number of restatements. The reason

for the large number of restatements is that the GAO database includes all restatements relating

to accounting irregularities regardless of managerial intent, materiality and economic

significance. Consequently, it includes a large number of economically insignificant

restatements. In addition, the results in Plumlee and Yohn (2008) suggest that many

restatements are a consequence of misinterpreting accounting rules rather than intentional

misstatements. Another shortcoming of the GAO database is that it specifies only the year in

which the restatement was identified in the press and not the reporting periods that were required

to be restated.4

2. Stanford Law Database on Shareholder Lawsuits. Shareholder lawsuits typically result

from material intentional misstatements. However, shareholder lawsuits can also arise for a

number of other reasons that are unrelated to financial misstatements. Shareholder lawsuits

alleging misstatements are also very common after a stock has experienced a precipitous price

decline, even when there is no clear evidence supporting the allegation. In contrast, the SEC

issues an enforcement action only when it has established intent or gross negligence on the part

of management in making the misstatement.

4
For example, while Xerox is included in the GAO database in 2002, the restatements in question relate to Xerox’s
financial statements for 1997, 1998, 1999, 2000 and 2001.

12
 Using the SEC’s Accounting and Auditing Enforcement Releases as a sample of

misstatement firms has several advantages relative to other potential samples. First, the use of

AAERs as a proxy for manipulation is a straightforward and consistent methodology. This

methodology avoids potential biases induced in samples based on researchers’ individual

classification schemes and can be easily replicated by other researchers. Second, AAERs are

also likely to capture a group of economically significant manipulations as the SEC has limited

resources and likely pursues the most important cases. Relative to other methods of identifying a

sample of firms with managed earnings, such as the modified Jones abnormal accruals model,

using misstatements identified in AAERs as an indicator is expected to generate a much lower

Type I error.

Despite the advantages of using AAERs to identify accounting misstatements, there are

caveats. We can investigate only those firms identified by the SEC as having misstated earnings.

The inclusion of the misstatements that are not identified by the SEC in our control sample is

likely to reduce the predictive ability of our model. Therefore, our analyses can be interpreted as

joint tests of engaging in an accounting misstatement and receiving an enforcement action from

the SEC. If it is assumed that the SEC selection criteria are highly correlated with our prediction

variables, then another criticism is that identified variables could reflect SEC selection.

However, as noted above, the SEC identifies firms from a variety of sources and not just from its

own internal reviews, and many cases are brought to its attention because the firm itself either

restates or take a large write-off. Thus, selection choices are unlikely to be a complete

explanation for our findings. In addition, from a firm’s perspective, being subject to an SEC

enforcement action brings significantly negative capital market consequences (Dechow et al.,

13
1996; Karpoff et al., 2008). Therefore, avoiding these characteristics could be useful and thus

affect firm and market behavior.

3.2 Datasets

We catalog all the AAERs from AAER 1 through AAER 2261 spanning May 17th, 1982

through June 10th, 2005. We next identify all firms that are alleged to have violated GAAP by at

least one of these AAERs (we describe this procedure in more detail in the next section). We

then create three data files: the Detail, Annual and Quarterly files. The Detail file contains all

AAER numbers pertaining to each firm, firm identifiers, a description of the reason the AAER

was issued, and indicator variables categorizing which balance sheet and income statement

accounts were identified in the AAER as being affected by the violation. There is only one

observation per firm in the Detail file. The Annual and Quarterly files are compiled from the

Detail file and are formatted by reporting period so that each quarter or year affected by the

violation is a separate observation. Appendix 1 lists the variable names and description for each

file in the database.

3.3 Data Collection

The original AAERs are the starting point for collecting data. Copies of the AAERs are

obtained from the SEC website and the LexisNexis database. Each AAER is separately

examined to identify whether it involves an alleged GAAP violation. In cases where a GAAP

violation is involved, the reporting periods that were alleged to be misstated are identified.

The data coding was completed in three phases. In the first phase, all releases were read

in order to obtain the company name and period(s) in which the violation took place. The

AAERs are simply listed chronologically based on the progress of SEC investigations. To

14
facilitate our empirical analysis, we record misstatements by firm and link them back to their

underlying AAERs in the detail file. Note that multiple AAERs may pertain to a single set of

restatements at a single firm. Panel A of Table 1 indicates that we are unable to locate 30 of the

2,261 AAERs, because they were either missing or not released by the SEC. A further 41

AAERs relate to auditors or other parties and do not mention specific company names. This

leaves us with 2,190 AAERs mentioning a company name.

Panel B of Table 1 reports that in the 2,190 AAERs, the SEC takes action against 2,614

different parties. Note that one AAER can be issued against multiple parties. In 49.2 percent

(1,077) of the cases the party was an officer of the company (e.g., CEO or CFO), in 15.1 percent

(331) of the cases both an officer and the company were charged by the SEC, in 14.1 percent

(308) of cases the party was the firm itself, in a further 15.9 percent (348) of cases the party was

an auditor, in 3.1 percent (68) the party was a combination of various parties (e.g., auditor and

officer), and in 2.65 percent (58) cases the party was classified as “other,” which includes

consultants and investment bankers.

Table 1 Panel C provides the distribution of the 2,190 AAERs across years based on the

AAER release date. Relatively few AAERs were released prior to 1990. However, the number

of AAERs increased particularly after 2000, when over one hundred AAERs were released per

year. The number of AAERs in 2005 falls to 94 because our sample cutoff date is June 10th,

2005, so our sample does not include the full year. Table 1 Panel D reports that in many cases

there are multiple AAERs referring to the same firm. This is because the SEC can take action

against multiple officers as well as the firm itself. The number of releases ranges from one per

firm (370 firms) to a high of 24 per firm (Enron). From our reading of the AAERs we obtain a

list of 896 firms mentioned in the 2,190 releases.

15
 In phase two, we created the Annual and Quarterly files. All releases were reread in

order to identify the year and/or quarter-end when the misstatements occurred. Panel E of Table

1 indicates that of the 896 original firms identified, 220 firms involved either wrongdoing

unrelated to financial misstatements (such as bribes or disclosure related issues) or financial

misstatements that were not linked to specific reporting periods. This leaves us with 676 firms

with alleged financial misstatements. We lose a further 132 firms because we are unable to

obtain a valid Cusip identifier.5 For each firm that is in the Detail file but excluded from both the

Annual or Quarterly files, we create indicator variables in the Detail file to categorize why it was

excluded. Panel E of Table 1 indicates that for 544 firms, the misstatement involved one or

more quarters. We provide the number of firms with assets and share price data since firms can

have a Cusip but no data. In 92 firms the misstatement involved only quarterly financial

statements and was corrected by the end of the year. Therefore the annual file contains

misstatements of annual data for 451 firms. Among these 451 firms, 387 firms have total assets

listed on Compustat during the misstatement period.

For each annual/quarterly period that was misstated, an additional field was added to the

Annual/Quarterly file. If an understatement of earnings or revenues occurred during the quarter

or year of the violation, we code the understatement variable 1. Since most AAERs involve the

overstatement of earnings or revenues, this flag is helpful in conducting earnings management

and other discretionary accruals tests. In our empirical analyses in Tables 4-8, we delete firm-

year observations that understated earnings. We also exclude banks and insurance companies

since many accruals related variables are not available for these firms. The Annual file contains

5
Further investigation revealed that, among these 132 firms, 33 were traded on non-major exchanges or over the
counter but had no Cusip, 12 were IPO firms that never went public, 12 were sanctioned when registering securities
under 12 (g), and 13 were subsidiaries of parent firms already included in the sample or private companies that
helped a public company commit the misstatement. The rest of the firms are brokerage firms, have unregistered
securities traded, or simply do not have sufficient detail to identify a Cusip.

16
837 firm-year observations with Cusips, and the Quarterly file contains 3,612 firm-quarter

observations with Cusips.

Phase three involves reading the AAERs a final time in order to obtain additional details

on the misstatements. For each firm, we summarize the reason(s) for the enforcement action(s)

in one or two sentences in the “explanation” column of the Detail file. We then create eleven

indicator variables to code the balance sheet and income statement accounts that the AAER

identified as being affected by the misstatements. Table 1 Panel F reports the frequency of

misstatement accounts for various samples based on available data. The patterns are quite

similar across the four samples. For example, Column (2) indicates that 770 accounts were

affected across the 435 misstating firms that have stock price data. Most misstatements relate to

revenue recognition, which occurs in 59.5 percent of firms. Types of revenue misstatements

include the following: front-loading sales from future quarters (e.g., Coca Cola, Computer

Associates), creating fictitious sales (e.g., ZZZZ Best), incorrect recognition of barter

arrangements (e.g., Qwest), shipping goods without customer authorization (e.g., Florafax

International). Revenue misstatements also frequently involve a misstatement of the allowance

for doubtful debts. Other accounts frequently affected by misstatements include costs of goods

sold and inventory (13.1 percent and 14.5 percent, respectively). Other types of misstatements

include capitalizing expenses or creating fictitious assets (e.g., WorldCom). This occurs in about

20% of the firms. The AAERs do not provide consistent information on the magnitude of the

misstatements. Therefore, there is insufficient detail to provide a consistent analysis of the

magnitude of the misstatements.

[Table 1]

4. EMPIRICAL RESULTS

17
 Our empirical results first discuss the characteristics of misstatement firms. We then

develop our logistic model and associated F-Score.

4. 1 Characteristics of Misstating Firms

Table 2 Panel A presents information on size for misstating firms. To calculate size

deciles, we rank firms based on their market capitalization of equity in each fiscal year. We then

determine the decile rankings of misstating firms in their first misstatement year. The results in

bold identify the size deciles that are overrepresented in the misstatement firm population. The

results indicate that 14.7 percent of firms that misstate their earnings are from the top size decile

(decile 10). There are several explanations for why larger firms appear to be relatively more

likely to misstate their earnings. First, large firms have greater investor recognition and are

under more scrutiny by the press and analysts; therefore, when an account appears suspicious

there is likely to be more commentary that alerts the SEC to a potential problem (analyst and

press reports are potential triggers for an SEC investigation). Second, the SEC is likely to review

large firms on a more regular basis than other firms, so misstatements are more likely to be

identified. Note also that only 5.1 percent of misstating firms are in decile 1. Recall that 132

firms are excluded from our analysis because we could not obtain their firm identifier. These are

likely to be smaller firms.

Panel B of Table 2 reports the industry distribution of both misstatement firm-years and

all available firm-years on Compustat. We follow Frankel, Johnson, and Nelson’s (2002) SIC-

based industry classification scheme. The bolded results highlight industries that are

significantly overrepresented for misstating firms. Over twenty percent of misstating firms are in

the computer industry, whereas only 11.1 percent of firms in the general population are in this

industry. The computer industry includes software and hardware manufacturers. This industry

18
is relatively new and has exhibited substantial growth. It is also characterized by substantial

investment in intangible assets. Misstating firms frequently overstate their sales to meet

optimistic business expectations (e.g., Computer Associates), ship goods without authorization

(e.g., Information Management Technologies Corp.), or create fictitious sales (e.g., Clarent

Corporation and AremisSoft Corporation). Retail is also overrepresented among misstating

firms (12.9% versus 9.9%). Examples of retail firms in our sample include Crazy Eddie, Kmart,

and Rite Aid. Services are also overrepresented (12.7% versus 10.4%). Examples of service

firms include Tyco International, ZZZZ Best, Healthsouth Corporation, Future Healthcare Inc.,

and Rent-Way, Inc. These firms typically capitalized expenses as assets and misstated sales.

Note also that the SEC could systematically review more firms from growth industries and so

identify a relatively greater proportion of manipulators in those industries.

Panel C of Table 2 provides the distribution of misstatements over calendar time.

AAERs are not timely and are often released several years after the manipulation takes place.

Our sample covers misstatements in fiscal years beginning in 1971 and ending in 2003. The

years 1999 and 2000 have by far the most misstatements (7.89% and 7.17% respectively). This

may be because growth in technology stocks slowed around this time, providing incentives for

managers to misstate earnings in order to mask declining performance.

[Table 2]

4. 2 Variables Analyzed

In this section we discuss the motivation and the selection of the financial statement

variables that we hypothesize to be associated with misstatements. Each variable is briefly

discussed, with more detailed definitions provided in Table 3. The variables we analyze focus

19
on accrual quality, financial performance, non-financial performance, off-balance sheet variables

and stock market performance.

Accrual Quality
Starting with Healy (1985), a large body of literature hypothesizes that earnings are

primarily misstated via the accrual component of earnings. We therefore investigate whether

misstatement years are associated with unusually high accruals. The first measure, termed WC

accruals, focuses on working capital accruals and is described in Allen, Larson and Sloan

(2009). Prior research typically includes depreciation expense as part of working capital

accruals. We exclude depreciation because, as discussed by Barton and Simko (2002), managing

earnings through depreciation is more transparent because firms are required to disclose the

effects of changes in depreciation policies (Beneish 1998). Our next measure, which we term

RSST accruals, is from Richardson, Sloan, Soliman, and Tuna (2005). This measure extends the

definition of WC accruals to include changes in long-term operating assets and long-term

operating liabilities. This measure is equal to the change in non-cash net operating assets. We

also examine two accrual components. The first is change in receivables. Misstatement of this

account improves sales growth, a metric closely followed by investors. The second is change in

inventory. Misstatement of this account improves gross margin, another metric closely followed

by investors.

