Executive Overconfidence and the Slippery Slope to …...Even in 1937, Adam Smith recognized the...
Transcript of Executive Overconfidence and the Slippery Slope to …...Even in 1937, Adam Smith recognized the...
*University of Pennsylvania. ** University of Chicago. Correspondence to Catherine Schrand, e-mail [email protected]. The authors thank Gavin Cassar for helpful discussions. We thank Ray Ball, Ed Maydew and Greg Miller and audiences at MIT, Penn State University, the Stanford 2007 Summer Camp, and the University of Minnesota Mini-Conference on Empirical Accounting for helpful comments. Sarah Zechman thanks the Deloitte Foundation and the University of Chicago Booth School of Business for providing financial support.
Executive Overconfidence and the
Slippery Slope to Fraud
Catherine M. Schrand* and Sarah L. C. Zechman**
First Draft: July 2007
This draft: November 2008
Abstract We propose that executive overconfidence increases the likelihood that a firm commits financial reporting fraud. A manager that faces an earnings shortfall is more likely to manage earnings to overcome it if he believes the shortfall is temporary and, hence, the earnings management will be a one-off event that likely will go undetected. If performance does not improve, however, the manager, faced with reversals of prior-period earnings management and continuing poor performance, may choose to engage in the type of egregious financial reporting that the SEC prosecutes. Overconfident managers with unrealistic beliefs about future performance are more likely to find themselves in this situation. Using industry-level proxies for executive overconfidence, we find industries that attract overconfident executives have a greater proportion of frauds. Our analysis that uses firm-level proxies for overconfidence suggests that there are two types of frauds: Those associated with moderate levels of overconfidence, perpetrated by executives who ex post fall down the slippery slope, and those perpetrated by executives with extreme overconfidence that commit fraud for opportunistic reasons ex ante. Analysis of individual executives supports the notion that there are two types of overconfident executives that engage in fraud. Those with opportunistic motives are more likely to be from a founding family, have greater commitment to the firm, earn more total and have a higher percent of variable cash compensation, and are less likely to have accounting experience. Finally, we document that a matched sample of non-fraud firms do not have stronger governance mechanisms that prevent fraud. This result mitigates the possibility that it is weak governance rather than executive overconfidence that is a significant determinant of fraud.
1
1. Introduction
In the mid-1990s, papers such as DeBondt and Thaler (1995) and Daniel, Hirshleifer, and
Subrahmanyam (1998) generated renewed academic interest in the impact of overconfidence on
decision making, especially in corporate finance decisions.1 Since then, empirical studies have
documented that executive overconfidence is associated with what are viewed as distorted
financing decisions (e.g., Ben-David, Graham, and Harvey, 2007) and corporate investment (e.g.,
Malmendier and Tate, 2005). In an attempt to explain the observation that humans, in general,
and corporate executives and entrepreneurs, in particular, are overconfident despite its negative
impact on decision-making,2 a burgeoning literature models overconfidence as an optimal
endogenous behavior choice by individuals, specifically in a corporate setting (e.g., Bénabou and
Tirole, 2002; Compte and Postlewaite, 2004; Brunnermeier and Parker, 2005; Goel and Thakor,
2000; Gervais and Goldstein, 2007; Gervais, Heaton, and Odean, 2007). The upshot of this
literature is that overconfidence has benefits that make it an optimal trait overall, despite its
negative effects on particular decisions.
Our purpose in this paper is to examine whether overconfidence is associated with a
greater likelihood of one particular corporate decision – financial reporting fraud. This
prediction recognizes that financial reporting fraud often represents the escalation of minor
earnings management infractions. Earnings management in minor amounts in a given period is
likely to go undetected. If performance does not improve in the next period, however, the
manager is forced either to manage earnings in an increasing amount to cover reversals and to
1 The literature that examines behavioral biases as they relate to economic decision making under uncertainty has a long history (e.g., Kahneman and Tversky, 1979). Even in 1937, Adam Smith recognized the bias of overconfidence: “The chance of gain is by every man more or less over-valued and the chance of loss is by most men under-valued.” (Cited in Arabsheibani, de Meza, Maloney, and Pearson, 2000). 2 See Englmaier (2004), Heaton (2002), and Schultz and Zaman (2001) and the references therein, for surveys of this literature.
2
keep up the earnings expectations he has created or to reveal the poor performance. Eventually,
the level of earnings management required to hide a series of bad performance realizations can
be obtained only if the manager “cooks the books” and makes the kinds of accounting
misstatements that are prosecuted by the SEC.
We propose that an overconfident manager with unrealistic beliefs about future
performance is more likely to underestimate the need for more egregious earnings management
and thus is more likely to start down the slippery slope to fraud. As a result, he is more likely to
be in the position that egregious earnings management is the optimal choice, conditional on
getting a bad draw on realized earnings.3 This explanation for fraud assumes a moderate level of
overconfidence that leads to a naïve misperception of the earnings distribution. It does not
envision a manager who exhibits a more extreme level of overconfidence associated with
narcissism or sensation-seeking (e.g., Grinblatt and Keloharju, 2001; Puri and Robinson, 2005;
Rosenthal and Pittinsky, 2006).
Descriptive evidence on a subsample of frauds used in the paper indicates patterns in the
timing of unmanaged earnings and the associated earnings management that are consistent with
this explanation. At the firm level, we document that a subsample of fraud firms associated with
moderate levels of overconfidence (or non-opportunistic frauds) have a similar level of earnings
in the year prior to the fraud as that of a matched sample. However, going forward into Year 1
and then Year 2 of the fraud there is a decreasing pattern in unmanaged performance and an
3 The Waste Management fraud is consistent with this scenario: “Defendants' improper accounting practices were centralized at corporate headquarters. Each year, Buntrock, Rooney, and others prepared an annual budget in which they set earnings targets for the upcoming year. During the year, they monitored the Company's actual operating results and compared them to the quarterly targets set in the budget. To reduce expenses and inflate earnings artificially, defendants then primarily used "top-level adjustments" to conform the Company's actual results to the predetermined earnings targets. The inflated earnings of prior periods then became the floor for future manipulations. The consequences, however, created what Hau (the Waste Management vice president, corporate controller, and chief accounting officer) referred to as a "one-off" problem. To sustain the scheme, earnings fraudulently achieved in one period had to be replaced in the next.” (AAER 1532)
3
increasing percent of loss firms in the fraud firms that is not evident in the matched sample. In
addition, the fraud firms increase the amount of earnings management in the second year of the
fraud relative to the first year. This pattern is consistent with the increasing amount of earnings
management required to fill the growing gap between unmanaged earnings and a benchmark,
which is consistent with the slippery slope explanation.
Our evidence on the association between overconfidence and fraud comes from three sets
of analyses. First, we use industry-level and firm-level characteristics as proxies for executive
overconfidence and we predict industry level and firm level fraud propensity (“fraud prediction”
tests). Our fraud sample includes firms accused of frauds in SEC Accounting and Auditing
Enforcement Releases (AAERs) in the 1990s and 2000s.
We find that industries with high sales growth and that face significant idiosyncratic risk
have a higher proportion of frauds. These results are consistent with the locus-of-control
literature (e.g., Rotter, 1966) which predicts and finds that overconfident executives who exhibit
control-seeking behavior are attracted to work in risky, dynamic, high growth environments.
At the firm level, we create proxies for overconfidence based on prior literature that
documents that executive overconfidence is associated with corporate financing, investing, and
compensation choices (e.g., Heaton, 2002; Ben-David, Graham, and Harvey, 2007). Assuming
that an overconfident manager will be overconfident with respect to all decisions, not just their
financial reporting decisions, we use these firm characteristics to identify overconfident
executives. We find that firms with lower dividend yields, greater tax avoidance, more total cash
compensation, and a greater percent of variable compensation are more likely to commit fraud,
which is consistent with the joint hypothesis that these firm characteristics are associated with
overconfidence and that over confidence is associated with fraud. However, we find little or no
4
evidence in support of an association between the probability of fraud and capital structure or
investment decisions that have been linked to overconfidence.
In our second type of test, we select two industries – software and hardware – and create
an executive-level OC-SCORE for the CEOs of all firms in the industry. The score is a function
of the CEO’s prevalence in photographs in the firm’s annual report and the CEO’s cash and non-
cash pay relative to that of the second highest paid executive at the firm. The sample and
overconfidence score follow Chatterjee and Hambrick (2007). We find strong evidence that
higher overconfidence scores are associated with a greater propensity for fraud.
In our third analysis, we use a subsample of the AAERs identified in Erickson, Hanlon,
and Maydew (2006), and the executives named in those frauds. We compare characteristics of
the executives from the fraud sample to those of the executives from the matched sample of non-
fraud firms. The smaller sample allows for hand collection of data and a more contextual
analysis of the nature of the frauds. The characteristics of the executives at the fraud firms
relative to the matched executives are fairly similar. However, within the fraud firms there are
significant differences between the characteristics of the executives as a function of whether the
SEC suggests the motives are based on opportunism or not. Opportunistic executives are more
likely to be from a founding family and have a longer tenure with and commitment to the firm,
where commitment is measured as the number of years the executive was at the firm prior to the
fraud relative to the total number of years the firm was public prior to the fraud. Each of these
characteristics is consistent with ex ante predictions that overconfidence is associated with a
greater sense of control, personal investment in the outcomes, and self-attribution. In addition,
these executives receive more total cash compensation and a greater percent of variable
compensation. These compensation patterns are consistent with predictions that extreme
5
overconfidence is associated with higher incentive pay due to the distortion in value of such
compensation, but low or moderate levels of overconfidence are associated with lower executive
pay as less is necessary to induce effort (de la Rosa, 2007).
Further analysis suggests that frauds that the SEC does characterizes as opportunistic are
distinct. These frauds are perpetrated by executives that the psychology literature might
characterize as narcissists, or at the extreme end of the overconfidence continuum. Extremely
OC fraud-firm executives, or those that commit opportunistic frauds, are more likely than
moderately OC fraud-firm executives to be from founding families and have a greater time
commitment at the firm. These executives are less likely to be CPAs and have audit experience.
They have significantly higher total compensation and a significantly greater percentage of their
total pay is variable (bonus scaled by bonus plus salary).
Finally, we examine the association between governance mechanisms and fraud. This
analysis recognizes that overconfidence is not an undesirable trait of an executive, and may even
be an optimal trait, when considering its effects on “net” performance (Bénabou and Tirole,
2002; Compte and Postlewaite, 2004; Brunnermeier and Parker, 2005; Gervais and Goldstein,
2007; Gervais, Heaton, and Odean, 2007). If the Board or higher-level executives recognize
overconfidence and its potential impact on earnings management, then mitigating forces in the
form of better governance could be used to control the effects of overconfidence on this
particular behavior. Thus, we explore the hypothesis that all executives are equally
overconfident, but that better governance was in place to control certain executives, such that
they did not commit fraud.
In summary, we observe no significant differences between the fraud firms and the
matched sample firms with respect to commonly studied governance mechanisms including
6
block ownership, board size, board composition, and measures of the entrenchment and busy-
ness of the inside, outside, and gray board members. Combined with our earlier results, the lack
of correlation between governance and fraud may suggest that fraud firm executives are more
overconfident than those at the non-fraud firms and that better governance mechanisms were not
in place to mitigate their tendency to commit fraud. These results are not consistent with the
alternative hypothesis that fraud firm and non-fraud firm executives are equally overconfident,
but that better governance mitigated the adverse effects of their overconfidence on earnings
management decisions.
The notion that manager behavioral characteristics may be related to fraud is not new.
The fraud literature recognizes “attitudes” as a potential risk factor, with specific indicators
including an aggressive, evasive, or disrespectful manner; high turnover; undue emphasis on
meeting projections or maintaining stock price; and a poor reputation, prior irregularities, or a
history of violations (e.g., Loebbecke, Eining, and Willingham, 1989; SAS 99, 2002). These
attitudes represent outcomes of fraudulent behavior, however, rather than more fundamental firm
or executive characteristics that might be associated with it. The COSO report (1999) on
fraudulent financial reporting also concludes that characteristics such as “tone at the top” and
managerial “integrity” are important, but it provides no evidence on this issue. Our analysis
attempts to examine the relation between a more fundamental behavioral trait and fraud.
