U

UNIVERSITY OF CINCINNATI
Date:

05/14/2009
,

I,

Buhui Qiu

hereby submit this original work as part of the requirements for the degree of:

Doctor of Philosophy
in

Business Administration with Concentration in Finance

It is entitled:

Two Essays on Corporate Fraud

Student Signature:

Buhui Qiu

This work and its defense approved by: Committee Chair:

Steve Slezak (co-chair)

Michael Ferguson (co-chair)

Hui Guo Weihong Song Yan Yu

Approval of the electronic document: I have reviewed the Thesis/Dissertation in its final electronic format and certify that it is an accurate copy of the document reviewed and approved by the committee. Committee Chair signature:

Steve Slezak

Two Essays on Corporate Fraud

A dissertation submitted to the Graduate School of the University of Cincinnati in partial fulfillment of the requirements for the degree of

Doctor of Philosophy

In the Department of Finance and Real Estate of the College of Business Administration by Buhui Qiu

M.A. Sun Yat-Sen University (China)

May 2009

Committee Chair: Michael Ferguson, Ph.D. and Steve Slezak, Ph.D.

Dissertation Abstract
The dissertation consists of two essays. Essay I proposes a theory which explicitly models the strategic fraud detection behavior of the regulator (e.g., the SEC), and studies the strategic interaction between corporate fraud commission and detection. The model generates several new testable empirical implications. Essay II empirically studies whether the SEC strategically responds to fraud commission and how effective Sarbanes-Oxley is in reducing fraud commission using detected fraud data of corporate America in the last 10 years. Essay I: The paper considers an agency model of fraudulent misreporting which implies a rich set of relationships between the commission of fraud, the observation or detection of fraud, economic performance, and the compensation policy of the firm. The paper develops a number of testable empirical implications and highlights several interesting phenomena, including implications on exogenous variables that can cause an increase in the amount of fraud committed but a decrease in the amount of fraud being observed (and visa versa). Thus, empirical studies that seek to identify the firm or managerial characteristics associated with the commission of fraud cannot infer a relationship by simply examining how the amount of observed fraud varies with these characteristics. In addition, the paper also shows that an increase in an industrys growth potential can cause that industry to fall from a high-productivity pooling equilibrium (with high levels of incentive compensation and effort and, as a result, many high-productivity firms) to the lower-productivity mixed-strategy equilibrium (with lower levels of incentive compensation and effort and, as a result, fewer high-productivity firms), resulting in a drop in economic performance. Essay II: In the wake of recent financial scandals, there are heated debates over whether the SEC is effective in combating fraud as well as over the costs and benefits of the Sarbanes-Oxley Act. This paper investigates two research questions empirically: 1. Does the SEC strategically respond to fraud commission? 2. How effective is SOX in reducing fraud commission? Using a sample of firms subject to SEC litigations for fraud and employing the bivariate probit with partial observability technique, we find strong evidence in favor of theoretical predictions with the assumption that the regulator is strategic in combating fraud, but contradicting to theoretical predictions assuming that the fraud detection environment is exogenous or mechanical (i.e., without a strategic regulator). We also find that SOX has been very effective and decreased fraud commission by two thirds after its enactment. Our finding should provide some useful insight to policy makers in light of the current debates.

iii

iv

Acknowledgements
I thank God, my heavenly father, for blessing me with wisdom, strength and health everyday to complete my dissertation and the doctoral study at the University of Cincinnati. I thank Drs. Michael Ferguson and Steve Slezak, my advisors, for their invaluable mentoring and guidance in the past five years. I thank my dissertation committee members, Drs. Hui Guo, Weihong Song, and Yan Yu for their outstanding service on my committee and their precious guidance, comments and encouragement. I am grateful for the financial support from the Department of Finance and Real Estate, the College of Business Administration and the University of Cincinnati in the forms of assistantships, scholarships and travel grants. The faculty, staff, and students of the Department of Finance and Real Estate have always been very helpful. I thank my wife, Enjia, who is always there to support me and make my life full of love and color. I thank my parents, who raised me and taught me how to be a good person. I thank all brothers and sisters of the UCC fellowship and Cincinnati Chinese Church, who make my familys stay in Cincinnati so enjoyable and pleasant. I also thank the fellow students at the finance PhD program, who gave me friendship and help during the past years.

Table of Contents Essay 1: Page

Abstract.1 1. Introduction..2 2. The Model8 3. Equilibrium.13 4. Empirical Implications...25 5. Conclusion..35 Essay 2: Abstract42 1. Introduction43 2. Related Literature...46 3. Empirical Hypotheses.52 4. Methodology...54 5. Data and Variables..57 6. Empirical Results64 7. Conclusions72 Bibliography84 Appendix.87

vi

Essay1: The Strategic Interaction between Committing and Detecting Fraudulent Misreporting*
ABSTRACT The paper considers an agency model of fraudulent misreporting which implies a rich set of relationships between the commission of fraud, the observation or detection of fraud, economic performance, and the compensation policy of the firm. The paper develops a number of testable empirical implications and highlights several interesting phenomena, including implications on exogenous variables that can cause an increase in the amount of fraud committed but a decrease in the amount of fraud being observed (and visa versa). Thus, empirical studies that seek to identify the firm or managerial characteristics associated with the commission of fraud cannot infer a relationship by simply examining how the amount of observed fraud varies with these characteristics. In addition, the paper also shows that an increase in an industrys growth potential can cause that industry to fall from a high-productivity pooling equilibrium (with high levels of incentive compensation and effort and, as a result, many high-productivity firms) to the lower-productivity mixed-strategy equilibrium (with lower levels of incentive compensation and effort and, as a result, fewer high-productivity firms), resulting in a drop in economic performance.

* I would like to thank Sandra Betton, Mathijs van Dijk , Alan Douglas, Mike Ferguson (my advisor), Hui Guo, Young Koan Kwon, Albert (Pete) Kyle (FMA doctoral seminar session chair), Gregory Lipny, Carolina Salva, Raj Singh (WFA discussant), Steve Slezak (my advisor), Weihong Song, James Thomson, Marno Verbeek, Kenneth Vetzal, Tracy Wang, seminar participants at Concordia University (Montreal), Rotterdam School of Management (Erasmus University), Singapore Management University, the University of Cincinnati, the University of Waterloo, and Vlerick Leuven Gent Management School, session participants at the 2008 Western Finance Association annual meeting in Waikoloa, Hawaii, session participants at the 2008 Finance Management Association annual meeting and FMA doctoral seminar in Dallas, Texas, and session participants at the 14th Conference on the Theories and Practices of Securities and Financial Markets (SFM) in Kaohsiung, Taiwan for helpful comments and suggestions. I also gratefully acknowledge the financial support from the 14th SFM conference through their Taiwan Stock Exchange best paper award for an earlier version of the paper. All errors are mine.
1

1. Introduction The revelation of fraudulent misreporting in numerous high-profile cases in the United States around the start of the twenty-first century (e.g., Adelphia, Enron, Global Crossing, Tyco, Waste Management Inc., and Sunbeam) resulted in a substantial loss in market value.1 It is

unclear, however, whether these cases represent isolated instances of lapses in corporate ethical judgment or whether they indicate a general degradation in corporate morality and/or an increase in the incentive to commit fraud. Clearly indicating a belief in a systemic source, the U.S.

Congress enacted the Sarbanes-Oxley Act in 2002 in an effort to rein in managers in what was feared to be a pervasive fast and loose with the facts opportunistic corporate culture. Yet, it is

still unclear what social or economic forces changed to cause the increase in fraudulent misreporting. In addition, without knowing the cause, it is also unclear whether

Sarbanes-Oxley will be an effective counter-measure (especially given the time-series and cross-sectional variation in the economic conditions firms face). In fact, the recent arrests of

two Bear Sterns hedge fund managers and the Securities and Exchange Commission investigations of dozens of corporate fraud cases related to sub-prime mortgage securities would seem to raise doubt. The above issues are difficult to address because the amount of fraud committed is not directly observable; we only observe the amount of fraud that is detected, which is jointly determined by the amount of fraud actually being committed and the probability of getting caught given the extent to which fraudulent activities are investigated. To the extent that

environmental influences may affect the commission and investigation of fraud differently, there
1

According to Cornerstone Research, 231 fraud lawsuits in the year 2002 alone resulted in a total disclosed dollar loss of $203 billion in market capitalization. From 1996 to 2004, on average there were 195 lawsuits per year with a total disclosed dollar loss of $127 billion per year; typically around 80% of these lawsuits involve misrepresentation of financial statements. When looking into these misrepresentation cases, one can find that almost all of them involve earning inflation of some sort.

may not be a one-to-one correspondence between the amount of fraud observed and the amount committed. Thus, the fact that fraudulent behavior is not observable makes it difficult to discover and document links between potential environmental influences and fraudulent behavior. Instead, we must develop theoretical models which provide testable implications with respect to observable phenomena and, when these models are supported by the data, infer relationships among non-observable variables based on the implications of such models. In order to provide such structure, this paper develops a theoretical model of fraud with two critical features: (1) there is a strategic interaction between the commission and detection of fraud (which allows us to develop conditional statements on what can and cannot be inferred about the commission of fraud from the observation of fraud), and (2) the extent of fraud committed and investigated varies with the economic environment (which allows us to develop time-series and cross-sectional implications on the amount of fraud and the effectiveness of regulation). Specifically, we develop an agency model in which managers are induced via an equity-based compensation (EBC) contract to exert personally costly effort that increases the expected returns of the firm. In the model, the realized return of the firm is not observable by

the market; rather the manager must report its value. Similar to Goldman and Slezak (2006), EBC provides the manager the incentive to exert effort but also the incentive to upwardly bias reports. The regulatory agency, seeking to minimize the deadweight loss associated with fraud, is responsible for detecting fraud and imposing penalties; it bases its investigation strategy on the managers equilibrium fraud commission strategy in order to optimally trade off the benefit of reducing fraud against the agencys detection cost. To capture both cross-sectional and time-series variation (high growth versus old-economy firms and recession versus expansion), we assume that after the initial stage, a

potential new investment project arrives with a certain probability (proxying for growth potential). After privately observing the projects expected profitability, the manager either adopts or rejects it. In order to obscure past fraud, fraudulent managers have the incentive to over-invest (i.e., invest in new projects that have negative expected NPV). overinvestment results in the deadweight loss associated with fraud.2 Depending upon the parameters characterizing the regulatory and contracting environment, three potential types of equilibrium may obtain: truthful separating equilibrium in which each manager truthfully reports their realized return, pooling equilibrium in which all poorly-performing managers mimic the reports of highly-performing managers but the regulatory agency does not monitor to verify reports, and a mixed strategy equilibrium in which poorly-performing managers commit fraud with an equilibrium probability while the regulatory agency randomly audits those firms that report high earnings. The paper provides a number of empirical implications regarding the commission and detection of fraud, incentive contracts, and economic performance. First, the model implies This result is This

that fraudulent reporting activities will be concentrated in high-growth industries.

similar to the theory predictions of Wang (2006) and consistent with the empirical evidence in Wang (2005) and Johnson, Ryan and Tian (2003).3 Second, the model implies that while the

fraud incentive is strongest in good times, fraud commission and detection are more likely to

The exact nature of the deadweight loss is not critical for most of our results; all that is needed is that there be some benefit to reducing fraud (in terms of more efficient production) so that the regulator must balance the benefit of reduced fraud against the implementation costs associated with monitoring fraud. However, the paper does develop some implications related to the specific form of fraud inefficiency we assume.
3

Wang (2005) provides empirical evidence that firms engaging in fraudulent reporting tend to overinvest relative to their peers. Johnson, Ryan and Tian (2003) find that their sample of exposed fraudulent reporting firms are not random draws from all possible industries, but rather demonstrates a statistically significant industry concentration, with the concentrated industries having significantly higher than average growth potential.

occur when the high growth industries fall into downturns.4

Specifically, when the parameters

are such that the pooling equilibrium obtains, then there will be relatively high levels of EBC, high average short-term performance, and no fraud being exposed (although fraud is committed). In contrast, when the parameters are such that the mixed strategy equilibrium obtains, the equilibrium will have low EBC, low average short-term performance, and fraud will be exposed. Thus, consistent with the empirical evidence in Johnson, Ryan and Tian (2003), these results imply that exposed fraud will occur in periods with relatively weak economic performance.5 Third, the model implies that an increase in growth potential, which is typically good news (i.e., implies higher future profitability and, as a result, increased firm value), can strikingly have a negative impact on value and economic performance. Specifically, we show that, by

altering the incentives to commit and investigate fraud, an increase in growth potential alone can cause an industry to fall from a high-productivity pooling equilibrium (with high levels of EBC and effort and, as a result, many high-productivity firms) to the lower-productivity mixed-strategy equilibrium (with lower levels of EBC and effort and, as a result, fewer high-productivity firms), resulting in a drop in economic performance. We show that, given the

strategic interaction between the incentives to commit and investigate fraud, this drop in economic performance will be accompanied by an increase in the amount of exposed fraud. These results imply that, while innovation may be beneficial to economic growth by generating increased future growth opportunities, innovation can also have a dark side when fraud is possible. scandals:
4

In fact, there are many examples of innovations that were accompanied by fraud financial innovation in mortgage derivatives, product innovation in

Its primary focus, the model in Povel, Singh, and Winton (2007) generates similar boom-and-bust results. discussed below, the two models and the mechanisms by which these boom-and-bust results obtain differ.
5

As

Their results show that both the exposed fraud firms and their (industry- and size-matched) control firms significantly underperformed the overall stock market during the fraudulent reporting periods.

telecommunications, and innovation created by the deregulation of energy markets, to name a few. Fourth, the model implies that the extent of detected fraud need not be indicative of the extent of fraud committed. This is in contrast to the signal jamming models of fraud (see, for example, Goldman and Slezak (2006)) in which the equilibrium probability of committing fraud is one. In these types of models, an increase in the probability of detection will necessarily In our model, however, the equilibrium

result in an increase in the incidence of observed fraud.

probability of committing fraud can be inversely related to the probability of observing detected fraud (in the mixed strategy equilibrium), leading to potential ambiguity in the statistical relationship between the amount of fraud detected and committed. Similar to our model, there is a strategic interaction between the commission and detection of fraud in Povel, Singh, and Winton (2007), hereafter PSW, and Wang (2006); in both of these models the extent to which managers are monitored depends upon the information content of the managers equilibrium reports and the extent to which managers commit fraud (i.e., bias and reduce the informativeness of reports) depends upon the likelihood of being monitored. In both models, monitoring serves to reduce adverse selection caused by fraud. In Wang (2006),

managers commit fraud on behalf of current equity holders who benefit from a lower cost of capital stemming from inflated equity prices caused by fraud. In PSW, managers seek outside

funding for their projects and provide (potentially fraudulent) information on the prospects of these projects to potential investors who face adverse selection in deciding whether or not to provide funding. In their model, managers receive non-contractible control benefits from any

(even negative NPV) investment and, as a result, they commit fraud in an effort to mislead investors into funding negative NPV projects so that they can obtain these benefits of control.

In both models, the cost of fraud derives from an over-investment problem similar to Myers and Majluf (1984). In contrast to these models, fraud in our model stems from an agency problem between managers and shareholders, with managers seeking to manipulate prices upward in order to increase their equity-based compensation. As in Goldman and Slezak (2006), equity-based

compensation is a double-edged sword in that it provides both the incentive for the managers to exert costly effort in improving the profitability of the firm and the incentive to commit fraud. Given this dual role, the possibility of fraud alters the incentive contract, which alters the equilibrium level of effort and the productivity of firms. Thus, in contrast to PSW and Wang

(2006), which take the distribution of firm productivity as given and consider how fraud affects the allocation of resources among the fixed set of firms, our model endogenously determines the productivity of the set of firms via the incentive contract. In contrast to Goldman and Slezak

(2006), which takes the investigation of the regulatory agent as given, our model considers the strategic interaction between the regulatory agent and the manager in the context of this agency problem. We show that this combination of elements generates new insights.

Another key difference between our model and the model in Wang (2006) is that we consider the behavior of a regulatory agency that is concerned with social welfare.6 In Wang

(2006), the monitor chooses whether or not to investigate by trading off the investigation cost against the penalties the monitoring agency earns when fraud is detected. That is, the

monitor in Wang (2006) seeks to maximize the expected profit from monitoring, with the penalties -- set exogenously -- representing revenue to the monitor. Thus, since the

In PSW, firms are not monitored by a regulatory agency such as the SEC. Rather, the potential investors decide whether or not to investigate the claims of firms further prior to investing. Although there is no regulatory agency in PSW, their potential investors decision to investigate depends upon trade-offs that are analogous to those considered by the RA in our model.

exogenously set penalty is not tied to the endogenously-determined cost of fraud, the behavior of the monitor in Wang (2006) is not motivated by social welfare. behavior is motivated by social welfare.7 The remainder of the paper is organized as follows. Section 3 discusses the equilibria. Section 2 describes the model. In contrast, our monitors

Section 4 discusses the empirical implications of the model.

Section 5 concludes. All proofs, as well as a numerical example, are provided in the appendix. 2. The Model The model consists of a large number of competitive firms (in a variety of different industries) owned by atomistic risk-neutral investors and managed by risk neutral agents. Every firm within a given industry has exactly the same characteristics. The sequence of events unfolds over four periods as following. 2.1. Period 1: The Contracting Stage In period t = 1, an entrepreneur with an idea starts a firm consisting of real and intellectual assets whose initial value is normalized to 1. The entrepreneur has limited expertise at managing the on-going operations of the firm and, thus, hires a wealth-constrained professional manager from a competitive managerial labor market to manage the firm for her. Because the manager is wealth constrained, the first-best contracting solution, in which the entrepreneur sells the firm to the manager, is not feasible. Instead, the entrepreneur offers the manager a compensation contract ( w , ) , where w is a nonnegative fixed wage (paid to the manager at t = 1) and is

the percentage of the firms shares offered to the manager in the form of a stock option with a zero strike price; the option vests in period t = 2. The manager either accepts or rejects the contract,

Both the investors in PSW and the regulatory agency in our model trade off the benefit of reduced over-investment against the investigation cost (which includes the cost of investigating truthful firms). In both models, the deadweight loss associated with fraud-induced resource misallocation and the costs incurred to limit fraud are minimized.

based on a comparison of the managers expected utility under the contract and his reservation utility, which for simplicity (and without a loss of generality) is assumed to be zero. Once a manager has been hired and the terms of the contract are set, the entrepreneur sells her ownership stake in the firm at an initial public offering. The risk-neutral entrepreneur chooses the contract in order to maximize her expected wealth, given that the value of the firm will depend, via rational expectations, on the incentives embodied in the contract and on other features of the market, especially the regulatory environment and the behavior of the regulatory agency (hereafter referred to as the RA). After the manager is hired and the IPO is complete, the manager exerts an unobservable amount of costly effort e , which affects the return on the firms assets realized in period t = 2 (described further in the next section). The manager chooses the amount of effort to exert given the trade off between its beneficial effect on his compensation (via its effect on firm value) and its detrimental effect on his utility via a disutility of effort given by

e2 , where is a positive

constant. That is, the managers objective function is U(.) = E[W] -

e2 , where E[W] is his

expected wealth conditional on the contract and the economic/regulatory environment.

2.2. Period 2: The Reporting and Investigating Stage

In period t = 2, the return on the firms initial assets is realized and privately observed by the manager. For simplicity, we assume that the gross return of the firm is either a H or a L < a H .8 The probability that the gross return is a H is P (e) = e , where (i.e., the marginal
In the real world, the returns of firms are likely to be continuous. However, all that is required for our results is that the return support be bounded above. If the return distribution is bounded from above, then there is a limit to the amount that higher-type managers can exaggerate their return. As a result, since the higher-type managers will not always report returns exceeding those of lower-type managers (at some point their claims cease to be feasible), there will exist situations in which lower-type managers will pool (with some positive probability) with higher-type managers.
8

productivity of effort) is a positive constant and e [0,1 / ] ; while the probability that the gross return is a L is 1 - P (e) . Once the manager privately observes the realized return on assets, he must make an earnings report (denoted r) to the market. Since the realized return is not directly observable by anyone other than the manager, the manager can chose to either report truthfully or fraudulently. We assume the manager can either truthfully disclose or inflate his earning. Specifically, a manager with a realized return a H reports earning truthfully thus r( a H ) = a H . However, a manager with a realized return of a L may report truthfully (i.e., r( a L ) = a L ) or fraudulently (i.e.,

r( a L ) = a H ). 9

Given the equilibrium information content of the managers equilibrium

reporting strategy, market investors rationally value the firm conditional on the firms reported earnings. The manager then exercises his stock option and sells all his vested shares to the market.10 The RA, which seeks to protect the interests of investors (including the entrepreneur), is responsible for investigating and detecting fraud. For simplicity, we assume that if the RA chooses to investigate fraud, it will always detect fraud when it exists and will never detect fraud when it does not exist. That is, the RA does not make ex-post Type I or Type II errors. In order

Since there are only two return values possible, the manager will report either r = a H or r = a L . reports will not be credible as other values of r are not feasible.

Any other

Here we assume the manager has a short horizon. We do this for two reasons. First, there is some evidence that this is consistent with the situations in many real-world fraud cases. (Bergstresser and Philippon (2006) present evidence that CEOs exercise unusually large amount of options during periods of high accruals (which indicate intensive earning manipulation).) Second, it markedly simplifies the analysis without affecting the nature of the results. Even if the manager is given a multi-period contract, as long as he receives some compensation based on intermediate value of the firm, he will still have an (albeit mitigated) incentive to manipulate earnings reports to raise the intermediate price of the firm. In a setting in which the manager receives part of his compensation based on long-term value, the weight placed on long-term value will reduce the incentive to commit fraud and will complement the incentives created by penalties for fraud. Since our model has both incentives (EBC) and disincentives (penalties) to commit fraud, adding an addition disincentive (by placing weight on the terminal value) will not change the nature of our results. See Wang (2006) for an analysis of the case where the manager has a long horizon.