We also examine the % Soft assets. This is defined as the percent of assets on the

balance sheet that are neither cash nor PP&E. Barton and Simko (2002) provide evidence

consistent with firms with greater net operating assets having more accounting flexibility to

report positive earnings surprises. In their Table 5 they decompose net operating assets into

working capital assets, long-term assets, and other assets. Their results suggest that the level of

20
working capital has a much stronger effect (the coefficient is 9 to 28 times larger) on the odds of

reporting a predetermined earnings surprise than on the level of PP&E. We therefore assume

that when firms have more soft assets on their balance, there is more discretion for management

to change assumptions to meet short-term earnings goals.6

We examine several ‘discretionary accrual’ models commonly used in the accounting

literature. Our comprehensive sample of misstatements provides a unique opportunity to

investigate whether these models enhance the ability to detect earnings misstatements. First, we

employ the cross-sectional version of the Modified Jones model discretionary accruals (see

Dechow et al. (1996) for modified Jones model, and DeFond and Jiambalvo (1994) for the cross-

sectional version). We also investigate the effect of adjusting discretionary accruals for financial

performance as suggested in Kothari et al. (2005). We term this Performance-matched

discretionary accruals. Finally, we employ two variations of the accrual quality measure

described in Dechow and Dichev (2002). The Dechow and Dichev measure is based on the

residuals obtained from industry-level regressions of working capital accruals on past, present,

and future operating cash flows. Our first variation on this measure takes the absolute value of

each residual and subtracts the average absolute value of the residuals for each industry. We

term this the mean-adjusted absolute value of DD residuals. Our second variation scales each

residual by its standard error from the industry-level regression. This measure leaves the sign of

the residual intact and provides information on how many standard deviations the residual is

above or below the regression line. We term this variable the Studentized DD residuals. We

predict a positive association between all accrual variables and misstatement years.

6
PP&E is subject to discretion in the sense that managers can over-capitalize costs and delay write-offs. Changes in
the level of PPE that reflect such adjustments and choices will be reflected in the RSST accrual measure.

21
Performance

Our next set of variables gauges the firm’s financial performance on various dimensions

and examines whether managers misstate their financial statements to mask deteriorating

performance (Dechow et al. 1996; Beneish 1997, 1999b). The first variable we analyze is change

in cash sales. This measure excludes accruals-based sales, such as credit sales, and we use it to

evaluate whether sales that are not subject to accruals management are declining. We also

analyze change in cash margin. Cash margin is equal to cash sales less cash cost of goods sold.

This performance measure abstracts from receivable and inventory misstatements. We anticipate

that when cash margins decline, managers are more likely to make up for the decline by boosting

accruals. Change in return on assets is also analyzed since managers appear to prefer to show

positive growth in earnings (Graham, Harvey and Rajgopal 2005). Therefore, during

misstatement periods managers could be attempting to provide positive increases in earnings.

Change in free cash flows is a more fundamental measure than earnings since it abstracts from

accruals. We predict that managers are more likely to misstate when there is a decrease in free

cash flows. We also investigate whether deferred tax expense increases during misstatement

periods. Larger accounting income relative to taxable income is reflected in the deferred tax

expense and could indicate more misstatement of book income (Phillips, Pincus, and Olhoft-

Rego 2003).

Non-financial Measures

Economics teaches us that a firm trades off the marginal cost of labor against the

marginal cost of capital to maximize profits. Investments in both labor and capital should lead to

increases in future sales and profitability. However, unlike capital expenditures, most

expenditure on labor must be expensed as incurred (the primary exception being direct labor that

22
is capitalized in inventory). We therefore conjecture that managers attempting to mask

deteriorating financial performance will reduce employee headcount in order to boost the bottom

line. Moreover, if managers are overstating assets, then the difference between the change in the

number of employees (which is not likely overstated) and the change in assets (which is

overstated) might be a useful measure of the underlying economic reality. Brazel et al. (2009)

make a similar argument for the use of non-financial measures for detecting misstatements. In

their discussion of Del Global Technologies they state, “it is improbable that the company would

double in profitability while laying off employees, and it is even less probable that employee

layoffs would correspond with a significant increase in revenue.” We measure abnormal change

in employees as the percentage change in the number of employees less the percentage change in

total assets. We predict a negative association between abnormal change in employees and

misstatements.

Greater order backlog is indicative of higher future sales and earnings (Rajgapol, Shevlin,

and Venkatachalam 2003). When a firm exhibits a decline in order backlog, this suggests a

slowing demand and lower future sales. We measure abnormal change in order backlog as the

percentage change in order backlog less percentage change in sales. We predict a negative

association between abnormal change in order backlog and misstatements.

Off-Balance Sheet Activities

The most prevalent source of off-balance sheet financing is operating leases. The

accounting for operating leases allows firms to record lower expenses early on in the life of the

lease (because the interest charge implicit in capital lease accounting is higher earlier in the life

of the lease). Therefore, the use of operating leases (existence of operating leases) and unusual

increases in operating lease activity (change in operating lease activity) could be indicative of

23
managers who are focused on financial statement window-dressing. We predict that change in

operating lease activity is positively associated with misstatements. Change in operating lease

activity is measured as the change in the present value of future non-cancelable operating lease

obligations following Ge (2007).

Another off-balance sheet activity is the accounting for pension obligations and related

plan assets for defined benefit plans. Firms have considerable flexibility on the assumptions that

determine pension expense. The expected return on plan assets is an assumption that is relatively

easy for managers to adjust. Management can increase the expected return on plan assets and so

reduce future reported pension expense. Comprix and Mueller (2006) provide evidence that such

income-increasing adjustments are not filtered out of CEO compensation. Therefore, similar to

leases, such adjustments could be indicative of managers who are focused on financial statement

window-dressing. For the subset of firms that have defined benefit plans, we obtain the expected

return on pension plan assets and calculate the change in expected return on pension plan assets.

We predict that in misstatement years, firms will assume larger expected returns on their plan

assets.

Market-related Incentives

One incentive for misstating earnings is to maintain a high stock price. We investigate

whether managers who misstate their financial statements are particularly concerned with a high

stock price. We examine two motivations:

First, if the firm needs to raise cash to finance its ongoing operations and growth plans,

then a high stock price will reduce the cost of raising new equity. We use four empirical

constructs to capture a firm’s need to raise additional capital. (i) We use an indicator variable

identifying whether the firm has issued new debt or equity during the misstatement period

24
(actual issuance). (ii) We identify the net amount of new financing raised, deflated by total

assets (CFF). (iii) We construct a measure of ex ante financing need. Some firms may have

wished to raise new capital but did not because they were unable to secure favorable terms; our

ex ante measure of financing need provides a measure of the incentive to raise new capital.

Following Dechow et al. (1996) we report an indicator variable that equals one if the firm is

estimated to have negative free cash flows over the next two years that exceed its current asset

balance. (iv) We expect that managers of firms with higher leverage will have incentives to

boost financial performance both to satisfy financial covenants in existing debt contracts and to

raise new debt on more favorable terms (leverage).

A second motivation for why managers may be particularly dependent on a high stock

price is that a significant portion of management compensation is typically tied to stock price

performance. This can cause managers to become overly concerned with maintaining or

increasing their firm’s stock price, since it affects their wealth. Such managers can become

focused on managing ‘expectations’ rather than managing the business. We expect that

managers whose firms have had large run-ups in their stock prices and have high prices relative

to fundamentals are more prone to ‘expectations’ management. Managers of such firms are

predicted to be more likely to misstate earnings to hide diminishing performance. We identify

firms with optimistic expectations built into their stock prices using market-adjusted stock

return, earnings-to-price, and book-to-market.

[Table 3]

4.3 Time-series Analysis of Misstating Firms

In section 4.3.1 we analyze misstating firms and compare years that are identified as

misstated by the SEC to all non-misstatement years. We calculate means at the firm level for

25
misstatement years versus all non-misstatement years and conduct tests of pair-wise differences

in means. Therefore, each firm is directly compared to itself during manipulation years versus

other years. This approach reduces the number of observations used to calculate t-statistics and

could lower the power of our tests, but its advantage is that it accurately weights the observations

used to calculate means. In section 4.3.2 we follow the same approach and compare

misstatement years to years prior to the misstatement. This analysis provides insights into the

predictive nature of the variables analyzed, since hindsight does not affect the calculations. One

issue of concern for the power of our tests is the proportion of firms that end up restating their

financial statements and filing an amended 10-K (versus taking a write-down and/or reporting

the restated numbers for prior years in future 10Ks). According to discussions with Standard &

Poor’s, COMPUSTAT will backfill misstated numbers when a company files an amended 10-K.

In order to determine whether filing amended 10Ks is common among manipulation firms, we

randomly select nine companies that provide detailed information on misstated numbers in 2000

and 2001. We find that only one of the nine firms' financial data on COMPUSTAT has been

backfilled with restated numbers. In addition, many of the misstatements are discovered and

revealed more than a year after they occur, and so firms are less likely to file amended financial

statements. Thus backfilling, although a concern for the power of our tests, does not appear to be

highly prevalent in the sample.

4.3.1 Comparisons using all available years

Table 4 provides results for our comparisons of misstating years versus other non-

misstating years. We predict and find that accruals are larger in misstatement years. The results

indicate that WC accruals has a slightly larger t-statistic than the RRST accruals measure. The

reason is likely to be the greater weighting of receivables in WC accruals. In Table 1 we

26
document that almost half of the misstating firms are alleged to have misstated sales, and Change

in receivables has the highest t-statistic of all accrual variables (6.12). The % Soft assets is also

significantly different, suggesting more reporting flexibility in misstating years.

The next set of accrual variables relates to various models of accruals. The objective of

these models is to provide more powerful measures of earnings management by controlling for

“nondiscretionary” or “normal” accruals that are required under GAAP. However, interestingly,

the t-statistic on the Modified Jones discretionary accrual is lower than that on either the WC

accruals or RSST accruals, and performance-matched discretionary accruals is lower still. This

suggests that estimating the modified Jones model at an industry level appears to add greater

error to estimates of discretionary accruals. Performance adjusting appears to add further error.

The next measure is a variant of the Dechow and Dichev measure of quality. This measure uses

the absolute value of residuals and is therefore unsigned. However, our sample consists only of

income increasing earnings manipulations, which are likely to be reversed in future years.

Therefore, the use of the absolute value reduces the power of the test because the reversal is also

likely to have a high absolute value (the t-statistic is 1.44). Our next measure, signed Studentized

DD residuals, does not suffer from this problem. The residuals are almost ten times larger in

manipulation years than in non-manipulation years. This measure has a t-statistic of 5.92 and is

second only to receivables in terms of statistical significance.

We next examine measures of financial performance. We predict that misstatements are

made to mask deteriorating financial performance. Our first measure is change in cash sales.

Contrary to our expectations, cash sales significantly increase (rather than decline) during

misstatement years. A reading of the AAERs helps to explain why. We find that many firms

engage in transactions-based earnings management. That is, they front-load their sales and

27
engage in unusual transactions at the end of the quarter (e.g., Coca Cola, Sunbeam, Computer

Associates). Cash sales increase with this type of misstatement, providing an explanation for the

finding. Management of cash sales could also play a role in the low power of the modified Jones

Model. If cash sale manipulation is positively correlated with other types of accrual

manipulation, then the modified Jones model could incorrectly classify part of this discretion as

non-discretionary. The other performance measures, change in cash margin, return on assets,

and free cash flows, are all statistically insignificant in our time-series tests. Deferred tax

expense is also not significantly different. For a small sample of 27 firms subject to SEC

enforcement actions, Erickson, Hanlon, and Maydew (2004) show that firms pay substantial

taxes on overstated earnings. For example, misstating cash sales boosts both accounting and tax

income. If their findings generalize, then this could explain why deferred taxes are not unusually

high during misstatement years.

We next examine non-financial variables, abnormal change in employees and abnormal

change in order backlog. Both variables show insignificant changes during misstatement years.

For our off-balance sheet variables, we find an increase in both the magnitude of operating lease

commitments and the percentage of firms that use operating leases during misstatement years. It

appears that misstating firms exploit the financial reporting flexibility afforded by operating

leases. For defined benefit pension plans we have only a small sample size. We find that both

the expected return on pension plan assets and change in expected return on pension plan assets

are significantly greater in misstatement years.

The final set of variables relates to market incentives. As predicted, we find that during

misstating years more firms are issuing either debt or equity (93.8% versus 86.9%) and cash

from financing is significantly higher (24.7% versus 15.1%). We argue that incentives to

28
misstate are higher during issuing periods. However, the SEC is probably more likely to perform

a review when a firm is raising capital, and hence detect the misstatement. Therefore, it is

possible that SEC bias could also explain this result. Both Leverage and Ex ante financing need

are insignificantly different for misstatement years. Market-adjusted stock return is higher

during misstatement years (23.4% versus -0.4%). We analyze this finding in more detail in

Figure 3, discussed in the next section. Book to market ratios are in the opposite direction from

what we predicted while earnings to price ratios are lower in misstating periods, consistent with

our prediction that misstating firms have optimistic future earnings growth expectations built into

their prices.

In summary, the results are generally consistent with our expectations. Misstating firms

have high accruals, show declining performance, are raising financing, and have high growth

expectations embedded in their stock prices.