The paper is organized as follows. Section 2 develops the hypotheses and predictions.
Sections 3 summarizes the various analyses in the paper and provides descriptive evidence on the
frauds. Section 4 presents results of the fraud prediction tests using both industry-level and firm-
level proxies for overconfidence. Sections 5 and 6 present the analyses of the smaller
subsamples of overconfident executives. Finally, Section 7 concludes.
7
2. Hypothesis development
The study requires a careful definition of overconfidence. The term is used inconsistently
in the academic literature, and it is often used interchangeably with optimism. In some studies,
optimism is unrealistic (positive) beliefs about the levels of outcomes (i.e., cash flows) whereas
overconfidence is unrealistic (positive) beliefs about some aspect of the distribution of an
uncertain outcome (e.g., underweighting the likelihood of negative outcomes or underestimating
the range). Other studies use optimism to describe a dispositional trait whereas overconfidence
is used to describe a characteristic that derives from the decision-maker’s experience (e.g.,
education) or that is related to the task (e.g., the controllability of the outcome). The first
perspective dominates the economics-based literature while the second is more prominent in the
psychology-based literature.4 Throughout the paper, we refer to the behavioral bias of having
unrealistic (positive) beliefs about any aspect of the distribution of an uncertain outcome, such
that the mean is overstated, as overconfidence. Overconfidence can derive from the decision-
maker’s innate temperament or his experience.
The psychology literature and the emerging corporate finance literature on
overconfidence identify characteristics of overconfident individuals. The recent studies of
financial decisions assess overconfidence using questionnaire responses (e.g., self life
expectancy assessments in Puri and Robinson, 2005; forecasts of stock market performance in
Ben-David, Graham, and Harvey, 2007) or unique datasets (e.g., results of required
psychological tests for Finnish males in Grinblatt and Keloharju, 2001).
4 Clearly, this is a broad generalization. Weinstein and Klein (1996), for example, describe “optimism” as one-sided – underestimating the likelihood of negative outcomes (as opposed to also overestimating the likelihood of positive). Grinblatt and Keloharju (2001) attempt to distinguish two behavioral biases that may lead to similar observed decision outcomes: overconfidence (miscalibration of risk) and sensation-seeking (innate desire for greater risk).
8
One of the most studied phenomenon and robust findings in the psychology literature is
that the greater the perceived controllability of an event, the greater the tendency to exhibit
optimistic bias (Weinstein and Klein, 1996). Another common theme is that individuals are
more likely to be overconfident about the outcomes or the risk of a project when they have a
personal stake in it. Psychologists suggest that optimistic bias is a mechanism individuals use to
preserve self-esteem and the danger to one’s self-esteem is greatest if the event is one in which
the individual has a personal stake. A final common theme is that overconfident individuals
exhibit self-attribution bias: They attribute successes, especially recent successes, to ability,
while attributing failures to bad luck.
Our prediction that overconfidence is associated with fraud results because the manager
makes an earnings management decision that incorporates expectations about future earnings
management.5 Assume the manager’s expected utility is the sum of his expected payoffs in time
t and t + 1. The manager’s expected payoffs depend on current period reported earnings and
expected future reported earnings. The manager faces an expected cost associated with earnings
management – which varies with the amount of managed earnings during the period. The cost
function embeds expectations about both the probability of detection and the penalty.6
The manager makes, and anticipates making, an earnings management decision at each
financial reporting date. At any given date t, the manager chooses the amount of period t
earnings management after observing unmanaged earnings for the period. He has expectations
over the next period’s earnings and he understands the rate at which his period t earnings
management will reverse in period t+1.
5 The idea that overconfidence becomes relevant in the two period setting is based on ideas in Brunnermeier and Parker (2005) who model optimism as an endogenous choice that maximizes average “felicity” over two periods. 6 Several studies indicate that fines for fraud and restatements are large, executives face jail time, and reputational penalties are substantial (COSO, 1999; Karpoff, Lee, and Martin, forthcoming; and Karpoff and Lott, 1993).
9
At time t, if the manager has unrealistic expectations about future period unmanaged
earnings, he anticipates greater slack in period t+1 earnings. Thus, he anticipates less (or no)
earnings management at t+1, which increases the expected utility of managing earnings at t. A
manager that has unrealistic expectations at t about t+1 earnings is more likely to underestimate
the likelihood of needing to manage earnings in period t+1, and thus is more likely to be in a
position at t+1 in which egregious earnings management is the optimal choice. A similar
prediction holds if the manager has unrealistic (overconfident) expectations about the likelihood
of detection.
The assumed utility function that underlies our predictions is based on the definition of
overconfidence stated above. However, the economics and psychology literature view
overconfidence as a continuum. The economics-based literature that models self-confidence
yields an optimal level (Bénabou and Tirole, 2002; Compte and Postlewaite, 2004; Brunnermeier
and Parker, 2005), which implies that there can be a greater-than-optimal level. Empirical
studies have characterized extremely overconfident executives as “sensation seeking” (e.g.,
Grinblatt and Keloharju, 2001) or as making “imprudent decisions” (Puri and Robinson (2005)).
The psychology-based literature links excessive confidence and unrealistically optimistic
perceptions to narcissism (Rosenthal and Pittinsky, 2006; Post, 1993). Narcissism in leaders is
characterized by self-serving attitudes and the choice of actions “principally motivated by their
own egomaniacal needs and beliefs” (Rosenthal and Pittinsky, 2006).
Our slippery slope explanation for fraud envisions the moderately overconfident
executive and predicts that fraud is the outcome. An extremely overconfident executive may
increase utility by beating the system or may believe that he deserves to opportunistically cheat
shareholders. While this sort of extreme overconfidence may also be associated with a greater
10
likelihood of fraud, it is not the interest of this study. Therefore, throughout the analysis we
identify individuals that the SEC indicates engaged in the fraud based on self-serving motives
and we identify frauds for which the AAER suggests that the primary motive is insider trading
profits, bonuses, or other compensation. This distinction is important because these types of
frauds ex ante do not fit the slippery slope explanation; self-serving opportunism rather than an
unrealistic expectation is the explanation for the fraud according to the SEC.
3. Research design, samples, and descriptive evidence
We test the basic prediction that overconfidence is associated with a greater propensity
for fraud using three types of analyses, and appropriate samples, as described in the following
three sections.
3.1 Fraud prediction tests
We use industry-level and firm-level proxies for executive overconfidence to predict the
proportion of frauds within industries, defined at the three digit level, and within firms,
respectively. We refer to these tests as the “fraud prediction” tests. In the industry-level
analysis, we assume that overconfident managers will self-select to certain industries. Hence,
such industries are more likely to have overconfident managers, and we predict a greater
proportion of frauds in such industries. By estimating the proportion of frauds in the industry,
these tests control for motives and opportunities for fraud that vary with industry characteristics.
In the firm-level analysis, we model the likelihood that a firm will engage in fraud as a
function of firm characteristics that proxy for the overconfidence of the firm’s executives. We
assume that executives that are overconfident will exhibit overconfidence with respect to all of
their decisions, not just their financial reporting decisions. Empirical and survey literature links
11
overconfidence to corporate finance and investment decisions, such as capital structure. Thus,
we use these other firm characteristics as a proxy for the likelihood that the firm has
overconfident executives.
In both the industry-level and firm-level fraud prediction tests, we use the sample of
misstating firms from Dechow, Ge, Larson, and Sloan (DGLS, 2008). Dechow et al. create a
sample of firms subject to SEC AAERs between 1982 and 2005. They find 447 firms subject to
at least one annual period of manipulation, of which 350 have stock price data available on
Compustat. Of these 350 firms, 121 have all the required data for our analysis.
3.2 Matched sample tests
The subsample of frauds that we analyze for the more contextual analysis in a matched
sample research design is 49 of the 50 sample firms from Erickson, Hanlon, and Maydew (EHM,
2006) that were subject to AAERs from January 1996 to November 2003.7 EHM do not
eliminate firms in particular industries or impose other data constraints that induce obvious
selection bias in this sample, but it exhibits some industry clustering. Eleven firms (22%) are in
the 2-digit DNUM covering computer programming, data processing, and miscellaneous
business services and six firms are in the computer and office equipment category.8
Table 1 provides a summary of the smaller subsample of frauds. Revenue recognition is
the most frequent fraudulent activity. Over 50% of the revenue recognition cases involve
premature revenue recognition. This pattern in our subsample is consistent with evidence in
7 We eliminate Thor Industries from their sample. The AAER accuses a subsidiary-level controller of managing earnings to hide his theft of cash from corporate-level executives. The SEC does not accuse the firm of securities law (10b-5) violations. 8 Prior research on fraud has noted industry concentrations but using different benchmarks to identify a concentration. For example, there is over-representation in banking, high-tech, and savings and loans and under-representation in education, government and other not-for-profit among a particular set of audit clients (Bell and Carcello, 2000); high frequencies (raw counts) in computer hardware and software, other manufacturing, financial services, and healthcare (COSO, 1999) and high-tech industries (Dechow et al., 1996). The studies speculate that greater opportunities for fraud given the nature of the activities in these industries could explain the patterns.
12
larger samples in Dechow, Sloan and Sweeney (1996), Farber (2005), and COSO (1999). The
most frequently alleged goal (84%) is overstatement of earnings. These two facts are important
because they are consistent with the types of earnings management activities and objectives that
are envisioned in the slippery slope explanation for fraud. The extreme firms engage in
fraudulent behavior with respect to multiple accounts and they are more often accused of
reporting fictitious as opposed to just premature revenue. They also are more likely to have as a
goal hiding debt off the balance sheet.
The “fraud period” is from the first periodic report in which the SEC alleges fraudulent
reporting to the last period in which the SEC alleges fraudulent behavior.9 The average fraud
period is 3.78 years with a minimum of one year and a maximum of eight years.
To create a matched sample of non-fraud firms, each fraud firm is matched to a firm in
the same industry on the basis of firm size, measured as total assets (ASSETS) as of the end of
the year before the earnings management began (CLEANYR). Foreign firms, firms missing
DNUM classifications, and firms that are not included in the Compustat and CRSP databases
from the CLEANYR to the final year of the fraud are excluded from the potential pool of match
firms. We insure that each matched firm is itself not the subject of an AAER during the fraud
period.10
Matching on industry attempts to control for the firm’s opportunity to manage earnings
given that revenue recognition and asset and expense over and understatement are the primary
fraud vehicles. Matching on size is important because the sample represents firms selected by
the SEC to be accused of fraud and size may be correlated with SEC scrutiny. Beneish (1999)
also discusses this issue and matches on firm age to control for SEC scrutiny.
9 For all but one firm, the periodic reports are the 10-K or 10-Q. For Microstrategy, the fraudulent reports are IPO and SEO registration statements. 10 Two of the original matched sample firms were replaced as a result of this procedure.
13
This procedure generates a match at the four-digit level with total assets within 10% for
27 of the 49 firms. We match at the three-digit level (two-digit level) for 15 (7) of the firms.11
The AAER suggests that the primary motive is insider trading profits, bonuses, or other
compensation for 13 frauds, and we set an indicator variable OPP_FRAUD equal to 1. Of the
remaining 36 frauds, the AAER suggests that the firm was attempting to meet either internal
targets such as budgets or external targets such as Wall Street forecasts or investor expectations
for 26. The AAER emphasizes external financing transactions as a motive for six firms. Other
motivations include influencing merger transactions, moving exchanges, or hiding details that
would reveal bad business decisions (e.g., credit losses/store closings).
