Auditors’ Ability to Detect Financial Deception: The Role of Auditor Experience and

Management Cognitive Dissonance

Jessen L. Hobson*
William J. Mayew†
Mark Peecher*
Mohan Venkatachalam†

*Department of Accountancy, University of Illinois at Urbana-Champaign

†
Duke University – Fuqua School of Business

Preliminary. Please do not cite, distribute, or post without permission.

May 2015

We are grateful to the accounting professionals from various audit firms who participated in the
study. We appreciate helpful comments from Spencer Anderson, Scott Asay, Clara Chen, Jon
Davis, Brooke Elliott, Scott Emett, Kevin Jackson, Joseph Johnson, Tracie Majors, Terence Ng,
Matt Pickard, Mike Ricci, Chad Simon, Jason Smith, Hun-Tong Tan, Mark Zimbelman,
reviewers and participants at the Auditing Section Midyear Meetings, the Brigham Young
University Accounting Symposium, the 2014 Accountancy Conference the University of São
Paulo, the 2015 EDEN Doctoral Seminar in Audit Research (London), the Financial Accounting
and Reporting Section Midyear Meeting, and the International Symposium on Audit Research,
and workshop participants at Lehigh University, Nanyang Technological University, and the
University of Illinois at Urbana-Champaign. We thank EB Altiero and Chris Calvin for their
helpful research assistance.
 Auditors’ Ability to Detect Financial Deception: The Role of Auditor Experience and
Management Cognitive Dissonance

We examine how extensive audit experience and a prompt to attend to the CEO’s cognitive
dissonance individually and jointly influence auditors’ detection of financial deception. We
predict and find that experienced auditors outperform both chance and inexperienced auditors,
especially when prompted to attend to managers’ cognitive dissonance. This is encouraging, as
meta-analyses from psychology find that experts generally outperform neither chance nor
novices in detecting deception. Also as predicted but more worrisome, unprompted experienced
auditors’ performance edge over novices arises predominantly from fewer false positives. While
false positives about fraud could strain an auditor’s relationship with management, false
negatives jeopardize audit effectiveness and increase the risk that financial statement users will
suffer loss from fraud. In supplemental process analysis, we find that experience enables auditors
to identify more, and more accurate, red flags in CEO’s narratives. This red flag advantage,
when experienced auditors are prompted to attend management’s cognitive dissonance, translates
into more accurate fraud detection. Finally, in exploring whether adding audio to written
transcripts improves accuracy, we observe it does so only for inexperienced auditors.
 Auditors’ Ability to Detect Financial Deception: The Role of Auditor Experience and
Management Cognitive Dissonance

1. Introduction

We use an experiment to examine whether and how extensive audit experience (over 20

years on average) and a prompt to attend to the CEO’s cognitive dissonance individually and

jointly influence auditors’ detection of financial deception from CEO question and answer

(Q&A) portions of earnings conference calls. We also explore whether the medium used to

convey these Q&A portions, transcript alone or transcript plus audio, improves auditors’

detection accuracy.

Our experimental examination is important because detection of material financial

statement fraud is a long-standing responsibility of public-company financial statement audits.

Despite their responsibility to provide high assurance that management’s financial statements are

free of material fraud, auditors seldom have been the first party to detect frauds, especially prior

to the Sarbanes-Oxley Act (Dyck et al. [2010]). Further, there is concern that considerably more

material financial statement fraud goes undetected than is ever detected (Dyck et al. [2013]).