[Table 4]

4.3.2 Comparisons using only years before the misstatement

Using all firm-years, as in Table 4, provides more powerful tests because we capture

information in accrual reversals and declines in performance, as well as changes in market

expectations in the years after the misstatements. In many settings such information is readily

available (e.g., for a researcher or regulator). However, it is important also to determine the

predictive nature of the variables. Table 5 replicates the analysis in Table 4 but uses only the

years prior to the misstatement as non-misstatement years. In Table 5 we do not report the

results for accrual models. The modified Jones model pools across firms in each industry and

over time and so is not typically used by researchers in a predictive setting. The Dechow and

29
Dichev model uses future cash flows in the regression and so is not applicable to a predictive

setting.7

The first thing to note in Table 5 is that in contrast to Table 4, the accrual variables are

not significant; in fact, RSST accruals is the wrong sign and significant. This suggests that in

years prior to the manipulation, accruals are the same as or higher than in the manipulation years

identified by the SEC. Only the % Soft asset is significantly higher in misstating years relative to

prior years, suggesting a build-up of assets whose values are more subject to manipulation in the

misstating years.

There are several potential explanations for the lack of accrual results. First, managers

could use “within GAAP” accounting flexibility to overstate earnings before resorting to the

accounting techniques viewed as “outside” of GAAP by the SEC. Thus the high prior accruals

are indicative of more estimation error and reliability concerns (e.g., Richardson et al. 2005).

Second, the high prior accruals could reflect manipulation that was too difficult for the SEC to

identify or prove and so was not reported in the AAER. Third, the high prior accruals could

indicate over-investment by managers. Many of the AAER firms are raising financing, and both

managers and investors could believe that greater investment will lead to more growth.

Managers could therefore be optimistic about the value of inventory, credit sales, PPE, and other

assets. However, when growth slows, managers may not wish to reveal the decline in sales or

their over-investment and so resort to the aggressive accounting techniques identified by the

SEC.

Turning to the performance variables, we see that most variables are not significantly

different at conventional levels. The two exceptions are return on assets, which, as in Table 4, is

7
Both models could be adjusted in various ways to make them predictive. We chose not to do this since it would
detract from the main focus of our paper (i.e., the use of variables easily identified in the financial statements).

30
the wrong sign (i.e., manipulating firms’ ROA is lower in misstating years) and deferred tax

expense (also the wrong sign). Turning to the non-financial measures, we see that Abnormal

change in employees is higher in manipulation years. We predicted that employee growth would

be slower than asset growth (due to the inflation of assets) in manipulation years. Thus the

results are inconsistent with our expectations. Recall that net operating assets are increasing at a

faster rate in earlier years (RSST accruals is the wrong sign), so the inconsistent results could be

due to the change in assets rather than change in employees. For off-balance sheet variables, the

proportion of firms with operating leases is higher in manipulation years than in earlier years

(0.857 versus 0.685), consistent with firms switching to off-balance sheet assets. Operating

leases tend to have lower expenses than capitalizing the lease, early in the asset’s life, so

switching to leasing will improve earnings in growing firms. For market-related variables, we

find that the need for financing (ex ante financing need) is higher in earlier years, while Actual

Issuance is higher in manipulation years. This suggests that the firm needed to raise capital in

earlier years and did so (CFF is also higher in earlier years) but that more firms actually issued

debt or equity financing in manipulation years. This is consistent with managers manipulating

the financial statements to obtain the financing. Market to book ratios and stock price

performance are similar in earlier years and manipulation years. However, the earnings-to-price

ratio is lower in manipulation years, consistent with managers being able to maintain investor

expectations during manipulation years even though earnings declined.

Figure 1 provides a graphical timeline of annual market-adjusted stock returns for

misstating firms before and after the misstatement years. For the firms misstating for multiple

years, we take the average of their stock returns during the misstatement period. The graph

reveals that returns are increasing in the three years leading up to the misstatement. In the

31
misstatement years, on average, the firms are able to maintain positive stock returns. However,

in the first year after the misstatement years, the stock prices decline and returns are negative.

The negative returns likely result from the revelation of the misstatement (Feroz et al. 1991;

Karpoff et al. 2008). Note that the strong positive stock price performance during and prior to

manipulation years, combined with the negative stock price performance that followed, could be

one factor that triggered the SEC investigation.

Taken together, misstating firms appear to be engaging in more off-balance sheet

financing through leases, with relatively more of their asset bases being composed of soft assets

whose values are more subject to manipulation. Moreover, misstating firms could be concerned

about investor expectations. Their stock price performance is very high (around 20%) prior to

and in the year of misstatement, and they are issuing equity and raising financing. They also

appear to be either over-investing in their assets or engaging in accrual manipulation prior to and

at the time of the misstatement. Misstating firms have high accruals followed by significant

declines in return on assets during misstatement years.

[Table 5 and Figure 1]

4.4 Cross-sectional Analysis of Misstating Years

Our next test reported in Table 6 compares misstating firm-years to all firms listed on the

Compustat Annual File between 1979 and 2002. We limit the sample to these years since the

first AAER release occurred in 1982, and very few firms are identified as misstating prior to

1979. The objective of this test is to identify unusual characteristics of misstating firms relative

to the population. As mentioned earlier, we do this broad comparison (rather than an industry

size match comparison) because when we develop our F-Score model we want a measure that

can be easily applied to all firms. Industry adjustments would add complexity to this objective,

32
although we acknowledge that industry adjustment could improve the model’s predictive ability.

We provide some insights into this issue later in our robustness tests, reported in Table 9.

The first set of variables reported in Table 6 relates to accrual quality. We exclude

performance matched discretionary accruals, since this adjustment is redundant when using the

entire population. The results indicate that all accrual variables are unusual and in the predicted

direction in misstating years, relative to the population. For example, in misstating years the

RSST accrual measure is 12.6 percent of assets, whereas for the population, this measure is 3.2

percent of assets. Similarly, change in receivables is 6.1 percent for misstating firms and only

1.7 percent for the population. The % Soft assets is also higher for misstating firms. The

studentized DD measure indicates that misstating firms’ residuals are on average 0.428

deviations from the regression line in the positive direction.

For the performance variables, change in cash sales for misstating firms is about twice as

large as for the population (0.492 versus 0.208). However, the change in return on assets is

significantly lower for misstating firms (-0.024 versus -0.010). The results for non-financial

variables and off-balance sheet variables are significant in most cases and all in the predicted

direction. Notably, for firms with pension plans we find that misstating firms assume

significantly higher expected returns on their plan assets (8.20% versus 7.30%). However, the

change in expected return on plan assets is insignificant.

Finally, for the market-related variables, the results are all in the predicted direction and

significant in all but two variables (Ex ante finance need and Leverage are insignificant). We

report market-adjusted stock returns in the misstatement year and the prior year. Compared to

the average firm, misstating firms have significantly greater returns in both years. In addition,

misstating firms have high valuations relative to fundamentals when compared to the Compustat

33
population. In contrast to Table 4, book-to-market now loads in the correct direction. Thus both

book-to-market and earnings-to-price are significantly lower for misstating firms (i.e., they have

high valuations relative to fundamentals).

[Table 6]

4.5 Prediction Analysis and Development of the F-Score

In this section we provide multivariate analysis of variables discussed in Table 6.

Misstatements resulting in SEC Enforcement Actions are rare events. Our misstatement sample

represents less than half of one percent of the firm-years available on Compustat. However,

misstatements are extremely costly to auditors (in terms of lawsuits), to investors (in terms of

negative stock returns), to regulators like the FASB and SEC (in terms of reputation for quality

and enforcement of accounting rules) and to capital markets (in terms of lost investor confidence

and reduced liquidity). Therefore, even though misstatements are rare, a model that can help

identify misstatements can be used as a first-pass screen to identify firms that warrant further

investigation.

Table 7 provides our logistic models for the determinants of misstatements. Our

dependent variable is equal to one for firm-years involving a misstatement, and zero otherwise.

We estimate logistic regressions to determine whether the variables we examined in univariate

tests are jointly significant in predicting misstatement firm-years. We build three models for

predicting misstatement. Model 1 includes only financial statement variables as predictors,

Model 2 adds non-financial statement and off-balance sheet variables, and Model 3 incorporates

market-based measures. We form our models in this way so we can see the incremental benefit

from including information beyond the financial statements for predicting misstatement. We use

a backward elimination technique to arrive at our prediction models. The backward elimination

34
technique begins with all of our selected variables.8 We then use the computational algorithm of

Lawless and Singhal (1978) to compute a first-order approximation of the remaining slope

estimates for subsequent variable eliminations. Variables are removed based on these

approximations. We set the significance level for elimination at the 15% level.9

Model 1 begins with our accruals quality measures, the performance measures, and the

market-related measures that are computed from variables in the financial statements. After

performing backward elimination, we retain the following variables: RSST accruals, change in

receivables, change in inventory, % Soft assets, change in cash sales, change in return on assets,

and actual issuance. For Model 2, we retain the variables from Model 1 and add the non-

financial variables and off-balance sheet variables. After backward elimination, we retain

abnormal change in employees and existence of operating leases. For Model 3, we add our

market-based variables (our two return measures and book to market). The two return measures,

lagged market-adjusted stock return and market-adjusted stock return in the current year, are

retained in the model after backward elimination. Table 7 Panel A provides the resulting

coefficient estimates for the models.

Next, we examine the quality of our models. As we suggested above, it is important to

remember that because we cannot identify firms that misstate and are not subsequently caught

and investigated by the SEC, our analysis of errors is based on the joint fact that a firm misstated

and received an enforcement action from the SEC. To examine the quality of our models, we

analyze the predicted probabilities that the model assigns to each observation. It is not possible

8
We exclude from the selection process discretionary accrual measures because we want variables that can be
relatively easily calculated from the financial statements, variables calculated using the statement of cash flows
(CFF and Ex ante financing need) because these variables would restrict our analysis to observations after 1987,
variables that are not significantly different in Table 6, and order backlog and pension variables since these are
available for a limited set of firms.
9
We run the logistic procedure in SAS, with the model selection equal to BACKWARD and FAST. Other model
selection procedures produce similar results.

35
to test how well our model performs at predicting firms that misstate but are not subsequently

caught by the SEC. Predicted values are obtained by plugging each firm’s individual

characteristics into the model and using the estimated coefficients to determine the predicted

value. The predicted probability is derived as:

e ( PredictedValue)
Probability =
(1 + e ( PredictedValue) )

We then divide the probability by the unconditional expectation of misstatement to

calculate our F-Score. The unconditional expectation is equal to the number of misstatement

firms divided by the total number of firms. Below is an example of how this is done for Model 1

for Enron in 2000.

Enron in 2000

Predicted Value:
=-7.893+0.790 x (rsst_acc)+ 2.518 x (ch_rec)+ 1.191 x (ch_inv)+ 1.979 x (soft assets)+ 0.171 x (ch_cs)+ -0.932
x (ch_earn)+ 1.029 x (issue)
Predicted Value:
=-7.893+0.790 x (0.01659)+ 2.518 x (0.17641)+ 1.191 x (.00718) +1.979 x (0.79975)+ 0.171 x (1.33335)+
-0.932 x (-0.01285)+ 1.029 x (1)
Predicted Value=-4.575
Probability =e (-4.575) / (1+e(-4.575) )
e= 2.71828183
Probability =0.01020
Unconditional probability = 494/(132,967 + 494) = 0.0037
F-Score = 0.01020/0.0037
F-Score for Enron = 2.76

An F-Score of 1.00 indicates that the firm has the same probability of misstatement as the

unconditional expectation. F-Scores greater than one indicate higher probabilities of

misstatement than the unconditional expectation. Enron has an F-Score of 2.76. This suggests

that Enron has more than twice the probability of having misstated compared to a randomly

selected firm from the population.

Table 7 Panel B ranks firm-years into five portfolios based on the magnitude of their F-

Score. We report the frequency with which misstating and non-misstating firms fall into each

36
quintile and the minimum F-Score required to be included in each quintile. If our models do a

good job of identifying misstatement firms, then we expect misstatement firms to be clustered in

the fifth portfolio. The results for Model 1 indicate that 51.01 percent of misstatement firms are

in Quintile 5, compared to the expected level of 20 percent. The cut-off to be included in

Quintile 5 (i.e., the minimum value) is 1.397, so Enron’s score for 2000 of 2.76 easily places it in

Quintile 5. Model 2 adds non-financial and off-balance sheet variables to the variables in Model

1, and 51 percent of misstating firms are in Quintile 5. For Model 3, which includes market-

related variables, 48.02 percent of misstating firms are in Quintile 5.

In Panel C of Table 7, we evaluate the sensitivity of our models and determine Type I and

Type II error rates for an F-Score cut-off of 1.00. The results for Model 1 indicate that we

correctly classify 339 of the 494 firms (Sensitivity equal to 68.62%). A Type II error occurs

when our model incorrectly classifies a misstating firm as a non-misstating firm. The Type II

error rate is 31.38% (155 divided by 494). A Type I error occurs when our model incorrectly

classifies a non-misstating firm as a misstating firm. For an F-Score cut-off of 1.00, we

incorrectly classify 48,282 non-misstating years of a total of 132,967 as misstating years (our

Type I error is 36.31%). Similar results are found for Model 2 and Model 3.10

[Table 7]

10
In Table 6 we include all observations with available data for the calculation of variables included in model
selection. This results in the number of observations declining across models and makes direct comparisons across
the models difficult. We re-estimated Model 1 and 2 using only observations available for Model 3 and find that
variable selection does not change, with the exception of abnormal change in employees dropping from the model
(not tabulated). When using the variables in Table 6 and setting an F-Score cut-off of 1.00, we find that with a
consistent set of observations Model 3 correctly classifies as above, 66.95% (237/354) of misstatement firms and
63.86% (55,225/88,032) of non-misstatement firms. For Model 1 the correct classification of misstatement firms
increases slightly to 67.23% (238/354), while the correct classification of non-misstatement firms decreases slightly
to 63.55% (55,947/88,032). Model 2 correctly classifies 65.53% (232/354) of misstatement firms and 63.51%
(55,908/88,032) of non-misstatement firms. This suggests that when considering only firms with stock price
information, Model 3 offers a slight improvement in classifying non-misstatement firms over Model 1 and Model 2
and a slight deterioration in classifying misstatement firms relative to Model 1.