Table 2 shows that the fraud firms are not significantly different from the matched
sample with respect to size measured by total assets (ASSETS) at the end of CLEANYR, which
was the matching criterion. They also are not different with respect to size measured by net sales
(SALES) or the market value of the firm (MVSIZE), which is the sum of the market value of
equity, the book value of long-term debt, and the book value of preferred stock.
Table 2 also reports characteristics of the samples that prior literature has identified as
determinants of fraud. We include these comparisons to assess the ability of the matched sample
to control for opportunities, incentives, and motives to commit fraud that are not related to
overconfidence. All variables are measured during or at the end of CLEANYR. We draw the
control variables from three primary (recent) sources that study fraud determinants: Beasley
(1996), the COSO report (1999), and EHM (2006). (Appendix A provides details on
construction of the variables defined in this and subsequent sections.)
The first set of variables relates to pre-fraud performance, broadly speaking, which may
11 The Compustat DNUM for Enron Corp. is 5172 (Petroleum, Ex Bulk Statn-Whsl); the DNUM for Enron Corp.-Old is 4923 (Natural Gas Transmis and Distr). Enron Corp. is matched to Public Service Enterprise Group, Inc., in DNUM 4923, because there were no reasonable size matches to Enron in DNUM 5172.
14
be associated with greater incentives to commit fraud. We measure operating performance using
return on assets; net income scaled by sales; and income before extraordinary items scaled by
sales. Following EHM (2006), we measure financial health using a market-based debt-to-equity
ratio. Following Beasley (1996), we measure financial health as a binary variable that equals 1 if
the firm had at least three loss years during the six year period preceding the fraud and equals
zero otherwise.12 We also include four measures of financial health that have been used in other
contexts: S&P bond ratings; the current ratio; the quick ratio; and the interest coverage ratio. As
summary measures of operating performance and financial health, we use the book-to-market
ratio and the earnings/price ratio following EHM (2006). Consistent with COSO (1999), we
measure pre-fraud performance based on stock returns.
The overall conclusion from the comparison of these variables is that the matched sample
effectively controls for performance-related earnings management incentives. The median
values of NISALES and IB4XSALES are not significantly different. The means are significantly
lower for the fraud sample due to several extreme negative observations at high-tech startups (-1,
0, and 3 years old) who are matched to more established firms. The fraud firms’ coverage ratios
also are marginally lower, although their debt-equity ratios are not higher, which is consistent
with the earnings deficiencies relative to the match firm sample. The returns are greater for the
opportunistic fraud firms than their matched sample, though not for the non-opportunistic firms
relative to their matches. Finally, the volatility of returns does not differ between either of the
fraud groups and their respective match samples.
The second set of control variables for fraud determinants are proxies for external
financing demands. The non-opportunistic fraud firms are significantly younger than their
12 Results for similarly constructed TROUBLE variables over shorter windows (2 years and 4 years) to maximize data availability are similar.
15
matched sample,13 and they have significantly lower free cash flow consistent with the
predictions of EHM (2006) and Beasley (1996). However, their merger activity is not
significantly different. These results suggest that the fraud sample firms may have greater
external financing demands, and thus greater incentives to commit fraud. The opportunistic
fraud firms do not differ significantly from their matched sample on any of these dimensions.
The third set of control variables measures the nature of the firm’s operations and growth
to assess potential differences in the motivation and opportunity to manage earnings.14 The non-
opportunistic fraud firms have significantly higher asset and sales growth than their matched
sample, especially in recent periods. This difference is worth noting because Ben-David et al.
(2007) find a positive association between overconfidence and 5-year sales growth, which they
use to represent past performance. In summary, the comparison indicates that the non-
opportunistic sample may have greater financing needs and growth than their matched sample,
which suggests that differences in incentives and motives to commit fraud may still be a concern
in the subsequent analysis.
3.3 Time-series patterns in earnings management and unmanaged earnings
The slippery slope explanation for fraud assumes that the realization of unmanaged
earnings in the second year of the fraud is more likely to be lower than expected for firms with
overconfident managers. Thus, for the slippery slope explanation to hold, it should be the case
that unmanaged earnings in Year 2 of the fraud deviate more from the benchmark than did Year
1 earnings. Assuming a fixed benchmark, this leads to a prediction of a decrease in unmanaged
13 For three fraud firms, AGEPUB equals -1, which indicates that the first year of the fraud was the year of the IPO. For one fraud firm (Finehost), AGEPUB = -5; the fraud started prior to their 1996 IPO. 14 We also examine external auditors as a potential determinant of (deterrent to) fraud, consistent with the early literature on fraud determinants, however, there is little variation in either sample. All but four of the fraud firms use Big-8 auditors, two use named second tier firms. All but five of the fraud firms use Big-8 auditors, two use named second tier firms.
16
earnings from Year 1 to Year 2. It should also be the case that the dollar amount of earnings
management increases from Year 1 to Year 2 as more egregious earnings management is
required to fill the gap. These predictions are unique to the slippery slope story. We do not
expect such a pattern for the frauds associated with opportunism.
Table 3 first reports the time-series changes in the dollar amounts of managed revenues
and net income (Panel A) for the subsample of EHM frauds. These amounts are hand collected
from reported restatements in 10-K filings and cannot be observed for the full sample or the non-
fraud sample. The dollar amount of the earnings management escalates from one period to the
next. The mean (median) amount of revenue and earnings restatement of the non-opportunistic
fraud firms increases by 221% (97%) and 40% (18%) between the first year and the second year
of the fraud. While the increasing pattern in earnings management is not direct evidence of the
slippery slope explanation for fraud, it is nonetheless consistent with the story. The same change
in revenue restatements in not observed for the firms with opportunistic frauds.
Panels B and C report patterns in unmanaged performance-related variables.
Performance is measured as unmanaged net income in dollars. For all years for the match
sample firms, and for CLEANYR for the fraud firms, performance equals net income (data item
#172) from Compustat. For the fraud firms in Years 1 and 2 of the fraud, performance equals to
net income less restatement amounts if restatement amounts are available and it is missing
otherwise. %LOSS FIRMS is an indicator equal to 1 for firms with negative unmanaged
performance and equal to 0 otherwise.
The non-opportunistic fraud firms have a median decline in unmanaged earnings from
CLEANYR to Year 1 of 33%, which is consistent with the need to manage earnings initially.
Unmanaged earnings decline by a greater amount (86%) from Year 1 to Year 2. A similar time-
17
series pattern is evident from the relative percents of loss firms in the samples (Panel C).
The opportunistic fraud sample, however, does not exhibit the same deterioration in
unmanaged earnings in Year 2. While they show a decrease in median unmanaged earnings of
5% from CLEANYR to Year 1, there is an increase of 65% from Year 1 to Year 2 and the
percent of loss firms declines in Year 2. Thus, the opportunistic firms continue the fraud, and
even elevate it as noted in Table 1, despite the fact that performance was improving. The
matched samples also do not exhibit a pattern of deterioration in performance during this period.
4. Fraud prediction tests
4.1 Industry-level proxies for overconfidence
We hypothesize that overconfident executives self-select to work in certain industries,
and thus there will be a greater likelihood of fraud in these industries. The locus-of-control
literature (e.g., Rotter, 1966) provides a basis for identifying industries that attract overconfident
executives. This literature distinguishes “internal” individuals who perceive control over the
outcomes in their lives from “external” individuals who believe outcomes are beyond their
control. Broadly speaking, the studies suggest that internals – or control-seeking individuals –
are attracted to innovation and risk-taking, dynamic/heterogeneous environments, and jobs that
require proactive rather than passive management. Based on the link between overconfidence
and control, we predict that overconfident executives are more likely to cluster in industries that
require innovation and risk-taking, have dynamic/heterogeneous environments, and jobs that
require proactive rather than passive management.
We use the following proxies to measure these industry characteristics. We measure
industry-level innovation by intangibles intensity and capital expenditures. We proxy for risk-
18
taking using return volatility. The concept of a growing or more dynamic work environment is
measured by growth in sales and in operating cash flows, and by whether growth is attributed to
merger activity. We also measure the length of the operating cycle, consistent with the notion
that internals are attracted to “futurity” (Miller, Kets de Vries, and Toulouse, 1982).15 Finally,
we measure heterogeneity and complexity based on the number of business (or geographic)
segments. We measure each variable at the firm level on an annual basis and compute the
median and mean of the observations at the 3-digit industry level. Values for industries with less
than five annual observations are set to missing. We use the industry medians in the year 1995,
the middle of the period.16
Finally, we predict that overconfident executives, who seek control, are attracted to
industries in which the returns are driven by idiosyncratic rather than macroeconomic factors.
Our proxy for this industry-level characteristic is the average variance about a market model
regression estimated using monthly returns over the period 1988-2002 for firms in each three-
digit industry (NONSYNC). This industry-level return synchronicity measure follows the
country-level return synchronicity measure in Li, Morck, Yang, and Yeung (2003). Higher
values of NONSYNC indicate less synchronicity in firm stock returns within the industry, which
suggests more individual control over outcomes for the executive. Thus, we predict a positive
association between NONSYNC and fraud propensity. Anecdotally, a notable feature of the fraud
sample and earlier studies is that frauds by firms in commodity-based industries, in which
15 Miller, Kets de Vries, and Toulouse (1982) test the internal/external theories using proprietary proxies for firm characteristics that would attract internals. Their proxies include a “technocratization” score, which reflects the prevalence and importance of engineers, scientists and the like; a “scanning” score, which reflects the need to understand shifts in customer tastes (a “dynamic” environment); and a “futurity” score which reflects a longer planning horizon. 16 Results for grand averages of the annual medians or means over the period 1990-2004 for all variables except NUMSEG, for which we compute the grand averages pre and post-SFAS 131, are similar.
19
managers are subject to price/demand/supply shocks that are outside their control, are rare.17
We estimate the relation between these industry-level characteristics and industry level
fraud concentration. Fraud concentration in industry i ( %FRAUDi ) equals the total number of
number of firms in industry i subject to enforcement action by the SEC scaled by the average
number of firms in the industry from 1990 – 2004. We also estimate the relation between the
industry-level characteristics and restatement concentration based on the number of firms in the
3-digit industry with financial restatements between January 1997 and September 2005, as
identified by the Government Accountability Office (GAO). %GAOi equals the total number of
restatements in industry i scaled by the average number of firms in the industry from 1997 –
2005. Because the dependent variable is the proportion of firms in the industry that misstate, our
primary model specification uses an arc sine-square root transformation of the dependent
variable.18
Table 4 reports the results. Industries with less return synchronicity (higher values of
NONSYNC) have greater percentages of AAERs and restatements (NONSYNC is significant at
p-value<0.10 and <0.01, respectively). Assuming that low-synchronicity industries attract more
overconfident managers, this positive relation is consistent with more overconfident managers
having a greater propensity for fraud. Industries with high sales growth have a higher proportion
17 Bell and Carcello (2000), based on SAS 53, predict that a high “sensitivity of operating results to economic factors” is a risk factor for fraud because such firms have greater incentives for earnings management, which is opposite to our conjecture. They find no evidence, however, that auditors view sensitivity to economic factors as an audit risk factor. 18 The arc sine transformation is one of two commonly suggested transformations for models of proportions data with multiple “trials” across a treatment group (e.g., Draper and Smith, 1998). The other commonly suggested transformation is the log-odds transformation of the dependent variable, and a WLS estimation of the model; observations are weighted by the reciprocal of the variance of the log-odds ratio. The log-odds transformation is less appropriate for our sample given the large number of industries with no frauds or no restatements. In the log-odds transformation, such observations are assigned an arbitrarily small proportion > 0. We estimate the models using the log-odds transformation and it provides similar inferences. We also estimate separate Tobit regression models of %FRAUDi and %GAOi and a logistic regression model using a logistic transformation of the proportions data. Industries with no frauds or no restatements during the sample period are assigned an arbitrarily small proportion = 0.0001. The results are consistent from these alternative model specifications.