Investors, regulators, standard setters, and auditors all agree that it would be helpful if auditors

were more effective in detecting fraud (e.g., Christensen, et al. [2015], Hogan et al. [2008]),

particularly given recent increases in accounting-related class action lawsuits (Heller [2015]).1

Our examination is also important for three, more specific reasons. One, it helps build

theory about the determinants of auditors’ ability to detect fraud from relatively unscripted
1
In 2007, the Public Company Accounting Oversight Board (PCAOB) issued a “Rule 4010” report describing
deficiencies in the overall approach and tests used by some auditors to assess and respond to fraud risk factors
(PCAOB [2007]). In 2010, the Center for Audit Quality (CAQ) issued a report on behalf of its member firms, stating
that fraudulent reporting continues to undermine investor confidence and calling for the identification of ways to
better detect or deter material fraud (CAQ [2010]). Still more recently, the PCAOB noted that the number and
magnitude of financial fraud cases “demonstrate that heightened auditor scrutiny is warranted” (PCAOB [2013b])
and appointed a task force to explore ways to improve auditor effectiveness in detecting fraud (PCAOB [2013a]).

1
management narratives about their financial statements. We view Q&A portions of

management’s earnings conference calls as a prominent example of relatively unscripted

narratives. Other examples pervade audits, including management responses to formal and

informal inquiries to help auditors gain an understanding of and test internal controls as well as

to plan and conduct substantive tests (see, e.g., PCAOB AS No. 12, No. 13, and No. 15).

Second, this study of auditor deception detection contributes to the larger psychological

literature on deception detection. Prior studies using participants ranging from experienced

judges and police officers to lay persons and college students across numerous contexts only

rarely attain above-chance performance in detecting deception. Meta-analyses in psychology

research indicate even experts rarely detect deception at better than chance levels (Bond and

DePaulo [2006]; Bond and DePaulo [2008]), and that experts in general are no better than non-

experts at detecting deception (Vrij et al. [2006]). As described below, our “experts” have

somewhat unique incentives in that they are detecting deception from a party (management) who

can strongly influence their own compensation. Additionally, our potential deceivers (CEOs and

CFOs) can be very experienced and even coached in communicating in a self-interested manner.

Third, it has a pragmatic benefit of improving our understanding of whether and how

auditors can detect fraud using earnings conference calls, per se. Recent evidence suggests

automated analysis of conference call speech can assist in detecting financial misreporting at

better than chance levels (Larcker and Zakolyukina [2012], Hobson et al. [2012]). However, we

know of no research examining whether and when auditors judgmentally can do so. Assessing

auditor performance in this regard is important given current audit standards recommend

reviewing earnings conference calls to help assess risk of misstatement (AS No. 12, PCAOB

[2010]). AS 12 neither requires nor recommends that auditors use conference calls to help detect

2
warning signs of fraud, however, and the results of our study add to mounting evidence

suggesting that audit standard setters may want to consider doing so.2

The theory we rely on warrants predicting that many years of experience will help

auditors avoid false positives (i.e., predicting fraud when it is not present) but will help

significantly less, if at all, in avoiding false negatives (i.e., predicting no fraud when it exists).

An implication is that very experienced auditors will outperform inexperienced auditors in

evaluating non-fraud companies, but that they will do so to a lesser extent when evaluating fraud

companies. It also predicts that a prompt to attend to management’s cognitive dissonance will

help very experienced auditors overcome this imbalance, enabling them to use knowledge

acquired over many years to avoid false negatives.

These theory-based predictions build on prior psychology and auditing research. Over

time, decision makers adaptively attend to and selectively interpret information so as to avoid

committing what they experientially learn to be the “primary error” in their natural decision

environments (Friedrich [1993]; Friedrich et al. [2005]). That is, rather than just learning to be

accurate, decision makers–arguably quite rationally–learn to avoid more costly types of errors. In

the language of motivated reasoning theory, decision makers adaptively learn to pursue

directional goals. For auditors, commitment to directional goals biases them towards acceptance

of management’s accounting methods, even more so when auditors explicitly are asked to

objectively assess the quality of these methods (Kadous, et al. [2003]).

2
In contemplating new standards on how auditors should consider using conference calls to help assess the risk of
misstatement due to fraud, standard setters may also want to consider the stream of research comparing accuracy
rates of statistical linear models unaided by human judgment, statistical linear models of human judgment, and
human judgments themselves (e.g., Dawes, et al. [1989]), but we are focused on unaided auditors’ professional
judgments in this paper. Swets et al. [2000, 5] observe that, in complex professional judgment contexts (e.g.,
medicine), a “prevailing practice” is to supply output from statistical linear models to human decision makers who
then make final judgment calls.