10

10

to detect fraud, however, the RA must choose to investigate fraud. Thus, a manager who has committed fraud can get away with it if the RA chooses not to investigate the managers firm. We assume that it costs C I > 0 to investigate fraud and that the RA decides whether to investigate fraud to minimize the deadweight loss associated with fraudulent reporting, taking into consideration both the expected benefits (i.e., the deterrent effect) and costs associated with its investigation policy. If the RA detects fraud, the manager is assessed a penalty

fa D f (a H a L ) proportional to the extent of the misreporting, where f is the constant of proportionality.11

2.3. Period 3: The Investment Stage

At t = 3, new investment opportunities may become available. The probability that a new investment opportunity arrives is . For simplicity, all new investment opportunities require an additional investment of capital equal to I , which is raised by issuing equity. If a new investment opportunity arrives, its gross return will be + , where is the mean of the gross return and is a white-noise term following N (0, 2 ) . We assume that the manager privately observes the realized value of mean , but that, with respect to the investors information set, is a random variable distributed as Uniform( , ) with 0 < < 1 < and E ( ) 1 . 12 The white noise

error is realized at the end of the economy at t = 4. The distributions of and are common knowledge. The manager makes an optimal investment decision to maximize his own payoff; if he is indifferent between investing and forgoing the project, he will invest in any positive

11

As we will see below, the only type of fraud committed is when the manager reports a H when in fact a L has occurred. Thus, the extent of the fraud is a D (a H a L ) and f is the marginal penalty. For simplicity, we do not model the optimal choice of f by the RA. We assume E ( ) 1 to reflect the fact that it is not easy to find positive NPV projects in the real world.

12

11

NPV project to maximize the ultimate shareholders value.13

2.4. Period 4: The Liquidation Stage

At t = 4, the firm is liquidated and the gross returns from old and (if existing) new projects are distributed to shareholders. Once the firm is liquidated, the truth in past reports may be revealed. For example, consider the situation in which there is undetected fraud (i.e., the gross return was aL, the manager reported aH but the RA did not investigate at t = 2). If there is no new investment at t = 3, the realized terminal cash flow of the firm, aL, will make it apparent that the manager committed fraud when he reported aH. The RA has no discretion and has to investigate such cases. If, however, there is new investment at t = 3, the terminal cash flow will be aL + + . As a result, if the new investment is taken, there will be no direct evidence of fraud since the support of aL + + overlaps with the support of aL + + . We assume that if the cash flow is sufficiently low, such that the probability that the t= 2 gross return was aL is sufficiently high, then the RA will investigate.14 That is, there is a set critical value K such that if the terminal cash flow is at or below this critical value, the RA will investigate fraud. Since the realized cash flow of a non-fraudulent firm may also fall below K, non-fraudulent firms may be investigated. However, again we assume that the RA does not make any ex-post Type I or Type II errors once it decides to investigate. That is, if it investigates a fraudulent firm, the manager is caught and assessed the penalty; if it investigates a non-fraudulent firm, the manager is exonerated and no penalty is assessed.
These assumptions are employed to abstract away from the Myers-Majluf type underinvestment problem associated with high-return firms. These assumptions simplify our model and are consistent with the manager receiving some performance-based compensation based on the overall terminal value of the firm. If we instead assume the manager acts to maximize old shareholder value as is in Myers and Majluf (1984), the fraudulent firm manager will have additional incentive to overinvest (i.e., to exploit the overvaluation of his firms stock price). Thus, the nature of our results will be the same. Wang (2006) provides a detailed justification for this behavior by showing how mingling cash flows from multiple projects negatively affects the information content of realized returns with respect to fraud. We adopt a simple abstraction of her model to capture this feature.
14 13

12

Given that the RA investigates at t = 4 whenever the reported gross return from new investment is lower than K , if prior undetected fraud exists and there is new investment at t = 3, the firm will be investigated if its gross return from the new investment is lower than
aD +K. I

The first term is the difference in the return that the new investment must make up for the final return to be consistent with the managers prior earnings report. The value of K is an addition term that requires the reported return from new investment to be sufficiently unusual to warrant investigation. We assume that K is small (K<<1) and its value is set prior to t = 1.15,16 The investigation cost of the RA is again C I per case.
3. Equilibrium

The model is solved by backward induction.

Section 3.1 determines the optimal

investment rule of the manager at t = 3 as a function of the exogenous parameters as well as the endogenous variables determined prior to t = 3 (i.e., the compensation contract ( w , ) and the managers reporting strategy). Section 3.2 derives the optimal reporting and auditing strategies, anticipating the investment strategy derived in section 3.1 and taking as given the compensation contract ( w , ) . Section 3.3 characterizes the optimal compensation contract ( w , ) offered to

15

Since the objective of the RA is to minimize deadweight loss, if the RA chooses K at t = 4, it will set K equal to zero. This is true because (as can be seen in the next section) the only benefit of fraud detection in our model is to prevent the cost associated with over-investment, which, at t = 4, is sunk. Thus, there is no benefit to investigating at t = 4 and, given the positive investigation cost, the optimal amount to investigate is zero.

We assume that K is small and exogenously set (rather than endogenously determined) in the model. We do this because there are likely to be many factors outsider the bounds of this model that will affect the setting of K in the real world. For example, in the model, for simplicity we assume that the RA makes no ex-post type I and type II errors, and that the investigation cost is fixed per case. However, in the real world, corporate executives are insiders and understand the complicated businesses of their firms much better than outsiders. When a firm is big and there exist multiple investment projects, it is very difficult for outsiders to distinguish cash flows from different projects of the firm. The investigation and litigation process will become very lengthy and very costly, and ex-post type I and type II errors, which are very costly to stakeholders of the firm, are also possible to be made. Thus, although ideally an optimal K should be set through trading off the expected benefits of deterring fraud and the investigation costs, there will be other important factors outside the boundary of this model that will affect the optimal K, which justifies K to be set exogenously and be small.

16

13

managers at t = 1, anticipating all of the optimal strategies in the subsequent sub-games. To fully characterize the equilibrium, Section 3.4 describes the conditions under which a specific type (e.g., separating, pooling, or mixed) of equilibrium obtains. The equilibrium concept we adopt is the rational expectation perfect Bayesian equilibrium (PBE) characterized by: (a) common belief of the RA and market investors (regarding the investment behavior of the manager at t = 3, the probability that a firm that reports a high return is truly an a H -type firm at t = 2, and the managerial effort choice at t = 1) is reasonable (derived from the managers effort choice and reporting and investing strategies with rational expectation and using Bayes rule whenever possible); and (b) given the reasonable common belief, the effort choice and reporting strategy of the manager, the evaluation strategy of market investors and the detecting strategy of the RA are sequentially rational.17
3.1. The Investment Decision of the Manager at t = 3

In the liquidation process (i.e., t = 4) prior undetected fraud (if any) will be investigated and the fraudulent-reporting ( a L -type) manager will be subsequently penalized if either new investment does not occur at t = 3, or the realized return from new investment is too low (i.e.,

+ <

aD + K ). I

At t = 3, if there is no prior undetected fraud, the manager will invest in the expansion opportunity when 1 . That is, a non-fraudulent manager only invests in opportunities with an expected NPV greater than or equal to zero. If there is prior undetected fraud, however, the manager will rationally invest in any project (even a negative expected NPV project) that arrives.

The PBE concept here should be equivalent to the sequential equilibrium concept of Kreps and Wilson (1982) since there are only two types of firms in the model the sequential equilibrium concept will be stronger if the types of firms are more than two. See Fudenberg and Tirole (1991) for a proof.

17

14

Thus, fraud causes an overinvestment problem. 18

The reason for this overinvestment is

straightforward. If the manager does not invest in the expansion opportunity, evidence of his prior fraud will be exposed in liquidation and he will be penalized with probability 1. If, however, he invests in the project, the probability of his getting caught for prior fraud will be
aD +K aD Pr ob( < + K ) = ( I ) < 1 . Since this probability is less than 1 for all projects, I the manager with undetected fraud will rationally invest in any available project in order to reduce his probability of being penalized at t = 4. Consequently, the ex ante probability of a manager with new investment (and undetected fraud) being prosecuted at t = 4 is aD +K I ( ) d .

(1)

Moreover, at t = 4 the RA is also likely to investigate a non-fraudulent a H -type firm with new investment at t = 3, if the firms gross return from new investment happens to be lower than
K . The ex-ante probability of this (ex-ante) type I error is

1 K ( )d ,

(2)

which is less than H. The expected value of future investments (including the possibility that a new project will arrive) from non-fraudulent managers is given by

One way to solve the overinvestment problem is to decouple the investment decision at t=3 with the (fraudulent) reporting decision at t=2 so that the person making the investment decision is not interested in covering up fraudulent reports. One way to do this is to always require the manager at t=2 be replaced. But, this policy is likely to be very costly (especially in non-fraud situations) as continuity in management/leadership is important and the expertise of the initial manager is likely to be useful as old projects come to fruition.

18

15

I ( 1) 2 > 0. 2( )

(3)

While the expected value of future investments from fraudulent manager is simply

I ( + 2)
2

0.

(4)

The difference between the expected value of optimal future investment (G) and sub-optimal future investment associated with fraud (B) is the expected over-investment (deadweight) loss associated with fraud:
J GB

I (1 ) 2 19 . 2( )

(5)

3.2. The Reporting and Auditing Strategies at t = 2

The following proposition describes the optimal reporting strategy of the manager and the auditing strategy of the RA as a function of the contract and the parameters of the model. Intuition is provided following the proposition. First, we define:
Crit PRA

J C I (1 H ) , J C I (1 H ) + C I (1 L)

(6)

Crit PManager

(1 + H ) fa D / + J , aD + J

(7)

and

Crit

f . 1 H

(8)

In what follows, we assume that f 1 (i.e., penalties such as loss of reputation and a prison term
19

So that the problem is non-trivial, we assume that the expected overinvestment loss J is larger than (1 H )C I ; otherwise the RA would never audit at t = 2.

16

are substantial) so that / f 1 . Further define P * to be the probability of realizing the high return a H .
Proposition 1: For given values for and P * (and the exogenous parameters):

a. If Crit , all managers truthfully report and the RA does not audit (i.e., separating). b. If > Crit , the exact reporting and audit strategies depend on the following conditions. i.
Crit Crit When P * Max[ PRA , PManager ] , the optimal reporting strategy is pooling (i.e.,

r (a L ) = a H and r (a H ) = a H ) and the RA does not audit. ii. When

Crit Crit P * < Max[ PRA , PManager ] , and Crit Crit PRA PManager

(or, equivalently,

f (1 + H )aD / VDM ), the optimal reporting strategy is mixed.

Specifically, a H managers always truthfully report, but a L managers fraudulently report a H with probability m and truthfully report a L with probability 1 m ,where m= C I P * (1 L) . (1 P * )[ J C I (1 H )] (9)

In this case, the RA investigates any claimed a H -type firm with probability n :
M V D 1 n = 1 (1 ), fa D (1 H )

(10)

where
M VD = a D

C I (1 L)(a D + J ) . J C I (1 H ) + C I (1 L)
Crit Crit PRA < PManager

(11) (or, equivalently,

iii.

When

Crit Crit P * < Max[ PRA , PManager ] , and

M < f (1 + H )a D / V D ), the optimal reporting strategy is mixed, with a H

managers always truthfully reporting, but a L

managers fraudulently with

reporting a H with probability m and truthfully reporting a L probability 1 m , where m= P * a D [1 (1 + H ) f / ] . (1 P * )[ J + (1 + H ) fa D / ] (12)

17

In this case, however, the RA never investigates.

Proof. See the appendix.

Henceforth we refer to situations in which the condition specified in part a holds as either the truthful or separating equilibrium. Similarly, situations in which the condition in part b holds will be referred to as fraudulent equilibria; within the set of fraudulent equilibria, we will refer to pooling (as in part b.i), fully mixed (as in part b.ii), or partially mixed (as in part b.iii) equilibrium. When the firms growth potential is sufficiently low (i.e., Crit as in part a.), the probability that the manager will be able to mask his fraud via new investment will be low enough to prevent fraudulent reporting. Furthermore, since there is no fraudulent reporting, the RA does not need to audit at t = 2. The proposition also shows that the higher the EBC, the lower the threshold growth potential. This is true because the higher the EBC, the greater the managers gain from committing fraud and selling his shares at the fraudulently inflated price. Thus, for a given distribution of potential growth levels within an industry, the larger the value of , the larger is the set of firms in that industry that will commit fraud.20 When the industrys growth potential is sufficiently high (i.e., > Crit ), low-return firms may misreport their earnings, knowing that it is very likely that they will receive a new investment opportunity at t = 3 that will allow them to hide their prior fraud. When P * is sufficiently high (as in part b.i), the market will believe that any firm reporting a high earning is very likely telling the truth, and thus will give it a value close to that given to a true high earning type. In this case, an aL manager has much to gain by reporting a H and selling his shares at the relatively high pooled price. Thus, in addition to the higher likelihood (due to high ) of being able to obscure

20

This is consistent with recent empirical studies. In particular, Johnson, Ryan and Tian (2003) find that their sample of exposed fraud firms uses a significantly higher level of EBC than their (size- and industry-matched) control sample. Peng and Rell (2004) find that incentive pay in the form of vested options increases the probability of securities class action litigation.

18

past fraud with new investment (which lowers the expected penalty), the expected benefit to misreporting is also greater due to the higher pooled price. In addition, the RA will not

investigate any claimed high earning firm (even if all low return firms misreport their earnings) since, in this case, the deadweight cost C I of investigating potential fraud is large relative to the low expected gain from preventing the infrequently occurring aL firms from committing fraud and over-investing. In such a favorable environment, all low-earnings firms will naturally misreport their earnings to pool with the high-return firms. When P * is small (as in parts b.ii and b.iii), the optimal reporting strategies are mixed. To understand this result, suppose that all low-earning firms chose to pool with the high earning firms by reporting r (a L ) = a H . The RA would rationally investigate any firm claiming to be a
a H type, since the expected gain from preventing the deadweight overinvestment loss of a

potential fraud firm will outweigh the investigation cost. It then would not be profitable for the aL firm managers to misreport earnings. However, if aL firms misreport earnings with a certain equilibrium probability (less than 1), the RA will choose not to investigate all claimed high earnings firms. The proposition shows that there exists a mixed fraudulent reporting probability that makes the RA indifferent between investigating any claimed high-earnings firm and not investigating such a firm. Similarly, there exists a mixed strategy investigation probability that makes low-earnings managers indifferent between truthful disclosure and fraudulent reporting. Hence, the proposition characterizes the mixed strategy probabilities m and n such that the mixed strategies of low-earnings managers and the RA are rational reactions to each other. When
Crit Crit the parameters are such that PRA < PManager , the manager optimally mixes even though the RA

never investigates at t = 2 because the mixed fraudulent reporting probability m is low enough

19

(due to low ) such that the RA cannot justify bearing the investigation costs at t = 2. Corollary 1 is self evident given Proposition 1.
Corollary 1. If the equity-based executive compensation is small enough such that fH ,

that is, Crit 1 , it will be impossible for fraud to exist in any equilibrium; if is large enough such that f , that is, Crit 0 , fraud will exist in any equilibrium.

It is clear from Corollary 1 that if the probability of a fraudulent-reporting manager with new investment being prosecuted at t = 4, H , is big, then it will be difficult for fraud to exist in equilibrium. However, if is small, K is small and I is relatively large compared with a D such that a D / I + K , then H 0 , which we assume in the rest of the paper to ensure the existence of equilibrium fraud.21
3.3. The Contracting Problem at t = 1

In this section, we examine the contracting problem faced by the entrepreneur who anticipates the equilibrium strategies that will occur in the subsequent sub-games. For each of the potential equilibria, we solve for the managerial compensation contract that maximizes the IPO price, which reflects the effect of the contract on managerial effort and the assumed subsequent strategies. In Section 3.4, we then refine the set of potential equilibria by including only those for which the assumed reporting/auditing strategies are optimal given the optimal contract under those assumed reporting/auditing strategies. The next proposition specifies the optimal contracts for each of the potential equilibria.
Proposition 2: Let denote the set of all possible collections of the exogenous parameters.

Let T denote the set of parameters for which the truthful equilibrium obtains at t = 2. Similarly denote P , M , and N as the sets of parameters for which, respectively,
Corollary 2 presented in the appendix provides partial comparative static results on how the mixing probabilities in the fully mixed case (characterized in b.ii of Proposition 1) vary with growth potential and marginal fraud penalty given values for and P * . These partial comparative static results are useful in the proofs of subsequent propositions and corollaries.
21

20

the pooling, the fully mixed, and the partially mixed equilibrium (in which No monitoring occurs) obtain. For each type of t=2 equilibrium, the optimal compensation contract ( w * , * ) and the managerial effort e * induced by that optimal contract are as follows: a. Separating: For any T ,

(a L + G ) 1 * wT* = 0 , T = [1 ]; 2 (a D ) 2
* * eT = T

(13) (14)

a D .

b. Pooling Equilibrium: For any P ,

(1 + H ) fa D (a + B) 1 * * 2 L wP = 0 , P = [1 ]; aD + J 2 (a D + J ) 2
* * eP = P

(15)
(16)

(a D + J ) (1 + H ) fa D . +

c. Fully Mixed: For any M ,

(a + G ) 1 * * wM = 0 , M = [1 2 L M ] ; 2 VD
* * eM = M M VD ,

(17)

(18)

where
M VD [1 +

C I (1 L) C I (1 L) J M ] = aD > VD , J C I (1 H ) J C I (1 H )

M and VD is as defined in Proposition 1.

d. Partially Mixed: For any N , Optimal contract does not exist; e* = N

(1 + H ) fa D .

(19)

Proof. See the appendix.

Since no optimal compensation contract exists when agents anticipate the partially mixed strategy equilibrium at t = 2, the partially mixed strategy equilibrium does not exist for the overall game. Thus, this case is not analyzed further below.

21

Corollary 3: The following comparative static relationships hold.

a. Separating Equilibrium: for any strictly inside T : i.

* * * * d T d T d T d T = 0, > 0, < 0, < 0; d df d d
* * * deT deT deT < 0, = 0, > 0, d df d * deT < 0; d

ii.

iii. b.

dPT* dPT* dPT* < 0, = 0, > 0, d df d

dPT* < 0. d

Pooling Equilibrium: for any strictly inside P :

i.

* * * * d P d P d P d P >0, < 0, >0, < 0; d d df d * de * de * deP P P < 0 if f is sufficiently large , > 0, > 0, d df d

ii.

de * P < 0; d dPP* < 0. d

iii. c.

dPP* dPP* dPP* if f is sufficiently large, > 0, > 0, <0 d df d

Fully Mixed Equilibrium: For any strictly inside M :22

i.

* * * * d M d M d M d M is ambiguous, = 0, > 0, < 0; d d df d * * * deM deM deM is ambiguous, = 0, >0, d df d * * * dPM dPM dPM is ambiguous, = 0, > 0, d df d * deM <0; d * dPM < 0. d

ii.

iii.

Proof: These results can be easily verified using Proposition 2.23 3.4. The Equilibrium of the Overall Game

22

If we take into consideration the fact that market investors can rationally infer the probability n that the RA adopts in investigating fraud (i.e., market investors can rationally expect the RA to correct the overinvestment M M problem of a portion n of the fraudulent reporting firms) at t = 2, V H thus V D will be greater, and the changes M M in V D and n will always reinforce each other. Since n / f < 0 , we will have VD / f < 0 , and, as a result * * * d M / df < 0 , deM / df < 0 , and dPM / df < 0 . See also footnote 28.

Since the proofs are fairly tedious but straightforward, we do not provide proofs in the appendix for brevity. However, they are available upon request from the authors.

23

22

In this section we show that for each of the three types of equilibrium (separating, pooling, and fully mixed) there exists a non-empty set of parameters for which a particular type of equilibrium obtains. Proposition 3 is obvious given Propositions 1 and 2:

Proposition 3: For sufficiently small, the equilibrium to the overall game is such that the

truthful separating equilibrium obtains in the t = 2 reporting/auditing sub-game. Thus, the set
T is non-empty and includes firms/industries that have low growth opportunities.

Proof: See the Appendix.

In the event that there are high growth options, Proposition 4 states that, depending upon the collection of exogenous parameters, either the pooling or the fully mixed equilibrium obtains.

Proposition 4: For sufficiently high, the equilibrium to the overall game is such that either

the pooling or the fully mixed equilibrium obtains in the t = 2 reporting/auditing sub-game. Thus, the sets P and M are non-empty and include firms/industries that have high growth opportunities.
Proof: See the appendix.
Figure 1 illustrates how the equilibrium depends upon the parameters. The figure plots
* * * T , P , and M as a function of . Corollary 3 and the proof of Proposition 4 justify the

ranking and the shape of each curve. In addition, figure 1 also plots Crit =

f as a function of 1 H

. That is, the line labeled Crit denotes, for a given , the specific value of (i.e., the value

on the Crit line associated with that ) that is such that if is smaller (greater) than that value,
the separating (a fraudulent) equilibrium obtains. The line Crit denotes, for a given , the
Crit Crit M boundary where PRA = PManager (or equivalently = f (1 + H ) a D / V D ); any strictly to the

M right of this line corresponds to a value such that f (1 + H )a D / V D , which is the condition

for the fully mixed equilibrium. As implied by Proposition 1, in order to obtain a separating equilibrium, and must

23

be below (to the left of) the Crit line. Similarly, in order for a pooling equilibrium to obtain, and must be above (to the right of) the Crit line. A fully mixed equilibrium obtains if and
are above (to the right of) the Crit line. For both the pooling and fully mixed potential

equilibria, Proposition 1 imposes an additional condition on the equilibrium level of P* which will further refine which equilibrium obtains. Finally, if, for a specific , all of the corresponding values of are such that none of the above rankings is satisfied, then no equilibrium exists. We next consider some specific ranges. First consider the equilibrium for any value of 1 . Specifically, consider = 0 .
* For 0 , the optimal under the assumption that the separating equilibrium obtains is T (0 )

* (i.e., the value of on the T line that corresponds to 0 ). For this combination of and ,

the conditions required for the separating equilibrium are satisfied:

* 0 < Crit (T (0 )) . None of

the other potential equilibria obtain for < 1 because the optimal for that under other potential equilibria do not satisfy the conditions required for those equilibria (since
* * 0 < Crit ( P (0 )) and 0 < Crit ( M (0 )) ).