20
of frauds. At the same time, there is a negative and statistically significant relation between
operating cash flow growth and both AAERs and restatements. Taken together, one
interpretation of these findings is that industries that are dynamic and have high sales growth are
attractive to overconfident executives, but the current low operating cash flow growth provides a
greater motivation to manage earnings.
A weakly significant positive coefficient on the number of geographic segments suggests
that the frauds are clustered in more complex industries. We find no support for the notion that
frauds or restatements are clustered in industries that provide more innovation, risk, or “futurity”
as measured by the operating cycle.
4.2 Firm-level proxies for overconfidence
We hypothesize that overconfident executives are likely to be consistently overconfident
in all of their corporate finance and investment decisions, and thus we can use firm
characteristics to proxy for overconfidence.19 We use five firm characteristics to identify firms
that are likely to have overconfident executives: Capital structure, dividend policy, tax
avoidance, capital expenditures, and executive compensation. Recent studies suggest that
overconfidence is associated with these firm-characteristics. If the firm’s managers are
overconfident with respect to these other corporate decisions, then we expect they also will be
overconfident with respect to their financial reporting decisions.
Related to capital structure choice, we hypothesize that overconfident executives are
associated with a stronger pecking order preference and riskier debt. Studies that link capital
structure choice to managerial overconfidence suggest that overconfidence can have two effects
on inputs to a firm’s optimal decision about financing corporate investment (Heaton, 2002;
19 Ideally, we would predict only those frauds in which the CEO or CFO was involved. We are in the process of getting these data.
21
Hackbarth, 2007). An optimistic manager overestimates cash flows from investment projects.
An optimistic manager also believes external financing costs (both debt and equity) are too high.
Predictions about the net impact of these two forces on capital structure choice depend on the
relative degree of optimism about each. Heaton (2002) generates standard pecking order
preferences by optimistic managers. Hackbarth (2007) generates predictions of a pecking order
preference if the first effect dominates and a reverse pecking order preference if the second effect
dominates.
Our predictions of a stronger pecking order preference and riskier debt for the fraud firms
are consistent with the survey results of Ben-David et al.(2007) who find that overconfident
executives are at firms with higher debt ratios and longer debt duration, which are their proxies
for riskier debt. We examine two capital structure variables. The first proxy is the debt-to-
equity ratio which is long-term debt scaled by the market value of the firm. The second proxy is
an indicator variable equal to 1 if the firm uses either convertible debt or preferred stock.
Related to dividend policy, we use a low dividend yield as a proxy for overconfident
executives.20 Ben-David et al. (2007) find that overconfident executives are less likely to pay
out dividends. Their explanation is that overconfident executives preserve cash because they
expect to have valuable investment opportunities.
Related to investment policy, we examine capital expenditures. Ben-David et al. (2007)
expect overconfident managers to invest in more projects than non-overconfident manager.
Overconfident managers underestimate the risk of the project, thus estimate a lower discount rate
and perceive a greater number of projects to have positive net present values. Consistent with
this prediction, they find firms with overconfident managers invest more in capital expenditures.
We measure capital expenditures as the log of (capital expenditures plus 0.001). 20 Our results are robust to using an indicator variable that equals 1 if the firm pays dividends and 0 otherwise.
22
Related to tax aggressiveness, we hypothesize that more aggressive executives are more
overconfident. We measure tax avoidance following Dyreng, Hanlon, and Maydew (2008) who
suggest that a low cash-based effective tax rate represents aggressive tax avoidance. CASHETR
is the ratio of cash taxes paid during the first year of the fraud to pretax income excluding special
items. Following Dyreng et al., CASHETR is set to missing for firms that have a negative
denominator, and values less than zero (greater than one) are truncated to zero (one).
Related to compensation, the optimal contract with an overconfident manager trades off
two forces: overconfident managers “overvalue” success-based incentive pay, but less incentive
pay is necessary to induce effort from them (de la Rosa, 2007). This model predicts that extreme
overconfidence is associated with higher incentive pay because of the manager’s distorted
preference for success-based pay, but low or moderate levels of overconfidence are associated
with lower incentive pay.21 We examine two compensation variables: The natural log of salary
plus bonus (CASHCOMP) and the ratio of bonus to salary plus bonus as a measure of variable
pay (VAR$PAY).
While we include compensation proxies in these large sample fraud prediction tests, the
predictions are ambiguous because the sign of the relation depends on the degree of the
manager’s overconfidence. We are able to test the compensation hypothesis more directly in the
matched sample. However, inclusion of the compensation variables in the fraud prediction tests
is nonetheless important as a control variable. If overconfident executives have more variable
compensation and if variable compensation creates greater incentives for earnings management,
then compensation is a potential omitted variable in our analysis. The severity of this potential
problem is not clear, however, because empirical evidence on the relation between compensation
21 Gervais, Heaton, and Odean (2007) predict that moderately overconfident executives require less convexity in an optimal contract as some overconfidence reduces the classic principal/agent conflict and better aligns manager’s incentives with shareholders. However, for extremely overconfident executives, more convexity is optimal.
23
and fraud suggests a weak relation, if any.22
The capital structure, dividend, investment, and compensation variables are measured for
year t - 1 for non-fraud firms and in the year prior to the first year of the alleged fraud for the
fraud sample. The CASHETR is computed for the first year of the fraud for the fraud sample
and for year t for the non-fraud sample, as we expect that the earnings management and tax
avoidance are concurrent.
In summary, we predict that the likelihood of fraud will be increasing with debt-to-equity
ratios, risky debt, aggressive tax practices (or low CASHETR), capital expenditures, and total
and variable compensation and decreasing with dividends, because these firm-level
characteristics are associated with overconfidence. We test these predictions using a logistic
regression model. The dependent variable equals one in the first year of the fraud for firm i and
equals zero for all firms that are not the subject of an AAER in any year from 1989-2001. We
control for firm size in the model because the AAER sample represents firms selected by the
SEC to be accused of fraud and size may be correlated with SEC scrutiny. We control for pre-
fraud performance using return on assets and book-to-market ratios because it may be associated
with greater incentives to commit fraud. We control for firm growth, in sales and assets, because
the nature of the firm’s operations and growth may be associated with differences in motivation
and opportunities to commit fraud. Finally, we include free cash flow as a proxy for external
financing demands.23 Standard errors are clustered by firm and fiscal year.
Table 5 reports the results. Consistent with the notion that overconfident executives are 22 For example, Dechow, Sloan, and Sweeney (1996) do not find a relation between earnings-based bonus plans and fraud, and EHM (2006) do not find evidence of any relation between fraud and the sensitivity of equity to stock price. Johnson, Ryan, and Tian (2008), however, find that fraud executives receive greater total compensation in the fraud years and that the sensitivity of their unrestricted holdings is higher than that of the matched sample. Related studies of restatements find a positive association with option-compensation delta (Burns and Kedia, 2006) and levels of in-the-money option holdings (Efendi, Jap, Srivastava, and Swanson, 2006). 23 Results are robust to alternative controls for firm size (SALES or LOGMVE), growth (ASSETGRO4, PPE, or RDINTENSE), and external financing demands (MERGE).
24
less likely to pay out dividends, we find that the fraud firms have a lower dividend yield than
non-fraud firms. When the sample is restricted to firms with compensation data available in the
ExecuComp database (columns 3-4), we find that fraud firms are more likely to be aggressive on
taxes. Finally, fraud firms pay more total cash compensation and a great percent of variable
compensation than the firms without AAERs. Capital expenditures and capital structure are not
related to fraud propensity. Contrary to the findings of Ben-David et al. (2007), we find a
negative relation between DERATIO and fraud propensity. However, when we use their metric
of the ratio of total debt to total assets, this negative relation disappears. Further, in untabulated
univariate tests between the EHM fraud sample and match firms, there is no difference in either
of these variables across the samples.
5. Subsample analysis of overconfident executives
In this section we focus on a small subsample of executives for which we compute an
executive-specific measure of overconfidence. We select a sample of executives in the software
(primary SIC 737) and hardware (SIC 357) industries between 1992 – 2004 for which we are able to
create a proxy for their overconfidence following Chatterjee and Hambrick (CH, 2007).24 The
focus on a single industry is necessary because of the significant hand-data collection for the
photoscore, and the restriction on the sample period is necessary because we require a digital form of
the annual report. CH focus on these two industries because they had a large number of publicly held
firms to aid in data collection and the industries are not subject to a high degree of external
constraints or regulations to allow for greater variation in managerial characteristics. The CH
industries happen to be appropriate for our study because of the cross-sectional variation in fraud
24 Malmendier and Tate (2008) use another ex ante measure of overconfidence based on the text of news sources that describes a CEO as “optimistic” or “confident.”
25
frequency within them. The industries have nine firms accused of fraud in an AAER during our
sample period and 98 firms not accused. Thus, given that the measures used by CH have
established validity for this industry, we focus on the same industries.
Our sample contains 111 executives from 107 companies over the sample period.25
Similar to CH, we required that a CEO had to have begun their tenure in 1991 or later and have a
minimum of four years tenure with the firm.
For each of the 111 executives in the sample, we create a proxy for overconfidence based
on two inputs: 1) the CEO’s prevalence in photographs in the firm’s annual report, and 2) the
CEO’s cash and non-cash pay relative to that of the second highest paid executive at the firm.
The photoscore is specified as follows: four points if the CEO’s photo in the annual report
included no other individuals and was at least equal to one half page in size; three points if the
CEO’s photo included no other individuals and was less than one half page in size; two points if
there were other individuals pictured with the CEO; and one point if there was no photograph of
the CEO. Relative cash compensation is equal to the ratio of the CEO’s salary plus bonus to the
salary plus bonus of the second highest paid executive. Relative non-cash compensation is
similarly measured using total compensation on Execucomp (TDC1) less cash compensation.
The combined score “OC-SCORE” is equal to the sum of the standardized values of the
photoscore and relative compensation variables. Each variable is equal to the average across the
CEO’s second and third years of tenure with the firm.26 Data are available to compute an OC-
SCORE for 102 of the 111 executives. We estimate the probability of fraud at the executive’s
firm as a function of the OC-SCORE in a logit regression model. The model includes control
variables consistent with the firm-level fraud prediction analysis.
25 This is similar to the sample of 111 CEO’s in 105 firms used by CH. 26 Results are qualitatively consistent using an alternative relative non-cash proxy measured using the Black Scholes value of options granted.
26
Table 6 presents the results. While all standardized components of OC-SCORE are
positively correlated with the occurrence of fraud, only the relative non-cash compensation
variable based on total compensation less cash compensation is significant at traditional
significance levels (Column 1). When the variables are combined into OC-SCORE, the results
are consistent with overconfidence increasing the propensity for fraud, as defined by the
existence of an AAER (Columns 2 and 3). The OC-SCORE is positively associated with fraud
(p-value < 0.05). The sample size is less than the original 102 firms due to missing data.
6. Subsample analyses of the EHM frauds
6.1 Analysis of accused executives in the EHM subsample
We compare the accused executives of the EHM subsample of fraud firms to their
counterparts in the matched sample firms. The AAERs and related litigation releases identify
224 respondents that are employees of the 49 fraud firms. Of these, we classify 107 as
executives and 117 as non-executive employees that do not appear in proxy statements. Forty of
the frauds involve at least one executive, and the average number of executives for these 40
frauds is 2.7.
We match each fraud firm executive to an executive at its matched firm. The matching is
not straightforward because organizational structures/titles differ across firms. Our goal is to
match on the executive’s level of decision-making authority over a particular kind of decision
(e.g., financial vs. operational).
For the 107 fraud firm executives, we are able to find reasonable matches for 99.27 Table
27 The fraud sample contains 26 CEOs, 23 of which are matched to CEOs. The fraud sample contains 35 financial executives (CFOs, controllers, chief accounting officers, VP-finance, or other financial titles), 32 of which are matched to similarly titled financial executives. Board membership was not a requirement for matching, but ex post there is moderate correlation between the two samples. Of the 57 fraud firm executives that are on the BOD, 32
27
7 Panel A presents univariate comparisons of executive-level proxies for overconfidence across
the fraud firm and matched sample executives. Panel B presents comparisons within the sample
of 107 accused executives at the 49 EHM fraud firms. There are 32 accused executives at firms
in which the SEC characterizes the fraud as opportunistic (OPP_FRAUD = 1) and 75 accused
executives at firms for which OPP_FRAUD = 0.