3
 Evidence suggests experience increases auditor commitment to management-preferred

directional goals (Shaub and Lawrence [1999]). They predict and find that inexperienced

auditors (audit staff) are relatively aggressive skeptics while experienced auditors (audit

managers and partners) are relatively reluctant skeptics in terms of beliefs and actions. While

university courses on auditing likely amplify inexperienced auditors’ skepticism by covering

infamous frauds (e.g., WorldCom), even seasoned auditors rarely directly experience fraud. One

recent archival study (Dyck et al. [2013]) finds that the rate of detected frauds, by any party, is

less than 3% for public companies, even as it warns that detected frauds could be the “tip of the

iceberg”.3 By contrast, auditors regularly encounter fee pressure to not expand audit testing,

jeopardizing audit quality (Ettredge, et al. [2014]). While post-SOX regulatory inspections may

reduce fee pressure, Peecher et al. [2013], in reviewing auditor incentives, conclude that while

auditors are penalized for failing to find fraud, they are only ambiguously (if at all) rewarded by

regulators, management, or investors for actually detecting fraud. What is salient is that when

auditors detect fraud, their expected litigation costs immediately increase, and they often lose

that client (Doty [2014]). Thus, over many years of experience, false positives likely become

auditors’ “primary error” to avoid, predisposing them to explain away (embrace) cues suggestive

of fraud (no fraud), and reinforcing their susceptibility to false negatives.

We next examine a potential remedy to help experienced auditors overcome this learned

behavior: prompting attention to management’s cognitive dissonance. Motivated reasoning

theory holds that people pursue directional goals only if they reasonably can maintain an illusion

of objectivity (Pyszczynski and Greenberg [1987]; Kunda [1990]). Our prompt, randomly

assigned to half of our participants, states that research has shown cognitive dissonance to be a

3
Dyck et al. [2013] estimate that, if one considers undetected fraud, the ongoing fraud rate is about 13% in US
public companies.

4
correlate of fraudulent reporting, and that CEOs who say things they believe to be untrue likely

experience cognitive dissonance—a negative emotion that triggers unease, discomfort,

annoyance or botheredness. In addition, only prompt condition participants are asked to assess

how much cognitive dissonance the CEO felt during the conference call.

Because people are considerably more successful in detecting others’ emotional states

than in detecting deception (Ambady and Weisbuch [2010]), we expected that prompted auditors

would be able to detect unpleasant emotions in CEOs experiencing cognitive dissonance.

Detection of these emotions coupled with knowledge that dissonance is correlated with

deception, would make it unreasonable for very experienced auditors to rationalize away signs of

fraud. Thus, the prompt frees very experienced auditors to apply their knowledge to fraud

companies, instead of applying it predominantly to non-fraud companies. As a result, we

predicted that prompted, very experienced auditors would commit fewer false negatives than

unprompted, very experienced auditors. Regarding inexperienced auditors, we expected them to

have accuracy goals instead of management-preferred directional goals, so that the prompt would

yield less, if any, benefit for their evaluation of fraud companies.

Finally, we explore whether the medium—transcript only or transcript plus audio—used

to process the earnings calls matters. While PCAOB AS No. 12 encourages auditors to review

earnings conference calls, it does not identify a preferred medium. A priori, our theory does not

warrant predicting the medium, per se, to matter overall or to matter differentially for very

experienced versus inexperienced auditors. While an intuitively appealing possibility is that

audio facilitates conveyance of emotional cues correlated with deception, several psychology

studies on deception detection in which participants fail to outperform chance use live or

5
recorded audio. Exploring medium effects also serves as a robustness check for our main

findings.