Next consider any value of > 3 . In particular, consider = 4 . For 4 , the optimal
* * * under the various potential equilibria are T (4 ) , P (4 ) , and M (4 ) . For each combination of

and , only the conditions required for the pooling and fully mixed equilibria are satisfied.
* * * The separating equilibrium does not obtain in this case since 4 > Crit (T (4 )) . Let PM ( M )

denote the probability of the high return given the optimal and the optimal managerial effort under fully mixed and
* PP* ( P )

denote

that

probability

under

pooling.

If

* * Crit * Crit * * Crit * Crit * PM ( M ) < Max[ PRA ( M ), PManager ( M )] and PP* ( P ) < Max[ PRA ( P ), PManager ( P )] , then the only

* * Crit * Crit * equilibrium at = 4 is the fully mixed equilibrium. If PM ( M ) > Max[ PRA ( M ), PManager ( M )]

24

* Crit * Crit * and PP* ( P ) > Max[ PRA ( P ), PManager ( P )] , then the only equilibrium at = 4 is the pooling

equilibrium.

* * Crit * Crit * * Crit * Crit * If PM ( M ) < Max[ PRA ( M ), PManager ( M )] and PP* ( P ) > Max[ PRA ( P ), PManager ( P )] ,

then both pooling and mixed equilibria are possible, and which one occurs depends on the entrepreneurial utility in these two equilibria (as the entrepreneur is the Stackelberg leader in the overall game and can pick the equilibrium by setting ). A similar analysis implies that for values of such that 2 < < 3 , the only possible equilibrium is pooling while for values of such that 1 < < 2 , no equilibrium exists. Figure 2 is similar to Figure 1 and isolates the effect of f on the equilibrium. Specifically, it shows how and f are related for the separating (T), pooling (P), and mixed (M) equilibria. The line labeled Crit indicates the boundary between separation and

fraudulent equilibria, with values of f below (i.e., to the right) such that separation obtains (i.e.,
> Crit ).
The line labeled Crit denotes the boundary between fully mixed and partially

M mixed, with values of f to the left implying fully mixed (i.e., f (1 + H )a D / VD ). As can

be seen, for f greater than f3, only separating equilibria occur. For f between f2 and f3, either pooling or separating occur (depending on whether the condition on P is satisfied for pooling or not and on the entrepreneurs utility). Between f1 and f2 only pooling is possible. And for f < f1, either pooling or fully mixed occurs (again depending on the conditions on P and on the entrepreneurial utility).

4. Empirical Implications
In this section we examine the implications of the model with respect to (1) the commission, detection, and observation of fraud, (2) the impact of fraud on equity based executive compensation, and (3) public policy and economic performance.

4.1. The Incidence of Fraud

25

With respect to the incidence of fraud, the first empirical implication of the model is that fraudulent reporting will be concentrated in new economy industries for which growth opportunities are high (Propositions 1, 3, and 4). Conversely, fraud will not occur in old economy industries for which the expected arrival of new projects is not sufficient to provide enough potential masking of fraud to create the incentive to commit fraud in the first place. The second empirical implication of the model is that while fraud incentive is strongest in good times (i.e., when the marginal productivity of effort and the probability of realizing good earnings are high, and the pooling PBE occurs), a significant amount of detected fraud is more likely to occur when the new economy industries fall into downturns (i.e., when the marginal productivity of effort and the probability of realizing good earnings are low, and the fully mixed PBE occurs). To understand this point, consider a situation where a high growth industry is in the pooling PBE. If there is a decrease in the marginal productivity of effort (which usually is a result of industry downturns), the equilibrium probability of realizing good returns, PP* , will decrease.
Crit Crit * However, the threshold probability for pooling, Max[ PRA , PManager ( P )] , will weakly increase

* due to a reduction in P caused by the decrease in . If this decrease in is big enough, it

Crit Crit * * can result in PP < Max[ PRA , PManager ( P )] . The pooling PBE is no longer feasible and a drop

from pooling to fully mixed occurs (the proof of Proposition 4 shows that the fully mixed PBE is the only feasible equilibrium in this case since

Crit Crit * Crit Crit * * PM < PP* < Max[ PRA , PManager ( P )] Max[ PRA , PManager ( M )] ).

Thus we will observe

substantially worse economic performance and a significant amount of detected fraud in the industry. The third main empirical implication of the model concerns the relationship between the commission of fraud and the ex-post observation of fraud in the fully mixed strategy equilibrium.
26

Combining the results of Corollary 2 (in the Appendix) and Corollary 3, we have the following
* comparative statics regarding the probability of committing fraud (i.e., (1 PM )m ), the

probability of detecting fraud (i.e., n ), and the probability of observing firms being caught for
* fraud (i.e., (1 PM )mn ).

Corollary 4: If the set of parameter values are such that the fully mixed strategy equilibrium

obtains, the following comparative statics results hold for the probability of committing fraud, the probability of detecting fraud, and the probability of observing firms being caught for fraud: 1)
* * d (1 PM )m d (1 PM )mn dn f < 0, > 0, > 0 if * 2 ; d d d M

* * d (1 PM )m d (1 PM )mn dn 2) = 0, < 0, < 0 ;24 df df df * d (1 PM )m dn > 0, > 0, d d * d (1 PM )mn > 0; d

3)

4)

* * d (1 PM )m d (1 PM )mn dn < 0, < 0, < 0. d d d

Proof: Parts 2), 3), and 4) can be easily verified. Part 1) can only be verified numerically. We
examine a wide range of parameter values and find the results of Part 1) hold. Part 1 shows that an increase in an industrys growth potential has opposite effects on the probability of committing fraud and the probability of detecting fraud in the fully mixed equilibrium, with the increase in the detection probability dominating the decrease in the commission probability, resulting in a net increase in the amount of frauds that are exposed. A numerical illustration is provided in Panel D of Table 1. Thus, if the number of exposed frauds is taken as a (proportional) proxy for the amount of fraud being committed, cross-sectional
24

If we take into consideration the fact that market investors can rationally infer the probability n that the RA adopts in investigating fraud (i.e., market investors can rationally expect the RA to correct the overinvestment * problem of a portion n of the fraudulent reporting firms) at t = 2, we should also have d (1 PM )m < 0 . See also df footnotes 22 and 28.

27

comparisons of exposed frauds in industries differing only in their growth potential would inappropriately conclude that the industries with fewer exposed frauds have less fraud being committed. The basic intuition is as follows. When the industrys growth potential increases, the expected overinvestment loss of a potential fraud firm will increase and it will be more difficult for fraud to get exposed in future liquidation. Thus, the RA will choose to detect fraud at t = 2. Low-earnings managers hence need to commit fraud with lower probability in order to keep the RA indifferent between investigating and not investing a claimed high earning firm in equilibrium. When the industrys growth potential increases, however, fraudulently-reporting managers are less likely to suffer a penalty (from being caught) in the future (since fraud is now less likely to get exposed in liquidation), and the difference in market value between a claimed high earning firm
M and a claimed low earning firm (i.e., VD ) is larger (since the low earning firms now misreport

their earnings with a lower probability). Hence, earning misreporting becomes more attractive to the low-earnings managers. To keep the low-earnings managers indifferent between truthful disclosure and misreporting, the probability of detecting fraud at t = 2 must increase. When the fraud penalty is substantial, the effect of a reduction in personal cost to fraudulent reporting managers due to the increase in growth potential (thus less likelihood of future fraud exposure) will also be quite significant. Therefore, the probability of detecting fraud has to substantially increase in order to balance this reduction in personal cost to fraudulent reporting managers, resulted in (ex post) higher probability of any firm in the industry being caught for committing fraud. Since a marginal change in fraud penalty does not affect the (t = 2 or t = 4) payoff to the RA, it should not directly cause the probability of committing fraud to change, as this probability

28

should keep the RA indifferent between investigating and not investigating a claimed high earning firm.25 However, as fraud penalty increases, it will be more costly for the low-earnings managers to commit fraud. Thus, the probability of detecting fraud at t = 2 need to be reduced so that the low earning managers can still be indifferent between fraudulent reporting and truthful disclosure in equilibrium, resulted in lower probability of observing firms being caught for fraud. Similar to Goldman and Slezak (2006), our model also predicts that EBC is a double-edged sword. Higher EBC induces more managerial effort but also makes fraud easier to occur in the reporting/auditing sub-game, and increases both the probability of committing fraud and the probability of detecting fraud in the fully mixed equilibrium. Given a marginal increase in EBC (caused by either an increase in marginal productivity of effort or a decrease in managerial disutility of effort , or other reasons not incorporated in our simple model), ceteris paribus, in the fully mixed equilibrium the optimal managerial effort will increase, which in turn will increase the probability of realizing high earnings thus make the RA choose not to investigate a claimed high earning firm at t = 2. Consequently, the equilibrium probability of (any low-earnings firm) committing fraud will increase (so that the RA can still be indifferent between investigating and not investigating a claimed high earning firm). Furthermore, when EBC increases, earning misreporting will become more attractive to the low return manager as his personal benefit from misreporting increases while his personal cost remains unchanged. Therefore, in order to keep the manager still indifferent between truthful disclosure and misreporting in equilibrium, the probability of detecting fraud has to increase. Hence, this increase in both the probability of committing fraud and that of detecting fraud results in higher probability of observing firms being caught for fraud in the industry.
25
* However, as pointed out in footnote 22, a marginal increase in f reduces PM , thus it also reduces the probability of fraud commission.

29

4.2. Incentive Compensation, Effort, and Fraud

The model also generates empirical implications regarding the effects of growth options and fraud penalties on the equilibrium level of and effort (thus economic performance). Corollary 3 specifies the effects. In general, growth opportunities and fraud penalties will have different marginal effects on and effort, depending on the type of equilibrium that obtains for those parameters changes. As the growth potential increases, the equilibrium under separation decreases. When

increases, the firms growth value (thus the firm value) also increases. However, keeping the
compensation contract constant, the managers chosen effort levels remain unchanged given an increase in . The entrepreneur would pay the manager too much for his effort if she did not adjust the original compensation level accordingly the original would be too costly to the entrepreneur given the higher firm value associated with a higher . Therefore, the entrepreneur will lower the managers in response to an increase in . Since changes in the severity of fraud penalty f affect neither the managers effort choice nor the firm value in the truthful disclosure separating equilibrium, the optimal will not vary with f . As the growth potential increases, the equilibrium under pooling generally increases. In the pooling PBE, a marginal increase in will make fraudulent reporting less likely to be exposed in the future. If f is big enough, this reduced likelihood of future penalty will be quite substantial to the manager. The manager will optimally respond by reducing his costly effort. Since he can fraudulently report return as a H if the firm actually realizes a L , and he is now less likely to suffer penalty in the liquidation stage, he needs not exert the same level of costly effort as before. This reduction in managerial effort hurts the entrepreneur (as the firm now has more chance to realize low return and suffer overinvestment loss due to misreporting). Moreover,

30

when increases, the firms overinvestment loss (i.e., J ) will also increase. Therefore, it is meaningful for the entrepreneur to respond by optimally increasing in order to induce more managerial effort. On the contrary, a marginal increase in f will make the manager work harder in order to avoid becoming an a L type and suffer the intensified expected future penalty. This intensified penalty can partially substitute for the effort incentive provided by costly executive compensation. The entrepreneur will then respond by optimally reducing . As the growth potential increases, the change in equilibrium under fully mixed is ambiguous. In the mixed strategy equilibrium, keeping the compensation contract constant and given a marginal increase in , the manager will increase his effort. The reason is that a marginal increase in will enlarge the difference in market value between the claimed a H and
M a L types (i.e., VD ) through reducing the a L types probability of misreporting; thus the

manager will work harder to increase his probability of being an a H type. Thus, an increase in

can partially substitute for the effort incentive provided by , which becomes increasingly
costly to the entrepreneur given the increase in as the firms growth value is now higher. This effect will cause the entrepreneur to reduce the . However, since the increase in enlarges the difference in market value between the claimed a H - and claimed a L -type firms, it will increase the difference in personal benefits between the a H -type and a L -type managers. Therefore, will become more effective in providing managerial effort incentive, i.e., a marginal increase in will spur more incremental managerial effort. Furthermore, given the now increased market value difference between the claimed a H - and claimed a L -type firms, managerial effort becomes more important and valuable to the entrepreneur. Therefore, it makes sense for the entrepreneur to respond by increasing . The simultaneous working of these two

31

effects results in the overall effect of the increase in on being ambiguous. As pointed out by footnote 22, since a marginal increase in f reduces the detection probability n thus reduces
M VD (under fully mixed), it also reduces .

In the US, the old economy industries (e.g., the manufacturing industry) have low growth potential. Thus, they are usually in the truthful disclosure PBE. Our model then implies a negative effect of a marginal change in on equilibrium and no effect of a marginal change in f on for such industries. On the contrary, the new economy industries (e.g., the high-tech industry) have high growth potential. Thus they are usually in the fraudulent equilibria. Our model implies a positive effect of a marginal change in on especially when the marginal productivity of effort is high thus these industries are booming (i.e., under pooling PBE), and a negative effect of f on for such industries.26,27

4.3. Growth Opportunities, Fraud, and Economic Performance

This section examines a situation in which a marginal increase in growth potential can cause a regime switch from a high-performance, no-exposed-fraud equilibrium (i.e., the high EBC and high effort pooling equilibrium) to a low-performance, exposed-fraud equilibrium (i.e., the lower EBC, lower effort fully mixed PBE). Thus, paradoxically, an increase in growth potential, which typically implies increased future prosperity, can, via its impact on fraud and EBC, result in worse economic performance and a significant increase in the amount of fraud exposed. That is, there is a dark side to innovation. For example, following innovations in financial instruments

A numerical illustration of the situation where the marginal productivity of effort is high and the pooling PBE dominates the fully mixed PBE (in terms of entrepreneurial utility) for the new economy industries can be obtained from the authors.
27

26

This implication can potentially explain the empirical finding of Murphy (2003), Itner, Lambert and Larcker (2003), and Anderson, Banker and Ravindran (2000). These papers document that the high growth new economy sector substantially more and more rely on EBC to provide managerial incentive, while the phenomenon is not as pronounced in the low growth old economy sector.

32

designed to increase efficiency in risk sharing and capital allocation (for example, from the financial innovation during the roaring twenties to the recent wide-spread use of mortgage-backed securities), there have been episodes of significant declines in economic performance (the great depression following the twenties and the credit/liquidity crisis following the mortgage meltdown in 2007) combined with the revelation of significant amounts of fraudulent behavior (widespread abuse during the twenties and over dozens of SEC investigations of corporate fraud in connection with the sub-prime mortgage meltdown). Similarly, deregulation which implies increased growth opportunities can also lead to increased exposure of fraud. For example, deregulation of energy markets lead to increased growth opportunities associated with efficiency gains from production and demand smoothing but is also associated with the Enron scandal.
* Figure 3 illustrates this regime shift. The figure plots PP* and PM as a function of .

Crit * Crit * Crit It also plots PRA , PManager ( P ) and PManager ( M ) as a function of . Equations 6 and 7,

Corollary 3 and the proof of Proposition 4 justify the ranking and the shape of each curve. Since
* Crit PP is decreasing in (due to the reduction in managerial effort) and PRA is increasing in

(due to the strategic reaction of the RA to higher deadweight overinvestment loss and lower likelihood of fraud exposure at t = 4), the pooling equilibrium is not feasible beyond 1 (where
* Crit PP = PRA ). Consider an increase in from 0 to 2 . If the high growth industry is in the

pooling PBE at 0 , this increase in growth potential alone will cause the industry to drop from the pooling PBE to the fully mixed PBE (the fully mixed PBE is the only feasible equilibrium at
* * Crit = 2 since PM < PP < PRA ).

We will then observe significantly worse economic A numerical illustration

performance and a significant amount of detected fraud in the industry. of such a fall is presented in Panel B and Panel C of Table 1.

When the growth potential

33

increases from 0.9 to 0.93, the pooling PBE is no longer feasible and a drop from pooling to fully mixed PBE occurs. The economic performance of the industry substantially deteriorates (the probability of realizing good earnings falls from 0.819 to 0.242) and the amount of exposed fraud cases in the industry significantly increases (from 0 percent to 0.52 percent of the firms in the industry).

4.4. Public Policy, Fraud, and Economic Performance

Next we examine how an increase in the penalty for fraud will affect the extent of fraud and the economic performance of the firm/industry. Proposition 1 shows that, holding fixed the

compensation contract, an increase in f will lead to less fraud being committed in the economy as a whole since an increase in f raises Crit ; as Crit increases, more firms will fall into the parameter space in which separation obtains, thus reducing fraudulent activity. presumably the intended effect of increased penalties. Proposition 2, however, implies that as penalties increase, the compensation contract may change. Firms already in T are obviously unaffected by an increase in f. There is no Thus effect is

* impact on the compensation contract, and, as a result, no change in the managers effort ( eT ) or

productivity ( PT* ).
* P .

Firms that are in P , however, will respond to the increase in f by lowering Holding

This happens because the fraud penalties act as substitutes for under pooling.

fixed , the manager will exert more effort when f increases so as to raise P*, which results in a lower probability of realizing the low return state in which the manager knows he will commit fraud. An increase in the penalties make the manager want to avoid the low return state more

and, thus, raises his effort level. Given the increased effort, the entrepreneur can scale back , thus allowing shareholders to retain more of the firms cash flow, which raises the entrepreneurs wealth since shareholders are willing to pay more for the firm at the IPO. Proposition 2 shows

34

that an increase in f on net increases e and P*, with the increase in effort due to an increase in f dominating the drop in effort due to the resulting fall in . Thus, for the pooling case, an

increase in f results in both a drop in fraud (as intended) and the spillover effect of an increase in economic performance as the probability of realizing the high return P increases. On the other hand, if the firm is in M , penalties generate mixed results. Although an increase in f will be successful at reducing the amount of fraud committed and observed (see Corollary 4 and footnote 24), an unfortunate side-effect of the increase in f is that both e and P* fall (footnote 22), resulting in declining economic performance. While there are clearly other

forces at work, this result indicates that the increased penalties imposed by Sarbanes-Oxley may have contributed to the weak performance of the U.S. economy following its enactment.

5. Conclusion
In contrast to existing models of fraud, this paper considers an agency model in which the auditing/investigation strategy of the regulatory agency is strategically determined by considering the equilibrium fraud commission strategy of managers and investigation costs. We show that the strategic interaction of the managers and the regulators strategies under an agency framework generates a rich set of implications on (1) the commission, detection, and observation of fraud, (2) the impact of growth opportunities and fraud penalties on managerial equity based compensation, (3) the surprising effect of growth opportunities on short-run economic performance and exposed fraud, and (4) the relationship between public policy with respect to fraud and economic performance. For example, we show that in some cases, an increase in growth potential alone can result in worse economic performance and a significant amount of detected fraud for a high growth industry. We also show that when growth opportunities increase, the amount of fraud being committed may fall, but that the amount of fraud observed (i.e., committed fraud that is detected)

35

will always increase. Thus, cross-sectional and time-series comparisons in the amount of fraud observed may not be indicative of differences in the amount of fraud being committed. The model highlights the existence of an equilibrium link between the level of EBC, economic performance, and the extent of fraud committed and observed.

36

Table 1: A Simple Numerical Illustration of the Model

We assume the following parameter values: the high gross return from the firms assets in place is aH=2, and the low return is aL=0.5; the scale of new investment is I=5, its mean gross return is ~Uniform(0.7, 1.2), and its white noise return error is ~N(0, 0.052); the critical reported level from new investment below which the RA will have to investigate fraud at t = 4 is K=0.3; the marginal fraud penalty is f=1.2; the managers effort disutility coefficient is =0.45; the RAs investigation cost is CI=0.08 per case; the marginal productivity of effort =0.62, thus the probability of realizing a high return is P(e) = 0.62e . Based on these parameter values, it can be calculated that the ex ante probability of a manager with prior undetected fraud and new investment being prosecuted at t = 4 is H=8.48*10-4 and the ex ante probability of the RA investigating a non-fraudulent firm (with new investment) in period 4 is virtually zero, i.e., L=0.