Panel C distinguishes individual executives that the SEC characterizes as opportunistic.
If an AAER reports that the accused was required to disgorge funds directly linked to fraud-
related trading or specifically alleges self-serving compensation motives for the fraud, the
indicator variable OPP_EXEC equals 1, and equals zero otherwise.28 This person-level variable
is distinct from the firm-level variable (OPP_FRAUD), but they are related. Eleven of the 13
firms that primarily have an opportunistic motivation have at least one respondent with insider
trading allegations against them. However, 11 of the 36 firms not classified as primarily
opportunistic also have at least one opportunistic respondent. There are 40 accused executives
defined as opportunistic (OPP_EXEC = 1) based on insider trading allegations and 67 accused
executives for which OPP_EXEC = 0.
Finally, we separately examine the characteristics of the 73 EHM sample firm executives
that the SEC indicates “orchestrated” the fraud, the 23 that “participated” in the fraud, and the
seven that were “wreckless in not knowing” about the fraud. 29 Due to small numbers, we do not
present the results of this analysis, but the distinction provides additional context for interpreting
some of the tabulated results. match to executives on the match firm’s BOD. Of the 20 fraud firm executives that are BOD chairs, 13 match to executives that are BOD chairs. 28 If an AAER states that the executive is required to disgorge funds but does not specifically link the disgorgement to the repayment of ill-gotten gains from insider trading, then the executive is not classified as opportunistic. 29 Based on discussion in the AAER, we are able to determine the role for all but four executives. This classification is less subjective than it appears. The terms in quotations (orchestrated, participated, and wreckless in not knowing) are frequently used in the AAERs. Even when these terms are not used, the classification is fairly straightforward. For example, AAER 1749 states that executive x, “acting at the direction of executive y, …”
28
We measure several individual characteristics of the executives to proxy for
overconfidence. Broadly speaking, the proxies are observable individual characteristics, such as
education, that prior literature has linked to overconfidence. As discussed previously, this
literature indicates that overconfidence is associated with controllability, a sense of personal
investment in outcomes, and self-attribution. Before we describe the specific proxies, it is worth
noting that all of the proxies for overconfidence are endogenous at some level. For example,
there are documented patterns between education and overconfidence, but it is not clear whether
overconfidence leads individuals to attain greater education levels or whether education leads to
overconfidence. While this endogeneity is emphasized in the psychology literature, it is not
relevant to this study as long as the proxy variable measures the degree of an executive’s
overconfidence at the time of the earnings management decision, and its determinants are not
correlated with determinants of fraud.
Our first executive-level characteristic is founding family status. FOUNDER equals 1 if
the executive is a founder or co-founder of the firm, or is part of the founding family, and equals
zero otherwise.30 Consistent with the finding that entrepreneurs tend to be overconfident (e.g.,
Bernardo and Welch and the cites therein, 2001; Puri and Robinson, 2005), we propose that
executives that are founders or members of founding families are more likely to be
overconfident. Founders, however, also may have greater private benefits of control.
Table 7 Panel A reports that fraud firm executives are marginally more likely to be from
30 Earlier studies of fraud/restatements/earnings management have considered founding family status, however, they measure this variable at the firm-level – whether the top executives (or CEO) at the firm are founders – not at the level of the executive involved in the fraud. That specification confounds interpretation of the results because founders may have different incentives to manage earnings than non-founders, but they also may have different incentives to monitor earnings management by others. Dechow, Sloan, and Sweeney (1996) suggest that a founder-CEO implies weak governance (greater influence combined with less accountability) and find a positive association between fraud and CEO-founders, and Loebbecke, Eining, and Willingham (1989) find that closely held firms are more likely to be charged with fraud. Agrawal and Chadha (2005), however, find that the probability of an accounting restatement is lower for firms in which the CEO is part of a founding family, and Wang (2006) finds firms with founding family members in positions of influence are associated with higher quality earnings.
29
founding families (p-value of 15%). Further analysis suggests that executives that orchestrate
the fraud are significantly more likely to be part of a founding family relative to executives who
were merely wreckless (not tabulated). Founding family influence is more related to
opportunistic frauds than non-opportunistic frauds (Panel B) and to self-serving fraud behavior
(Panel C). These results, however, are due to the four founding family members involved in the
Adelphia fraud. Excluding these executives, the percentages of opportunistic frauds involving
founding family members is higher than for the non-opportunistic frauds, but the difference is
not significant in either classification for opportunism.
Our second executive-level characteristic is the executive’s tenure at the firm. We
predict that an executive with longer tenure is more likely to have pet projects about which he
would be overconfident. We measure the executive’s “commitment” to the firm two ways: 1) as
a continuous variable equal to the number of years between the executive’s start date at the firm
and the first year of the fraud (TENURE) and 2) as the ratio of the number of years between the
executive’s start date at the firm and the first year of the fraud scaled by the number of years the
firm has been public at the first year of the fraud (COMMIT).31 COMMIT equals 1 for
executives who are with the firm at the time of listing, for which COMMIT would have been
greater than 100%. When we cannot determine the exact start year, we substitute the earliest
year that we could determine the executive was with the firm.32
Table 7 Panel A indicates that the fraud firm executives have significantly shorter tenures
overall (TENURE) and as a percentage of the firm’s life (COMMIT).33 This pattern for the
31 When we cannot determine the exact start year, we substitute the earliest year that we could determine the executive was with the firm. We made this substitution, which will create an understatement of tenure levels, to maximize available observations. The results are not sensitive to this substitution. 32 We made this substitution, which will create an understatement of tenure levels, to maximize available observations. The results are not sensitive to this substitution. 33 Three executives have negative tenures. Two arrived and escalated existing earnings management (revenue recognition) practices; one arrived at the firm and participated in an already egregious situation. None of them were
30
general sample is contrary to the prediction that the greater overconfidence is associated with
fraud, if overconfidence is correlated with having a personal stake in projects. However,
untabulated analysis indicates that the shorter tenure and lower commitment of the fraud firm
executives is only significant when the non-opportunistic executives are compared to their
respective matches. Consistent with this interpretation, the executives that orchestrate the fraud
have significantly longer tenures than those that are simply wreckless.
Our third executive level characteristic is the executive’s position measured by binary
variables that equal one (and zero otherwise) if the executive is on the board or is chair of the
board. We predict that board members and chairs are more overconfident than their counterparts
who are matched based on title. The fraud firm executives are no more likely to serve on the
board than the match firm executives. However, the orchestrators, relative to those who simply
participated or were wreckless in not knowing of the fraud, are significantly more likely to be on
the board and to be the board chair (not tabulated).
The psychology literature suggests that experts tend to be more overconfident than
novices (e.g., Griffin and Tversky, 1992). One might imagine that experts would be more
realistic (i.e., less biased) because of their expertise, but the offsetting force is that their expertise
causes them to believe they are better than average. Our first proxy for perceived expertise is
education. More educated executives are likely to be overconfident, consistent with evidence in
Puri and Robinson (2005) and Ben-David et al. (2007). We measure education using a discrete
variable that measures the highest level of education attained: No college (0); Bachelor’s degree
only (1); Master’s degree (2); two Master’s degrees (2.5); JD or PhD (3). In addition, we have
separate indicator variables if the executive’s Master’s degree is an MBA or Master’s of Finance
at opportunistic firms or were considered to be opportunistic executives. Without these three executives, the tenure results in Table 7 are unchanged.
31
as business education may create a sense of overconfidence related to financial reporting.34
We examine two additional proxies for expertise that we predict make an executive
overconfident, specifically in the fraud context, about his ability to avoid detection. CPA is a
binary variable equal to 1 if the manager is CPA and equal to zero otherwise.35 AUDITOR is a
binary variable equal to 1 if the executive has external audit experience and equal to zero
otherwise. External audit experience includes work at a Big 8 or second-tier firms or with a bank
regulator (one case).
In the full sample, education and financial expertise are not associated with the likelihood
of fraud. The orchestrators are less likely to be CPAs or auditors (untabulated), which suggests
that the “accountant” does not direct the fraud, but his participation is necessary to achieve it.36
The opportunistic executives are significantly less likely to be CPAs and have audit experience
(Panel C). This pattern suggests that ex-auditors (all of which are CPAs) may engage in fraud,
but they do not appear to engage in self-serving frauds.
The next executive level characteristic is gender. Men are more overconfident than
women, in general contexts, and empirical evidence documents this phenomenon specifically for
investment trading decisions (e.g., Barber and Odean, 2001; Estes and Hosseini, 2001). An
indicator variable (FEMALE) equals 1 if the executive is a woman, and equals zero otherwise.
Fraud firm executives are significantly more likely to be male. The opportunistic executives also
34 Englmaier (2004) similarly suggests that military service is associated with overconfidence. We were able to identify only two fraud firm executives (John J. Rigas of Adelphia Communications and “Chainsaw” Al Dunlap from Sunbeam, who was a paratrooper at Westpoint) and one matched sample executive (Kenneth E. Hyatt from Walter Industries, Inc.) with military service. Because of the small numbers, and our uncertainty about whether our data on this variable are complete, we do not analyze military service. 35 The CPA variable is also coded = 1 if the manager is a member of the ACCA (Association of Chartered Certified Accountants) in the United Kingdom or is a Chartered Accountant in Canada. 36 The executives that are cited as “wreckless” are high level executives (board chairs, CEOs or CFOs), and CPAs and executives with audit experience. These patterns are almost tautological; one can only be considered wreckless in not knowing about the fraud if he should have known. The SEC appears to expect board chairs, CEOs and CFOs, and individuals with financial expertise to detect at least the significant types of financial reporting defalcations in this sample.
32
are marginally more likely to be female (p-value = 12%), but this result is misleading given the
overall low number of female executives. The finding is based on one female executive (i.e., a
member of the Rigas family at Adelphia) accused of fraud who is defined as opportunistic.
We also examine executive age. The median age of the fraud firm executives is lower,
although only significant in a one-sided test and the result is tenuous given that we are measuring
age coarsely in years, not months. An executive’s age can be correlated positively or negatively
with factors that affect overconfidence. Empirical evidence on the relation between age and
decisions/forecasts that reflect overconfidence is context-specific with results of mixed sign,
linearity, and significance (e.g., Puri and Robinson, 2005; Arabsheibani, de Meza, Maloney, and
Pearson, 2000; Grinblatt and Keloharju, 2001; Malmendier and Tate, 2005; Ben-David et al.
2007). Thus, the results on the relation between age and fraud is descriptive.
Consistent with the regression results in Table 5, the compensation for the fraud sample
executives does not differ from that of the matched sample. However, within the fraud sample, a
greater percent of variable compensation and more total compensation are associated with
opportunistic frauds (Panel B) or opportunistic executives (Panel C). This result is consistent
with theoretical work that predicts that extreme overconfidence is associated with higher
incentive pay because of the distorted preference for such pay, but that lower levels of
overconfidence are associated with lower incentive pay because it is not necessary to induce
effort (de la Rosa, 2007).
6.2 Analysis of governance mechanisms in the EHM subsample
In this section, we examine the role of governance in monitoring the overconfident
executive. We compare the fraud firms to the matched sample firms with respect to commonly
studied governance mechanisms including block ownership, board size, board composition, and
33
measures of the entrenchment and business of the inside, outside, and gray board members. The
purpose of the analysis is to provide additional evidence about whether differences in executive
overconfidence are responsible for the observed differences in earnings management between the
EHM fraud sample firms and the matched sample. If all executives are equally overconfident,
but better governance mitigates the tendencies of the executives at the matched sample firms to
commit fraud, then we should observe better governance at the non-fraud firms.