37
 We provide further insights into sensitivity, Type I and Type II error rates in Figure 2.

Figure 2 (a) provides the cumulative distribution of F-Scores for misstating firm-years. Figure 2

(b) reports the cumulative distribution for all non-misstating firms. These Figures can be used to

assess sensitivity and Type I error rates for any F-Score cut-off. For example, for an F-Score of

2.45, Figure 2 (a) reveals that the sensitivity is 18.8% (93 of 494 misstating firms have F-Scores

above this cut-off) and the Type II error rate is 81.2% (401 of 494 misstating firms are classified

as “clean”), while the Type I error rate is 5% (6,665 of 132,967 non-misstating firms have F-

Scores above this cut-off), while 95% of non-misstating firms are correctly classified. As a

simple rule of thumb, we classify an F-Score greater than 1 as “above normal risk” and an F-

Score greater than 2.45 as “high risk”.

Figure 2 can also provide insights into the likelihoods of Type I and Type II errors. The

cost of these errors is not the same. From an auditor’s perspective the cost of Type I and Type II

error is likely to differ, with Type II errors being more costly. When a misstatement goes

undetected (and is later revealed), the auditor is likely to be sued by investors, to be sanctioned

by regulatory bodies such as the SEC and the PCAOB and to suffer a loss of reputation. A Type

I error (a non-misstating firm is suspected of misstatement) is not costless and may result in lost

fees, as the auditor may choose to drop a client. Since Type II errors are more costly to the

auditor, an auditor is likely to prefer an F-Score cut-off that makes more Type I errors than Type

II errors.

Figure 2 (c) provides insights into how a user might make the trade-off between Type I

and Type II errors and provides a relative cost of errors ratio. For each F-Score, Figure 2 (c)

provides the total number of firms with an F-Score equal to or greater than that F-Score divided

by the total number of misstating firms with an F-Score equal to or greater than that F-Score.

38
This is calculated as number of Type I errors (incorrectly classified non-misstatement firms) plus

the sensitivity (the number of correctly classified misstating firms) divided by the sensitivity.

Assume that the cost of investigating a firm is $1 and an investigation always detects

misstatements if they exist. At an F-Score cut-off of 1.00, the relative cost ratio is 143

((48,282+339)/339). A cost of $48,621 is incurred to avoid the 339 misstating firms. Therefore,

if the cost of missing a fraud firm is $143 or more, then an F-Score cut-off of 1.00 should be

used by the auditor. If the cost is over $250, then all firms should be investigated (since the F-

Score cut-off is equal to zero). If the cost is less than $50, it is cheaper not to do the

investigation and just pay the extra cost of the fraud firms as they are identified.

[Figure 2]

Next, we examine the time-series properties of the F-Score for misstating firms. Figure 3

plots the mean F-Score for misstating firms for the three years prior to misstatement years,

during misstatement years and the three years following misstatement years. Figure 3 indicates

that the average F-Score for misstating firms is 1.5 prior to the misstatement, peaks at around 1.9

during misstatement years and declines to approximately 1.0 following misstatement years. It is

interesting to note the high average F-Scores prior to the misstatement years in light of our Table

5 results indicating that accruals and many other variables do not differ significantly in

misstatement years relative to years before the misstatement. If we assume that firms first use

the flexibility to manage earnings within GAAP before resorting to more aggressive earnings

management identified by the SEC, then the elevated F-Scores in years t-3 to t-1 indicate that the

F-Score reflects earnings management within GAAP. Further, the significant number of Type I

errors reported in Table 7 could include a disproportionate number of “non-misstating” firms that

39
actually are managing earnings. To the extent that this is true, the number of actual Type I

errors in our model is overstated.

[Figure 3]

4.6 Marginal Analysis and Robustness Tests of F-Score Models

In this section we evaluate the influence of each of the variables in the models for

determining the magnitude of F-Scores (marginal effect analysis). We then provide a number of

tests to examine how sensitive the variables classification is to time periods and industry

clustering.

Table 8 provides our marginal effect analysis. In this test we (i) calculate the value of the

F-Score when all variables are held at their mean values; (ii) recalculate the F-Score after

moving one independent variable to its lower quartile value, holding all other variables at their

mean value; (iii) recalculate the F-Score moving the independent variable to its upper quartile

value; (iv) calculate the change in the F-Score across the inter-quartile range for that variable

(for indicator variables such as actual issuance, the marginal impact is the difference in F-Score

when the variable equals one versus zero); (v) repeat steps (ii) through (iv) for the next

independent variable.

Table 8 Panel A reports the mean, upper and lower quartile values of the variables

included in the models. Panel B provides the marginal effect analysis for Models 1 through 3.

The first thing to note is that the average F-Score for Model 1 is 0.728.11 When RSST accruals

are at their lower quartile value and all other variables are at their mean values, the F-Score

changes from 0.728 to 0.694. Moving RSST accruals to their upper quartile value changes the F-

11
The mean F-Score differs from 1.00 (the unconditional expectation) because predicted values are determined
using the exponential function that gives weights different from those obtained using a linear estimation technique
such as ordinary least squares regressions (see the Enron example included in the text).

40
Score to 0.771, giving an inter-quartile range of 0.077. The results for Model 1 indicate that the

three variables with the greatest impact are change in receivables (0.099), soft assets (0.595), and

issuing securities (0.559). Since 82.5 percent of the sample firms issue securities in a given year,

a firm that does not issue has a far lower risk of being a misstating firm. In Table 9 Panel C

(discussed next) we investigate whether industry factors could be influencing the importance of

% Soft assets. The joint marginal effect (when moving all independent variables in the predicted

direction between the 1st and 3rd quartiles or between zero and one for indicator variables)

increases the F-Score from 0.169 to 1.547. For Model 2, existence of operating leases

(leasedum) has a relatively large marginal impact on the F-Score (0.264); note that 73.4 percent

of firms have leases. For Model 3, the return variables appear to have little marginal impact on

the F-Score. Finally, in Panel C we provide the Spearman and Pearson correlations of the

variables with the F-Score. The correlations are all significant and are generally consistent with

the marginal effect analysis in the sense that variables with a higher correlation have higher

marginal effects.

Overall, the results in Table 8 suggest that while accrual variables are important

contributors to the F-Score they are not unduly influencing the F-Score; other variables in the

model also play an important role. This suggests that the F-Score is likely to provide

incremental information beyond accruals for researchers investigating earnings management.

[Table 8]

Table 9 provides a variety of robustness tests related to our models. Here we focus on

Model 3 since this provides all the variables we analyze. Similar results are obtained for Models

1 and 2.

41
 We first investigate the sensitivity of our models to the time period examined. In Table 7

we develop our prediction model and evaluate its effectiveness using the same sample.

Therefore, the models suffer from a hindsight bias and could over-represent our predictive

ability. To evaluate the importance of this concern, we test the sensitivity of variable selection

by estimating models using the backward elimination technique during the 1979 to 1998 time

period. We follow a similar procedure of first including only financial statement variables, then

adding off-balance sheet variables to Model 2 and market-related variables to Model 3. We

find that all and only the original variables load in Model 1 in the earlier time period. However,

for Model 2, we find that abnormal change in employees no longer loads, and for Model 3,

book-to-market loads, but market-adjusted return does not. We report the results for 1979-1998

Model 3 in Table 9 Panel A Column (1). We use the new estimates from this model to predict

the F-Scores for a hold-out sample of firm-years from 1999 to 2002. We rank the hold-out

sample firms into quintiles and report the frequency and mean probabilities for misstating and

non-misstating firms by quintiles. The results are reported in Table 9 Panel B. Compared to the

results in Table 7 Panel B for Model 3, the model shows a slight improvement in the percent of

misstating firms classified in Quintile 5 (51.4 versus 48.02% in Table 7). However, generally

the Model does not appear to be too time specific.12

12
We also conduct a rolling-window out-of-sample test. For Model 1 (financial statement variables), starting in
1990, we conducted backward elimination using all available data up until that point beginning with the significant
variables from Table 5. Then we used the model to select variables and estimate coefficients. Using the results, we
applied the model to the following year's data. We then did this for every year (the last year is estimated using the
data through 2001 and applying this to 2002). The variables selected are quite consistent over time. For the 12
years, the variables differ from the model in Table 6 as follows: 1990-RSST accruals and change in inventory are
excluded while WC accruals is included; 1991-RSST accruals and change in ROA are excluded; 1992- identical
variables, 1993- identical variables, and 1994- change in inventory are excluded while WC accruals is included;
1995- change in inventory is excluded; 1996- change in inventory is excluded while WC accruals is included; 1997-
2001 – identical variables. Out of 12 years, the model is identical for 7 years. The changes in other years primarily
involve substituting working capital accruals for total accruals or the elimination of inventory. This should not be
surprising as the three variables are very correlated. An analysis of the predicted out of sample F-Scores for
misstatement and non-misstatement firms shows results similar to the out of sample results presented in Table 9.

42
 Another concern is that the internet boom years (1998 to 2000) represent a large

proportion of misstatements and so could unduly affect variable selection. We therefore rerun

our backward elimination technique for our models excluding these years. The results are

reported in column (2) of Table 9. We find that abnormal change in employees no longer loads,

while book-to-market loads. In Panel B and C we find similar results to those reported in Table

7. Therefore, the models do not appear to be overly sensitive to internet firms and their specific

accounting problems.

A final concern is related to the undue impact of certain industries. Table 2 Panel B

documents that the computer, retail and service industries appear to be over-represented in the

population of misstating firms. In addition, since leasing is used extensively in retail, another

concern is that our leasing results could be due to the over-representation of retail firms in our

sample. Further, we find that % Soft asset is an important influencing variable, and this is likely

to vary with industry composition and be higher in computer and service related industries.13

Therefore, our next test investigates whether our models are unduly influenced by these

industries and the existence of operating leases. We create industry dummies and an interactive

dummy (retail x existence of operating leases) and add these variables to the estimation of

Model 3 in Table 7. The results in Table 9 Panel A column (3) indicate that the coefficient on

% Soft assets declines slightly from 2.265 in Table 7 to 2.214 in Table 9 but remains highly

statistically significant. In addition, the coefficients on the service industry and the computer

industry are significant. Table 9 Panel B indicates that the number of firms classified in Quintile

13
% Soft Assets is likely to vary by industry because the need for “soft” versus “hard” assets such as PPE varies with
output technologies. The average % Soft Assets is the following for each industry: Textile and Apparel 0.687; Retail
0.594; Durable Manufacturers 0.589; Computers 0.56; Lumber, Furniture, & Printing 0.551; Chemicals 0.543; Food
& Tobacco 0.537; Services 0.503; Agriculture 0.466; Pharmaceuticals 0.46; Mining & Construction 0.453;
Transportation 0.385; Utilities 0.277; Refining & Extractive 0.243. Note that in Table 5, each firm acts as its own
control and we find significantly higher % Soft Assets in manipulation years relative to earlier years. This suggests
that industry effects are not the only determinant of the higher Soft Asset ratio for AAER firms.

43
5 is 176, which is only six more firms than the number (170) in Table 7 Panel B. In addition,

using a cut-off F-Score of 1.00, we find that the number of correctly classified misstatement

firms is similar in Table 9 (237 firm-years) and Table 7 (238 firm-years). Overall, the model

does not appear to be unduly driven by characteristics of the computer, retail, or service

industries.

[Table 9]

5. CONCLUSION

This paper provides a comprehensive sample of firms investigated by the SEC for

misstating earnings. We conduct a detailed analysis of 2,190 Accounting and Auditing

Enforcement Releases available between 1982 and 2005 and identify firms with misstated

quarterly or annual earnings. We document the most common types of misstatements and find

that the overstatement of revenues, misstatement of expenses, and capitalizing costs are the most

frequent types of misstatements. We also identify the industries and time periods in which

misstatements are most common.

We investigate the characteristics of misstating firms on various dimensions, including

accrual quality, financial performance, non-financial performance, off-balance sheet activities,

and market-related variables. We find that at the time of misstatements, accrual quality is low

and both financial and non-financial measures of performance are deteriorating. We also find

that financing activities and related off-balance sheet activities are much more likely during

misstatement periods. Finally, we find that managers of misstating firms appear to be sensitive

to their firm’s stock price. These firms have experienced strong recent earnings and price

performance and trade at high valuations relative to fundamentals. The misstatements appear to

44
be made with the objective of covering up a slowdown in financial performance in order to

maintain high stock market valuations.