Our experiment is conducted online. Thirty-one auditors from multiple large public

accounting companies with an average audit experience of 24 years participated, with 21 (68%)

being current or retired U.S. partners. In addition, 180 inexperienced auditors participated (i.e.,

students from a large state university). Each participant provides deception judgments for four

publicly traded companies, using excerpted CEO responses to analyst questions during quarterly

conference calls. Programmed software randomly draws excerpts from a population of five fraud

and five non-fraud companies (total of 10), and we inform participants of this 50% fraud rate.

We classify excerpts as fraudulent if the company’s quarterly financial statements later were

restated and linked to fraud, regulator investigation, or class-action litigation. We also

dichotomously measure participants’ experience and manipulate, between subjects, the presence

of a cognitive dissonance prompt (absent versus present) and the medium of the conference call

(transcript only versus transcript plus audio). The experiment begins with instructional videos,

examples, and practice. For each company, participants receive background information and

financial statements. Using CEO answers to analyst questions, participants decide whether or not

they think the results being discussed are fraudulent. This decision is our primary dependent

measure.

We find that, while experienced auditors’ fraud accuracy rates (67%) exceed those of

students (53%) and chance overall, experienced and inexperienced auditors are equally prone to

false negatives. That is, we observe no evidence that many years of experience alone helps

auditors identify fraud companies. In fact, despite being told they would most likely evaluate two

fraud companies and two non-fraud, very experienced auditors pick fraud only 40% of the time

6
(relative to 46% for inexperienced auditors). Though consistent with a predicted tendency to

minimize false positives, this finding is disconcerting, as false negatives increase the risk of

investor loss due to undetected financial statement fraud.

Encouragingly, however, our prompt to attend to management’s cognitive dissonance

significantly improves how accurately very experienced auditors assess fraud companies, from

43% to 70%. Further, very experienced auditors pick fraud only 31% of the time without the

cognitive dissonance prompt, but 50% of the time with the prompt. Thus, a simple prompt to

consider management’s cognitive dissonance significantly improves experienced auditors’

sensitivity to fraud.

In exploratory findings, we observe that adding audio to conference call transcripts

improves the accuracy rates of inexperienced auditors (from 49% to 57%), but has no effect on

very experienced auditors. Finally, we perform several robustness tests and examine process

measures underlying the auditors’ fraud judgments. During training, participants were asked to

identify specific CEO sentences that they perceived to include a red flag, before rendering their

final judgment. We find that very experienced auditors are more accurate at identifying

fraudulent statements in conference calls as red flags and that our cognitive dissonance prompt

increases this accuracy for very experienced auditors.

In summary, this study contributes to the scholarly audit literature as well as to practice.

Practitioners likely care that earnings conference calls can help auditors judgmentally assess

fraud risk, and that many years of experience helps auditors more accurately interpret CEO

narratives during conference calls, particularly when prompted to consider management

cognitive dissonance. Practitioners also may be interested to learn that inexperienced auditors’

7
performance is improved to rates matching those of unprompted, very experienced auditors when

inexperienced auditors are provided with audio.

From a theory perspective, we develop and empirically test two new hypotheses about the

individual and joint effects of many years of experience and a prompt to attend to management’s

cognitive dissonance on auditors’ ability to detect deception. We observe new evidence

consistent with very experienced auditors’ adaptation to their natural decision environments

causing them to outperform both chance and inexperienced auditors in avoiding false positives,

but also leading them to be significantly more concerned about false positives than false

negatives. However, the significant interaction between experience and our cognitive dissonance

prompt is encouraging and advances theory on how to improve auditors’ fraud detection

capabilities.

2. Prior Research and Hypotheses

2.1 AUDITOR JUDGMENTS ABOUT THE PRESENCE AND ABSENCE OF DECEPTION

IN EARNINGS CONFERENCE CALLS

As part of the requirement to understand conditions that might cause material

misstatement in management’s financial statement, Auditing Standard No. 12 asks auditors to

consider observing or reading the earnings conference call (PCAOB [2010]).4 The Q&A portion

of the earnings conference call is a relatively unscripted narrative about current company

operations, and recent archival evidence suggests that unique information exists in the Q&A

dialogues between analysts and management (Blau, Delisle, and Price [2015], Hollander et al.