In summary, the results in Table 8 do not indicate significant differences in governance
mechanisms across the fraud firms and the matched sample firms. None of the variables that
measure blockholdings are significantly different, in contrast to existing evidence that non-fraud
firms are more likely to have a blockholder or a larger percentage of shares held by blockholders
(DSS, 1996, Farber, 2005). The samples do not differ with respect to board size or whether the
CEO is the chair, in contrast to existing evidence that fraud firms have larger boards (Beasley,
1996), are less likely to have audit committees (DSS, 1996), and are more likely to have a CEO
as the chair (DSS, 1996; Farber, 2005; EHM, 2006).37 The fraud firms do not have significantly
different monitoring by inside, gray, or outside directors, as measured by their relative
representation on the board, their tenures, or their outside directorships. These results differ
from those of other matched sample studies in which fraud firms have a larger percentage of
outside and/or gray directors with greater total percentage share ownership and longer tenures
(Beasley, 1996; DSS, 1996; Farber, 2005); and have less busy outside directors (Beasley, 1996).
For the variables that we measure similarly to the prior literature, differences in the results are
likely due to our matching (and possibly smaller sample size).
These (non) results are consistent with the explanation that the fraud firm executives are
37 Efendi, Srivastava, and Swanson (2006) similarly find a positive association between the CEO being the board chair and accounting restatements.
34
more overconfident than those at the non-fraud firms and that better governance mechanisms
were not in place to mitigate their tendency to commit fraud. That is, the overconfident manager
was able to engage in the fraud because he was not better monitored, although he should have
been. The non-results are not consistent with the notion that the matched sample executives are
equally overconfident, but that better governance mitigated the adverse effects of their
overconfidence on earnings management decisions. This final interpretation is subject to the
caveat that our measures capture the sorts of governance mechanisms that would detect and
prevent fraud.
7. Conclusion
This paper examines whether overconfidence is associated with one particular executive
decision – earnings management. While accountants have long recognized the influence of risk-
taking preferences or behavioral biases on individual decision making, the analysis of the impact
of these biases has been limited mainly to auditor decisions, analyst forecasting, and to a lesser
extent, management forecasting. Even in these cases, studies seem to insist on a traditional
economic rationale for the behavioral bias, such as direct monetary incentives like restricted
stock, or indirect incentives such as Institutional Investor competitions and a tournament-style
promotion process in the analyst community. With an expanded view of the role of
overconfidence along the lines of Bénabou and Tirole (2002) and Compte and Postlewaite
(2004), one can observe biased decision making in the absence of direct or indirect monetary
incentives. This broader view of decision making implies interesting applications to external
reporting decisions, such as voluntary disclosure, or earnings management as in this study.
We also examine whether internal and external monitoring mitigates the predicted
35
adverse effects of overconfidence on earnings management decisions. The question of whether
individuals are aware of the impact of their overconfidence on decision-making is common in the
psychology literature. For example, Weinstein and Klein (1996) state: “One must wonder
whether these people recognize when they are at risk and whether their comforting beliefs
interfere with taking precautions to reduce their risk.” Similarly, the recent economics-based
literature characterizes overconfidence as an optimal behavioral choice despite self-awareness
(Compte and Postlewaite, 2004) or as a form of self-deception or “endogenous lack of
awareness” (Bénabou and Tirole, 2002; Brunnermeier and Parker, 2005).38 The examination of
governance mechanisms is related to these ideas in that it addresses whether principals recognize
the potential for biased financial reporting decision making by their agents.
38 See also Schelling (1978) for a non-rigorous discussion of similar ideas.
36
References
Agrawal, Anup and Sahiba Chadha, 2005, Corporate governance and accounting scandals, The Journal of Law and Economics 48: 371–406.
AICPA, 2002, Statement on Auditing Standards (SAS) No. 99: Consideration of fraud in a financial statement audit, New York: AICPA.
Arabsheibani, Gholamreza, David de Meza, John Maloney, and Bernard Pearson, 2000, And a vision appeared unto them of a great profit: Evidence of self-deception among the self-employed, Economics Letters 67: 35-41.
Barber, Brad M. and Terrance Odean, 2001, Boys will be boys: Gender, overconfidence, and common stock investment, Quarterly Journal of Economics (February): 261-292.
Beasley, Mark S., 1996, An empirical analysis of the relation between the board of director composition and financial statement fraud, The Accounting Review 71 (4): 443-465.
Bell, Timothy and Joseph V. Carcello, 2000, A decision aid for assessing the likelihood of fraudulent financial reporting, Auditing: A journal of practice and theory 19 (1): 169-184.
Ben-David, Itzhak, John Graham, and Campbell Harvey, 2007, Managerial overconfidence and corporate policies, Working paper, Duke University.
Bénabou, Roland and Jean Tirole, 2002, Self-confidence and personal motivation, Quarterly Journal of Economics 117 (3): 871-915.
Beneish, Messod D., 1999, Incentives and penalties related to earnings overstatements that violate GAAP, The Accounting Review 74 (October): 425-457.
Bernardo, Antonio E. and Ivo Welch, 2001, On the evolution of overconfidence and entrepreneurs, Journal of Economics & Management Strategy 10 (3): 301–330
Brunnermeier, Markus K. and Jonathan A. Parker, 2005, Optimal expectations, American Economic Review 95 (4): 1092-1118.
Burns, Natasha and Simi Kedia, 2006, The impact of performance-based compensation on misreporting, Journal of Financial Economics 79: 35-67.
Chatterjee, Arijit and Donald C. Hambrick, 2007, It's all about me: Narcissistic chief executive officers and their effects on company strategy and performance, Administrative Science Quarterly 52 (3): 351-386.
Coles, Jeffrey L., Naveen D. Daniel, and Lalitha Naveen, 2006, Managerial incentives and risk-taking, Journal of Financial Economics 79: 431-468.
Committee of Sponsoring Organizations (COSO) of the Treadway Commission, 1999, Fraudulent financial reporting: 1987-1997, An analysis of U.S. public companies, COSO.
37
Compte, Olivier and Andrew Postlewaite, 2004, Confidence-enhanced performance, The American Economic Review 94(5): 1536-1557.
Daniel, Kent, David Hirshleifer, and Avanidhar Subrahmanyam, 1998, Investor pyschology and security market under- and overreactions, Journal of Finance 53 (6): 1839-1885.
de la Rosa, Leonidas Enrique, 2007, Overconfidence and moral hazard, Working paper, Danish Center for Accounting and Finance (D-CAF).
Dechow, Patricia M, Weili Ge, Chad R. Larson, and Richard G. Sloan, 2008, Predicting material accounting manipulations, Working paper.
Dechow, Patricia M., Richard G. Sloan, and Amy P. Sweeney, 1996, Causes and consequences if earnings manipulation: An analysis of firms subject to enforcement actions by the SEC, Contemporary Accounting Research 13 (1): 1-36.
De Bondt, Werner F.M. and Richard H. Thaler, 1995, Financial decision making in markets and firms, In Finance. Edited by R. Jarrow, V Maksimovic, and W. T. Ziemba. Elsevier-North Holland: Series of Handbooks in Operations Research and Management Science, 1995
Draper, Norman R. and Harry Smith, 1998, Applied Regression Analysis, Third Edition, John Wiley & Sons, Inc.
Dyreng, Scott, Michelle Hanlon, and Edward L. Maydew, 2008, Long-run corporate tax avoidance, The Accounting Review 83 (1): 61-82.
Efendi, Jap, Anup Srivastava, and Edward P. Swanson, 2006, Why do corporate managers misstate financial statements? The role of in-the-money options and other incentives, Journal of Financial Economics (forthcoming).
Englmaier, Florian, 2004, A brief survey on overconfidence, Working paper, University of Munich.
Erickson, Merle, Michelle Hanlon, and Edward L. Maydew, 2006, Is there a link between executive equity incentives and accounting fraud?, Journal of Accounting Research 44 (1): 113-143.
Estes, Ralph and Jinoos Hosseini, 2001, The gender gap on Wall street: An empirical analysis of confidence in investment decision making, Journal of Psychology 122 (6): 577-590.
Farber, David B., 2005. Restoring trust after fraud: Does corporate governance matter? The Accounting Review 80, 539–561.
Gervais, Simon and Itay Goldstein, 2007, The positive effects of biased self-perceptions in firms, Review of Finance 11:453-496.
Gervais, Simon, J.B. Heaton, and Terrance Odean, 2007, Overconfidence, investment policy, and manager welfare, Working paper, Duke University.
38
Goel, Anand M. and Anjan V. Thakor, 2000, Rationality, overconfidence and leadership, Working paper, Washington University.
Griffin, Dale and Amos Tversky, 1992, The weighing of evidence and the determinants of confidence, Cognitive Psychology 24 (3): 411-435.
Grinblatt, Mark and Matti Keloharju, 2001, What makes investors trade?, Journal of Finance 56 (2): 589-616.
Hackbarth, Dirk, 2007, Managerial traits and capital structure decisions, Working paper, Washington University.
Heaton, J.B., 2002, Managerial optimism and corporate finance, Financial Management (Summer): 33-45.
Johnson, Shane A., Harley E. Ryan, Jr., and Yisong S. Tian, 2008, Managerial incentives and corporate fraud: The sources of incentives matter, Review of Finance 0: 1-31.
Kahneman, Daniel and Amos Tversky, 1979, Prospect theory: An analysis of decision under risk, Econometrica 47 (2): 263-292.
Karpoff, Jonathan M., D. Scott Lee, and Gerald S. Martin, forthcoming, The cost of cooking the books, Journal of Financial and Quantitative Analysis, forthcoming.
Karpoff, Jonathan M. and John R. Lott, Jr., 1993, The reputational penalty firms bear from committing criminal fraud, Journal of Law and Economics 36: 757-802.
Li, Kan, Randall Morck, Fan Yang, and Bernard Yeung, 2003, Time varying synchronicity in individual stock returns: A cross-country comparison, Working paper, University of Alberta Business School.
Loebbecke, James K., Martha M. Eining, and John J. Willingham, 1989, Auditors’ experience with material irregularities: Frequency, nature, and detectability, Auditing: A Journal of Practice and Theory 9 (1): 1-28.
Malmendier, Ulrike and Geoffrey Tate, 2005, CEO overconfidence and corporate investment, Journal of Finance 60 (6): 2661-2700.
Malmendier, Ulrike and Geoffrey Tate, 2008, Who makes acquisitions? CEO overconfidence and the market’s reaction, Journal of Financial Economics 89: 20-43.
Miller, Danny, Manfred F. R. Kets de Vries, and Jean-Marie Toulouse, 1982, Top executive locus of control and its relationship to strategy-making, structure, and environment, The Academy of Management Journal 25 (2): 237-253.
Post, Jerrold M., 1993, Current concepts of the narcissistic personality: implications for political psychology, Political Psychology 14 (1): 99-121.
39
Puri, Manju and David T. Robinson, 2005, Optimism and economic choice, Journal of Financial Economics 86: 71-99.
Rosenthal, Seth A. and Todd L. Pittinsky, 2006, Narcissistic leadership, The Leadership Quarterly 17: 617-633.
Rotter, Julian B., 1966, Generalized expectancies for internal versus external control of reinforcement, Psychological Monographs 80, Whole No. 609.
Schelling, T. C., 1978, Ergonomics, or the art of self-management, American Economic Review 68 (2): 290-294.
Schultz, Paul and Mir Zaman, 2001, Do the individuals closest to internet firms believe they are overvalued?, Journal of Financial Economics 59: 347-381.
Wang, Dechun D., 2006, Founding family ownership and earnings quality, Journal of Accounting Research 44 (3): 619-656.
Weinstein, Neil D. and William M. Klein, 1996, Unrealistic optimism: Present and future, Journal of Social and Clinical Psychology 15 (1): 1-8.