Panel A: Separating (Truthful Disclosure) Equilibrium

0.1 0.2 0.3 0.4 0.5 0.6 0.7

f(1-+H)
1.080 0.960 0.840 0.720 0.601 0.481 0.361

G
0.020 0.040 0.060 0.080 0.100 0.120 0.140

* T

* eT

* P( eT )

0.365 0.360 0.354 0.349 0.344 0.339 0.334

0.754 0.743 0.732 0.722 0.711 0.700 0.689

0.467 0.461 0.454 0.447 0.441 0.434 0.427

Principals Utility 0.776 0.788 0.801 0.814 0.827 0.841 0.854

Panel B: Pooling Equilibrium

0.8 0.85 0.9 0.93 0.95 0.97 0.99 1

f(1-+H) J
0.241 0.181 0.121 0.085 0.061 0.037 0.013 0.001 0.360 0.383 0.405 0.419 0.428 0.437 0.446 0.450

Crit PRA

Crit PManager

G
0.160 0.170 0.180 0.186 0.190 0.194 0.198 0.200

B
(0.200) (0.213) (0.225) (0.233) (0.238) (0.243) (0.248) (0.250)

* P

* eP

P( e * ) P
0.868 0.844 0.819 0.804 0.794 0.784 0.774 0.769

0.787 0.797 0.806 0.811 0.815 0.818 0.821 0.822

0.745 0.582 0.446 0.375 0.331 0.290 0.251 0.232

0.352 0.380 0.408 0.424 0.435 0.445 0.456 0.461

1.400 1.361 1.321 1.297 1.281 1.265 1.249 1.241

Principals Utility 1.240 1.162 1.086 1.042 1.013 0.985 0.957 0.943

37

Table 1 (Continued)

Panel C: Mixed Strategy Equilibrium

0.85 0.9 0.93 0.95 0.97 0.99 1

f(1-+H)
0.181 0.121 0.085 0.061 0.037 0.013 0.001

Crit PRA

Crit PManager

f(1-+H)
M * aD / VD

M VD

1.403 1.403 1.403 1.403 1.403 1.403 1.403

* M

* eM

* P( e M )

0.797 0.806 0.811 0.815 0.818 0.821 0.822

0.780 0.595 0.485 0.413 0.341 0.270 0.234

0.243 0.160 0.112 0.080 0.048 0.017 0.001

1.118 1.131 1.138 1.143 1.147 1.151 1.153

0.250 0.249 0.248 0.248 0.248 0.247 0.247

0.385 0.388 0.390 0.390 0.391 0.392 0.392

0.239 0.241 0.242 0.242 0.243 0.243 0.243

Principals Utility 1.005 1.018 1.025 1.030 1.034 1.039 1.041

Panel D: Probabilities of Fraud Commission and Detection in Mixed Strategy Equilibrium 0.85 0.9 0.93 0.95 0.97 0.99 1
* P( e M ) 0.2388 0.2407 0.2416 0.2421 0.2426 0.2429 0.2431

m 0.0798 0.0761 0.0740 0.0727 0.0714 0.0701 0.0694

n 0.0054 0.0620 0.0929 0.1124 0.1310 0.1488 0.1574

* (1 -P( e M ))*m 0.0607 0.0578 0.0562 0.0551 0.0541 0.0530 0.0526

* (1 -P( e M )*m*n 0.0003 0.0036 0.0052 0.0062 0.0071 0.0079 0.0083

38

Figure 1: Equilibrium in EBC-Growth Potential Space

Crit =
1/2

1 / f 1 H

Crit =

M 1 VD /( faD ) 1 H

Separating Equilibrium
* P (4 ) * T (0 )

Pooling Equilibrium

* T

* M
* T (4 )

Fully Mixed Equilibrium

* P

* M (4 )

* 0 Crit (T (0 )) 1 2

39

Figure 2: Equilibrium in EBC-Fraud Penalty Space

Crit

Pooling Equilibrium

Crit

* P

* T

Separating Equilibrium
* M

Fully Mixed Equilibrium

f1

f2

f3

40

Figure 3:

Discrete Drop from Pooling to Mixed Given an Increase in Growth Potential

P
Pooling Equilibrium

Crit PRA

PP*
Discrete Drop from Pooling to Mixed
* PM
Crit * PManager ( M )

Fully Mixed Equilibrium

Crit * PManager ( P )

0 1 2

41

Essay 2: Strategic Regulator, Sarbanes-Oxley and Fraud*

ABSTRACT
In the wake of recent financial scandals, there are heated debates over whether the SEC is effective in combating fraud as well as over the costs and benefits of the Sarbanes-Oxley Act. This paper investigates two research questions empirically: 1. Does the SEC strategically respond to fraud commission? 2. How effective is SOX in reducing fraud commission? Using a sample of firms subject to SEC litigations for fraud and employing the bivariate probit with partial observability technique, we find strong evidence in favor of theoretical predictions with the assumption that the regulator is strategic in combating fraud, but contradicting to theoretical predictions assuming that the fraud detection environment is exogenous or mechanical (i.e., without a strategic regulator). We also find that SOX has been very effective and decreased fraud commission by two thirds after its enactment. Our finding should provide some useful insight to policy makers in light of the current debates.

* I would like to thank Mathijs van Dijk, Mike Ferguson (my advisor), John Glascock, Hui Guo, Young Koan Kwon, Albert (Pete) Kyle (FMA doctoral seminar session chair), Filippos Papakonstantinou (FMA discussant), Steve Slezak (my advisor), Weihong Song, Marno Verbeek, seminar participants at Rotterdam School of Management (Erasmus University), Singapore Management University, the University of Cincinnati, and Vlerick Leuven Gent Management School, session participants at the 2008 Finance Management Association annual meeting and FMA doctoral seminar in Dallas, Texas for helpful comments and suggestions. All errors are mine.

42

1. Introduction
Recently, in the wake of the Bernard Madoff and Allen Stanford scandals, there are heated debates on the media regarding whether the U.S. Securities and Exchange Commission (the SEC) is effective in combating fraud. Many criticize the regulator as being mechanical and slow in reacting to suspicious fraud cases. For example, a recent Businessweek article claims that with each new scandal, regulators are taking heat for not moving fast enough to protect investors. Madoffs operation first raised red flags more than a decade before the SEC brought charges. There were warning signs about Stanford as far back as the early 1990s.28 Another Businessweek article claims that in hindsight, the (Madoff) episode should have been remembered a decade later when regulators started receiving anonymous tips about potential improprieties at Madoffs operations; but the resulting inquiries yielded nothing significant. That they did not raises serious questions about the SECs ability to combat and prevent fraud.29 Congress also held hearings in January 2009 to discuss why the SEC failed to uncover the Madoff scandal earlier. This debate is also reflected to some degree in academic literature, as different fraud commission and detection theories give different portraits of the U.S. fraud detection environment. For example, in the models of Goldman and Slezak (2006), Kadan and Yang (2005), Robinson and Santore (2005), and Wang (2005a), there is no role of a regulator to strategically detect fraud, thus fraud detection is assumed to be either exogenous or mechanical. In the model of Qiu and Slezak (2008), the fraud detection behavior of a strategic regulator is explicitly modeled the regulator bases its detection strategy on the equilibrium fraud commission strategy adopted by firm managers. Due to the different views of the fraud detection environment, these two groups of models

28 29

Goldstein, Matthew and Peter Wilson. Stanford: Where was the SEC? Businessweek, March 2, 2009. Burrows, Peter and Matthew Goldstein. The SECs Madoff Misery. Businessweek, January 12, 2009.

43

generate very different empirical predictions. For example, according to the first group of models, ceteris paribus, higher growth potential will reduce the (conditional) probability of fraud detection (conditioning on fraud has been committed) thus encourage fraud commission (since future growth opportunities can help noise up cash flows thus help hide past fraud); however, according to the Qiu and Slezak (2008) model, given the existence of a strategic regulator, higher growth potential will actually arouse the regulators attention thus unambiguously increase the (conditional) probability of fraud detection, while its effect on the probability of fraud commission will be ambiguous. Similarly, according to the models without a strategic regulator, higher equity-based executive compensation (EBC hereafter) will encourage fraud commission but has no effect on the probability of fraud detection; while a ceteris paribus increase in EBC, according to the model of Qiu and Slezak (2008), will also unambiguously increase the probability of fraud detection. In this paper, using a sample of 233 firms charged by the SEC for fraudulent information manipulation within the last 10 years, we want to investigate this interesting question of whether the SEC strategically responses to fraud commission or not, by examining whether greater growth opportunities and EBC, after proper controls, arouse the SECs attention and increase the (conditional) probability of fraud detection or not. As correctly pointed out by a New York Times article, because not every fraud is detected, it is hard to know whether there are more instances of fraud or the enforcement level is increasing 30 To achieve our goal, we need to empirically disentangle the unobservable probability of fraud commission and the also unobservable (conditional) probability of fraud detection from the only thing observable from data: the probability of detected fraud. We employ the bivariate probit with partial observability

Labaton, Stephen. Downturn and Shift in Population Feed Boom in White-Collar Crime. The New York Times, June 2, 2002.

30

44

methodology of Poirier (1980) to accomplish this task. This paper makes several contributions to the literature. First, we find that, in the past decade, detected fraud has mainly been concentrated in high-growth, new economy industries and fraud has mainly been detected during the economic downturn in those industries. More importantly, we document strong evidence that, ceteris paribus, higher growth potential (measured by Tobins Q and average past sales growth) is strongly positively associated with the (conditional) probability of fraud detection, while its effect on the probability of fraud commission is ambiguous. We also find that, after controlling for the endogeneity of EBC, a ceteris paribus increase in EBC is strongly associated with both a higher probability of fraud commission and a higher (conditional) probability of fraud detection. We view the documented empirical evidence as being consistent with the notion that the SEC is a strategic regulator in terms of fraud detection. The second contribution of the paper is that, we quantitatively measure the effectiveness of Sarbanes-Oxley Act (SOX) in reducing fraud commission. Due to the fact that our sample period covers both a pre-SOX subperiod and a post-SOX subperiod, our bivariate probit (with partial observability) models allow us to quantify the effect of SOX on unobservable fraud commission. We find that, after the enactment of SOX, the (unobservable) probability of fraud commission significantly decreases by two thirds (from above 3% to only above 1%). Therefore, SOX has been very effective in reducing fraudulent misreporting by firms. Although many studies have documented various costs of SOX, to our knowledge, this is the first study that empirically documents how effective SOX is in reducing (unobservable) fraud commission.31 Thus, our study
31

For example, the Finance Executives International (FEI)s annual survey on SOX Section 404 costs shows that in year 2007, for 168 companies with average revenues of $4.7 billion, the average annual compliance costs were $1.7 million per firm (0.036% of revenue). By measuring changes in market value around key SOX legislative events (assuming that SOX was the cause of related short-duration market value changes), Zhang (2005) estimates SOX compliance costs as high as $1.4 trillion. Piotroski and Srinivasan (2008) examine a comprehensive sample of international companies that list onto U.S. and U.K. stock exchanges before and after the enactment of SOX in 2002. They find that the likelihood of a U.S. listing among small foreign firms choosing between the Nasdaq and LSE's

45

provides useful insight to the continuing debate over the costs and benefits of SOX. Our third contribution is we document evidence that, in the relatively fraud-prone high-growth hightech industries, the usage of EBC is significantly positively associated with growth potential during good times, and the usage of EBC in those industries significantly decreases following the enactment of SOX; while there is no such relationship in the relatively fraud-free low-growth manufacturing industries. This evidence is consistent with the empirical implication of the Qiu and Slezak (2008) model, which indicates that given the existence of a strategic regulator, the high-growth industries and low-growth industries fall into different equilibria, and shareholders strategically take the equilibrium probabilities of fraud commission and detection into consideration in setting the EBC contracts. This, in turn, can be viewed as indirect evidence that the SEC is a strategic regulator in fraud detection. Overall, our empirical results highlight the importance of treating the SEC as a strategic player in the fraud commission and detection game. The interaction between the firm manager and the regulator determines the two unobservable probabilities that determine the level of observed fraud. In turn, the firms shareholders (or BOD) strategically incorporate the implied fraud commission and detection probabilities into the optimal EBC contract awarded to the manager. The remainder of the paper is organized as follows. Section 2 reviews the related literature. Section 3 develops the main hypotheses. Section 4 introduces the empirical methodologies. Section 5 describes the data and variables. Section 6 presents the empirical results. Section 7 concludes.

2. Related Literature
Our study is informed by recent theoretical models of fraud. Povel, Singh, and Winston (2007) and Wang (2005a) each employ non-agency models to study the time-series and
Alternative Investment Market decreased following SOX.

46

cross-sectional distributions of fraud. Povel, Singh and Winton (2007) develop an adverse selection model to study the time-series properties of fraud. In their model, managers seek outside funding for their projects (with the distribution of project quality being exogenously determined), and potential investors face adverse selection in deciding whether or not to provide funding. Managers receive non-contractible control benefits from any (even negative NPV) investment and, as a result, they manipulate information (an interim signal) on the prospects of their projects in an effort to mislead investors into funding negative NPV projects. PSW show that the incentive to commit fraud is high towards the end of a boom but that fraud tends to be revealed in the ensuing bust. However, in PSW, firms are not monitored by a regulatory agency such as the SEC. Rather, the potential investors (upon receiving the interim signal) decide whether or not to investigate the quality of firms further prior to investing. Thus, there is no fraud detection in the PSW model. Wang (2005a) develops a Myers-Majluf type model to show that fraud firms have higher growth potential but experience negative profitability shocks. They also over invest after committing fraud. Her model predicts that firms with higher growth potential are more likely to commit fraud but are less likely to be detected. In Wangs model, the fraud detection environment is mechanical in the sense that fraud detection is not based on the managers equilibrium strategy of fraud commission, but is based on a pre-set detection rule (i.e., fraud detection occurs when the expected difference between realized cash flow and reported cash flow is greater than an exogenous threshold). Thus, there is no role of a strategic regulator in Wangs model. Goldman and Slezak (2006), Kadan and Yang (2005), and Robinson and Santore (2005) endogenize the managerial compensation contract. Each paper develops an agency model illustrating that EBC is a double-edged sword it induces both managerial effort and fraud. These papers generate the same important prediction higher level of EBC usage will lead to a greater

47

amount of both committed and exposed fraud. However, in these papers, the fraud detection environment (i.e., the conditional probability of fraud detection) is exogenous. Due to the exogenous probability of fraud detection, these models generate a similar signal-jamming equilibrium, in which all firms commit fraud and bias reports to some extent while market investors rationally expect the degree of bias in the reported earnings thus are not fooled. These models imply that ceteris paribus a change in any factor that makes it less costly to commit fraud, such as an increase in growth opportunities that decreases the probability of fraud being detected, or a reduction in fraud penalty, will result in a lesser reliance on EBC by shareholders to compensate the firm manager. Qiu and Slezak (2008) explicitly model the behavior of a strategic regulator in detecting fraud. Specifically, they develop an agency model in which managers are induced via an equity-based compensation contract to exert personally costly effort that increases the expected returns of the firm. The realized return of the firm is not observable by the market; rather the manager must report its value. Similar to Goldman and Slezak (2006), EBC provides the manager the incentive to exert effort but also the incentive to upwardly bias reports. The regulatory agency, seeking to minimize the deadweight loss associated with fraud, is responsible for detecting fraud and imposing penalties; it bases its investigation strategy on the managers equilibrium fraud commission strategy in order to optimally trade off the benefit of reducing fraud against the agencys detection cost. Similar to Wang (2005a), the QS model assumes that after the initial stage, a potential new investment project arrives with a certain probability (proxying for growth potential). After privately observing the projects expected profitability, the manager either adopts or rejects it. In order to obscure past fraud, fraudulent managers have the incentive to over-invest (i.e., invest in new projects that have negative expected NPV), which results in the deadweight loss

48

associated with fraud. Due to the existence of a strategic regulator, three potential types of equilibrium may obtain in the QS model (depending upon the parameters characterizing the regulatory and contracting environment): truthful separating equilibrium in which each manager truthfully reports their realized return, pooling equilibrium in which all poorly-performing managers mimic the reports of highly-performing managers but the regulatory agency does not monitor to verify reports, and a mixed strategy equilibrium in which poorly-performing managers commit fraud with an equilibrium probability while the regulatory agency audits those firms that report high earnings with an equilibrium probability. The QS model predicts that, ceteris paribus, higher growth potential has an ambiguous effect on the probability of fraud commission, but it unambiguously increases the probability of fraud detection and results in an overall increase in the net probability of fraud being exposed; greater reliance on EBC increases both the probability of fraud commission and that of fraud detection. This differs from the prediction of models without a strategic regulator. Also, an increase in the fraud penalty reduces both the probabilities of fraud commission and detection. Moreover, the QS model implies that the equilibrium usage of EBC is positively associated with the firms growth potential in high-growth fraud-prone industries during good times, but the equilibrium usage of EBC decreases with the firms growth opportunities in the low-growth fraud-free industries. An increase in the fraud penalty (e.g., due to SOX) reduces the reliance on EBC for firms in the high-growth fraud-prone industries, but has no impact on the use of EBC for firms in the low-growth fraud-free industries. These predictions are also in contrast with the predictions of agency models without a strategic regulator, which imply that when firms have access to more growth projects or face lower fraud penalties (either of which make it less costly to

49

commit fraud), the result will be less reliance on EBC to compensate firm managers. The extant theoretical literature is consistent with several recent empirical findings. Johnson, Ryan and Tian (2007), Burns and Kedia (2006) and Peng and Rell (2004) all document a positive relation between EBC and detected fraud. Murphy (2003), Itner, Lambert and Larcker (2003), and Anderson, Banker and Ravindran (2000) all find that the high-growth new economy sector relies more and more on EBC to provide managerial incentives, while the phenomenon is not as pronounced in the low-growth old economy sector. Consistent with this, Johnson, Ryan and Tian (2007) report that firms prosecuted for fraudulently misreporting are concentrated within industries having significantly higher than average growth potential. Additionally, both the (detected) fraud firms and their (industry and size matched) control firms have significantly negative market-adjusted returns during the fraudulent misreporting periods, suggesting the industries in which the (detected) fraud firms compete did not perform well relative to the overall market during the fraud period. Erickson, Hanlon and Maydew (2004) find that firms accused of fraud have greater external financing needs. Although these empirical papers address different dimensions of the environment in which fraud is committed, there is a common problem associated with them: they do not disentangle the probability of fraud commission and that of fraud detection from the probability of detected fraud. Therefore, these studies implicitly assume that the probability of detected fraud equals that of fraud commission (in other words, the conditional probability of fraud detection is implicitly assumed to be one), which is certainly not the case. The importance of this has been established empirically in Li (2007). She finds that failing to model the interdependence between fraud and monitoring by the SEC leads to significant downward biases in estimating the effects of factors such as EBC and corporate governance on the probability of fraud commission.

50

The only empirical work we are aware of that accounts for the two unobservable probabilities are Wang (2005b) and Li (2007). Wang (2005b) employs the Poirier (1980) bivariate probit with partial observability methodology to back out the two unobservable probabilities. She focuses on whether the firms investment policy can affect the probability of committing and/or detecting fraud. Her main findings are that different types of investment can have different effects on these two probabilities, and that a higher probability of fraud commission leads to overinvestment. In contrast, our interest is on whether the SEC is a strategic regulator in terms of fraud detection, thus our study focuses on how EBC and growth potential affect these two unobservable probabilities, especially the (conditional) probability of fraud detection. Li (2007) uses the (more complex) Feinstein (1990) model specification to back out the two unobservable probabilities. In addition to establishing the importance of using the bivariate probit with partial observability model, her main finding is that the marginal benefits of increased SEC enforcement appear to far outweigh the marginal cots of increased enforcement. As noted in the Introduction, we employ the bivariate probit with partial observability methodology proposed by Poirier (1980) to empirically disentangle the probability of fraud commission and the (conditional) probability of fraud detection from the probability of detected fraud. We believe that the Poirier (1980) model specification is more suitable for our purposes.32 Moreover, to our knowledge, this paper is the first to examine the effectiveness of SOX in reducing the (unobservable) probability of fraud commission. Neither Wang (2005a) nor Li (2007) studies the post-SOX period. In addition, neither study controls for the endogeneity of EBC and neither
In Li (2007), the (perceived) probability of detecting fraud enters the fraud equation and the (perceived) probability of committing fraud enters the detection equation in the bivariate probit model specification. However, according to the Qiu and Slezak (2008) model, although the firm manager takes the equilibrium probability of detecting fraud into account when making his fraud decision and the regulatory agency takes the equilibrium probability of committing fraud into account when making its detection decision, both probabilities are simultaneously determined by exogenous factors. We hence believe that the Poirier (1980) specification is more suitable for our purpose. The Poirier (1980) specification also is simpler, employing fewer restrictions to achieve a full and unique identification of the model parameters.
32

51

examine how the usage of EBC is affected by exogenous parameter changes (especially, growth potential and fraud penalty) across the new economy and old economy industries, while these issues are all central to our study.

3. Empirical Hypotheses
As discussed in the last two sections, extant theoretical models predict that, ceteris paribus, higher growth potential will provide more opportunities for firm managers to hide past fraud through overinvestment, thus lead to higher fraud-induced deadweight loss; higher usage of EBC will provide stronger incentive for managers to commit fraud to inflate stock prices. If the fraud detection environment is exogenous or mechanical, in other words, if the SEC does not strategically respond to fraud commission, ceteris paribus higher growth potential will reduce the (conditional) probability of fraud detection, and higher EBC will have no effect on the probability of fraud detection. However, if the SEC does indeed respond strategically to fraud commission (i.e., the SEC is a strategic regulator in combating fraud), ceteris paribus higher growth potential and higher EBC will both capture the attention of the SEC, thus both will increase the (conditional) probability of fraud detection. Therefore, we develop the first two hypotheses as follows: H1: Ceteris paribus, a higher growth potential increases the conditional probability of fraud detection. H2: Ceteris paribus, a higher level of EBC increases the conditional probability of fraud detection. As stated earlier, another purpose of the paper is to study the effect of SOX on fraud commission. We thus have the following hypothesis: H3: SOX is effective in reducing the probability of fraud commission. Under the agency framework, EBC is an endogenous choice variable of shareholders. As

52

stated earlier, in the agency models without a strategic regulator, due to the exogenous probability of fraud detection, a signal-jamming equilibrium will occur, which implies that ceteris paribus a change in any factor that makes the manager less costly to commit fraud, such as an increase in growth opportunities (which decreases the probability of fraud being detected), or a reduction in fraud penalty, will result in less usage of EBC by shareholders to compensate the manager. On the contrary, in the agency model with a strategic regulator (the QS model), the high-growth fraud-prone industries fall into the pooling or mixed strategy equilibrium, while the low-growth fraud-free industries fall into the truthful separating equilibrium. The model predicts that in the high-growth fraud-prone industries, ceteris paribus the usage of EBC is positively associated with growth potential during good times (i.e., in the pooling equilibrium),33 while in the low-growth fraud-free industries the usage of EBC is negatively associated with growth potential.34 The model also predicts that ceteris paribus an increase in the fraud penalty should lead to lower EBC levels in the high-growth fraud-prone industries, while changes in the fraud penalty have no effect on the EBC usage in the low-growth fraud-free industries.35
33 The intuition is that, ceteris paribus, given more growth opportunities and, thus, a lower likelihood of future fraud exposure, in the pooling equilibrium the manager will optimally reduce his costly effort and relies more on misreporting (if the firm realizes poor performance). This reduction in managerial effort is directly costly to the principal (shareholders or board of directors) and it increases the expected overinvestment losses to shareholders. Therefore, the shareholders will optimally offer the manager a higher level of EBC to induce more managerial effort. In the QS model, ceteris paribus an increase in growth potential has an ambiguous effect on EBC usage in the mixed strategy equilibrium. 34 The intuition is that when growth potential increases, the firms growth value (thus the firm value) also increases. However, the managers chosen effort does not change with growth potential in the truthful separating equilibrium. The shareholders would pay the manager too much for his effort if the original compensation level was not decreased accordingly. Thus, as the growth potential increases, the equilibrium EBC under the separating equilibrium decreases. 35 The intuition is that ceteris paribus, a higher fraud penalty induces the manager to work harder in the pooling equilibrium in order to reduce the chance of realizing poor performance, as it is the occurrence of poor performance that causes the manager to resort to fraudulent misreporting. Thus, the shareholders will respond by reducing the costly EBC level as the increased fraud penalty can partially substitute for costly EBC in inducing managerial effort. In the mixed strategy equilibrium, ceteris paribus an increase in the fraud penalty will reduce the market value difference between the (reported) high earnings firms and the (reported) low earnings firms thus reduce the marginal incentive of EBC, which also causes a reduction in the usage of EBC. Since there is no fraud in the truthful

53

Therefore, we will also test the following two hypotheses regarding the equilibrium EBC in the high-growth fraud-prone and low-growth fraud-free industries. H4: A higher growth potential increases the EBC usage for firms in the high-growth fraud-prone industries during good times, but decreases the EBC usage for firms in the low-growth fraud-free industries. H5: An increase in the penalties of fraud due to the enactment of SOX reduces the EBC usage for firms in the high-growth fraud-prone industries, but has no effect on the EBC usage for firms in the low-growth fraud-free industries.