4
Very experienced practitioners from several audit firms with whom we have spoken noted that, for that last several
years, it has been common for experienced audit team members to read or listen to earnings conference calls to
gather evidence about misstatement risks and business risks. However, they indicated that generally there were no
formal audit procedures about conference calls in their audit methodologies during the same time span. Of course,
even if auditors cannot use conference calls to detect the absence or presence of fraud, these calls still may help
them assess client business risks, misstatement risks for complex accounting estimates, and update their audit plans.

8
[2010], Matsumoto et al. [2011], Mayew and Venkatachalam [2012], Price et al. [2012]).

Research extending this line of work suggests that conversations and discussions during earnings

conference calls contain information useful for identifying financial misreporting. For example,

Larcker and Zakolyukina [2012] estimate linguistic-based classification models and document

that CEO and CFO narratives during conference calls help identify deceptive discussions

significantly better than chance levels. Hobson et al. [2012] use CEO speech samples from

earnings conference call Q&As and document that vocal cues predict financial restatements

beyond the predictive ability of financial accounting and linguistic-based predictors.

However, judgment based attempts to detect the presence of deception are quite

challenging for individuals (DePaulo et al. [2003], Bond and DePaulo [2006]). Prior theoretical

and empirical work in social psychology has shown that the accuracy rates of deception

judgments made by experienced professionals are rarely better than those of inexperienced

professionals or laypersons (Bond and DePaulo [2006] and Bond and DePaulo [2008]; Vrij

[2008]).5 In auditing, Jamal et al. [1995] (see also Johnson et al. [2001]) find that few of the audit

partners they examine detect seeded fraud in case materials. Lee and Welker [2007, 2008] find

that upper-level accounting majors detect deception from inquiries about real-estate property

values at rates only slightly better than chance.

Nevertheless, some evidence suggests that very experienced auditors may outperform

inexperienced auditors in the domain of financial fraud detection (e.g., Bonner and Lewis [1990];

Knapp and Knapp [2001]).6 Brazel et al. [2010] show greater success in fraud brainstorming

5
Experts examined include police officers, detectives, judges, interrogators, customs officials, mental health
professionals, polygraph examiners, job interviewers, etc. One exception to this finding is that of so called “lie
detection wizards” (O'Sullivan and Ekman [2004]), who do systematically outperform novices.
6
We use the terms very experienced and inexperienced rather than expert and novices because part of what we test
is whether experience helps auditors move from novice to expert (or to at least better) performance levels.

9
sessions when audit partners or forensic specialists, versus lower level staff, lead the session.

Carcello and Nagy [2002] find a negative relationship between auditor industry specialization

and undetected fraud. Knapp and Knapp [2001] show a positive effect for audit experience on

the effectiveness of analytical procedures in detecting fraud. Similarly, Bernardi [1994] shows

that managers outperform seniors in fraud detection, but only when managers have relatively

high moral development.

This conflicting evidence is at least partially resolved, however, if one considers that

auditors’ experience may improve fraud detection accuracy only conditionally. Shaub and

Lawrence [1999] find that experienced auditors are somewhat reluctant skeptics while

inexperienced auditors are aggressive skeptics. Thus, audit experience could bias auditors away

from predicting fraud for at least two reasons. First, auditors generally experience low rates of

fraud occurrence during their careers. Even when auditors discover a material misstatement, that

misstatement is usually due to unintentional error. While the true rate of financial statement

fraud is unknown, detected fraud occurs at less than 3% of U.S. public companies (Dyck et al.

[2013]). Prior psychology research finds that when individuals make decisions from experience

they tend to underweight rare occurrences (Hertwig et al. [2004]).

Second, auditors’ incentives are structured such that there are disincentives to find fraud.