40
Appendix A: Variable Definitions OPP_EXEC = executive level variable equal to 1 if an AAER states that the accused was required to
disgorge funds directly linked to fraud-related trading or alleges that the fraud was engaged in for self-serving compensation motives, 0 otherwise
OPP_FRAUD = firm level variable equal to 1 if the discussion of fraud in the AAER indicates opportunistic reasons as the primary motive for the fraud, 0 otherwise
Size SALES = total sales (data12) ASSETS = total assets (data6) MVSIZE = total market value of the firm, equal to the sum of the market value of equity
(data199*data25) plus the book value of long-term debt (data9+data34) plus the book value of preferred stock (data130)
Performance/Financial Health: ROA = return on assets, equal to net income (data172) divided by total assets (data6) NISALES = net income (data172) divided by total sales (data12) IB4XSALES = net income before extraordinary items (data18) divided by total sales (data12) DERATIO = debt to equity ratio, equal to long-term debt (data9) plus current portion of long-term
debt (data34) divided by MVSIZE TROUBLE = 1 if the firm had at least three loss years during the six years prior to the fraud, 0
otherwise SPBOND = S&P long-term domestic credit rating (data280), renumbered to be consecutive from 1
(AAA) to 24 (D) CURRENT = current ratio, equal to current assets (data4) divided by current liabilities (data5) QUICK = quick ratio, equal to cash and short-term investments (data1) divided by current
liabilities (data5) COVRATIO = interest coverage ratio, equal to the sum of pretax income (data170) plus interest
expense (data15) divided by the sum of interest expense plus capitalized interest (data147)
BKMKT = book-to-market ratio, equal to the book value of equity (data60) divided by MVSIZE E-P = earnings/price ratio, equal to net income (data172) divided by share price (data199) 1YR_RET = cumulative return in the twelve months prior to the year end of the last clean fiscal year
using CRSP monthly returns with distributions When the fraud firm and its matched pair do not have the same FYE, we accumulate returns over the appropriate 12 months prior to the fraud firm fiscal year end and for the same 12 calendar months for the matched firm, regardless of its fiscal year end.
1YR_RETV = volatility of the monthly returns with distributions used in 1YR_RET Need for external financing: AGEPUB = the number of years the firm has been public, equal to CLEANYR less the start year in
CRSP. FCF = free cash flow (data308) in year t less average capital expenditures (data128) of the
three years prior to year t, scaled by current assets (data4) at t-1. MERGE = 1 if evidence of significant acquisition activity (data249>0 or data129>0), 0 otherwise Growth: ASSETGRO4 (2) = growth in total assets (data6) from four (two) years before the fraud) SALESGRO = one-year percentage change in sales (data12) for the year prior to the fraud PPE = property, plant and equipment (data187) divided by MVSIZE OPCFGRO = One-year percentage change in operating cash flow (data308). OPCYCLE = Length of operating cycle in days (average days in inventory plus average days in
accounts receivable less average days in accounts payable). This variable is set to missing for financial institutions.
41
Appendix A (continued) Innovation: INTANGS = Intangibles intensity, equal to intangible assets (data33) / total assets CAPEX = Capital expenditures from the Statement of Cash Flows (data128) LOGCAPEX = log (CAPEX + 0.001) Risk-taking: RETVOL = equal to 1) the standard deviation of daily stock returns from CRSP (in descriptive
statistics) or 2) a transformed measure that equals the logarithm of the variance of daily stock returns consistent with Coles, Daniel, and Naveen (2006) (in regression analysis).
Complexity/heterogeneity: NUMSEG = Number of business (or geographic) segments. Idiosyncratic risk: NONSYNC = average in each industry I of the nI firm-specific estimates of variance from a market
model estimated for each firm j that has at least 50 monthly return observations from January 1988 through December 2002. The firm-specific estimate of the variance
about the regression is 22 )(1
1)( jj
j est
e−
=σ , where 2)( jes is the sum of squares about
the regression for firm j and tj is the number of return observations for firm j. We use multiple market indices (the results are similar)
Executive characteristics FOUNDER = 1 if the executive is a founder or co-founder of the firm, or is part of the founding
family, 0 otherwise TENURE = the number of years between the executive’s start date at the firm and the first year of
the fraud COMMIT = TENURE scaled by the age of the firm, using the firm start year on CRSP. If the
executive was with the firm prior to the public listing (and start year on CRSP), this variable is 1.
BOD = 1 if the executive is a member of the board, 0 otherwise BODCHAIR = 1 if the executive is the chair of the board, 0 otherwise CEO = 1 if the executive is the chief executive officer (CEO) of the firm, 0 otherwise CFO = 1 if the executive is the chief financial officer (CFO) of the firm, 0 otherwise EDUCATION = 0 if the executive did not attain a college degree, 1 Bachelor’s degree, 2 Master’s
degree, 2.5 two Master’s degrees, or 3 JD or PhD, based on the highest level of education obtained.
MBA = 1 if executive holds a MBA or Master’s of Finance, 0 otherwise AGE = executive’s age in the first year of the fraud CPA = 1 if the executive holds a CPA (or CPA equivalent from another country), 0 otherwise AUDITOR = 1 if the executive worked as an external auditor, 0 otherwise FEMALE = 1 if the executive is a female, 0 if male Compensation: CASHCOMP = Natural log of salary plus bonus VAR$PAY = Ratio of bonus to salary plus bonus
42
Appendix A (continued) Overconfidence score: PHOTOSCORE = four points if the CEO’s photo in the annual report included no other individuals and
was at least equal to one half page in size; three points if the CEO’s photo included no other individuals and was less than one half page in size; two points if there were other individuals pictured with the CEO; and one point if there was no photograph of the CEO.
REL_CASHCOMP = the ratio of salary plus bonus for the CEO to that of the second highest paid executive, from ExecuComp.
REL_NONCASH = the ratio of non-cash compensation for the CEO to that of the second highest paid, where non-cash compensation is equal to total compensation (TDC1) less cash compensation from ExecuComp.
OC-SCORE = equal to the sum of the standardized values of three proxies for executive overconfidence based on those used by Chatterjee and Hambrick (2007): PHOTOSCORE, REL_CASHCOMP, and REL_NONCASH. Each proxy is the average value over the second and third years of the CEO’s tenure.
Capital structure/Dividend policy/Tax avoidance: RISKYDEBT = An indicator variable = 1 if either the ration of convertible debt (data39) to assets or
preferred stock (data130) to assets is > 0, and = 0 otherwise. DIVYLD = dividend yield, equal to dividends per share (data26) divided by share price (data199)
for the firms that pay dividends (and missing otherwise) CASHETR = cash effective tax rate, equal to the ratio of cash taxes paid (data317) during the first
year of the fraud to pretax income excluding special items (data170-data17) Blockholders: BH_IND = An indicator variable = 1 if the proxy identifies at least one blockholder, and = 0
otherwise BH_NUM = Number of reporting persons identified in the proxy statement who are the beneficial
owners of more than 5% of the common stock outstanding as defined under Section 13(d) of the Securities Exchange Act of 1934, as amended
BH_PCT = Percent of common shares held by identified blockholders Board characteristics: BDSIZE = Number of directors on the board CEOCHAIR = An indicator variable = 1 if the CEO is the chair and = 0 otherwise AUDITCOMM = An indicator variable = 1 if the board has an audit committee and = 0 otherwise PERCENT = Percentage of directors of the noted type ENTRENCH = An indicator variable = 1 if the director served more than 5 years or 100% of the firm’s
life (AVGTEN% = 1) and = 0 otherwise AVGTEN% = Average tenure on the board in years from the first year of the directorship to the
meeting date (as per IRRC) scaled by firm age (truncated at 1) for directors of the noted type
AVGBUSY = Average number of other directorships held by directors of the noted type
43
Table 1. Summary of the subsample of 49 EHM fraud firms Table 1 first summarizes the effect of the fraud on the financial statements. The column labeled “Occurred” indicates the percent of AAERs in which the SEC alleges that a particular type of earnings management occurred. The column labeled “Primary” indicates which of the earnings management activities we identified as the most significant activity based on the AAER discussion. The second section describes the goals of the fraud. The third section reports the average length of the frauds.
EHM sample firms Non-opportunistic frauds (OPP_FRAUD=0)
Opportunistic frauds (OPP_FRAUD=1)
Occurred Primary Occurred Primary Occurred Primary Financial statement effect: Revenue recognition Premature 51.0% 28.6 58.3 33.3 38.5 15.4 Fictitious 40.8 18.4 36.1 16.7 53.9 23.1 Unclear 18.4 - 19.4 - 15.4 - Any type 63.3 44.9 63.9 47.2 61.5 38.5 Overstate assets 26.5 8.2 22.2 8.3 38.5 7.7 Other income increasing 46.9 16.3 41.7 16.7 61.5 15.4 Capitalize expenses 20.4 6.1 19.4 5.6 30.8 7.7 Create (or use) hidden reserves 18.4 8.2 22.2 11.1 15.4 - Off-balance sheet financing 6.1 6.1 2.8 2.8 15.4 15.4 Improper income statement classification 18.4 2.0 19.4 2.8 15.4 - Illegal transactions39 6.1 - 2.8 - 15.4 - Related party 16.3 4.1 8.3 - 38.5 15.4 Goals: Increase income 83.7% 86.1 76.9 Smooth income 6.1 8.3 - Hide debt off-balance sheet 6.1 2.8 15.4 Length of fraud period: Average number of years 3.78 3.58 4.31 Minimum/Maximum 1/8 1/7 2/8
39 We do not classify falsifying documents, which is alleged in many revenue recognition cases, as an illegal transaction.
44
Table 2. Comparison of characteristics of the 49 EHM fraud firms and the matched sample Descriptive characteristics of the fraud and matched samples. All variables, which are defined in Appendix A, are measured at the end of or during CLEANYR. * (**){***} in the “Matched sample” columns indicate that the mean [median] of the matched sample is significantly different from that of the corresponding fraud sample at the 10% (5%) {1%} level. * (**){***} in the “Fraud firms” columns for the extreme firms indicates significant differences from the moderate fraud firms. The p-values of paired t-tests of differences in means across the samples assume equal variances unless equality is rejected at 10% level. The p-values of differences in medians across the samples are for a two-sided Wilcoxon rank-sum test.
Non-opportunistic frauds (OPP_FRAUD = 0) Opportunistic frauds (OPP_FRAUD=1) Fraud firms Matched sample Fraud firms Matched sample Mean Median Mean Median Mean Median Mean Median ASSETS 5,050 204.8 4,747 200.0 4,198 284.3 3,822 286.4 SALES 4,305 243.7 4,681 293.5 2,813 249.0 2,308 365.5 MVSIZE 8,317 440.8 9,598 218.0 7,518 2,613.6 6,926 522.8
ROA -0.03 0.04 0.01 0.06 0.05* 0.06 0.03 0.07 NISALES -0.51 0.03 -0.02 0.04 -0.51 0.06 0.06 0.09 IB4XSALES -0.40 0.03 -0.02 0.04 -0.51 0.06 0.06 0.10 DERATIO 0.15 0.10 0.16 0.10 0.21 0.18 0.28 0.19 TROUBLE 0.19 0 0.15 0 0.29 0 0 0 SPBOND40 8.25 8.5 6.9 6 9.0 8.0 10.0 9.0 CURRENT 3.00 2.08 2.14 1.84 1.85* 1.67 2.89 2.66 QUICK 1.35 0.37 0.66 0.24 0.61 0.55 1.30 0.26 COVRATIO -2.42 4.51 38.31 5.81 14.42 3.56 67.26 4.92 BKMKT 0.42 0.32 0.42 0.38 0.32 0.36 0.59* 0.48 E-P 5.18 0.68 10.15 0.54 3.32 1.05 0.51 1.10 1YR_RET 0.38 0.20 0.42 0.23 0.57 0.50 0.02** 0.04* 1YR_RETV 0.14 0.12 0.12 0.11 0.13 0.11 0.14 0.13
AGEPUB 10.19 6 15.55 12* 13.92 11 12.25 8 FCF 0.03 0.04 0.04 0.10* 0.11 0.13 0.09 0.08 MERGE 0.40 0 0.28 0 0.55 1.00 0.55 1 Growth: ASSETGRO4 1.90 0.45 0.81 0.25 1.11 0.79 0.77 0.63 ASSETGRO2 0.41 0.18 0.15* 0.10** 3.38 0.36 1.02 0.22 SALESGRO 0.53 0.18 0.21* 0.13* 1.81 0.31 0.21 0.11* PPE 0.31 0.15 0.39 0.30 0.40 0.28 0.39 0.29
40 N = 8 (10) in the non-opportunistic fraud (matched firm) sample. N = 5 (4) in the opportunistic fraud (matched firm) sample.