4. Methodology
To test H1 to H3, we need to empirically disentangle the probability of fraud commission and that of fraud detection from the probability of detected fraud. We use the Poirier (1980) bivariate probit with partial observability technique to achieve this goal. Let Fi be a binary variable which equals 1 if firm i commits fraud and 0 otherwise; let Di be a binary variable which equals 1 if the regulatory agency detects firm i and 0 otherwise; let Gi be a binary variable which equals 1 if firm i gets caught of committing fraud ex post and 0 otherwise. We should have Pr ob(Gi = 1) = Pr ob( Fi = 1, Di = 1) = Pr ob( Fi = 1) Pr ob( Di = 1 | Fi = 1) If the incentive for firm i to commit fraud and that for the regulatory agency to detect firm i can be denoted by the latent variables Fi * and Di* respectively, with Fi * = X F ,i F + F ,i , F ,i ~ N (0,1) ( Di* | Fi = 1) = X D ,i D + D ,i , D ,i ~ N (0,1)
separating equilibrium, changes in the fraud penalty have no effect on the usage of EBC in the low-growth fraud-free industries.

54

Fi = 1 iff Fi * > 0 ( Di = 1 | Fi = 1) iff ( Di* > 0 | Fi = 1) we can then have the following expressions for the probability of fraud commission and the conditional probability of fraud detection
Pr ob( Fi = 1) PF = ( X F ,i F )

Pr ob( Di = 1 | Fi = 1) PD = ( X D ,i D ) In the above expressions, X F ,i and X D ,i are row vectors of explanatory variables for the probability of fraud commission and the conditional probability of fraud detection respectively, and F and D are the respective coefficient column vectors. Therefore, we have Pr ob(Gi = 1) = ( X F ,i F ) ( X D ,i D )
Pr ob(Gi = 0) = 1 ( X F ,i F ) ( X D ,i D )

The likelihood and log-likelihood functions of observing a (cross-sectional) sample of firms with some being caught committing fraud ex post and some not are then L = [ ( X F ,i F ) ( X D ,i D )] [1 ( X F ,i F ) ( X D ,i D )]
Gi =1 Gi = 0

log L =

Gi =1

log(( X

F ,i

F )( X D ,i D )) +

Gi = 0

log(1 ( X

F ,i

F ) ( X D ,i D ))

(1)

Estimates of F and D can then be obtained through the maximum-likelihood estimation (MLE). T-tests can then be performed to test H1 to H3 based on the parameter estimates and their respective standard errors. One might notice that there exists a potential endogeneity problem in the above bivariate probit model. EBC i (i.e., the equity-based executive compensation level of firm i ) is one of the explanatory variables in both the PF and PD equations (i.e., EBC enters both X F ,i and X D ,i ),

55

and we implicitly treat all variables in X F ,i and X D ,i as exogenously given to obtain parameter estimates for F and D . However, in the real world EBC i is an endogenous choice variable, and it is also endogenous in the aforementioned agency models. Ignoring this endogeneity problem, the coefficient estimates of the above bivariate probit model may be inconsistent and biased. We use the two-stage probit least squares (2SPLS) technique introduced in Maddala (1983) to account for this endogeneity problem. Basically, we run the first-stage reduced form regression of EBC i on a group of instrumental variables (we use lag explanatory variables as instruments), and use the predicted value of EBC i to replace the actual value in the second stage maximum-likelihood estimation. Alvarez and Glasgow (2000) use Monte-Carlo simulation and show that this technique is capable of dealing with endogeneity problems associated with probit models. The disadvantage of the 2SPLS technique, however, is that the standard errors (thus the T-statistics) from maximum likelihood estimation will be biased, and there is no easy way to correct them. In order to avoid this bias, we use bootstrapped standard errors instead i.e., we bootstrap 500 samples from our original sample, get the coefficient estimates for each bootstrapped sample through MLE, calculate the standard deviation for each coefficient based on the 500 estimates and use the calculated standard deviation as the standard error for the coefficient. As can be seen later, based on the industry distribution of our fraud sample, the high-tech industry (Fama-French ten industry classification code 5) is relatively fraud-prone, while the manufacturing industry (Fama-French ten industry classification code 3) is relatively fraud free. To test H4 and H5, we run panel regressions on a panel of non-fraud firms (i.e., firms which have not been accused of fraud by the SEC during in the sample period) from the high-tech industry and on a panel of non-fraud firms from the manufacturing industry respectively, with EBC i being the dependent variable.

56

5. Data and Variables 5.1. Data 5.1.1. Data Source

We manually collected our fraud sample from SEC litigation releases. We reviewed the SEC litigation releases dated between 01/01/1997 and 06/30/2007, and were able to identify 596 unique information manipulation cases of US public companies.36
37

For the 596 cases in our initial sample, we imposed the following conditions to select firms for our final sample: 1. The fraud is committed by top corporate executives (not mid-level managers or heads of subsidiaries); 2. The fraud is intentional (not accounting errors); 3. The firm is covered by Compustat; 4. Initial fiscal year of fraud is from 1996 through 2003;38 5. If a firm is associated with more than one fraud case, only the first case enters our sample. Among the 596 cases, 233 unique firms (cases) satisfy the above five conditions simultaneously. Therefore, our final fraud sample includes the 233 firms. Among them, 79 are covered by Execucomp.

5.1.2. Sample Description Cross-sectional (Industry) Distribution of the Fraud Sample

Table 1 reports the industry distribution of the fraud sample. We use Fama-French ten industry classification.
36

39

Firms prosecuted for fraud demonstrate a significant industry

The nature of these information manipulation cases includes earning inflation, material misrepresentation in SEC filings and false press release. As shorthand, we refer to these information manipulations as fraudulent misreporting, misreporting, or simply fraud throughout the paper. 37 That is, we do not include information manipulation cases related to mutual funds, hedge funds, pension funds, investment trusts, investment companies, investment advisors, broker-dealers, REITs, ADRs, limited liability partnerships and private companies etc. 38 We do not include fraud firms with the first fiscal year of fraud after 2003. The reason is that including fraud firms with the initial year of fraud after 2003 may produce bias, as some fraud that actually occurred after 2003 may have not yet been detected by the SEC by June 2007 (based on our sample, the average time period between fraud initiation and SEC enforcement is about 4 years). These missing, but to-be-detected, incidences of fraud are the problem. 39 We thank Prof. Kenneth French for providing this information on his website.

57

concentration effect, primarily in the high-tech industry (Fama-French ten industry classification code 5); 81 of the 233 firms (or 34.76% of the fraud sample) fall into this category. Among all industries, the hightech industry also has the highest percentage of detected fraud firms within the industry 2.04% of Compustat firms in the industry are caught of committing fraud; as a comparison, only 0.53% of Compustat firms in the manufacturing industry (Fama-French ten industry classification code 3) are caught of committing fraud. Pearsons, likelihood ratio and Mantel-Haenszel Chi-square statistics all strongly reject the null hypothesis that there is no industry concentration effect in the fraud sample.40

Time-Series Distribution of the Fraud Sample

Table 2 reports the time-series distribution of the fraud sample. It shows that most firms in the fraud sample committed fraud initially in year 1999, 2000, or 2001. These years overlap with the years when the technology sector slowed down and fell into downturn. There is a sharp decline in the number of firms in the fraud sample initiating fraud in year 2002 or 2003, which are the years when SOX is in effect. For most firms in the fraud sample, the first year of SEC investigation occurred in 2000, 2001 or 2002. These years also coincide with the time period when the technology sector fell into downturn. This finding is consistent with the theory implications of Povel, Singh and Winton (2007) and Qiu and Slezak (2008). From Table 2, we can see that, on average, a fraud affects 2.5 fiscal years accounting data. The time period between fraud initiation and investigation initiation is 2.2 years (2.5-1+0.7) and the time period between fraud initiation and SEC enforcement is 4.4 years. The investigation lasts for 2.2 years on average.

5.2. Variables

This result is consistent with the empirical finding of Johnson, Ryan and Tian (2007) and the theoretical prediction of Qiu and Slezak (2008).

40

58

Variables in the PF Equation

The row vector of variables X F in the PF equation of the bivariate probit with partial observability model includes EBC, growth potential, the SOX dummy, and control variables. For EBC, following Li (2007) we use the pay performance sensitivity (PPS) averaged among the top five executives of the firm as a proxy. PPS is calculated as PPS due to unrestricted stocks plus PPS due to vested stock options. PPS due to unrestricted stocks is calculated as the number of unrestricted shares owned by an executive divided by total shares outstanding; while PPS due to vested stock options is calculated as the number of shares in vested options granted to an executive multiplied by the Black-Scholes hedge ratio then divided by total shares outstanding (Yermack (1995)). We follow the Core and Guay (2002) procedure to calculate the Black-Scholes hedge ratio. This measurement of PPS based on unrestricted stocks and vested options is most consistent with the Qiu and Slezak (2008) model, in which the manager has a short horizon and therefore must be able to exercise options or sell shares freely. Bergstresser and Philippon (2006) also provide empirical evidence that CEOs exercise unusually large amount of options during periods of high accruals (a proxy for earnings manipulation), which suggests that firm management incentive is short-horizoned. Data used to calculate PPS are obtained from Execucomp and CRSP. For firms in our fraud sample which are not covered by Execucomp, we manually collect the related information from proxy statements. As a result, we are able to calculate PPS for all of the 233 firms in our fraud sample. We use two proxies to measure a firms growth potential. First of all, following Smith and Watts (1992), Yermack (1995) and others, we use Tobins Q to proxy for growth potential. Tobins Q is formally defined as the ratio of a firms market value of assets to its replacement value of assets. Since it is difficult to get market value and replacement value of a firms assets, following

59

the existing literature we proxy the market value of assets by:

[Book Value of Assets Book Value of Common Equity Deferred Taxes + Mkt Value of Common Equity].

We use the book value of assets to proxy for the replacement value of assets. Data used to calculate Q are obtained from Compustat. The second proxy, SG, is the average percentage sales growth rate over the last two years, which serves as a direct measure of a firms growth potential. Since SOX is in effect in year 2002, in order to measure the impact of SOX on fraud commission, we use a SOX_COMMIT dummy, which equals one if the first fiscal year of fraud is 2002 or later and equals zero otherwise. We also include firm size (SIZE), age (AGE), return on assets (ROA), total debt ratio (DEBT), Altmans Z score (Z), and percentage of institutional shareholdings (INST) as controls in the row vector of variables X F . SIZE is the calendar-year end log market value of equity (in $MM). AGE is the age of the firm since its IPO. ROA is calculated as operating income after depreciation divided by book value of assets. DEBT is calculated as book value of debt divided by book value of assets. A higher total debt ratio may lead to a higher probability of fraud commission, as firm management may try to satisfy accounting-based debt covenants through earning manipulation (Dechow et al. (1996)). Following Li (2007) and Graham et al. (1998), we use a modified version of Altmans Z score as follows: Z = 1.2 working capital retained earnings EBIT sales + 1.4 + 3.3 + 1.0 total assets total assets total assets total assets

A lower Z-score reflects a higher degree of financial distress, which may lead to a higher probability of fraud commission. INST is calculated as the ratio of total institutional shareholdings to total shares outstanding, which can be viewed as a proxy for the strength of corporate governance (Shleifer and Vishny (1986), Bertrand and Mullainathan (2001) and Hartzell and Starks (2003)). Stronger corporate governance may be associated with a lower probability of fraud

60

commission. The data used to calculate these variables are obtained from CRSP, Compustat, and CDA Spectrum of Thomson Financials. Later, as a robustness check, we also incorporate the size of the firms board of directors (BOARD) and the percentage of directors on the board who are independent directors (INDEP_DIR) as additional corporate governance controls in both the PF and PD equations. We manually collected these board data from proxy statements for all of the 233 fraud firms. The board data of the non-fraud control sample are collected from the IRRC database whenever they are available. To control for possible industry effect, we include the industry dummies constructed according to Fama-French ten-industry classification. For each firm in our fraud sample, all the above variables except SOX_COMMIT and industry dummies are measured as of the year prior to the initial year of fraud to assure that our measurement is not affected by misreporting.41 Our non-fraud control sample consists of all firms which are covered by Compustat between 1995 and 2003 and which are not the subject of SEC litigation between 01/01/1997 and 06/30/2007. 42 Similar to Li (2007), for each firm in our non-fraud sample, we randomly select a year between 1996 and 2003 and assign it as the firms initial fiscal year of fraud, and use the firms data in the fiscal year prior to the (assigned) initial year of fraud to match the cross-sectional data of the fraud firms. In this way we construct a full cross-sectional sample of fraud and non-fraud firms for the bivariate probit estimation.

Variables in the PD Equation

The row vector of explanatory variables X D in the PD equation of the bivariate probit model includes the aforementioned variables measuring EBC and growth potential, which are our

Since the IRRC does not have data coverage for the years before 1996, we use the 1996 data to calculate BOARD and INDEP_DIR instead if a firms initial year of fraud is 1996. 42 Whenever the variable PPS is used in our bivariate probit model, the non-fraud control sample is constrained to firms that are covered by Execucomp. When the variables BOARD and INDEP_DIR are included (in the robustness test), the non-fraud control sample is further reduced to the intersection of Execucomp and IRRC.

41

61

main interest. Similar to the SOX_COMMIT dummy in the PF equation, we also include a SOX_DETECT dummy in the PD equation, which equals one if the year of initial fraud investigation is 2002 or later and equals zero otherwise. We also include the percentage of institutional shareholdings (INST), and later as a robustness check, size of the board (BOARD) and percentage of independent directors on the board (INDEP_DIR) in the PD equation, as the strength of corporate governance may affect the (conditional) probability of fraud detection. In addition, we include four more explanatory variables in the PD equation as controls. The first control variable is the total market value of fraud firms prosecuted in the same industry one year prior to the investigation of a fraud case (IndusPriorDec). IndusPriorDec is calculated as the total market value of all prosecuted fraud firms in the firms industry one year prior to the investigation year, scaled by the industrys total market value in that year (the year prior to the investigation year). Data used to calculate this variable are from Compustat. A higher intensity of prosecuted fraud in an industry may catch the attention of the SEC (or indicates that the industry has already caught the SECs attention) and increase the probability of fraud detection. The second control variable, AUDIT, is the auditors opinion toward the firm one year prior to the year of SEC investigation. Adverse opinion from the auditor may catch the SECs attention and increase the probability of fraud detection. AUDIT is obtained from Compustat.43 We also include another control variable, INVESTMENT, which is the sum of capital expenditures (Compustat data 128), acquisitions (data 129) and increase in investments (data113) divided by book value of assets measured as of the year prior to the year of SEC investigation, as the manager may try to affect the probability of fraud detection through making new investment (Wang (2005a) and Qiu

43

Audit code 0 means Unaudited, 1 means Unqualified, 2 means Qualified, 3 means No Opinion, 4 means Unqualified with additional language and 5 means Adverse Opinion. Generally speaking, the bigger the auditor code, the worse is the auditors opinion toward the firm.

62

and Slezak (2008)). The fourth control variable, STDDEV, is the standard deviation of the firm's daily stock returns in the year prior to the year of SEC investigation, and is obtained from CRSP. Higher stock return volatility may arouse the SECs attention thus increase the probability of fraud detection. For each non-fraud firm, these four control variables are measured as of the year after the (randomly selected and assigned) initial year of fraud, since the average time between fraud initiation and investigation initiation is about 2 years in our fraud sample. Similar to the PF equation, we also include industry dummies to control for possible industry effect in the PD equation.44

Variables in the EBC Panel Regressions

To test H4 and H5, we perform panel regressions of PPS on growth potential (Q or SG) and a SOX dummy, which equals one if the fiscal year of financial data is 2002 or later and equals zero otherwise, for the (relatively) fraud-prone high-tech industry (Fama-French industry code 5) and the (relatively) fraud-free manufacturing industry (Fama-French industry code 3). Since Qiu and Slezak (2008) predict that in the high-growth fraud-prone industries, EBC is positively associated with growth potential during good times, we hence add an interaction variable BOOM*Q or BOOM*SG as an explanatory variable in the panel regressions. BOOM*Q is the product of a dummy variable BOOM and Q, with BOOM being zero if the fiscal year of financial data is 2001 or 2002 and one otherwise. Similarly, BOOM*SG is the product of BOOM and SG. We also add SIZE, AGE, DEBT, Z, ROA, STDDEV, INST, BOARD and INDEP_DIR as controls in the panel regressions. All firms used in the panel regressions must be free from fraud litigation during the
44

To remove coding errors and reduce the impact of outliers on our empirical results, we winsorize our full cross-sectional sample (of fraud firms and non-fraud firms) by replacing extreme values (below 2 percentile or above 98 percentile) of variables PPS, Q, SG, SIZE, AGE, ROA, DEBT, Z, INVESTMENT and STDDEV with their respective 2 percentile or 98 percentile values of the full sample.

63

sample period from 1994 to 2003. PPS is measured for each firm from 1995 to 2004 (given that data are available on Execucomp). The explanatory variables are matched with PPS with a lag in all panel regressions because any change of these variables will likely affect EBC with a lag.45

6. Empirical Results 6.1. Univariate Analysis

Univariate comparisons of variables for the prosecuted fraud firms and non-fraud firms are presented in Table 3. It is clear that firms in our fraud sample utilize significantly higher levels of equity based compensation than those in the non-fraud control sample. P-values from the Welsh two sample T-test and Mann-Whitney test suggest that the mean and median of PPS are both significantly higher for the fraud sample at 1% level. The average PPS of the fraud sample is 3.47% (the median is 1.82%) while that of the non-fraud sample is only 1.09% (the median is 0.39%). This finding conforms with the empirical results of Johnson et al. (2007) and Li (2007). We further find that firms in the fraud sample have significantly higher growth potential than those in the non-fraud control sample. The mean Tobins Q of the fraud sample is significantly higher than that of the non-fraud sample at 5% level, and the median difference in Q between the fraud sample and the non-fraud control sample is significant at 1% level. The mean and median of SG of the fraud sample are both significantly higher than their counterparts of the non-fraud control sample at 1% level. This again is consistent with the empirical result of Johnson, Ryan and Tian (2007) and supports the theoretical predictions of Qiu and Slezak (2008). Compared with firms in the non-fraud control sample, firms in our fraud sample are also

To remove coding errors and reduce the impact of outliers on our empirical results, we similarly winsorize the high-tech panel dataset and manufacturing panel dataset by replacing extreme values (below 2 percentile or above 98 percentile) of variables PPS, Q, SG, SIZE, AGE, ROA, DEBT, Z and STDDEV with their respective two-digit SIC 2 percentile or 98 percentile values of the two datasets.

45

64

significantly larger in SIZE,46 significantly younger in AGE, employ significantly lower levels of debt in their capital structures. They also have significantly higher levels of institutional shareholdings, significantly smaller board size, and significantly lower percentage of independent directors. They are significantly less likely to initiate fraud post-SOX (2002 or 2003).47 In the year prior to the year of SEC investigation, these firms make significantly more investment, and have significantly higher stock return volatility and worse auditors opinions than firms in the non-fraud control sample.