Many accounting researchers use measures of “discretionary accruals” as their proxy for

earnings management. Our paper contributes to this line of research by showing that financial

statement information beyond accruals is useful for identifying earnings manipulation. Our

composite measure of the likelihood of manipulation (the F-Score) offers researchers a

complementary and supplementary measure to discretionary accruals for identifying “low

quality” earnings firms. In addition, our paper shows that the modified Jones model tends to have

less power than unadjusted measures of accruals (such as working capital accrual or RSST

accruals) to identify misstatement years. We also find that growth in cash sales is unusually high

during misstatement years. An important avenue for future research is to better understand the

role of real transaction or cash flow management.

We emphasize that one unavoidable issue in developing models to detect misstatement is

that the revelation of a misstatement by the SEC is a rare event. Thus, similar to bankruptcy

prediction models, our models generate a high frequency of false positives (i.e., many firms that

do not have enforcement actions against them are predicted to have misstated their earnings).

Another limitation of our analysis is that we can identify only misstatements that were actually

identified by the SEC. There are likely many cases where a misstatement goes undetected, or is

at least not subject to an SEC enforcement action. An interesting avenue for future research

would be to investigate the characteristics of high F-Score firms that are not identified by the

SEC. For example, do high F-Score firms engage in earnings management within the realm of

GAAP? Do they experience declines in subsequent financial performance? Are they more likely

to record future asset write-offs or write-downs? In addition, can models be improved by

45
considering corporate governance arrangements, top executive characteristics, or industry

specific characteristics?

Finally, our analysis should provide useful insights to auditors, regulators, investors, and

other financial statement users about the characteristics of misstating firms. By better

understanding these characteristics, financial statement users should be in a better position to

identify and curtail misstatement activity in the future. The efficient functioning of capital

markets depends crucially on the quality of the financial information provided to capital market

participants. Curtailing misstatement activity should lead to improved financial information and

hence improved returns for investors and more efficient allocation of capital.

46
References

Allen, E., C. Larson and R. Sloan. 2009. Accrual reversals, earnings and stock returns. Working
paper, Washington University in St. Louis.
Altman, E. I. 1968. Financial ratios, discriminant analysis and the prediction of corporate
bankruptcy. The Journal of Finance 23: 589-609.
Barton, J. and P. Simko. 2002. The balance sheet as an earnings management constraint. The
Accounting Review 77, 1-27.
Bayley, L. and S. Taylor. 2007. Identifying earnings overstatements: A practical test. Working
paper, ABN Amro Sydney and University of New South Wales.
Beasley, M. 1996. An empirical analysis of the relation between the board of director
composition and financial statement fraud. The Accounting Review 71, 443-465.
Beneish, M. D. 1999b. Incentives and penalties related to earnings overstatements that violate
GAAP. The Accounting Review 74, 425-457.
Beneish, M. D. 1999a. The detection of earnings manipulation. Financial Analysts Journal,
September/October: 24-36.
Beneish, M. D. 1998. Discussion of “Are accruals during initial public offerings opportunistic?”
Review of Accounting Studies 3: 209-221.
Beneish, M. D. 1997. Detecting GAAP violations: Implications for assessing earnings
management among firms with extreme financial performance. Journal of Accounting
and Public Policy 16: 271-309.
Brazel, J. F., K. L. Jones, and M. F. Zimbelman. 2009. Using nonfinancial measures to assess
fraud risk. Journal of Accounting Research 47: 1135-1166.
Comprix, J., and K. A. Mueller III. 2006. Asymmetric treatment of reported pension expense
and income amounts in CEO cash compensation calculations. Journal of Accounting and
Economics 42: 385–416.
Dechow, P. M., W. Ge, and C. Schrand. 2009. Understanding earnings quality. Working paper,
University of California, Berkeley, University of Washington, and University of
Pennsylvania.
Dechow, P. M., and I. Dichev. 2002. The quality of accruals and earnings: The role of accrual
estimation errors. The Accounting Review 77 (Supplement): 35-59.
Dechow, P. M., R. G. Sloan, and A. P. Sweeney. 1995. Detecting earnings management. The
Accounting Review 70: 193-226.
Dechow, P. M., R. G. Sloan, and A. P. Sweeney. 1996. Causes and consequences of earnings
misstatement: An analysis of firms subject to enforcement actins by the
SEC. Contemporary Accounting Research 13: 1-36.
DeFond, M. L. and J. Jiambalvo. 1994. Debt covenant violation and misstatement of
accruals. Journal of Accounting and Economics 17: 145-176.
Erickson, M., M. Hanlon, and E. L. Maydew. 2004. How much will firms pay for earnings that
do not exist? Evidence of taxes paid on allegedly fraudulent earnings. The Accounting
Review 79, 387-408.
Ettredge, M., L. Sun, P. Lee, and A. Anandarajan. 2006. Do deferred tax data signal earnings
fraud? Working paper, University of Kansas School of Business.
Farber, D. 2005. Restoring trust after fraud: Does corporate governance matter? The Accounting
Review 80, 539-561.

47
Feroz, E., K. Park and V. Pastena, 1991. The financial and market effects of the SEC's
accounting and auditing enforcement releases. Journal of Accounting Research 29, 107-
142.
Frankel, R. M., M. F. Johnson and K. K. Nelson. 2002. The relation between auditors’ fees for
no audit services and earnings management. The Accounting Review 77: 71-103.
Ge, W. 2007. Off-balance-sheet activities, earnings persistence and stock prices: Evidence from
operating leases. Working paper, The University of Washington.
Graham, J., C. Harvey and S. Rajgopal. 2005. The economic implications of corporate financial
reporting. Journal of Accounting and Economics 40: 3-73.
Healy, P. 1985. The effect of bonus schemes on accounting decisions. Journal of Accounting
and Economics 7: 85-107.
Karpoff, J., S. Lee and G. Martin. 2008. The cost to firms of cooking the books. Journal of
Financial and Quantitative Analysis 43, 581-611.
Kothari, S. P., A. Leone, and C. Wasley. 2005. Performance matched discretionary accrual
measures. Journal of Accounting and Economics 39: 163-197.
Lawless J. F., and K. Singhal. 1978. Efficient screening of nonnormal regression
models. Biometrics 34: 318-327.
Miller, G. S. 2006. The press as a watchdog for accounting fraud. Journal of Accounting
Research 44 (5): 1001-1033.
Plumlee, M. and T. L. Yohn. 2008. Analysis of the underlying cause of restatements. Working
paper, University of Utah.
Phillips, J., M. Pincus, and S. O. Rego. 2003. Earnings management: New evidence based on
deferred tax expense. Accounting Review 78 (2): 491-521.
Rajgopal, S., T. Shevlin, and M. Venkatachalam. 2003. Does the stock market fully appreciate
the implications of leading indicators for future earnings? Evidence from order backlog.
Review of Accounting Studies 8: 461-492.
Rangan, S. 1998. Earnings management and the performance of seasoned equity
offerings. Journal of Financial Economics 50: 101-122.
Richardson, S., R. Sloan, M. Solimon, and I. Tuna. 2005. Accrual reliability, earnings
persistence, and stock prices. Journal of Accounting and Economics 39: 437-485.
Skousen, C. J., and C. J. Wright. 2006. Contemporaneous risk factors and the prediction of
financial statement fraud. Working paper, University of Texas at Arlington.
Skinner, D., and R. Sloan. 2002. Earnings surprises, growth expectations, and stock returns or
don’t let an earnings torpedo sink your portfolio. Review of Accounting Studies 7, 289-312.
Teoh, S. H., I. Welch, and T. J. Wong. 1998. Earnings management and the underperformance of
seasoned equity offerings. Journal of Financial Economics 50: 63-99.

48
Appendix 1: Variable Definitions of the Enforcement Releases Datasets
Panel A: DETAIL FILE detail.sas7bdat

Variable Name Description

coname Name from AAER
CIK Central Index Key
cnum 6-digit Cusip
ticker Compustat ticker
gvkey Compustat Gvkey
permno CRSP Permno
iticker IBES Ticker
eticker Exchange Ticker
explanation Two sentence explanation of the violation
Indicator variables (file
inclusion):
annual Equals 1 if the firm is in the Annual file, 0 otherwise
quarter Equals 1 if the firm is in the Quarterly file, 0 otherwise
reason Reason firm is not included in Annual or Quarterly files
Indicator variables (exclusion
from Annual or quarterly
files):
Equals 1 if the AAER was brought against the auditor and there was no
audit
misstatement, 0 otherwise
bribes Equals 1 if the AAER was for bribe charges, 0 otherwise
Equals 1 if related to disclosure issue only and not earnings misstatement, 0
disclosure
otherwise
Equals 1 if the time period of the financial misstatements cannot be determined
Nodates
from the AAER, 0 otherwise
Other Equals 1 if related to other issues not listed above, 0 otherwise
Indicator variable (Accounts
affected):
Rev Equals 1 if misstatement affected Revenues, 0 otherwise
rec Equals 1 if misstatement affected Accounts Receivables, 0 otherwise
Cogs Equals 1 if misstatement affected Cost of Goods Sold, 0 otherwise
Inv Equals 1 if misstatement affected Inventory, 0 otherwise
Res Equals 1 if misstatement affected reserves accounts, 0 otherwise
Debt Equals 1 if misstatement affected bad debts, 0 otherwise
mkt_sec Equals 1 if misstatement affected Marketable Securities, 0 otherwise
Pay Equals 1 if misstatement affected Accounts Payable, 0 otherwise
Equals 1 if misstatement affected an asset account but could not be classified in an
Asset
asset account above, 0 otherwise
Liab Equals 1 if misstatement affected liabilities, 0 otherwise
Equals 1 if misstatement could not be classified in an income, expense or equity
inc_exp_se
account above, 0 otherwise
Equals 1 if the actual amount of the misstatement can potentially be obtained from
Figure
the AAER, 0 otherwise
AAER columns There are 24 columns that identify all AAERs related to the firm
Total AAERs Total number of AAERs for the firm
Reason for no Cnum 0 if firm has a Cusip or a number that identifies why the firm has no Cusip

49
Appendix 1: (continued)
Panel B: ANNUAL FILE ann.sas7bdat

Variable Name Description

coname Name from AAER
CIK Central Index Key
permno CRSP Permno
yeara Compustat convention year
fyr Fiscal month end
date Actual misstatement date collected from AAER (DD/MM/YYYY)
p_aaer Primary AAER used to collect data
understatement Equals 1 if earnings/revenues were understated in the year, 0 otherwise

Panel C: QUARTERLY qtr.sas7bdat

FILE

Variable Name Description

coname Name from AAER
CIK Central Index Key
permno CRSP Permno
yeara Compustat convention year
fyr Fiscal month end
qtr Quarter (1, 2, 3 or 4)
date Actual date collected from AAER (DD/MM/YYYY)
p_aaer Primary AAER used to collect data
understatement Equals 1 if earnings/revenues were understated in the quarter, 0 otherwise

50
Figure 1 Annual market-adjusted stock returns surrounding misstatement years

0.44

0.19
0.13
0.11 0.11
0.05 0.04 0.06

t-3 t-2 t-1 t t+1 t+2 t+3

-0.11 -0.06 Mean
Median
-0.25

-0.41
-0.44
-0.51

Year Relative to Misstatement Years

Note: For all firm-years with available returns data on CRSP. Returns include delisting returns. For year t-3 n=137,
for year t-2 n=166, for year t-1 n=188, for year t n=474, for year t+1 n=250, for year t+2 n=202, and for year t+3
n=153. Year t is the average return for all misstatement firm-years. Market-adjusted returns are calculated as the
difference between annual raw returns and value-weighted market returns.

51
 Figure 2: Evaluating the likelihood of a given F-Score
Figure 2A: Cumulative distribution of F-Scores for misstatement firms

100%

18.8%
90%
68.6% 31.8%
80%

70%

60%

50%

40%
68.2% 81.2%
30%

20%
31.4%
10%

0%

4+
5

5
0

3
25

75

25

75

25

75

25

75
0.

1.

2.

3.
0.

0.

1.

1.

2.

2.

3.

3.
Fscore 1 (N=494) Fscore 2 (N=449) Fscore 3 (N=354)

Figure 2B: Cumulative distribution of F-Scores for non-misstatement firms (Compustat

firm-years)

100% 5%
10%
90%
36.3%
80%

70%

60%

50%
90% 95%
40%

30%
63.7%
20%

10%

0%
4+
5

5
0

3
25

75

25

75

25

75

25

75
0.

1.

2.

3.
0.

0.

1.

1.

2.

2.

3.

3.

Fscore 1 (N=132,967) Fscore 2 (N=122,341) Fscore 3 (N=88,032)

Interpreting F-score 1 (Similar interpretation for F-score 2 and F-score 3)

F-score 1 greater than 2.45 5% of Non-Misstatement Firms 18.8% of Misstating Firms “High” Risk
F-score 1 greater than 1.85 10% of Non-Misstatement Firms 32.6% of Misstating Firms “Substantial” Risk
F-score 1 greater than 1 36.3% of Non-Misstatement Firms 68.6% of Misstating Firms “Above Normal” Risk
F-score 1 less than 1 63.7% of Non-Misstatement Firms 31.4% of Misstating Firms “Normal or Low” Risk

52
Figure 2 (continued)

Figure 2C: Relative cost ratio: An analysis of error rates across F-Scores

300

250
(Type I Errors+Sensitivity) / Sensitivity

200
Model 1: (48,282+339) /339=143

150

100

50

0
5

5
0

3
25

75

25

75

25

75
0.