45
Table 3 Comparisons of time-series changes in restatement amounts and unmanaged earnings Panel A shows comparisons of percent changes in restatements in Year 2 of the fraud relative to Year 1 for the non-opportunistic and opportunistic fraud firms. Panel B shows changes in performance for the non-opportunistic and opportunistic fraud firms and their matched samples. Performance is measured as pre-managed (or restated) net income. For all years for the match sample firms, and for CLEANYR for the fraud firms, performance equals net income (data item #172) from Compustat. For the fraud firms in Year 1 and Year 2, performance is restated net income if restatement amounts are available and it is missing otherwise. The percent change in restatement amounts and percent growth in performance (Panels A and B) are measured at the firm level. Averages across the firms are presented. Panel C shows the percent of firms with negative performance, where % LOSS FIRMS is an indicator equal to 1 for firms with negative performance and equal to 0 otherwise. * (**){***} in the “Matched sample” columns indicate that the mean [median] of the matched sample is significantly different from that of the corresponding fraud sample at the 10% (5%) {1%} level. There are no significant differences between the means or medians of any variables for the opportunistic fraud firms and the non-opportunistic fraud firms. The p-values of t-tests of differences in means across the samples assume equal variances unless equality is rejected at 10% level. The p-values of differences in medians across the samples are for a two-sided Wilcoxon rank-sum test. For binary variables, p-values are for a χ2 test.
Non-opportunistic frauds (OPP_FRAUD = 0) Opportunistic frauds (OPP_FRAUD=1) Fraud firms Matched sample Fraud firms Matched sample Mean Median Mean Median Mean Median Mean Median
Panel A: Firm-specific % change in restatement amounts Revenue: Year 1 – Year 2 221% 97 -54 26 Net income: Year 1 – Year 2 40% 18 58 71 Panel B: Firm-specific % growth in restated performance CLEANYR – Year 1 -87% -33 63 -8 327 -5 49 18 Year 1 – Year 2 -174% -86 -75 -1 408 65 -49 -36 Panel C: % LOSS FIRMS Year 0 26% 0 14 0 23 0 23 0 Year 1 42% 0 22 0 63 1 31 0 Year 2 53% 1 18** 0*** 43 0 31 0
46
Table 4. Industry effects on fraud propensity
Models of the determinants of the proportion of frauds in an industry during the period 1990 - 2004 and the proportion of restatements in an industry during period 1997-2005. For both samples, industry is defined at the 3-digit SIC code level. %FRAUDi equals the total number of AAERs against firms in industry i scaled by the average number of firms in the industry from 1990-2004. %GAOi equals the total number of restatements in industry i scaled by the average number of firms in the industry from 1997-2005. The dependent variable in the model is the arc sine-square root transformation: %FRAUDarcsin%ASFRAUD ii = and %AOGarcsin%ASGAO ii = . The model is estimated using weighted least squares (WLS); observations are weighted by the average number of firms in the industry. Independent variables are estimated using the median value of the variable for the respective industry group for the year 1995, with the exception of SYNCSSE. Significance levels are indicated by ***, **, *, and # representing 1%, 5%, 10%, and 15% levels, respectively, 2-tailed. FRAUD RESTATEMENT Intercept 0.0040 0.5669* INTANGS 0.0031 0.1470 CAPEX 0.0001 -0.0001 SALESGRO 0.2591** 0.2540 OPCFGRO -0.1018*** -0.0901* OPCYCLE 0.0002 -0.0009*** MERGE -0.0254 -0.0082 RETVOL -0.1814 -2.3341 NUMSEG (Business) -0.0937 -0.0313 NUMSEG (Geographic) 0.2009# -0.0619 NONSYNC 1.1513* 2.8225*** N 115 115 Adj R2 19.76% 14.72%
47
Table 5. Firm-level fraud prediction model Models of the determinants of fraud at the firm level during the period 1989 – 2001. The dependent variable is an indicator equal to 1 in the first year the firm was accused of fraud in an AAER and is equal to zero for all non-AAER firms. The model is estimated using a logit regression with standard errors clustered by firm and fiscal year. Variable definitions of the independent variables are provided in Appendix A. Significance levels are indicated by ***, **, and * representing 1%, 5%, and 10%, respectively, 1-tailed if predicted and two-tailed otherwise. Dependent variable: 1 = AAER firm (YR1 of fraud in year t)
0 = non-AAER firm (no fraud in year t) Intercept -6.210*** -7.128*** -10.694*** -7.291***
DERATIO + -1.195** -1.561** -3.369*** -3.370***
RISKYDEBT + 0.139 -0.078 -0.388 -0.369
DIVYLD - -38.382*** -32.789*** -21.653*** -21.461***
CASHETR - 0.355 0.271 -1.274** -1.374**
LOGCAPEX + 0.004 -0.194 0.061 0.079
CASHCOMP + 0.753***
VAR$PAY + 2.228***
LOG(ASSETS) 0.162** 0.448*** 0.093 0.202**
ROA -1.286 0.345 -3.124 -3.392
BKMKT -0.242 -0.286 0.822 0.870
ASSETGRO2 0.271*** 0.326*** 0.250 0.178
SALESGRO 0.547*** 0.405** 0.733** 0.593
FCF -1.378***
N 34,620 30,628 10,145 10,145 Log likelihood -765.87 -696.93 -258.85 -257.67
48
Table 6. Fraud prediction using a subsample of overconfident executives Estimation of fraud prediction models for a sample of executives of firms in the software (SIC 737) and hardware (SIC 357) industries. The executive’s OC-SCORE aggregates the standardized values of three variables capturing executive overconfidence. The three components capture the prominence of the CEO’s picture in the annual reports, the ratio of CEO cash pay to that of the second highest paid executive, and the ratio of non-cash pay. The dependent variable is an indicator equal to 1 if the firm was accused of fraud in an AAER and is equal to zero for all non-AAER firms. The model is estimated using a logit regression. Non-cash pay is defined as total compensation less cash pay. Control variables are winsorized at 5%. Definitions of the independent variables are provided in Appendix A. Significance levels are indicated by ***, **, and * representing 1%, 5%, and 10%, respectively, 1-tailed if predicted and two-tailed otherwise. 1=AAER firm
0=Non-AAER Intercept -5.697** -5.727** -6.611**
OC-SCORE components:
PHOTOSCORE + 0.325
REL_CASHCOMP + 0.360
REL_NONCASH + 0.735
OC-SCORE + 0.441** 0.586**
LOGASSETS 0.429 0.420 0.458
ROA 1.583 1.397 1.648
DERATIO -2.884 -2.966 -25.425
BKMKT 1.094 1.192 3.023
ASSETGRO2 -0.102 -0.091 0.215
SALESGRO 0.758 0.810 0.496
FCF 2.029
N 100 100 92 Pseudo R2 15.15% 14.78% 23.03%
49
Table 7. Comparison of executive characteristics Means (medians of continuous variables) of characteristics of accused executives at the EHM subsample of fraud firms and the matched sample firms. Panel A presents statistics for the 107 executives of the EHM sample firms and their matched sample counterparts. Panel B presents statistics for 32 of the 107 accused executives at firms in which the SEC characterizes the fraud as opportunistic (OPP_FRAUD = 1) and for 75 of the 107 accused executives at firms for which OPP_FRAUD = 0. Panel C presents statistics for 40 of the 107 accused executives defined as opportunistic (OPP_EXEC = 1) based on insider trading allegations and for 67 of the 107 accused executives for which OPP_EXEC = 0. For continuous variables, significance is based on p-values of paired t-tests of differences in the means across the samples assuming equal variances unless equality is rejected at 10% level or on p-values of two-sided Wilcoxon rank-sum tests of differences in the medians. For binary variables, significance is based on p-values of a χ2 test. (*){**}[***] indicate statistical significance at the (10%) {5%} [1%] level.
Panel A: EHM sample vs. matched sample
Panel B: EHM sample divided by the nature of the fraud
Panel C: EHM sample divided by executive overconfidence
N EHM firms N Match Opportunistic Non-opp Extreme Moderate Max N 107 99 32 75 40 67 FOUNDER 99 15.15% 99 10.10 28.13 10.67 ** 25.00 10.45 **
TENURE 88 5.97 88 9.53 *** 8.75 4.90 *** 7.95 5.02 ** (3) (7) *** (8) (3) *** (5) (3) **
COMMIT 71 60.60% 71 76.01 ** 82.15 55.63 *** 68.38 60.77 (76.92%) (100.00) ** (100.00) (58.57) *** (100.00) (80.00)
BOD 98 54.08% 98 46.94 58.06 52.00 57.50 51.52 BODCHAIR 98 20.41% 98 22.45 25.81 16.00 22.50 16.67
EDUCATION 35 1.43 35 1.61 1.73 1.50 1.82 1.46 * (1) (2) (1) (1) (2) (1)
MBA 35 20.00% 35 37.14 20.83 25.00 27.27 20.59
AGE 99 48.05 years 99 49.12 46.69 48.44 47.55 48.13 (47 years) (49)
CPA 61 29.51% 61 27.87 24.14 33.96 19.35 37.25 * AUDITOR 56 23.21% 56 19.46 14.81 23.40 10.71 26.09 FEMALE 99 1.01% 99 5.05 * 3.13 0 2.5 0
VAR$PAY 44 25.34% 44 31.29 30.69 24.10 34.00 18.32 *** (22.63%) (37.32) (48.63) (28.90) (41.52) (5.70) ***
CASHCOMP 44 5.98 44 6.14 6.40 5.73 *** 6.37 5.50 *** (5.79) (6.05) (6.26) (5.63) *** (6.26) (5.36) ***
50
Table 8 Comparison of governance mechanisms Comparison of governance mechanisms for the fraud sample and matched sample. (*){**}[***] indicate statistical significance at the (10%) {5%} [1%] level. The p-values of the t-test of the differences in the means across the samples assume equal variances unless equality is rejected at 10% level. The p-values of tests of differences in the medians across the samples are for a two-sided Wilcoxon rank-sum test. For binary variables, significance is based on p-values of a χ2 test.
Fraud firms Matched sample firms Test of dif Mean Median Mean Median Mean Median N = 32 N = 32 Blockholders: BH_IND 81.82% 1 71.88% 1
BH_NUM 1.94 2 1.81 1
BH_PCT 20.14% 16.87 21.59% 18.95
Board characteristics: BDSIZE 8.82 8 7.94 8
CEOCHAIR 78.79% 81.25%
AUDITCOMM 69.70% 84.38%
PERCENT: Inside 23.01% 20.00% 23.08% 20.00%
Gray 21.01% 17.75% 17.75% 12.50%
Outside 56.36% 60.00% 59.18% 64.58%
Board member characteristics: ENTRENCH: Inside 74.19% 80.65%
Gray 81.82% 76.19%
Outside 80.00% 72.41%
AVGTEN%: Inside 55.30% 68.33% 70.99% 93.75%
Gray 55.35% 81.60% 75.43% 91.67%
Outside 54.60% 50.00% 60.78% 53.33%
AVGBUSY: Inside 1.09 0.17 0.70 0
Gray 1.12 0.50 1.24 0.83
Outside 1.64 1.67 1.21 1.00 *