6.2. The Effects of Growth Potential and SOX on Fraud Commission and Detection
As mentioned earlier, PPS (the measure of EBC) is an endogenous choice variable. Also, as pointed out in Footnote 15, whenever PPS is used in our bivariate probit model, the non-fraud control sample is reduced from the Compustat universe to firms that are covered by Execucomp (which is a reduction of 86.6% of firms in our non-fraud control sample). In order to accurately study the effects of growth potential and SOX on the two unobservable probabilities, we first leave out PPS and fit the bivariate probit with partial observability models.48 Model 1 of Table 4 shows that Q has no effect on the probability of fraud commission, but is strongly positively associated with the (conditional) probability of fraud detection at 1% level. Similarly, Model 2 shows that SG is significantly positively associated with the probability of fraud detection at 1% level, while its effect on the probability of fraud commission is insignificant. Since Q and SG are both proxies for growth potential, the results support Hypothesis 1 that ceteris paribus, a higher growth potential captures the SECs attention and increases the conditional
This is different from the finding of Li (2007). She finds that firms in her fraud sample are significantly smaller in size. We believe that the difference is driven by the fact that our sample period (from 1996 to 2003) is significantly different from the sample period (from 1992 to 1999) in Li (2007). 47 Recall that in order to match the fraud sample and form the full cross-sectional sample, for each firm in the non-fraud control sample, we randomly select a year between 1996 and 2003 and assume it as the initial year of fraud. 48 Before fitting the bivariate probit models, we also replace missing Q, SG, SIZE, AGE, ROA, DEBT, Z, AUDIT, INVESTMENT, or STDDEV values with their respective two-digit SIC medians of the full sample.
46

65

probability of fraud detection. Thus, the results appear to support the notion that the SEC is a strategic regulator and is inconsistent with theoretical predictions based on an exogenous or mechanical fraud detection environment. In both Model 1 and Model 2, SOX_COMMIT is significantly negatively associated with the probability of fraud commission at 1% level. This is consistent with Hypothesis 3 that SOX is effective in reducing fraud commission. To further quantify the effect of SOX on fraud reduction, we let SOX_COMMIT equal zero and evaluate the other explanatory variables in the PF equation at their respective means of the full sample to calculate the probability of fraud commission in the pre-SOX period based on both Model 1 and Model 2. We also let SOX_COMMIT be one and evaluate the other variables at their means to calculate the probability of fraud commission in the post-SOX period based on the two models. There results are presented in Panel A of Table 5. The results based on Model 1 and Model 2 are very consistent with each other. Both show that in the pre-SOX sample period, the probability of fraud commission is above 3% (3.26% based on Model 1 and 3.24% based on Model 2), while in the post-SOX sample period the probability is reduced to a little over 1% (1.09% based on Model 1 and 1.10% based on Model 2). Both models imply that SOX reduces fraud commission by two thirds (66.65% by Model 1 and 65.98% by Model 2). As a robustness check, we also evaluate the other explanatory variables at their respective medians of the full sample (instead of their means) to calculate the effectiveness of SOX on reducing fraud commission. Panel B of Table 5 presents the results, which are consistent with the results in Panel A and again show that SOX reduces fraud commission by two thirds (65.91% by Model 1 and 68.16% by Model 2). As discussed in the Introduction, there are heated debates over the costs and benefits of

66

SOX. Although several studies document various costs associated with SOX, to our knowledge, few study (if any) documents the benefits of SOX. We thus believe that our empirical results could help contribute to the debates and provide useful insight to policy makers. The results in table 4 also show that AUDIT, INVESTMENT and STDDEV are all significantly positively associated with the (conditional) probability of fraud detection, which suggests that the regulator pays attention to worse auditor opinion, larger investment, or higher stock return volatility when combating fraud.

6.3. The Effects of EBC on Fraud Commission and Detection

Given that EBC (thus PPS) is an endogenous variable, we need to control for its endogenous nature when incorporating EBC in the bivariate probit model. We use the aforementioned 2SPLS technique in the first stage, PPS is regressed on some instrumental variables to get a predicted value for PPS (i.e., P.PPS); in the second stage, the bivariate probit (with partial observability) models are estimated using maximum likelihood estimation with P.PPS in the PF equation. To correct for the bias in the standard errors introduced by 2SPLS, we use bootstrapped standard errors to calculate the P-values of coefficient estimates. The results are reported in Table 6. In Model 3 of Table 6, P.PPS is the predicted value of PPS resulted from OLS regression of PPS on Lag_Q, Lag_SIZE, Lag_ROA, Lag_DEBT, Lag_Z, Lag_AGE, Lag_STDDEV, Lag_INST and the industry dummies, with the lag variables being measured two years before the initial year of fraud. The instrumental variables in Model 4 are the same as those in Model 3, except that Lag_SG is used instead of Lag_Q. We find that P.PPS is strongly positively associated with the two unobservable probabilities at 1% level in both Model 3 and Model 4. Based on Model 3, a one-percentage-point increase in the average pay- for-performance sensitivity of the top five executives, which translates

67

into an average increase in the value of personal holding of unrestricted stocks and vested stock options by $9.07 millions for each of the top five executives (as the average market cap of Execucomp firms in the full sample is $907 millions), will increase the probability of fraud commission by 10.24 percentage points but will also substantially increase the (conditional) probability of fraud detection by 45.63 percentage points. Similarly, based on Model 4, a one-percentage-point increase in the average pay-for-performance sensitivity of the top five executives will increase the probability of fraud commission by 7.49 percentage points but will also substantially increase the probability of fraud detection by 40.86 percentage points. Therefore, the prediction of Hypothesis 2, that ceteris paribus a higher level of EBC increases the conditional probability of fraud detection, is strongly supported by the data. The results again support the notion that the regulator is strategic in responding to the increased incentive to commit fraud. In Model 3, Q is significantly positively associated with the (conditional) probability of fraud detection and significantly negatively associated with the probability of fraud commission, both at 1% level. SG is significantly positively associated with the probability of fraud detection at 5% level, while its effect on the probability of fraud commission is insignificant. These results again support the prediction of Hypothesis 1 that ceteris paribus a higher growth potential increases the conditional probability of fraud detection. In both Model 3 and Model 4, SOX_COMMIT is again significantly negatively associated with the probability of fraud commission, supporting Hypothesis 3. However, compared with that based on Model 1 and Model 2, the magnitude of reduction in fraud commission post SOX is less conspicuous based on Model 3 and Model 4. For example, fraud commission post SOX is reduced by 57.72% based on Model3 and 60.29% based on Model 4 (evaluated at the means) both numbers are lower than the corresponding numbers based on Model 1 and Model 2. Due to the fact

68

that whenever PPS is included the non-fraud control sample is restricted to firms covered by Execucomp (while there is no such restriction in the fraud sample as we manually collected executive compensation data from proxy statements for the prosecuted fraud firms), we believe Model 1 and Model 2 provide more accurate estimates of the effect of SOX in reducing fraud commission. Model 3 and Model 4 also show that the profitability measure ROA is significantly negatively related to the tendency to commit fraud, and AUDIT, INVESTMENT and STDDEV is again significantly positively related to the (conditional) probability of fraud detection.

6.4. Robustness Checks

We performed several robustness checks on the empirical results. As the first robustness check, we incorporate the size of the board of directors (i.e., variable BOARD) and the percentage of independent directors on the board (i.e., variable INDEP_DIR) as additional corporate governance controls. Both variables are measured as of the year prior to the initial year of fraud to be included in the PF and PD equations.49 The results are reported in Table 7. One caveat needs to be fully aware of is that, as pointed out in Footnote 15, whenever the variables BOARD and INDEP_DIR (as well as PPS) are included, the non-fraud control sample is greatly reduced to the intersection of Execucomp and IRRC, which (based on our data) is an exclusion of 92.3% of firms in the non-fraud control sample used in Model 1 and Model 2 of Table 4. Nevertheless, Table 7 shows that our main findings (that higher growth potential and/or higher EBC increase the probability of fraud detection and SOX is effective in reducing fraud commission) remain unchanged. Larger board size indicates weak corporate governance, as it is more difficult for directors
49

Lag_BOARD and Lag_INDEP_DIR, which are the lag variables measured two years before the initial year of fraud, are also included as instrumental variables in the 2SPLS estimation.

69

of large board to organize in opposition to the CEO (Core, Holthausen and Larcker (1999)). Thus it is not surprising that BOARD is significantly positively associated with the probability of fraud commission. However, we also find that INDEP_DIR is significantly positively associated with the probability of fraud commission. This seems to support the argument of Bebchuk and Fried (2004) that independent directors have various incentives to support or at least go along with the decisions of the companys top executives.50 We view this as additional evidence of governance failure in corporate America. We also find that BOARD is significantly positively related to the (conditional) probability of fraud detection, which appears to suggest that the SEC pays attention to signs of weak corporate governance when combating fraud. Since the bivariate probit technique uses the probit link, one valid concern is that our results may be specific to the probit link function. As the second robustness check, we use the logit link and estimate the bivariate logit with partial observability models with the same data. The results remain qualitatively unchanged and are reported in Table 8. In the earlier analyses we use Tobins Q and past sales growth (SG) as proxies for the firms growth potential. As the third robustness check, we use the mean analyst expected five-year sales growth rate from I/B/E/S to replace Q or SG in the bivariate probit models. Untabulated results show that the main findings remain robust to this change. As mentioned earlier, for each firm in our non-fraud control sample, we randomly select a year between 1996 and 2003 to be the initial year of fraud in order to match the fraud sample and form the full cross-sectional sample. One concern is that this randomization process may contribute to some of our results. As the fourth robustness check, we use the same randomization

For example, the CEO has significant (often crucial) influence over director nominations and director compensations; many independent directors have prior social connections with the CEO; the CEO also serves as independent director at the companies where the independent directors are executives; the firm has business dealings with the independent directors firms, etc.

50

70

procedure to form different non-fraud control samples in order to match the fraud sample and form different full cross-sectional samples for the bivariate probit estimation. Untabulated results (based on different full cross-sectional samples) show that the main findings remain unchanged.

6.5. Panel Regressions of EBC on Growth Potential and SOX

Descriptive statistics for the panel datasets of the (relatively) fraud-prone high-tech industry and the (relatively) fraud-free manufacturing industry are reported in Table 9, and the firm fixed-effect PPS panel regression results for both samples are reported in Table 10.51 Table 9 shows that the mean and median of the dependent variable PPS in the high-tech sample are both higher than their counterparts in the manufacturing sample. However, the variance of PPS is roughly the same across the two samples (the manufacturing sample even has a slightly higher variance in PPS). Table 9 also reports that firms in the high-tech sample generally have higher growth potential (in terms of Tobins Q and SG), are younger in age, have less debt but are less profitable (in terms of ROA), and have smaller board size and higher stock return volatility than firms in the manufacturing sample. Table 10 shows that for the high-tech sample, the regression coefficients of BOOM*Q and BOOM*SG are both positive, and the coefficient of BOOM*SG is also significant at 5% level; while for the manufacturing sample, the regression coefficients of BOOM*Q and BOOM*SG are both negative and insignificant. This evidence seems to at least partially support the prediction of Hypothesis 4 that a higher growth potential tends to increase the EBC usage for firms in the high-growth fraud-prone industries during good times, but tends to decrease the EBC usage for firms in the low-growth fraud-free industries. Table 10 also shows that in both regression models for the high-tech sample, the regression coefficient of the SOX dummy is negative and significant
51

Before running panel regressions, we also replace missing Q, SG, SIZE, AGE, ROA, DEBT, Z, BOARD, INDEP_DIR and STDDEV values with their respective two-digit SIC medians of the high-tech sample or the manufacturing sample.

71

at 5% level. However, the coefficient of SOX is insignificant in both models for the manufacturing sample. This result is consistent with the prediction of Hypothesis 5 that an increase in the penalties of fraud due to the enactment of SOX reduces the EBC usage for firms in the high-growth fraud-prone industries, but has no effect on the EBC usage for firms in the low-growth fraud-free industries. As a robustness check, we also performed random-effect panel regressions. Untabulated random-effect results are virtually the same as the reported fixed-effect results.52 Therefore, our empirical evidence appears to contradict to the theoretical implications of the agency models without a strategic regulator, which suggest a signal-jamming equilibrium, but is consistent with the agency model with a strategic regulator, which suggests different equilibrium outcomes for the high-growth and low-growth industries. The result can also be viewed as additional evidence that the regulator is a strategic player in the fraud commission and detection game, and the firms shareholders (or board of directors) rationally take the dynamic interplay between the manager and the regulator into consideration when designing an equity-based compensation contract for the manager.

7. Conclusions
In the wake of recent financial scandals, there are heated debates on whether the SEC is effective in combating fraud. There are also continuing debates on the costs and benefits of the Sarbanes-Oxley Act. In this paper, we investigate two research questions: 1. Does the SEC strategically respond to fraud commission? 2. How effective is SOX in reducing fraud commission? Using a sample of firms subject to SEC litigations within the last ten years, and employing the bivariate-probit with partial observability technique to disentangle the unobservable
52

The Hausman test rejects the random effect assumption though.

72

probability of fraud commission and the equally unobservable (conditional) probability of fraud detection from the observable probability of detected fraud, we find strong evidence that higher growth potential and equity-based executive compensation (EBC), after proper controls, substantially increases the probability of fraud detection. We also find that in the (relatively) fraud-prone high-tech industries, a higher growth potential is positively associated with higher EBC usage during good times and the EBC usage decreases after the enactment of SOX; while there is no significant relationship between growth potential and EBC or SOX and EBC in the (relatively) fraud-free manufacturing industries. The documented evidence is consistent with theoretical predictions with the assumption that the regulator is strategic in combating fraud, but contradicts to theoretical predictions with the assumption that the fraud detection environment is exogenous or mechanical (i.e., without a strategic regulator). We also find that the probability of fraud commission has been reduced by two thirds after the enactment of SOX, in other words, SOX has been very effective and decreased fraud commission by two thirds after its enactment. In light of the current debates, this finding should provide some useful insight to policy makers.

73

Table 1 - Industry Distribution of Detected Fraud Firms The fraud sample is collected from SEC litigation releases dated between 01/01/97 and 06/31/07, and consists of 233 unique detected fraud firms covered by Compustat and with initial year of fraud between 1996 and 2003. During the same period, there are totally 22,633 unique firms covered by Compustat. The industry classification of each firm is defined by the Fama-French 10 industries as following: Industry 1 is Consumer NonDurables (incl. Food, Tobacco, Textiles, Apparel, Leather, Toys); industry 2 is Consumer Durables (incl. Cars, TV's, Furniture, Household Appliances); industry 3 is Manufacturing (incl. Machinery, Trucks, Planes, Chemicals, Office Furn, Paper, Commercial Printing); industry 4 is Energy (incl. Oil, Gas, and Coal Extraction and Products); industry 5 is HighTec (incl. Computers, Software, and Electronic Equipment); industry 6 is Telecom (incl. Telephone and Television Transmission); industry 7 is Shops (incl. Wholesale, Retail, and Some Services); industry 8 is Health (incl. Healthcare, Medical Equipment, and Drugs); industry 9 is Utilities; industry 10 is Other (incl. Mines, Constr, BldMt, Trans, Hotels, Bus Serv, Entertainment, Finance). Industry Code Number of Compustat firms Number of Detected Fraud Firms % of Industry % of Fraud Sample 5.58 3.00 4.72 2.15 34.76 5.15 12.88 9.87 2.58 19.31 100

1 1080 13 1.20 2 482 7 1.45 3 2348 11 0.47 4 941 5 0.53 5 3974 81 2.04 6 884 12 1.36 7 2068 30 1.45 8 1849 23 1.24 9 506 6 1.19 10 8501 45 0.53 Total 22633 233 1.03 Chi-Square Test of the Null that No Association Exists between Industry and Detected Fraud Chi-Square Statistic Pearson's Likelihood Ratio Mantel-Haenszel DF 9 9 1 Value 76.7624 74.1467 12.9183 Prob. <.0001 <.0001 0.0003

74

Table 2 - Time-Series Distribution of Detected Fraud Firms Time-series distributions of the 233 detected fraud firms based on the first fiscal year of fraud, the first year of SEC investigation and the year of SEC enforcement are presented in Panel A, B and C respectively. Panel D presents statistics regarding various fraud-related periods. Fraud period is defined as the last fiscal year of fraud minus the first fiscal year of fraud plus 1; undetected period is the first year of SEC investigation minus the last fiscal year of fraud; investigation period is the year of SEC enforcement minus the first year of investigation; total period is the year of SEC enforcement minus the first fiscal year of fraud. Panel A: Time-Series Distribution by the First Fiscal Year of Fraud First Fiscal Year of Fraud 1996 1997 1998 1999 2000 2001 2002 2003 Total Number of Detected Fraud Firms 21 29 32 47 50 34 12 8 233 Panel B: Time-Series Distribution by the First Year of Investigation First Year of SEC Investigation 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Total Number of Detected Fraud Firms 1 5 16 23 36 34 53 30 21 11 3 233 Panel C: Time-Series Distribution by the Year of SEC Enforcement Year of SEC Enforcement 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 Total Number of Detected Fraud Firms 2 5 13 10 41 48 41 24 28 21 233 Panel D: Summary Statistics of Various Fraud-Related Periods Mean Fraud Period Undetected Period Investigation Period Total Period 2.506 0.695 2.202 4.403 Std Dev 1.520 0.903 1.342 1.928 Min 1 -1 0 0 Max 9 6 6 10 N 233 233 233 233 % of Fraud Sample 0.86% 2.15% 5.58% 4.29% 17.60% 20.60% 17.60% 10.30% 12.02% 9.01% 100% % of Fraud Sample 0.43% 2.15% 6.87% 9.87% 15.45% 14.59% 22.75% 12.88% 9.01% 4.72% 1.29% 100% % of Fraud Sample 9.01% 12.45% 13.73% 20.17% 21.46% 14.59% 5.15% 3.43% 100%

75

Table 3 - Univariate Analysis Univariate comparisons of variables for detected fraud firms and non-fraud firms are presented in this table. For each variable, the mean, median (in parentheses), p-value of the Welch two sample T-test and p-value of the Mann-Whitney test are reported. AUDIT, INVESTMENT, STDDEV, IndusPriorDec and SOX_DETECT are measured as of the year prior to the year of SEC investigation. All the other variables are measured as of the year prior to the first fiscal year of fraud. PPS is the pay performance sensitivity averaged among top five executives of the firm, and is calculated as PPS due to unrestricted stocks plus PPS due to vested stock options, where PPS due to unrestricted stocks is calculated as the number of unrestricted shares owned by an executive divided by total shares outstanding and PPS due to vested stock options is calculated as the number of shares in vested options granted to an executive multiplied by the Black-Scholes hedge ratio then divided by total shares outstanding. Q = (book value of assets - book value of common equity - deferred taxes + market value of common equity) / book value of assets. SG is the avg. percentage sales growth of the last two years. SIZE is the calendar-year end log market value of equity (in $MM). AGE is the age of the firm since its IPO. DEBT is calculated as book value of debt divided by book value of assets. 1.2 workingcapital + 1.4 retained earnings + 3.3 EBIT + 1.0 sales . Z= total assets total assets total assets total assets ROA is calculated as operating income after depreciation divided by book value of assets. INST is calculated as the ratio of total institutional shareholdings to total shares outstanding. BOARD is the size of the firm's board of directors. INDEP_DIR is the percentage of directors on the board who are independent directors. AUDIT is the auditors opinion toward the firm on Compustat. INVESTMENT is the sum of capital expenditures, acquisitions and increase in investments divided by book value of assets. IndusPriorDec is calculated as the total market value of all detected fraud firms in the firms industry scaled by the industrys total market value. STDDEV is the standard deviation of the firm's daily stock returns. SOX_COMMIT is a dummy which equals 1 if the first fiscal year of fraud is 2002 or later and equals 0 otherwise. SOX_DETECT is a dummy which equals 1 if the year of initial fraud investigation is 2002 or later and equals 0 otherwise. Variables PPS (%) Q SG (%) SIZE AGE DEBT Z ROA INST (%) BOARD INDEP_DIR (%) AUDIT INVESTMENT IndusPriorDec STDDEV SOX_COMMIT SOX_DETECT Fraud Sample 3.468 (1.819) 3.827 (1.827) 82.84 (33.90) 5.588 (5.600) 3.103 (2.000) 0.550 (0.492) -0.587 (1.287) -0.103 (0.052) 22.71 (0.65) 7.094 (6.000) 51.93 (50.00) 2.442 (1.000) 0.168 (0.102) 0.754 (0.033) 4.552 (4.305) 0.086 (0.000) 0.506 (1.000) N 233 181 194 195 155 222 169 221 233 233 233 206 178 233 161 233 233 Control Sample 1.087 (0.393) 2.838 (1.434) 40.67 (12.08) 4.433 (4.420) 3.655 (3.000) 0.669 (0.595) -1.997 (1.155) -0.126 (0.034) 14.65 (0.00) 9.366 (9.000) 61.25 (63.64) 1.737 (1.000) 0.116 (0.063) 0.670 (0.000) 3.456 (2.892) 0.233 (0.000) 0.485 (0.000) N 1725 7031 8391 8570 6155 9441 6945 9407 12903 1194 1194 8474 7088 12903 6936 12903 12903 Mean Diff. (P-value) Median Diff. (P-value) 0.000*** 0.013** 0.000*** 0.000*** 0.094* 0.000*** 0.028** 0.489 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.627 0.000*** 0.000*** 0.517 0.000*** 0.000*** 0.000*** 0.000*** 0.041** 0.000*** 0.343 0.477 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.001*** 0.000*** 0.000*** 0.515

76

Table 4 - The Effects of Growth Potential and SOX on Fraud Commission and Detection The effects of growth potential and SOX on fraud commission and detection are reported in this table. The models are estimated using the aforementioned bivariate probit with partial observability technique. P(F) is the probability of fraud commission. P(D|F) is the probability of fraud detection conditioning on fraud having been committed. For each model, the coefficient estimates and Pvalues (in parentheses) are reported. The log likelihood, model Chi-square and its degree of freedom (in parentheses) are also reported, along with the number of observations used in estimation. Detailed explanation for each variable used in estimation is presented in the discription of Table 3. Model 1 Constant Q SG AGE SIZE DEBT Z ROA INST SOX_COMMIT IndusPriorDec AUDIT INVESTMENT STDDEV SOX_DETECT Industry Effects Log Likelihood Model Chi-square (d.f.) Number of Observations Number of Detected Fraud Firms Controlled -0.005 (0.291) 0.058 (0.001)*** 0.049 (0.253) 0.013 (0.023)** -0.171 (0.026)** 0.002 (0.118) -0.451 (0.000)*** P(F) -2.077 (0.000)*** 0.006 (0.237) P(D|F) -4.475 (0.000)*** 1.352 (0.000)*** P(F) -2.362 (0.000)*** Model 2 P(D|F) -5.255 (0.000)***

0.001 (0.446)

0.000 (0.170) 0.002 (0.456) 0.091 (0.001)*** -0.093 (0.268) -0.002 (0.440) -0.150 (0.200) 0.020 (0.000)*** -0.443 (0.002)***