Peecher et al. [2013] highlight that while auditors can be penalized for failing to find fraud, they

are not rewarded for work to detect and prevent fraud. Nor are there any public regulatory

rewards for performing audits of particularly high quality. Further, auditors who blow the whistle

for fraud on their own client not only lose that client 50% of the time (Doty [2014]; Dyck et al.

[2010]), but also are specifically excluded from the set of persons eligible to receive monetary

10
rewards from the whistleblowing provisions of the Dodd-Frank Wall Street Reform and

Consumer Protection Act.

Despite few rewards for finding fraud, both internal and external to the audit firm, the

costs of predicting fraud loom large, particularly since fraud is a relative low-probability event.

These costs include fee pressure, budget issues, and client relationship issues, as auditors are

motivated to meet budget and as audit fees are generally stable or decreasing (Audit Analytics

Staff [2014]; Doty [2014]). Prior research finds that fee pressure and budget issues are salient

motivators for audit teams (Willett and Page [1996]; Kelley and Margheim [1999]) that affect

audit quality (e.g., Houston [1999]; Asare et al. [2000]; Ettredge et al. [2014]) and, potentially,

fraud detection (Braun [2000]). Moreover, voiced suspicion of fraud likely will require

additional audit procedures, causing tension, delays, and budget overruns.

We posit that these incentives, coupled with the rarity of fraud have profound

psychological effects on experienced auditors, making it both desirable and seemingly likely that

any particular company they are investigating is not committing fraud. In particular, two

substantially related theoretical accounts exist with regard to the psychological processes

experienced auditors likely adaptively learn to employ over time. One, auditors likely follow a

primary error detection and minimization (PEDMIN) testing strategy (Friedrich [1993]). When

doing so, they strive to minimize errors that they experientially have learned to be more

important or salient. Since salient errors from making a “fraud exists” judgment are personally

costly, for the reasons discussed above, false positive error (Type I) minimization—mislabeling a

clean company as fraudulent—is likely the experienced auditor’s primary concern. Thus, very

experienced auditors likely learn to not notice or to explain away red flag indicators of fraud

since taking them at face value would increase the likelihood of making an experientially costly

11
error. Importantly, this strategy of minimizing false positives likely comes at the expense of

committing more false negatives, potentially creating disutility for regulators, investors, and

other financial statement users.

Two, motivated reasoning theory holds that when decision makers have preferred

outcomes, they (often subconsciously) activate directional goals, triggering skepticism of

contrary evidence and ready acceptance of supportive evidence (Kunda [1990]). Auditors use

motivated reasoning (Nelson [2009]), and it can impair their objectivity in assessing the quality

of management’s preferred accounting treatments (Kadous et al. [2003]). To the extent

experience enables auditors with directional goals to attend to more red flags, these auditors will

attempt to explain them away in benign terms (Ditto and Lopez [1992]).

In sum, very experienced auditors will be more successful at correctly classifying non-

fraud companies than fraud companies. Accuracy rates of inexperienced auditors, on the other

hand, are subject to two forces. First, they are less likely to pursue management-preferred

directional goals than are experienced auditors, having yet to experientially learn the relative

costs of false positives and false negatives. Second, extant literature on deception detection

suggests that their accuracy rates are unlikely to be better than chance (Zuckerman et al. [1981];

Bond and DePaulo [2006]; Vrij [2008]). Thus, inexperienced auditors’ accuracy rates will likely

be similar and relatively low across fraud and non-fraud companies.

Overall, very experienced auditors are likely to outperform inexperienced auditors when

classifying non-fraud companies but this accuracy difference is expected to dissipate when

classifying fraud companies. These predictions lead to an ordinal interaction with the following

two simple main effects: a simple main effect of experience given non-fraud companies and a

12
simple main effect of company type (i.e., non-fraud versus fraud) given very experienced

auditors.