1.

2.
0.

0.

1.

1.

2.

2.
Fscore 1 Fscore 2 Fscore 3

Type I errors = misclassified non-misstating firm; Type II errors = misclassified misstating firms;
Sensitivity = correctly classified misstating firms. At an F-Score cut-off of 1.00, the total number of non-
misstating firms is 132,967, of which 48,282 have F-Scores greater than 1.00 (Type I error) and 339 of
the 494 misstating firms have F-Scores greater than 1.00. At this F-Score cut-off the relative cost ratio is
143 ((48,282+339)/339). If the cost of investigating a firm is $1 and investigations accurately identify
misstating firms, then it is worth investigating all firms with F-Scores greater than or equal to one, if the
cost of a misstating firm is greater than $143.

53
Figure 3 Mean F-Scores surrounding misstatement firm years

2.2

1.8

1.6

1.4
F-score

1.2

0.8 Fscore1 Mean

0.6 Fscore2 Mean
Fscore3 Mean
0.4

0.2

0
t-3 t-2 t-1 t t+1 t+2 t+3

Year Relative to Misstatement Years

This figure plots the mean F-Score for misstatement firms for the three years prior to misstatement years, during
misstatement years, and the three years following misstatement years.

54
Table 1
Sample description

Panel A: Sample selection of AAERs

Number of AAERs Number
AAER No. 1- No. 2261 from May 1982 to June 2005 2261
Less: missing AAERs (30)
Less: AAERs that do not involve specific company names (41)
Total 2190
Note: Among 30 missing AAERs, eleven AAERs are intentionally omitted and nineteen AAERs are
missing.

Panel B: Percent of the 2,190 AAERs that are against various parties
Party Number Percentage
Officer of the company 1,077 49.18%
Auditor 348 15.89%
331
Officer and company 15.11%
308
Company 14.06%
58
Other 2.65%
68
Other combination of parties 3.11%
Total 2,190 100.00%

Panel C: Frequency of AAERs by year

AAER
release Number of AAER Number of
date AAERs Percentage release date AAERs Percentage
1982 2 0.10% 1994 120 5.50%
1983 16 0.70% 1995 107 4.90%
1984 28 1.30% 1996 121 5.50%
1985 35 1.60% 1997 134 6.10%
1986 39 1.80% 1998 85 3.90%
1987 51 2.30% 1999 111 5.10%
1988 37 1.70% 2000 142 6.50%
1989 38 1.70% 2001 125 5.70%
1990 35 1.60% 2002 209 9.50%
1991 61 2.80% 2003 237 10.80%
1992 78 3.60% 2004 209 9.50%
1993 76 3.50% 2005 94 4.30%
Total 2190 100.00%

55
Table 1 (continued)
Panel D: Frequency of AAERs by firm
Percent of
Number of AAERs for each firm Number of firms firms Total AAERs
1 370 41.3% 370
2 235 26.2% 470
3 108 12.1% 324
4 70 7.8% 280
5 40 4.5% 200
6 33 3.7% 198
7 13 1.5% 91
8 10 1.1% 80
9 3 0.3% 27
10 6 0.7% 60
11 2 0.2% 22
12 2 0.2% 24
13 1 0.1% 13
15 1 0.1% 15
20 1 0.1% 20
24 1 0.1% 24
Total 896 100.0% 2218
Note: There are 28 (2218 less 2190) AAERs involving multiple companies.

Panel E: Number of distinct firms

Number of distinct companies mentioned in the AAERs Number
AAER No. 1- No. 2261 from May 1982 to June 2005 896
Less: Enforcements that are unrelated to earnings misstatement (e.g., bribes, disclosure
etc.) or firms with misstatements that cannot be linked to specific reporting periods 220
Earnings misstatement firms 676
Less: firms without CUSIP 132
Firms with at least one quarter of misstated numbers 544
Firms with total assets on Compustat: 457
Firms with stock price data on Compustat: 435
Less: firms with quarterly misstatements corrected by the end of the fiscal year 92
Firms with at least one annual misstated number 451
Firms with total assets on Compustat: 387
Firms with stock price data on Compustat: 362

56
Table 1 (continued)
Panel F: Type of misstatements identified by the SEC in the AAER
Percent of 676 Percent of 435 Percent of 451 Percent of
misstatement firms with at least firms with at 387 firms
firms one quarterly least one annual with at least
misstatement and misstatement one annual
stock price data misstatement
and total
Type of Misstatement assets data
(1) (2) (3) (4)
Misstated revenue 54.0% 59.5% 58.3% 60.2%
Misstatement of other
expense/shareholder equity account
25.1% 25.1% 24.6% 24.8%
Capitalized costs as assets 27.2% 20.5% 21.7% 20.9%
Misstated accounts receivable 19.1% 20.0% 21.1% 20.7%
Misstated inventory 13.2% 14.5% 16.2% 16.3%
Misstated cost of goods sold 11.4% 13.1% 13.7% 14.2%
Misstated liabilities 7.4% 7.1% 8.4% 8.0%
Misstated a reserve account 5.9% 4.4% 5.1% 4.4%
Misstated allowance for bad debt 4.3% 4.1% 4.7% 4.4%
Misstated marketable securities 3.6% 3.4% 3.8% 3.4%
Misstated payables 1.6% 2.3% 2.2% 2.6%
Note: There are a total of 1272 misstatements mentioned in Column (1), 770 misstatements in Column (2), 826
misstatements in Column (3), and 709 misstatements in Column (4), so percentages sum to more 100 percent.

57
Table 2
Frequency of misstating firms by size, industry and calendar year
(both annual and quarterly misstatements)

Panel A: Frequency of the misstating firms by firm size (market capitalization) deciles
Decile rank of market value of
Compustat population Frequency Percentage
1 22 5.1%
2 36 8.3%
3 37 8.5%
4 44 10.1%
5 38 8.7%
6 53 12.2%
7 48 11.0%
8 55 12.6%
9 38 8.7%
10 64 14.7%
Total 435 100.0%

Panel B: Frequency of the misstating firms by industry

Industry Misstating Firms Compustat Population

Agriculture 0.2% 0.4%
Mining & Construction 2.7% 3.0%
Food & Tobacco 2.5% 2.1%
Textile and Apparel 2.5% 1.7%
Lumber, Furniture, & Printing 2.2% 3.1%
Chemicals 2.2% 2.0%
Refining & Extractive 1.0% 4.7%
Durable Manufacturers 19.4% 18.9%
Computers 20.4% 11.1%
Transportation 4.7% 5.8%
Utilities 1.6% 3.2%
Retail 12.9% 9.9%
Services 12.7% 10.4%
Banks & Insurance 12.0% 20.8%
Pharmaceuticals 3.1% 3.2%
Total 100.0% 100.0%
Note: There are 435 misstating firms in the annual and quarterly files that have data to calculate market value and 490 misstating
firms that have SIC codes. Industries are based on the following SIC codes: Mining: 1000–1299, 1400–1999; Food: 2000–2199;
Textiles: 2200–2799; Drugs: 2830–2839, 3840–3851; Chemicals: 2800–2829, 2840–2899; Refining: 1300–1399, 2900–2999;
Rubber: 3000–3499; Industrial: 3500–3569, 3580–3659; Electrical: 3660–3669, 3680–3699; Miscellaneous Equipment: 3800–
3839, 3852–3999; Computers: 3570–3579, 3670–3679, 7370–7379; Transportation: 4000–4899; Utilities: 4900–4999; Retail:
5000–5999; Banks: 6000–6999; Services: 7000–7369, 7380–8999.

58
Table 2 (continued)

Panel C: Distribution of misstated firm-years

Firm- Firm-
Year years Percentage Year years Percentage
1971 1 0.12% 1987 24 2.87%
1972 1 0.12% 1988 27 3.23%
1973 1 0.12% 1989 42 5.02%
1974 2 0.24% 1990 33 3.94%
1975 2 0.24% 1991 44 5.26%
1976 1 0.12% 1992 47 5.62%
1977 1 0.12% 1993 41 4.90%
1978 4 0.48% 1994 35 4.18%
1979 9 1.08% 1995 37 4.42%
1980 17 2.03% 1996 40 4.78%
1981 23 2.75% 1997 43 5.14%
1982 31 3.70% 1998 53 6.33%
1983 25 2.99% 1999 66 7.89%
1984 25 2.99% 2000 60 7.17%
1985 17 2.03% 2001 38 4.54%
1986 29 3.46% 2002 15 1.79%
2003 3 0.36%
Total 837 100.00%

Note: This table is calculated based on the sample of 451 misstating firms
(as shown in Table 1 Panel D) with at least one misstated annual financial statement.

59
Table 3: Variable definitions
Pred
Variable
Abbreviation
Sign* Calculation
Indicator variable equal to 1 for misstatement firm-years and 0
Misstatement flag misstate N/A
otherwise
Accruals quality
related variables
[[ΔCurrent Assets (DATA 4) – ΔCash and Short-term
Investments (DATA 1)]– [ΔCurrent Liabilities (DATA 5) –
WC accruals WC_acc +
ΔDebt in Current Liabilities (DATA 34) – ΔTaxes Payable
(DATA 71)] /Average total assets
(∆WC + ∆NCO + ∆FIN)/Average total assets, where WC =
[Current Assets (DATA 4) – Cash and Short-term Investments
(DATA 1)] – [Current Liabilities (DATA 5) – Debt in Current
Liabilities (DATA 34)]; NCO = [Total Assets (DATA 6) –
Current Assets (DATA 4) - Investments and Advances (DATA
RSST accruals rsst_acc + 32)] – [Total Liabilities (DATA 181) – Current Liabilities
(DATA 5) – Long-term Debt (DATA 9)]; FIN = [Short-term
Investments (DATA 193) + Long-term Investments (DATA
32)] – [Long-term Debt (DATA 9) + Debt in Current
Liabilities (DATA 34) + Preferred Stock (DATA 130)];
following Richardson et al. (2005).
Change in receivables ch_rec + ∆Accounts Receivables (DATA 2)/Average total assets
Change in inventory ch_inv + ∆Inventory (DATA 3)/Average total assets
(Total assets (DATA 6) - PP&E (DATA 8) – Cash and cash
% Soft assets soft_assets -
equivalent (DATA 1))/Total assets (DATA 6)
The modified Jones model discretionary accrual is estimated
cross-sectionally each year using all firm-year observations in
Modified Jones model the same two-digit SIC code: WC Accruals = α +
da +
discretionary accruals β(1/Beginning assets) +γ(∆Sales-∆Rec)/Beginning assets +
ρ∆PPE/Beginning assets + ε. The residuals are used as the
modified Jones model discretionary accruals.
The difference between the modified Jones discretionary
accruals for firm i in year t and the modified Jones
Performance-matched discretionary accruals for the matched firm in year t, following
dadif +
discretionary accruals Kothari et al. (2005); each firm-year observation is matched
with another firm from the same two-digit SIC code and year
with the closest return on assets.
The following regression is estimated for each two-digit SIC
industry: ΔWC = b0 + b1*CFOt-1 +b2*CFOt + b3*CFOt+1 + ε .
Mean-adjusted absolute
resid + The mean absolute value of the residual is calculated for each
value of DD residuals
industry and is then subtracted from the absolute value of each
firm’s observed residual.
^

The scaled residuals are calculated as ei where hii is

^
σ 1 − hii
the ii element of the hat matrix, X(X X) − 1XT and
T
Studentized DD
sresid + ^ 1 m ^2 where m is the number of parameters in the
residuals σ= ∑ε j
n − m j =1
model and n is the number of observations. SAS can output the
scaled residuals using the following code: proc reg data=
dataset; model Y=X; output data=temp student=studentresidual

60
 Performance variables
Percentage change in cash sales [Sales(DATA 12)-∆Accounts
Change in cash sales ch_cs -
Receivables (DATA 2)]
Percentage change in cash margin, where cash margin is
measured as 1- [(Cost of Good sold (DATA 41) - ∆Inventory
Change in cash margin ch_cm -
(DATA 3)+ ∆Accounts payable (DATA70))/(Sales(DATA
12)- ∆Accounts Receivable(DATA 2))]
Change in return on [Earningst (DATA 18)/Average total assetst ]-
ch_roa ?
assets [Earningst-1/Average total assetst-1]
Change in free cash
ch_fcf - ∆[Earnings (DATA 18)-RSST Accruals] /Average total assets
flows
Deferred tax expense for year t (DATA 50) / total assets for
Deferred tax expense tax +
year t-1 (DATA 6)
Non-financial
variables
Abnormal change in Percentage change in the number of employees (DATA 29) -
ch_emp -
employees percentage change in assets (DATA 6)
Abnormal change in Percentage change in order backlog (DATA 98) - percentage
ch_backlog -
order backlog change in sales (DATA 12)
Off-balance sheet
variables
Existence of operating An indicator variable coded 1 if future operating lease
leasedum +
leases obligations are greater than zero
The change in the present value of future non-cancelable
Change in operating
oplease + operating lease obligations (DATA 96, 164, 165, 166 and
lease activity
167) deflated by average total assets following Ge (2007)
Expected return on
pension +
pension plan assets (%) Expected return on pension plan assets (DATA 336)
Change in Expected
return on pension plan ch_pension + ∆Expected return on pension plan assets [DATA 336 at t) -
assets (%) (DATA 336 at t-1)]
Market Incentives
An indicator variable coded 1 if [(CFO-past three year
Ex ante financing need exfin +
average capital expenditures)/Current assets]<-0.5
An indicator variable coded 1 if the firm issued securities
Actual issuance issue + during year t (i.e., an indicator variable coded 1 if DATA
108>0 or DATA111>0)
CFF cff + Level of finance raised (DATA 313/Average total assets)
Leverage leverage + Long-term debt (DATA 9)/ Total assets (DATA 6)
Market-adjusted Stock Annual buy-and-hold return inclusive of delisting returns
rett +
return minus the annual buy-and-hold value-weighted market return
Previous year’s annual buy-and-hold return inclusive of
Lagged market-
rett-1 + delisting returns minus the annual buy-and-hold value-
adjusted Stock return
weighted market return
Book to market bm - Equity (DATA 60)/ Market value (DATA 25 x DATA 199)
Earnings (DATA 18)/ Market Value (DATA 25 x DATA
Earnings to price ep -
199)

*Predicted Sign shows the expected direction of the relations between various firm-year characteristics and misstatements.