0.026 (0.000)***

0.039 (0.000)***

-0.015 (0.396) 1.630 (0.000)*** 2.218 (0.062)* 0.118 (0.065)* 0.263 (0.174) Controlled

Controlled

-0.018 (0.267) 0.368 (0.000)*** 2.014 (0.000)*** 0.104 (0.003)*** -0.095 (0.262) Controlled

-1072.83 4588.66 (33) 13127 233

-2031.83 2670.66 (33) 13127 233

77

Table 5 - The Effectiveness of SOX in Reducing Fraud Commission The effectiveness of SOX in reducing fraud commission is presented in this table. In Panel A, the pre-SOX probability of fraud commission is calculated with SOX_COMMIT being set to zero and the other variables evaluated at their respective means of the full sample, while the post-SOX probability of fraud commission is calculated with SOX_COMMIT set to one and the other variables evaluated at their means. In Panel B, the only difference compared with Panel A is that the other variables are evaluated at their respective medians of the full sample instead of their means. Panel A: The Probability of Fraud Commission Calculated at the Mean Pre-SOX Post-SOX Reduction in Probability of Fraud Commission Fraud Commission Reduced by Model 1 3.26% 1.09% 2.17% 66.65% Model 2 3.24% 1.10% 2.14% 65.98%

Panel B: The Probability of Fraud Commission Calculated at the Median Pre-SOX Post-SOX Reduction in Probability of Fraud Commission Fraud Commission Reduced by Model 1 3.68% 1.26% 2.43% 65.91% Model 2 2.20% 0.70% 1.50% 68.16%

78

Table 6 - The Effects of EBC on Fraud Commission and Detection The effects of EBC, growth potential and SOX on fraud commission and detection are reported in this table. The models are estimated using the aforementioned bivariate probit with partial observability technique. P(F) is the probability of fraud commission. P(D|F) is the probability of fraud detection conditioning on fraud having been committed. For each model, the coefficient estimates and P-values (in parentheses) are reported. P-values are calculated using bootstrapped standard errors. The log likelihood, model Chi-square and its degree of freedom (in parentheses) are also reported, along with the number of observations used in estimation. In Model 3, P.PPS is the predicted value of PPS resulted from OLS regression of PPS on Lag_Q, Lag_SIZE, Lag_ROA, Lag_DEBT, Lag_Z, Lag_AGE, Lag_STDDEV, Lag_INST and the industry dummies, with the lag variables being measured two years before the initial year of fraud. The instruments in Model 4 are the same as those in Model 3, except that Lag_SG is used instead of Lag_Q. Detailed explanation for the other variables used in estimation is presented in the discription of Table 3. Model 3 Constant Q SG P.PPS AGE SIZE DEBT Z ROA INST SOX_COMMIT IndusPriorDec AUDIT INVESTMENT STDDEV SOX_DETECT Industry Effects Log Likelihood Model Chi-square (d.f.) Number of Observations Number of Detected Fraud Firms Controlled 0.312 (0.000)*** -0.011 (0.280) 0.046 (0.071)* 0.077 (0.162) 0.009 (0.349) -0.763 (0.000)*** -0.016 (0.000)*** -0.631 (0.000)*** 1.154 (0.000)*** P(F) -.625 (0.000)*** -0.039 (0.005)*** P(D|F) -4.610 (0.000)*** 0.170 (0.006)*** P(F) -0.397 (0.021)** Model 4 P(D|F) -4.664 (0.000)***

0.025 (0.000)***

0.000 (0.739) 0.246 (0.000)*** -0.037 (0.171) -0.104 (0.048)** 0.068 (0.413) -0.013 (0.338) -0.615 (0.000)*** 0.008 (0.424) -0.636 (0.000)***

0.024 (0.046)** 1.025 (0.000)***

0.012 (0.699)

-0.029 (0.167) 0.432 (0.013)** 1.380 (0.000)*** 0.106 (0.000)*** 0.180 (0.775) Controlled

Controlled -590.68 690.47 (35) 1958 233

-0.031 (0.559) 0.284 (0.332) 1.483 (0.000)*** 0.102 (0.000)*** 0.223 (0.684) Controlled

-522.44 826.95 (35) 1958 233

79

Table 7 - The Effects of EBC on Fraud Commission and Detection (Control for BOARD and INDEP_DIR) The models of Table 5 are reestimated with variables BOARD and INDEP_DIR added to the P(F) and P(D|F) functions. BOARD is the number of directors on the firm's board of directors. INDEP_DIR is the percentage of independent directors on the firm's board of directors. Both variables are measured as of the year prior to the first fiscal year of fraud. P.PPS is the predicted value of PPS resulted from OLS regression of PPS on all the other variables in the P(F) functions and the industry dummies. For each model, the coefficient estimates and P-values (in parentheses) are reported. P-values are calculated using bootstrapped standard errors. The log likelihood, model Chi-square and its degree of freedom (in parentheses) are also reported, along with the number of observations used in estimation. In Model 5, P.PPS is the predicted value of PPS resulted from OLS regression of PPS on Lag_Q, Lag_SIZE, Lag_ROA, Lag_DEBT, Lag_Z, Lag_AGE, Lag_STDDEV, Lag_INST, Lag_BOARD, Lag_INDEP_DIR and the industry dummies, with the lag variables being measured two years before the initial year of fraud. The instrumental variables in Model 6 are the same as those in Model 5, except that Lag_SG is used instead of Lag_Q. Detailed explanation for the other variables used in estimation is presented in the discription of Table 3. Model 5 Constant Q SG P.PPS AGE SIZE DEBT Z ROA INST BOARD INDEP_DIR SOX_COMMIT IndusPriorDec AUDIT INVESTMENT STDDEV SOX_DETECT Industry Effects Log Likelihood Model Chi-square (d.f.) Number of Observations Number of Detected Fraud Firms Controlled 0.604 (0.000)*** -0.007 (0.000)*** -0.061 (0.000)*** 0.222 (0.000)*** -0.032 (0.000)*** -0.742 (0.000)*** 0.002 (0.225) 0.044 (0.000)*** 0.022 (0.000)*** -0.901 (0.000)*** 1.445 (0.000)*** P(F) -2.430 (0.000)*** 0.039 (0.000)*** P(D|F) -8.707 (0.000)*** 0.054 (0.000)*** P(F) -1.620 (0.000)*** Model 6 P(D|F) -8.881 (0.000)***

0.001 (0.920) 0.160 (0.000)*** 0.039 (0.000)***

0.006 (0.005)*** 0.469 (0.000)*** -0.013 (0.162) -0.133 (0.000)*** 0.323 (0.000)*** 0.001 (0.928) -0.605 (0.000)*** -0.002 (0.556) 0.031 (0.001)*** 0.016 (0.000)*** -0.851 (0.000)***

0.021 (0.020)** 1.457 (0.000)***

0.007 (0.882) 0.152 (0.000)*** 0.045 (0.306)

-0.046 (0.000)*** 0.374 (0.000)*** 1.419 (0.000)*** 0.248 (0.000)*** -0.014 (0.000)*** Controlled -355.73 742.80 (39) 1223 233

Controlled

-0.045 (0.000)*** 0.348 (0.000)*** 1.446 (0.000)*** 0.210 (0.000)*** 0.159 (0.000)*** Controlled -319.87 814.52 (39) 1223 233

80

Table 8 - The Effects of EBC on Fraud Commission and Detection (Logit Link) The models of Table 6 are reestimated using logit link instead of probit link. For each model, the coefficient estimates and P-values (in parentheses) are reported. P-values are calculated using bootstrapped standard errors. The log likelihood, model Chi-square and its degree of freedom (in parentheses) are also reported, along with the number of observations used in estimation. In Model 7, P.PPS is the predicted value of PPS resulted from OLS regression of PPS on Lag_Q, Lag_SIZE, Lag_ROA, Lag_DEBT, Lag_Z, Lag_AGE, Lag_STDDEV, Lag_INST, Lag_BOARD, Lag_INDEP_DIR and the industry dummies, with the lag variables being measured two years before the initial year of fraud. The instruments in Model 8 are the same as those in Model 7, except that Lag_SG is used instead of Lag_Q. Detailed explanation for the variables used in estimation is presented in the discription of Table 3. Model 7 P(F) -4.664 (0.000)*** 0.066 (0.000)*** P(D|F) -8.707 (0.000)*** 0.054 (0.000)*** Model 8 P(F) -3.470 (0.000)*** P(D|F) -8.881 (0.000)***

Constant Q SG P.PPS AGE SIZE DEBT Z ROA INST BOARD INDEP_DIR SOX_COMMIT IndusPriorDec AUDIT INVESTMENT STDDEV SOX_DETECT Industry Effects Log Likelihood Model Chi-square (d.f.) Number of Observations Number of Detected Fraud Firms

1.175 (0.000)*** -0.021 (0.000)*** -0.108 (0.000)*** 0.278 (0.000)*** -0.072 (0.000)*** -1.032 (0.000)*** 0.000 (0.926) 0.088 (0.000)*** 0.039 (0.000)*** -1.568 (0.000)***

1.445 (0.000)***

-0.002 (0.308) 0.160 (0.000)*** 0.038 (0.000)***

0.022 (0.000)*** 0.965 (0.000)*** -0.023 (0.000)*** -0.226 (0.000)*** 0.454 (0.000)*** -0.019 (0.000)*** -0.906 (0.000)*** -0.002 (0.593) 0.066 (0.000)*** 0.037 (0.000)*** -1.496 (0.000)***

0.020 (0.000)*** 1.457 (0.000)***

-0.002 (0.571) 0.150 (0.000)*** 0.039 (0.000)***

Controlled

-0.046 (0.000)*** 0.374 (0.000)*** 1.419 (0.000)*** 0.248 (0.000)*** -0.014 (0.000)*** Controlled -323.19 971.43 (39) 1223 233

Controlled

-0.045 (0.000)*** 0.348 (0.000)*** 1.446 (0.000)*** 0.210 (0.000)*** 0.159 (0.000)*** Controlled

-303.91 1009.98 (39) 1223 233

81

Table 9 - Discriptive Statistics for the Panel Data Descriptive statistics for variables used in EBC panel regressions are presented in this table. Panel 1 presents discriptive statistics for the hightec sample (Fama-French industry code 5). Panel 2 presents statistics for the manufacturing sample (Fama-French industry code 3). Detailed explanation for the variables is presented in the discription of Table 3. Panel A - Hightec Sample
Variables PPS Q SG SIZE Age DEBT Z ROA INST BOARD INDEP_DIR STDDEV N 3100 2729 3034 3004 1957 3084 2578 3088 3100 1586 1586 2707 Mean 1.03 3.11 34.94 6.87 6.31 0.39 1.62 0.06 51.63 7.71 62.92 0.04 Median 0.50 2.20 17.38 6.72 5.00 0.36 1.94 0.09 57.65 7.00 66.67 0.04 Std. Dev. 1.35 2.54 84.10 1.57 5.41 0.20 1.91 0.17 29.89 2.28 18.14 0.02

Panel B - Manufacturing Sample

Variables PPS Q SG SIZE Age DEBT Z ROA INST BOARD INDEP_DIR STDDEV N 2838 2550 2823 2811 870 2835 2681 2837 2838 1857 1857 2689 Mean 0.78 1.69 9.52 6.95 7.20 0.58 2.10 0.11 55.45 9.66 67.23 0.02 Median 0.28 1.46 6.25 6.80 7.00 0.59 2.06 0.10 62.09 10.00 70.00 0.02 Std. Dev. 1.53 0.79 17.84 1.44 5.30 0.20 0.85 0.07 26.81 2.49 17.11 0.01

82

Table 10 - Panel Regressions of EBC on Growth Potential and SOX The results of fixed-effect panel regressions of PPS on growth potential and SOX are reported in this table. Panel regressions are performed for the (relatively) fraud-prevalent hightec industry (Fama-French industry code 5) and the (relatively) fraud-free manufacturing industry (FamaFrench industry code 3) respectively. All firms used in panel regressions must be free from fraud charge during the sample period (from 1994 to 2003). PPS is measured for each firm from 1995 to 2004 (given that data are available on Execucomp). The explanatory variables are lagged to match PPS in all panel regressions. For each model, the coefficient estimates and P-values (in parentheses) are reported, along with the Rsquare, the p-value of F-test for no fixed effect, and the number of observations used in regression. BOOM*Q is an interaction variable which is the product of a dummy variable BOOM and Q, with BOOM being 0 if the fiscal year is 2001 or 2002 and 1 otherwise. BOOM*SG is an interaction variable which is the product of BOOM and SG. For the ease of presentation, the regression coefficients of SG and BOOM*SG are both multiplied by 100. SOX is a dummy which equals 1 if the fiscal year of financial data is 2002 or later and equals 0 otherwise. Detailed explanation for the other variables used in regressions is presented in the discription of Table 3. Dependant Variable: PPS Constant Q BOOM*Q SG BOOM*SG SIZE AGE DEBT Z ROA STDDEV INST BOARD INDEP_DIR SOX R-square P-value of F-test for No Fixed Effect Number of Observations -0.130 (0.000)*** -0.073 (0.000)*** 0.238 (0.037)** 0.027 (0.063)* -0.418 (0.006)*** -0.039 (0.003)*** -0.002 (0.051)* 0.012 (0.352) 0.005 (0.001)*** -0.086 (0.041)** 0.780 0.000 3086 HighTec Industry Model 1 Model 2 2.404 (0.000)*** -0.005 (0.735) 0.013 (0.312) 2.406 (0.000)*** Manufacturing Industry Model 1 Model 2 0.685 (0.110) 0.058 (0.119) -0.025 (0.193) 0.638 (0.141)

-0.094 (0.107) 0.127 (0.028)** -0.124 (0.000)*** -0.072 (0.000)*** 0.251 (0.027)** 0.023 (0.119) -0.390 (0.010)*** -0.040 (0.002)*** -0.002 (0.052)* 0.010 (0.422) 0.005 (0.001)*** -0.099 (0.020)** 0.780 0.000 3086

-0.130 (0.000)*** -0.022 (0.020)** 0.183 (0.205) -0.002 (0.971) 0.817 (0.034)** 0.001 (0.664) -0.002 (0.005)*** -0.007 (0.488) 0.006 (0.000)*** -0.008 (0.824) 0.871 0.000 2832

0.011 (0.953) -0.093 (0.623) -0.119 (0.000)*** -0.023 (0.018)** 0.200 (0.169) -0.002 (0.970) 0.982 (0.010)*** 0.011 (0.563) -0.002 (0.007)*** -0.008 (0.472) 0.006 (0.000)*** -0.001 (0.968) 0.871 0.000 2832

83

Bibliography
Alvarez, R.M. and G. Glasgow, 2000. Two-Stage Estimation of Non-Recursive Choice Variables. Political Analysis 8, 147-165. Anderson, M., R. Banker, and S. Ravindran, 2000. Executive compensation in the information technology industry. Management Science 46, 530-47. Bebchuk, L.A. and J.M. Fried, 2004. Pay without performance: overview of the issues. Journal of Applied Corporate Finance 17, 8-23. Bergstresser, D. and T. Philippon, 2006. CEO incentives and earnings management. Journal of Financial Economics 80, 511-529. Bertrand, M. and S. Mullainathan, 2001. Are CEOs rewarded for luck? The ones without principals are. Quarterly Journal of Economics 116, 901-932. Braumoeller, B.F., 2003. Causal complexity and the study of politics. Political Analysis 11, 209-233. Bruner, D., D. Robinson, and R. Santore, 2005. Equity-based compensation and managerial fraud: an experimental investigation. Working paper, University of Tennessee. Burns, N. and S. Kedia, 2006. The impact of performance-based compensation on misreporting. Journal of Financial Economics 79, 35-67. Cho, I-K. and D.M. Kreps, 1987. Signaling games and stable equilibria. Quarterly Journal of Economics 102, 179-221. Core, J. and W.R. Guay, 2002. Estimating the value of executive stock option portfolios and their sensitivities to price and volatility. Journal of Accounting Research 40, 613-630. Core, J., R. Holthausen, and D. Larcker, 1999. Corporate governance, chief executive compensation, and firm performance. Journal of Financial Economics 51, 371-406. Erickson, M., M. Hanlon, and E.L. Maydew, 2006. Is there a link between executive compensation and accounting fraud? Journal of Accounting Research 44, 113-143. Feinstein, J.S., 1990. Detection controlled estimation. Journal of Law and Economics 33, 233-276. Fudenberg, D. and J. Tirole, 1991. Perfect Bayesian equilibrium and sequential equilibrium. Journal of Economic Theory 53, 236-60. Goldman, E. and S.L. Slezak, 2006. An equilibrium model of incentive contracts in the presence of information manipulation. Journal of Financial Economics 80, 603-26.

84

Graham, J.R., M.L. Lemmon, and J.S. Schallheim, 1998, Debt, Leases, Taxes, and the Endogeneity of Corporate Tax Status, Journal of Finance 53, 131-162. Hartzell, J. and L. Starks, 2003. Institutional investors and executive compensation. Journal of Finance 58, 2351-2374. Healey, P.M. and K.G. Palepu, 2003. The fall of Enron. Journal of Economics Perspectives 17, 3-26. Itner, C.D., R.A. Lambert, and D.F. Larcker, 2003. The structure and performance consequences of equity grants to employees of new economy firms. Journal of Accounting and Economics 34, 89-127. Jensen, M.C., 2005. Agency costs of overvalued equity. Financial Management 34, 4-19. Johnson, S.A., H.E. Ryan, and Y.S. Tian, 2007. Executive compensation and corporate fraud: The Sources of Incentives matter. Working paper, Texas A&M University. Kadan, O. and J. Yang, 2005. Executive stock options and earnings management: a theoretical and empirical analysis. Working paper, Washington University in St. Louis. Kreps, D.M. and R. Wilson, 1982. Sequential equilibrium. Econometrica 50, 863-94. Li, S., 2007. Corporate fraud and costly monitoring: An empirical analysis of a simultaneous system with partial observability. Working paper, Wilfred-Laurier University. Maddala, G.S., 1983. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge. Cambridge University Press. Mas-Colell, A., M.D. Whinston, and J.R. Green, 1995. Microeconomic Theory. New York: Oxford University Press. Murphy, K.J., 2003. Stock-based pay in new economy firms. Journal of Accounting and Economics 34, 129-47. Myers, S.C. and N.S. Majluf, 1984. Corporate financing and investment decisions when firms have information that investors do not have. Journal of Financial Economics 13, 187-221. Peng, L. and A. Rell, 2004. Executive pay, earnings manipulation and shareholder litigation. Working paper, City University of New York. Piotroski, J.D., and S. Srinivasan, 2008. Regulation and Bonding: The Sarbanes-Oxley Act and the Flow of International Listings. Working paper, Stanford University. Poirier, D.J., 1980. Partial Observability in Bivariate Probit Model. Journal of Econometrics 12, 209-217.

85

Povel, P., R. Singh, and A. Winton, 2007. Booms, Busts, and Fraud. Review of Financial Studies, Forthcoming. Robinson, D. and R. Santore, 2005. Managerial incentive and corporate fraud. Working paper, La Salle University. Shleifer, A. and R. Vishney, 1986. Large shareholders and corporate control. Journal of Political Economy 94, 461-488. Smith, C.W., Jr. and R.L. Watts, 1992. The investment opportunity set and corporate financing, dividend, and compensation policies. Journal of Financial Economics 32, 263-292. Wang, T.Y., 2005a. Securities fraud: An economic analysis. Working paper, University of Minnesota. Wang, T.Y., 2005b. Real investment and corporations incentive to misreport. Working paper, University of Minnesota. Yermack, D., 1995. Do Corporate Award CEO Stock Options Effectively? Journal of Financial Economics 39, 237-269. Zhang, I.X., 2005. Economic Consequences of the Sarbanes-Oxley Act of 2002. Working paper, University of Rochester.

86

Appendix Proof of Proposition 1, Part a: As is discussed in Section 3.1, truthful managers will only invest
in positive NPV projects. As a result, in a truthful revelation separating PBE to the

reporting/auditing sub-game (in which all managers are truthful at t = 2), market investors will rationally value firms reporting r = a H as VH = a H + G and firms reporting r = a L as VL = a L + G . Furthermore, since there is no auditing of claims made at t = 2 in the proposed separating equilibrium, a fraudulent manager only faces the penalty fa D if he is caught at t = 4, which occurs with probability (1 ) + H . Thus, a separating PBE for the reporting sub-game exists if every a L -type manager prefers the expected utility under truthful reporting r = a L to fraudulent reporting r = a H : a D [(1 ) + H ] fa D , where the left-hand side is the incremental benefit (in terms of extra compensation) from misreporting and the right hand side is the expected penalty associated with fraud. (Obviously, the a H -type manager never has an incentive to misreport when is non-negative.) This condition can be rearranged as follows: 1 f . 1 H

(A1)

If condition (A1) holds, then the a L -type manager has no incentive to mimic the a H type, and the conjectured market expectations (conditional on the reports) are rational and the conjectured auditing strategy of the RA is also optimal (there is no need to audit if all agents truthfully report even when no auditing occurs), thus the separating PBE in which both types truthfully disclose earnings and the RA never audits at t = 2 exists. Note that this equilibrium is also unique. When inequality (A1) is strictly satisfied, the a L -type manager will find reporting a H unprofitable, even when market investors believe that the firm is a real a H type and his fraudulent reporting

87

will never be detected by the RA at t = 2. When (A1) holds only weakly and the a L -type manager is indifferent between reporting a L or a H , if any a L -type managers commit fraud in (a different candidate) equilibrium, the benefit of committing fraud will fall (as the market rationally discounts firms with a H -reported returns), then the a L type managers will strictly prefer truthfully disclosing their returns. Thus, only the pure strategy separating equilibrium exists when condition (A1) holds.

Q.E.D.

Proof of Proposition 1, Part b.i: In the pooling PBE to the reporting/auditing sub-game with both
types reporting returns as a H and RA never detecting any firm at t = 2, market investors will rationally value a firm reporting a H as V pooling = P * (a H + G ) + (1 P * )(a L + B) , where P * denotes the probability of realizing the high return (anticipated by investors given the expected optimal effort level induced in the equilibrium) in the pooling equilibrium. When all firms claim returns as a H , the payoff to RA if it chooses to investigate a firm at t = 2 will be (1 P * ) J C I , where the first term is the expected overinvestment loss prevented by the investigation, and the second term is the investigation cost. If RA chooses not to investigate any firm at t = 2, its payoff will be [(1 P * )(1 + H ) + P * L ]C I . Thus, for such a pooling equilibrium to exist, we must have (1 P * ) J C I [(1 P * )(1 + H ) + P * L ]C I , which implies P* J C I (1 H ) Crit PRA < 1 . J C I (1 H ) + C I (1 L ) (A2)

Otherwise, RA would optimally choose to investigate any firm reporting return as a H at t = 2

88

(thus the a L -type manager would never want to pool with the a H type). For an a L -type manager to be willing to fraudulently report return as a H , his payoff from doing so must be no less than that from truthfully disclose , aL . Thus, which we must have is

V L V pooling (1 + H ) fa D

(a L + G ) [ P * (a H + G ) + (1 P * )(a L + B )] (1 + H ) fa D . This inequality can be

simplified to (1 + H ) fa D / + J P * (a D + J ) . We then have P* (1 + H ) fa D / + J Crit PManager (a D + J ) 1 Therefore, given > PBE to exist is
Crit Crit P * Max[ PRA , PManager ]

(A3)

f thus > f (1 + H ) , the necessary condition for the pooling 1 H

(A4)

This proves the only if part. Now let us prove the if part and the claim that the pooling outcome is the only possible outcome simultaneously. When inequality (A4) is strict, the RA will never investigate any firm in the industry at t = 2, and the a L -type manager will strictly prefer reporting return as a H (i.e., no a L -type manager would want to mix), therefore the pooling PBE exists and is the only PBE to the reporting/auditing sub-game. Now let us consider the case where
Crit Crit Crit Crit (A4) is weakly satisfied, that is P * = Max[ PRA , PManager ] . If P * = PManager > PRA , then again

RA will never audit at t = 2. Suppose in equilibrium some a L type chooses to truthfully disclose earning. Market value of firms reporting a H will be higher, then the a L type will strictly prefer fraudulently reporting a H . Again, the pooling PBE is the only PBE in this case. Finally, if
Crit Crit P * = PRA PManager , for the same reason the a L type cannot adopt mixed strategy when the a L

89

type mixes, the RA will never investigates at t = 2 and market value of firms reporting a H will be higher, thus all a L -type firms will strictly prefer reporting returns as a H . So the pooling PBE (in which RA never investigates fraud at t = 2 and all a L -type firms pool with the a H type) again exists and the pooling outcome is the only possible outcome.

Q.E.D.

Proof of Proposition 1, Part b.ii: Given the conditions Part b.ii, we know from prior proofs of
Part a and Part b.i that the only possible PBE to the reporting/auditing sub-game should be a mixed strategy PBE in which the a L type mixes between reporting a L and a H . If the RA believes that an a L -type manager fraudulently reports a H with probability m , then the RA will be indifferent between investigating a claimed a H -type firm and not investigating such a firm if m(1 P * ) m(1 P * ) P* J C I = [ * (1 + H ) + * L]C I , P * + m(1 P * ) P + m(1 P * ) P + m(1 P * ) where P * denotes the probability of realizing the high return (anticipated by investors given the expected optimal effort level induced in the equilibrium) in the mixed strategy equilibrium. The left-hand-side is the payoff to the RA if it chooses to investigate any claimed a H -type firm, and the right-hand-side is the (future) payoff to the RA if it chooses not to investigate such a firm. Rearrange the equation, we get m= C I P * (1 L ) . (1 P * )[ J C I (1 H )] (A5)

Crit Crit Probability m equals 1 when P * = PRA . Since we have 0 < P * < PRA , we should also have

0 < m < 1 . For such a mixed strategy equilibrium to exist, the a L -type manager has to feel

indifferent between reporting a L and a H . If the RA investigates a claimed a H -type firm with probability n , and market investors evaluate a claimed a L -type firm as VLM and a claimed

90

M a H -type firm as VH in at t = 2, then we should have M M V LM = n(VH fa D ) + (1 n)[VH (1 + H ) fa D ] .

Rearrange the above equation, we get n = 1

V M 1 M M (1 D ) , with VD V H VLM . fa D (1 H )

(A6)

A claimed a H -type firm should be valued by the market as53

M VH =

P* m(1 P * ) (a H + G ) + * (a L + B ) . P * + m(1 P * ) P + m(1 P * )

From equation (A5), we have C I (1 L) m(1 P * ) = , and * * P + m(1 P ) J C I (1 H ) + C I (1 L) J C I (1 H ) P* = . * * P + m(1 P ) J C I (1 H ) + C I (1 L) Therefore, we get
M VH =

J C I (1 H ) C I (1 L ) (a H + G ) + (a L + B ) . J C I (1 H ) + C I (1 L) J C I (1 H ) + C I (1 L)

A claimed a L type firm should be valued as VLM = a L + G . So we have

M M VD VH VLM =

J C I (1 H ) C I (1 L) aD J. J C I (1 H ) + C I (1 L) J C I (1 H ) + C I (1 L )

Rearrange the above equation, we get

53

In the current version of the model, when calculating market value of the claimed a H -type firm at t = 2, we do not take into consideration the fact that market investors can rationally infer the probability n that the RA adopts in investigating fraud (i.e., market investors can rationally expect the RA to correct the overinvestment problem of a portion n of the fraudulent reporting firms). In another version of the model, we take this rational expectation of market investors into consideration all main results remain the same (if not stronger) as those in the current version, but the mathematical formulation is substantially more complicated. The reason why our results will not be changed, however, is quite straightforward. If we take this rational expectation of market investors into account, M M M V H thus V D will be greater, and the changes in V D and n will always reinforce each other. Thus, the effects in our results will actually be stronger. We adopt the current version for the ease of mathematical presentation.

91

M VD = a D

C I (1 L )(a D + J ) . J C I (1 H ) + C I (1 L)

(A7)

M From equation (A7) we know that VD < a D , thus equation (A6) gives us n < 1 . To complete the

proof, we still need to make sure that n 0 , which is equivalent to require that
M VD fa D (1 + H ) / . M Substitute the expression of VD into inequality (A8), we get

(A8)

C I (1 L)(a D + J ) a D [1 f (1 + H ) / ] . J C I (1 H ) + C I (1 L) But from the conditions of the proposition we already have (1 + H ) fa D / + J J C I (1 H ) . aD + J J C I (1 H ) + C I (1 L) Therefore, a D (1 + H ) fa D / C I (1 L ) . aD + J J C I (1 H ) + C I (1 L)

(A9)

Rearrange the above inequality and we can get inequality (A9). Therefore, we are sure that n 0 . Up to this point, we complete the proof of the existence of the mixed strategy PBE to the reporting/auditing sub-game. This proof, combined with prior proofs of Part a and Part b.i, also show that the mixed strategy PBE is the only possible PBE given the conditions of Part b.ii.

Q.E.D. Proof of Proposition 1, Part b.iii: Given the conditions of Part b.iii, we first prove the existence
of such a mixed strategy equilibrium. If the a L -type fraudulently reports a H with probability m and the RA never detects any claimed a H -type firm, we must have

VL = VH (1 + H ) fa D ,
Where
92

VL = a L + G , and m(1 P * ) P* VH = * (a L + B) + * (a H + G ) . P + m(1 P * ) P + m(1 P * ) We then have (1 + H ) fa D / = V H VL = Simplify the above equation, we get P * [a D (1 + H ) fa D / ] = m(1 P * )[ J + (1 + H ) fa D / ] , and m= P * a D [1 (1 + H ) f / ] . (1 P * )[ J + (1 + H ) fa D / ] P* m(1 P * ) aD * J P * + m(1 P * ) P + m(1 P * )

The RA will never investigate any claimed a H -type firm if m(1 P * ) m(1 P * ) P* J C I < [ * (1 + H ) + * L]C I . P * + m(1 P * ) P + m(1 P * ) P + m(1 P * ) The above inequality is equivalent to m(1 P * )[ J C I (1 H )] < C I P * (1 L) . Thus, for such a mixed strategy equilibrium to exist, we must have m= P * a D [1 (1 + H ) f / ] P *C I (1 L) < . (1 P * )[ J + (1 + H ) fa D / ] (1 P * )[ J C I (1 H )]

However, from the condition of Part b.iii we have (1 + H ) fa D / + J J C I (1 H ) > , thus aD + J J C I (1 H ) + C I (1 L) a D (1 + H ) fa D / C I (1 L ) < , which is equivalent to aD + J J C I (1 H ) + C I (1 L) aD + J J C I (1 H ) . >1+ a D (1 + H ) fa D / C I (1 L)
93

Thus, aD + J J + (1 + H ) fa D / J C I (1 H ) . 1 = > a D (1 + H ) fa D / a D (1 + H ) fa D / C I (1 L ) Then we have a D [1 (1 + H ) f / ] C I (1 L) . < J + (1 + H ) fa D / J C I (1 H ) Therefore, inequality m < P *C I (1 L) is satisfied. (1 P * )[ J C I (1 H )]

To complete the proof of the existence of such a mixed strategy equilibrium, we still need to show 1 that 0 < m < 1 . m > 0 is obvious given > f (thus f (1 + H ) / < 1) ). We still need 1 H

to show that m =

P * a D [1 (1 + H ) f / ] < 1 , which is equivalent to require that (1 P * )[ J + (1 + H ) fa D / ] Thus we need to have

P * {1 +

a D [1 (1 + H ) f / ] aD + J } = P* <1 . J + (1 + H ) fa D / J + (1 + H ) fa D /

P* <

(1 + H ) fa D / + J , which is true given the conditions of Part b.iii. We complete the aD + J

proof of the existence of the specified mixed strategy PBE. Given the conditions of the Part b.iii, we know (from Part a, Part b.i and Part b.ii) that none of the truthful disclosure separating equilibrium, the pooling equilibrium or the mixed strategy equilibrium with both the a L type and the RA adopt mixed strategies can exist, thus the specified mixed strategy equilibrium is the only PBE to the reporting/auditing sub-game in this case.

Q.E.D.

Corollary 2.

Given values for and P * (and the exogenous parameters), in the mixed

strategy case characterized in b.ii of Proposition 1, the following (partial) comparative statics results with respect to growth potential and marginal fraud penalty hold:
94

m n (mn) f < 0, > 0 , and > 0 if 2; m n (mn) = 0, < 0 , and < 0. f f f

Proof.

m n < 0 is straightforward from equation (A5). To prove > 0 , we need to prove

M V D > 0 first. It is not difficult to show that M V D J [a D C I (1 L )] / C I (1 H )a D + LJ [ J C I (1 H ) + C I (1 L)] = . [ J C I (1 H ) + C I (1 L )]2

From equation (A9), we have C I (1 L)(a D + J ) a D [1 f (1 + H ) / ] < a D . J C I (1 H ) + C I (1 L) The above inequality gives us J [a D C I (1 L)] > C I (1 H )a D , thus J [a D C I (1 L)] / > C I (1 H )a D . Therefore, we have
M V D > 0 . Since

M M V M (V D / ) V D n 1 1 , = 2 + = 2 (1 D ) + (1 H ) fa D (1 H ) fa D fa D (1 H ) (1 H )

M M V D V D n > 0 is obvious due to the fact that (1 ) > 0 and > 0. fa D

Now let us prove that

(mn) f > 0 given 2.

M C I P * (1 L) C I P * (1 L) V D mn = (1 ). fa D (1 P * )[ J C I (1 H )] (1 P * )[ J C I (1 H )] (1 H )

95

It can be shown that

C I P [(2 L)(1 ) (1 H )] M fa D C I P * (1 L) VD (mn) = + . The (1 P * )[ J C I (1 H )]2 (1 H ) (1 P * )[ J C I (1 H )] (1 H ) fa D refore, a sufficient condition for

M V D

M V D

(mn) > 0 is (A10)

M V D

(2 L)(1

fa D

) > (1 H ) .
M V D

Given

2 , we have 2(1

fa D

) > 2(1

) 1 > {1 [ H L(1

fa D

)]} , thus inequality

(A10) is satisfied. So we have

(mn) f > 0 given 2.

m n = 0 and < 0 are self-evident from equations (A5) and (A6) respectively. Therefore, we f f should have (mn) < 0. f
Q.E.D.

Proof of Proposition 2: In our risk-neutral principal-agent setting, giving a positive fixed wage to

the agent (i.e., the manager) can be shown to be suboptimal for the principal (i.e., the entrepreneur), thus we should have w * = 0 in any equilibrium. The entrepreneur will then solve the following simplified optimal contracting problem if she anticipates the separating PBE in the subsequent sub-games. (1)
0 f (1 + H )

Max

* * (1 ){P (eT )[a H + G ] + (1 P(eT ))[a L + G ]}

1 *2 * * s.t. {P(eT )[a H + G ] + (1 P(eT ))[a L + G ] eT 0 2 1 * eT = arg max {P(e)[a H + G ] + (1 P (e))[a L + G ]} e 2 2 e

96

Solving the optimal effort choice problem of the manager (i.e., the second constraint) given
* constant the compensation contract , we have eT =

a D . Further notice that the first-order

derivative of the left-hand-side of the first constraint (i.e., the managers individual rationality constraint) with regard to is always positive. Therefore, the first constraint will be binding if
* = 0 thus eT =

a D =0 .

* Substitute eT =

a D into the objective function of the

(a L + G ) 1 * entrepreneur, then the first-order condition being zero will give us T = [1 ] . It is 2 ( a D ) 2

easy to see that there always exists some set of parameter values (e.g., the marginal productivity of effort, , being sufficiently large and/or managerial effort disutility, , being sufficiently small)
* such that T > 0 , i.e., it is worthwhile for the entrepreneur to hire the manager and give him some

portion of the firms shares.

* Since T < 1 / 2 , given that is sufficiently small,

* T f (1 + H ) can always be satisfied.

The entrepreneur will solve the following simplified optimal contracting problem if she anticipates the pooling PBE in the subsequent sub-games. (2)
> f (1 + H )

Max

* * (1 )[ P (e P )(a H + G ) + (1 P (e P ))(a L + B )]

1 *2 * s.t. V pooling (1 P(e P )(1 + H ) fa D e P 0 2 1 e * = arg max V pooling (1 P (e)(1 + H ) fa D e 2 P 2 e Given constant the compensation contract and solving the optimal effort choice problem of the manager, we have
* eP =

(a D + J ) (1 + H ) fa D . +
97

Substitute the expression of e * into the objective function of the principal, then the first-order P condition being zero will give us
* P = [1

1 2

(1 + H ) fa D (a + B) 2 L ]. aD + J (a D + J ) 2

There always exists some set of parameter values (e.g., being sufficiently large and/or
* being sufficiently small, and being sufficiently high) such that P > f (1 + H ) > 0

(recall that H 0 ). Notice that the managers individual rationality constraint will also be satisfied (the first-order derivative of the left-hand-side of the constraint with regard to is always positive; given that is sufficiently high, the individual rationality constraint will bind when is close to zero). The entrepreneur will solve the following simplified optimal contracting problem if she anticipates the first mixed strategy PBE (in which the RA adopts mixed strategy at t = 2) in the subsequent sub-games. (3)
M f (1 + H ) a D / VD

Max

* * M * (1 ){[ P (eM ) + (1 P(eM ))m)]V H + (1 P(eM ))(1 m)V LM }

1 *2 * M * s.t. [ P(eM )VH + (1 P (eM ))VLM ] eM 0 2 1 M * eM = arg max [ P(e)VH + (1 P(e))V LM ] e 2 2 e

* Solving the optimal effort choice problem of the manager given , we get eM =
M VD .

* Substitute the expressions of eM and m (from Part b.ii of Proposition 1) into the objective

function of the principal, and the first-order condition being zero will give us
* M = [1

1 2

(a L + G ) ] , where M 2V D

M VD [1 +

C I (1 L) C I (1 L) J M ] = aD > VD . J C I (1 H ) J C I (1 H )

98

Again, there always exists some set of parameter values (e.g., being sufficiently large and/or
* being sufficiently small) such that M > 0 . For the first mixed strategy PBE to occur, we * M need M f (1 + H )a D / VD , which will be satisfied given that is high enough. The

managers individual rationality constraint will also be satisfied (since the first-order derivative of the left-hand-side of the constraint with regard to is always positive, the individual rationality constraint will bind when = 0 ). The entrepreneur will solve the following simplified optimal contracting problem if she anticipates the second mixed strategy PBE (in which the RA never audits at t = 2) in the subsequent sub-games. (4)
M f (1 + H ) < < f (1 + H ) a D / VD

Max

(1 ){[ P(e * ) + (1 P(e * ))m)]V H + (1 P(e * ))(1 m)V L } N N N

2 1 s.t. [ P(e * )VH + (1 P (e * ))V L ] e * 0 N N N 2

1 e * = arg max [ P(e)VH + (1 P(e))V L ] e 2 N 2 e Solving the optimal effort choice problem of the manager given and utilizing the fact that

VH VL = (1 + H ) fa D /
e* = N

(from the proof of Part b.iii of Proposition 1), we get

(1 + H ) fa D , which is independent of . Thus, P(e * ) is also independent of . N

Substitute the expressions of m (from Part b.iii of Proposition 1) into the objective function of the principal, and it can be shown that the first-order derivative of the objective function with regards to is always negative. Thus the principal will want to minimize in this case (this is intuitive given that the managerial effort level is independent of ). It can also be verified that the left-hand-side of the individual rationality constraint of the manager can be simplified to

99

V L +

[ (1 + H ) fa D ]2 , which is always greater than zero. Therefore, the entrepreneur will 2

* pick to be as close to f (1 + H ) as possible as, but can never set N = f (1 + H ) * (recall that we must have N > f (1 + H ) , otherwise, the separating equilibrium will occur in * the subsequent sub-games). The optimal N hence does not exist.

Q.E.D.

Proof

of

Proposition

3:

For

given

set

of

exogenous

parameters ,

let

* * Crit T () = Crit ( T (), ) , where Crit ( T (), ) is the function defined in equation (8) * evaluated at = T () , conditional on all the relevant parameters in . From equation (13) it * Crit is clear that T () < 1 2 . From equation (8) we then have T () > 0 , thus there will always
Crit exist sufficiently low value of such that < T () and the separating equilibrium occurs.

Q.E.D.
* Proof of Proposition 4: There always exists some subset of such that T > 0 (e.g., given

that the coefficient of disutility of effort, , is small and/or marginal productivity of effort, , is
* Crit big), thus (since T > fH ) we will have T () < 1 . Therefore, for sufficiently high value of
Crit such that > T () , the separating equilibrium cannot exist, thus fraud must exist in any

alternative equilibrium. In that case, depending on different parameter values, either the pooling equilibrium or mixed strategy equilibrium obtains, or no equilibrium exists. Similarly, there
* always exist some subset of such that P > 0 (e.g., is small and/or is big and is * high), thus (since P > fH ) we will have Crit () < 1 . Therefore, for sufficiently high value of P
Crit Crit such that > Crit () , if P * () Max[ PRA (), PManager ()] , the pooling equilibrium P

* obtains. Also, there always exist some subset of such that M > 0 (e.g., is small and/or

100

* M * is big), thus (given that M > ( a D ) fH ) we will have 1 M VD ( fa D ) < 1 . M

VD

1 H

Thus, for

sufficiently

high

value

of

such

that

M * 1 M V D ( fa D ) 1 H

if

Crit Crit P * () < Max[ PRA (), PManager ()] , the fully mixed equilibrium obtains.

* * From Proposition 2, it is easy to verify that T > M . For sufficiently high value of , * * * we also have P > T > M , thus, * * Crit Crit Crit Crit Max[ PRA (), PManager ( M (), )] Max[ PRA (), PManager ( P (), )] .

It is also easy to verify that aL + B 2 M aL + G 2 * P ( ) = [a D + J + (1 + H ) fa D ] > PM () = VD . 2 2( a D + J ) 2 2

* P
* M * Therefore, for some subset of such that Crit () < 1 and 1 M VD ( fa D ) < 1 (i.e., P > 0 P

1 H

* and M > 0 ; recall that

H 0 ), for sufficiently high value of

such that

* * M 1 M V D ( fa D ) 1 P f , we should have either > 1 H 1 H

* * Crit Crit * Crit Crit * i) PP () > PM () > Max[ PRA (), PManager ( M (), )] Max[ PRA (), PManager ( P (), ) ,

or
Crit Crit Crit Crit * * * ii) Max[ PRA (), PManager ( M (), )] Max[ PRA (), PManager ( P (), ) > PP* () > PM () ,

or
* Crit Crit * Crit Crit * * iii) PP () > Max[ PRA (), PManager ( M (), )] Max[ PRA (), PManager ( P (), ) > PM () ,

or
* * * * Crit Crit Crit Crit iv) Max[ PRA (), PManager ( M (), )] > PP () > Max[ PRA (), PManager ( P (), ) > PM () ,

101

or
* Crit Crit * * Crit Crit * v) PP () > Max[ PRA (), PManager ( M (), )] > PM () > Max[ PRA (), PManager ( P (), ) ,

or
* * * * Crit Crit Crit Crit vi) Max[ PRA (), PManager ( M (), )] > PP () > PM () > Max[ PRA (), PManager ( P (), ) .

It is clear that pooling is the only possible equilibrium under i) and fully mixed is the only possible equilibrium under ii). Under iii) to vi), both pooling and fully mixed are possible. Which one obtains then depends on the entrepreneurial utility (since the entrepreneur is the Stackelberg leader in the overall game). Thus, when is sufficiently high, either the pooling equilibrium or the mixed strategy equilibrium obtains.
Q.E.D.

102