61
Table 4
Descriptive statistics of misstatement years versus all non-misstatement years for AAER firms
Misstatement years Non-misstatement years Misstate - Non-misstate
Pair-wise One
Pred Diff. in tailed P- Pair-wise
Variable N Mean Median N Mean Median sign Mean value t-statistics
Accruals quality variables
WC accruals 296 0.072 0.050 296 0.018 0.019 + 0.053 0.001 4.43
RSST accruals 294 0.135 0.090 294 0.044 0.043 + 0.091 0.001 3.85
Change in receivables 298 0.071 0.042 298 0.028 0.017 + 0.043 0.001 6.12
Change in inventory 294 0.046 0.020 294 0.023 0.011 + 0.022 0.001 4.22
% Soft assets 327 0.647 0.696 327 0.611 0.616 + 0.036 0.001 3.87
Modified Jones model
discretionary accruals 294 0.055 0.028 294 -0.009 -0.001 + 0.063 0.001 3.10
Performance-matched
discretionary accruals 294 0.061 0.044 294 0.008 0.003 + 0.053 0.006 2.55
Mean-adjusted absolute
value of DD residuals 184 0.022 -0.009 184 0.013 -0.010 + 0.010 0.076 1.44
Studentized DD
residuals 184 0.529 0.355 184 0.058 0.021 + 0.471 0.001 5.92
Performance variables
Change in cash sales 263 0.466 0.233 263 0.257 0.165 - 0.209 0.001 3.88
Change in cash margin 256 0.016 0.006 256 0.011 0.001 - 0.005 0.441 0.15
Change in return on
assets 263 -0.032 -0.014 263 -0.025 -0.002 + -0.006 0.354 -0.38
Change in free cash
flows 259 0.024 0.007 259 0.015 0.006 - 0.009 0.359 0.36
Deferred tax expense 318 0.002 0.000 318 0.001 0.000 + 0.001 0.202 0.83
Non-financial variables
Abnormal change in
employees 263 -0.221 -0.108 263 -0.163 -0.073 - -0.058 0.150 -1.04
Abnormal change in
order backlog 77 0.091 -0.049 77 0.056 0.040 - 0.036 0.355 0.37
Off-balance sheet variables
Change in operating
lease activity 298 0.020 0.006 298 0.013 0.005 + 0.007 0.007 2.37
Existence of operating
leases 327 0.841 1.000 327 0.752 0.900 + 0.089 0.001 5.36
Expected return on
pension plan assets (%) 40 0.084 0.090 40 0.081 0.089 + 0.004 0.017 2.21
Change in expected
return on plan assets
(%) 33 -0.032 0.000 33 -0.113 -0.100 + 0.081 0.007 2.58
Market-related variables
Ex ante financing need 224 0.179 0.000 224 0.177 0.000 + 0.002 0.473 0.07
Actual issuance 325 0.938 1.000 325 0.869 1.000 + 0.068 0.001 4.75
CFF 233 0.247 0.134 233 0.151 0.072 + 0.096 0.001 3.94
Leverage 327 0.187 0.161 327 0.192 0.163 + -0.006 0.276 -0.60
Market-adjusted stock
return 237 0.234 0.028 237 -0.004 0.022 + 0.238 0.021 2.06
Book to market 291 0.520 0.351 291 0.354 0.395 - 0.166 0.001 3.04
Earnings to price 185 0.056 0.039 185 0.072 0.057 - -0.016 0.001 -3.97

In this table, we calculate means at the firm level for misstatement years versus all non-misstatement years and conduct tests
of pair-wise differences in means. Each firm is directly compared to itself during manipulation years versus other years.
All variables are defined in Table 3. Each of the continuous variables (except stock return variables) is winsorized at 1%
and 99% to mitigate outliers.

60
 Table 5
Descriptive statistics on misstatement years versus FIRM-YEARS PRIOR TO MISSTATEMENT YEARS for AAER firms
Misstatement years Early years Misstate – Early Years
Pair- One
wise tailed Pair-
Pred Diff. in P- wise t-
Variable N Mean Median N Mean Median Sign Mean value statistics
Accruals quality variables
WC accruals 263 0.069 0.047 263 0.068 0.056 + 0.001 0.452 0.12
RSST accruals 261 0.120 0.076 261 0.173 0.100 + -0.053 0.025 -1.97
Change in receivables 265 0.069 0.040 265 0.073 0.047 + -0.004 0.328 -0.46
Change in inventory 262 0.043 0.019 262 0.049 0.027 + -0.006 0.168 -0.96
% Soft assets 299 0.649 0.692 299 0.595 0.605 + 0.054 0.001 5.07
Performance variables
Change in cash sales 209 0.410 0.206 209 0.473 0.287 - -0.063 0.126 -1.15
Change in cash margin 201 0.034 0.007 201 -0.001 0.002 - 0.035 0.210 0.81
Change in return on
+
assets 210 -0.026 -0.014 210 0.009 -0.001 -0.034 0.014 -2.23
Change in free cash flows 205 0.020 0.006 205 0.017 0.003 - 0.003 0.445 0.14
Deferred tax expense 268 0.002 0.000 268 0.004 0.000 + -0.002 0.085 -1.37
Non-financial variables
Abnormal change in
-
employees 223 -0.128 -0.102 223 -0.330 -0.120 0.202 0.001 3.32
Abnormal change in order
-
backlog 68 0.031 -0.063 68 -0.001 -0.013 0.032 0.375 0.33
Off-balance sheet variables
Change in operating lease
+
activity 266 0.020 0.006 266 0.024 0.009 -0.004 0.120 -1.18
Existence of operating
+
leases 299 0.857 1.000 299 0.685 0.833 0.172 0.001 9.41
Expected return on
+
pension plan assets (%) 32 0.084 0.090 32 0.081 0.091 0.002 0.190 0.89
Change in expected return
+
on plan assets (%) 25 -0.041 0.000 25 -0.076 0.000 0.035 0.167 1.01
Market-related variables
Ex ante financing need 191 0.143 0.000 191 0.186 0.000 + -0.043 0.053 -1.63
Actual issuance 296 0.941 1.000 296 0.905 1.000 + 0.035 0.008 2.44
CFF 199 0.210 0.121 199 0.266 0.129 + -0.056 0.025 -1.98
Leverage 299 0.189 0.164 299 0.181 0.149 + 0.008 0.232 0.73
Market-adjusted stock
return 188 0.226 0.008 188 0.208 0.107 - 0.018 0.451 0.12
Book to market 245 0.555 0.363 245 0.510 0.412 + 0.045 0.141 1.08
Earnings to price 168 0.053 0.039 168 0.063 0.051 - -0.010 0.008 -2.45

In this table, we calculate means at the firm level for misstatement years versus years prior to misstatement years and
conduct tests of pair-wise differences in means. Each firm is directly compared to itself during manipulation years versus
other years. All variables are defined in Table 3. Each of the continuous variables (except stock return variables) is
winsorized at 1% and 99% to mitigate outliers.

61
Table 6
Descriptive statistics on misstatement firm-years versus Compustat firm-years for the sample from 1979 to 2002.
Misstatement firm-years Compustat firm-years Misstate – Compustat
One-
tailed
Pred Diff. in P- t-
Variable N Mean Median N Mean Median sign Mean value statistics
Accruals quality variables
WC accruals 559 0.062 0.036 152170 0.007 0.006 + 0.055 0.001 7.55
RSST accruals 557 0.126 0.074 151862 0.032 0.026 + 0.094 0.001 6.30
Change in receivables 561 0.061 0.036 151928 0.017 0.008 + 0.044 0.001 8.97
Change in inventory 557 0.039 0.008 152741 0.011 0.000 + 0.028 0.001 7.30
% Soft assets 604 0.642 0.682 167982 0.509 0.535 + 0.133 0.001 14.65
Modified Jones model
discretionary accruals 524 0.057 0.027 145472 0.000 0.001 + 0.058 0.001 4.39
Mean-adjusted
absolute value of DD
residuals 338 0.017 -0.011 81369 0.000 -0.019 + 0.017 0.001 3.16
Studentized DD
residuals 338 0.428 0.284 81369 0.001 0.015 + 0.427 0.001 6.82
Performance variables
Change in cash sales 501 0.492 0.217 135333 0.208 0.079 - 0.283 0.001 6.38
Change in cash
margin 490 0.010 0.006 132352 0.013 0.001 - -0.003 0.462 -0.10
Change in return on
assets 506 -0.024 -0.012 140380 -0.010 -0.002 + -0.013 0.082 -1.39
Change in free cash
flows 501 0.041 0.007 137686 0.021 0.005 - 0.020 0.147 1.05
Deferred tax expense 579 0.001 0.000 166164 0.001 0.000 + 0.000 0.405 -0.24
Non-financial variables
Abnormal change in
employees 489 -0.223 -0.103 134837 -0.093 -0.049 - -0.130 0.001 -3.30
Abnormal change in
order backlog 139 0.006 -0.062 35766 0.088 -0.040 - -0.082 0.130 -1.12
Off-balance sheet variables
Change in operating
lease activity 561 0.018 0.004 154469 0.009 0.000 + 0.009 0.001 4.32
Existence of operating
leases 604 0.821 1.000 168481 0.710 1.000 + 0.111 0.001 7.11
Expected return on
pension plan assets 73 0.082 0.090 22617 0.073 0.085 + 0.009 0.002 3.00
Change in expected
return on plan assets 61 -0.001 0.000 18604 -0.041 0.000 + 0.040 0.143 1.08
Market-related variables
Ex ante finance need 422 0.185 0.000 100436 0.170 0.000 + 0.015 0.210 0.81
Actual issuance 599 0.932 1.000 166712 0.826 1.000 + 0.105 0.001 10.18
CFF 434 0.210 0.101 103471 0.140 0.007 + 0.069 0.001 4.17
Leverage 604 0.201 0.171 168161 0.199 0.140 + 0.002 0.408 0.23
Mkt-adj return 463 0.194 -0.113 110303 0.008 -0.114 + 0.185 0.006 2.55
Lagged mkt-adj return 393 0.332 0.031 99197 0.030 -0.099 + 0.301 0.001 3.86
Book-to-market 548 0.529 0.348 135946 0.605 0.527 - -0.076 0.006 -2.53
Earnings-to-price 343 0.064 0.043 84911 0.084 0.066 - -0.020 0.001 -5.18
All variables are defined in Table 3. Each of the continuous variables (except stock return variables) is winsorized at 1% and 99% to mitigate outliers.
Note that even though we restrict our sample to 1979-2002 in the cross-sectional analysis, for some variables, the number of observations appears to be
slightly larger in Table 6. This is because in the time-series analysis, we eliminate those observations with data available only in either misstatement
period or in the non-misstatement period, to make the comparison meaningful.

62
Table 7 Panel A: Logistic regressions (dependent variable is equal to one if the firm-year misstated earnings, zero
otherwise) examining the determinants of misstatements.
Model 1 Model 2 Model 3
Add Off-balance sheet
Financial Statement and Non-financial Add Stock Market-based
Variable Variables Variables Variables
Coefficient Estimate Coefficient Estimate Coefficient Estimate
(Wald Chi-square) (Wald Chi-square) (Wald Chi-square)
(P-value) (P-value) (P-value)
Intercept -7.893 -8.252 -7.966
1180.0 973.1 708.5
0.001 0.001 0.001
RSST accruals 0.790 0.665 0.909
24.1 12.5 11.1
0.001 0.001 0.001
Change in receivables 2.518 2.457 1.731
28.5 23.7 7.4
0.001 0.001 0.003
Change in inventory 1.191 1.393 1.447
4.4 5.4 4.0
0.019 0.010 0.023
% Soft assets 1.979 2.011 2.265
86.3 74.3 68.4
0.001 0.001 0.001
Change in cash sales 0.171 0.159 0.160
17.0 11.6 5.9
0.001 0.001 0.008
Change in return on assets -0.932 -1.029 -1.455
19.9 20.6 20.1
0.001 0.001 0.001
Actual issuance 1.029 0.983 0.651
28.5 22.4 8.2
0.001 0.001 0.002
Abnormal change in -0.150 -0.121
employees 4.2 1.5
0.020 0.113
0.419 0.345
Existence of operating leases 8.4 4.4
0.002 0.019
Market-adjusted stock 0.082
returns 8.4
0.002
Lagged market-adjusted
stock return 0.098
8.9
0.001
Misstating Firm-years: 494 449 354
Non-misstating Firm-years: 132,967 122,366 88,032
All variables are defined in Table 3. Each of the continuous variables (except stock return variables) is winsorized at 1%
and 99% to mitigate outliers. The reported p-values are based on one-tailed tests.

63
Table 7 (continued)
Panel B: Examination of the detection rates of misstating and non-misstating firms for each model reported in Panel A.
Model 1 Model 2 Model 3
Minimum % of Minimum % of Minimum % of
N F-Score Total N F-Score Total N F-Score Total
Quintile 1
Misstate Firms 16 0.053 3.24% 13 0.042 2.90% 10 0.052 2.82%
No-Misstate Firms 26,676 0.009 20.06% 24,550 0.008 20.06% 17,667 0.008 20.07%

Quintile 2
Misstate Firms 43 0.390 8.70% 37 0.393 8.24% 36 0.386 10.17%
No-Misstate Firms 26,649 0.385 20.04% 24,526 0.351 20.04% 17,641 0.382 20.04%

Quintile 3
Misstate Firms 84 0.615 17.00% 76 0.600 16.93% 61 0.631 17.23%
No-Misstate Firms 26,609 0.614 20.01% 24,487 0.598 20.01% 17,617 0.630 20.01%

Quintile 4
Misstate Firms 99 0.936 20.04% 94 0.926 20.94% 77 0.940 21.75%
No-Misstate Firms 26,593 0.933 20.00% 24,469 0.925 20.00% 17,600 0.935 19.99%

Quintile 5
Misstate Firms 252 1.397 51.01% 229 1.418 51.00% 170 1.415 48.02%
No-Misstate Firms 26,440 0.933 19.88% 24,334 0.925 19.89% 17,507 0.935 19.89%

Note: All observations are ranked based on their predicted probabilities (F-Scores) and sorted into quintiles. Minimum F-Score
is the minimum scaled predicted probability based on estimates in Panel A to enter each quintile.

Panel C: F-Score cut-off set at 1.00

Model 1 Model 2 Model 3
Predicted Predicted Predicted
No- No- No-
Observed Misstate Misstate Misstate Misstate Misstate Misstate
Misstate 339 155 494 305 144 449 238 116 354
No-
Misstate 48,282 84,685 132,967 43,996 78,370 122,366 31,934 56,098 88,032
48,621 84,840 133,461 44,301 78,514 122,815 32,172 56,214 88,386

Misstate 68.6% 31.4% 0.4% 67.9% 32.1% 0.4% 67.2% 32.8% 0.4%
No-
Misstate 36.3% 63.7% 99.6% 36.0% 64.0% 99.6% 36.3% 63.7% 99.6%

Correct classification 63.71% (1) 64.06% 63.74%

Sensitivity 68.62% (2) 67.93% 67.23%
Type I errors 36.31% (3) 35.95% 36.28%
Type II errors 31.38% (4) 32.07% 32.77%
Notes:
(1) Correct classification is calculated as ((339+84,685)/133,461)
(2) Sensitivity is calculated as (339/494)
(3) Type I errors are calculated as (48,282/132,967)
(4) Type II errors are calculated as (155/494)

64
Table 8 Marginal effect analysis on F-Scores for each model
Panel A: Descriptive statistics
soft
rsst_acc ch_rec ch_inv _assets ch_cs ch_roa issue ch_emp leasedum rett rett-1
Mean 0.019 0.014 0.009 0.518 0.210 -0.010 0.825 -0.071 0.734 0.017 0.036
Std Dev 0.267 0.085 0.068 0.250 0.766 0.205 0.380 0.466 0.442 1.039 0.868
Lower Quartile -0.043 -0.013 -0.006 0.315 -0.048 -0.043 1.000 -0.158 0.000 -0.396 -0.370
Median 0.024 0.007 0.000 0.545 0.080 -0.002 1.000 -0.047 1.000 -0.105 -0.091
Upper Quartile 0.091 0.041 0.024 0.719 0.245 0.026 1.000 0.063 1.000 0.208 0.219
1% -0.975 -0.293 -0.253 0.030 -0.987 -0.876 0.000 -2.087 0.000 -1.000 -0.961
99% 0.974 0.328 0.268 0.961 5.459 0.837 1.000 1.442 1.000 2.989 3.016

Panel B: F-Score when one variable is at the lower and upper quartile while all other variables at set at their mean values
(except indicator variables that are set at 0 or 1)
Model 1 rsst_acc ch_rec ch_inv soft_assets ch_cs ch_roa issue
Coefficient estimate 0.790 2.518 1.191 1.979 0.171 -0.932 1.029
b
F-Score at upper quartile 0.771 0.779 0.742 1.083 0.733 0.751 0.871
a
F-Score at lower quartile 0.694 0.680 0.716 0.488 0.697 0.704 0.312
Inter-quartile marginal
change in F-Score c 0.077 0.099 0.026 0.595 0.036 0.047 0.559
F-Score when all variables are at their mean values (0.728); lower quartile (0.169); upper quartile (1.547). Change in F-
Score for joint interquartile marginal effect (moving all variables from the lower to the upper quartile) (1.378)

rsst_ soft ch_ lease

Model 2 acc ch_rec ch_inv _assets ch_cs ch_roa issue emp dum
Coefficient estimate 0.665 2.457 1.393 2.011 0.159 -1.029 0.983 -0.150 0.419
F-Score at upper quartile 0.725 0.738 0.706 1.034 0.695 0.715 0.820 0.700 0.772
F-Score at lower quartile 0.663 0.646 0.677 0.460 0.663 0.666 0.307 0.677 0.508
Inter-quartile marginal
change in F-Score 0.062 0.091 0.029 0.574 0.032 0.049 0.513 0.023 0.264
F-Score when all variables are at their mean values (0.6911); lower quartile (0.120); upper quartile (1.651). Change in F-
Score for joint interquartile marginal effect (1.531).

rsst_ ch_ ch_ soft ch_ ch_ lease

Model 3 acc rec inv _assets ch_cs roa issue emp dum ret rett-1
Coefficient estimate 0.909 1.731 1.447 2.265 0.160 -1.455 0.651 -0.121 0.345 0.082 0.098
F-Score at upper
quartile 0.737 0.723 0.706 1.088 0.695 0.725 0.774 0.698 0.757 0.702 0.703
F-Score at lower
quartile 0.653 0.659 0.676 0.437 0.663 0.656 0.404 0.680 0.537 0.668 0.664
Inter-quartile
marginal 0.085 0.064 0.030 0.651 0.032 0.069 0.370 0.018 0.220 0.034 0.039
F-Score when all variables are at their mean values (0.6908); lower quartile (0.146); upper quartile (1.679). Change in F-
Score for joint interquartile marginal effect (1.533).
a
For indicator variables such as issue, we calculated F-Score when the indicator variable = 0 for lower quartile and 1 for
upper quartile. Therefore, for Actual issuance, the marginal effect on F-Score reflects changing issue from 0 to 1.
b
For variables with negative coefficient estimates (e.g., ch_roa, ch_emp), F-Score at lower quartile reflects the upper
quartile values of these variables, and F-Score at upper quartile reflects the lower quartile values.
c
Inter-quartile marginal change in F-Score reflects the difference in F-Score for interquartile change in the predicted
direction for each variable when holding other variables at their mean values.

65
Table 8 (continued)

Panel C
Correlation between F-Score and its components:
soft
rsst_acc ch_rec ch_inv _assets ch_cs ch_roa issue ch_emp leasedum rett rett-1
Spearman
correlation
F-Score 0.340 0.455 0.353 0.685 0.302 -0.043 0.534
(Model 1)
F-Score 0.302 0.430 0.344 0.693 0.282 -0.055 0.498 -0.130 0.433
(Model 2)
F-Score 0.312 0.395 0.354 0.776 0.262 -0.059 0.349 -0.110 0.382 0.194 0.106
(Model 3)
Pearson
correlation
F-Score 0.326 0.512 0.385 0.490 0.359 -0.066 0.296
(Model 1)
F-Score 0.300 0.486 0.381 0.484 0.328 -0.074 0.272 -0.238 0.259
(Model 2)
F-Score 0.306 0.397 0.361 0.486 0.259 -0.100 0.194 -0.193 0.218 0.303 0.190
(Model 3)
All the correlation coefficient estimates are significant at a less than 1% level. All variables are defined in Table 3. Each of
the continuous variables is winsorized at 1% and 99% to mitigate outliers.

66
Table 9 Panel A: Logistic regressions (dependent variable is equal to one if the firm-year misstated earnings, zero
otherwise) examining the determinants of misstatements estimated for Model 3
(2) Variables selected
(1) Variables for Model 3 (3) Adding industry
selected for Model 3 excluding boom variables to Model 3 in
Variable 1979-1998 period years (1998-2000) Table 7
Coefficient Estimate Coefficient Estimate Coefficient Estimate
(Wald Chi-square) (Wald Chi-square) (Wald Chi-square)
(P-value) (P-value) (P-value)
Intercept -7.560*** -7.597*** -8.005***
550.9 612.0 705.5
RSST accruals 1.049*** 0.885*** 0.849***
10.9 8.8 9.8
Change in receivables 1.227** 1.639*** 1.543***
2.8 5.5 6.0
Change in inventory 1.813** 1.797*** 1.680**
4.9 5.3 5.2
% Soft assets 1.768*** 1.963*** 2.214***
33.2 45.2 64.7
Change in cash sales 0.195*** 0.188*** 0.153**
8.2 8.2 5.2
Change in return on assets -1.280*** -1.317*** -1.328***
10.5 13.5 17.3
Actual issuance 0.521** 0.501** 0.639***
4.7 4.8 7.9
Abnormal change in employees -0.122
1.5
Existence of operating leases 0.406** 0.329** 0.274*
5.1 3.7 2.5
Book to market -0.157** -0.159***
4.1 5.4
Market-adjusted stock return 0.095***
8.1
Lagged market-adjusted stock 0.092*** 0.090*** 0.079***
return 9.5 9.9 7.4
Computer 0.515***
13.2
Retail 0.443
0.5
Services 0.305**
3.1
Retail X Existence of operating -0.259
leases 0.2
Misstating Firm-years: 269 299 354
Non-misstating Firm-years: 74,510 83,090 88,032
All variables are defined in Table 3. Each of the continuous variables (except stock return variables) is winsorized at 1%
and 99% to mitigate outliers. We do not report p-values because of space limitation. ***, **, * indicate statistical
significance at the 0.01, 0.05, 0.10 level, respectively, under one-tailed tests.

67
 Table 9 Panel B: Examination of detection rates for each model reported in Panel A
Out-of-sample test - using a
hold-out sample for the Excluding boom years Adding industry variables to
time period 1999-2002 (1998-2000) Model 3 in Table 6
Min. % of Min. % of Min. % of
N F-Score Total N F-Score Total N F-Score Total
Quintile 1
Misstate Firms 3 0.358 2.80% 12 0.060 4.01% 12 0.084 3.39%
No-Misstate Firms 3,553 0.012 20.10% 16,665 0.014 20.06% 17,665 0.009 20.07%

Quintile 2
Misstate Firms 8 0.481 7.48% 29 0.427 9.70% 29 0.382 8.19%
No-Misstate Firms 3,549 0.456 20.08% 16,649 0.415 20.04% 17,648 0.370 20.05%

Quintile 3
Misstate Firms 17 0.722 15.89% 53 0.657 17.73% 59 0.613 16.67%
No-Misstate Firms 3,540 0.697 20.03% 16,625 0.652 20.01% 17,619 0.609 20.01%

Quintile 4
Misstate Firms 24 1.019 22.43% 59 0.950 19.73% 78 0.930 22.03%
No-Misstate Firms 3,533 0.977 19.99% 16,619 0.938 20.00% 17,599 0.905 19.99%

Quintile 5
Misstate Firms 55 1.414 51.40% 146 1.388 48.83% 176 1.395 49.72%
No-Misstate Firms 3,501 0.977 19.81% 16,532 0.938 19.90% 17,501 0.905 19.88%

Note: All observations are ranked based on their predicted probabilities (F-Scores) and sorted into Quintiles. Minimum F-Score is
the minimum scaled predicted probability based on estimates in Panel A to enter each quintile.

Panel C: F-Score cut-off set at 1.00

Excluding boom Add industry
Out-of-sample test years (1998-2000) variables
Predicted Predicted Predicted
No- No- No-
Observed Misstate Misstate Misstate Misstate Misstate Misstate
Misstate 79 28 107 195 104 299 237 117 354
No-
Misstate 6,778 10,898 17,676 30,119 52,971 83,090 30,560 57,472 88,032
6,857 10,926 17,783 30,314 53,075 83,389 30,797 57,589 88,386

Misstate 73.8% 26.2% 0.6% 65.2% 34.8% 0.4% 66.9% 33.1% 0.4%
No-
Misstate 38.3% 61.7% 99.4% 36.2% 63.8% 99.6% 34.7% 65.3% 99.6%

Correct classification 61.73% (1) 63.76% 65.29%

Sensitivity 73.83% (2) 65.22% 66.95%
Type I errors 38.35% (3) 36.25% 34.71%
Type II errors 26.17% (4) 34.78% 33.05%
Notes:
(1) Correct classification is calculated as ((79+10,898)/17,783)
(2) Sensitivity is calculated as (79/107)
(3) Type I errors are calculated as (6,778/17,676)
(4) Type II errors are calculated as (28/107)

68