H1: Accuracy rates will be highest for very experienced auditors assessing non-fraud
companies, significantly lower for very experienced auditors assessing fraud
companies and similarly low for inexperienced auditors assessing both fraud and
non-fraud companies.

2.2 COGNITIVE DISSONANCE MANIPULATION

If auditors experientially learn from their decision environments that it is adaptive to

explain away ambiguous fraud indicators to avoid false positives, identifying remedies that

nudge auditors to reinterpret these cues likely will help prevent false negatives. In identifying a

remedy, we consider a key principle of motivated theory. Specifically, this theory holds that

decision makers use one-sided reasoning to attain directional goals only if they reasonably can

maintain an illusion of objectivity (Pyszczynski and Greenberg [1987], Kunda [1990], Kadous,

et al. [2003]).

To make it harder for experienced auditors to maintain this illusion, we devised a prompt.

This prompt informed auditors that prior research has shown that cognitive dissonance is

predictive of fraud and explained that people who say things that they believe to be untrue

usually experience cognitive dissonance. Participants receiving the prompt also were asked to

assess the degree to which managers felt cognitive dissonance on the calls that they analyzed.

Thus, viewed exhaustively, this prompt contains three parts—the definition of cognitive

dissonance (including its effect on individuals), a link from cognitive dissonance to deception,

and encouragement to observe cognitive dissonance in the conference call CEO.

This prompt was desirable for several reasons. One is that deception frequently leads to

cognitive dissonance (DePaulo et al. [2003], Harmon-Jones [2000]); Ekman [1985], [1992]).

Another is that prior research shows dissonance markers in the speech of the CEOs in our sample

13
were associated with financial misreporting (Hobson et al. [2012]), and a third reason is that

people generally are better at detecting states of emotion than at detecting deception. As Ambady

and Weisbuch [2010, 483] observe, “Whereas social perceivers do not exhibit particular

intelligence in deception detection, emotion recognition is a different story….” People often

accurately identify emotion at well above chance levels from vocal and linguistic cues, even

without observing facial expressions, which also improves emotion detection (Elfenbein and

Ambady [2002], Alm [2008]). Finally, psychology research finds that peak performance for the

ability to perceive emotion occurs between the ages of 40 and 60 (Hartshorne and Germine

[2015]), which is the likely age range of our experienced participants given that they have 24

years of audit experience on average.

We therefore conjecture that the cognitive dissonance prompt will help experienced

auditors better detect deception in fraud companies. We expect that very experienced auditors’

repeated exposure in casual and formal interview settings with management will enable them to

successfully detect the heightened emotions of CEOs experiencing cognitive dissonance, and that

once the cognitive dissonance prompt is given, experienced auditors will no longer be able to

reasonably ignore these cues. By contrast, we expect that for inexperienced auditors, the

cognitive dissonance prompt will be less, if at all, helpful in improving their accuracy in

detection of deception. While the prompt should make inexperienced participants more aware of

cognitive dissonance and the link from dissonance to deception, inexperienced participants have

not had repeated dialogs with management that facilitate detecting emotion and dissonance from

speech. Thus, for inexperienced auditors, a cognitive dissonance prompt is less likely to improve

the accuracy of their deception judgments. In combination, therefore, we predict the following

14
ordinal interaction in which the cognitive dissonance prompt is significantly more helpful for

very experienced than for inexperienced auditors.

H2: Accuracy rates for fraud companies will be highest for very experienced auditors
who are prompted about management cognitive dissonance, significantly lower for
very experienced auditors without a cognitive dissonance prompt and similarly low
for inexperienced auditors, regardless of whether or not they receive a cognitive
dissonance prompt.

2.3 EXPLORING THE EFFECTS OF CONFERENCE CALL MEDIUM

A priori, our theory does not compellingly warrant a directional prediction of how the

medium of the call will affect auditor accuracy rates at detecting deception. Nevertheless, we

explore the medium through which auditors consume the conference call in part because PCAOB

AS No. 12 suggests that auditors consider “observing or reading transcripts of earnings calls”.

Observing conference calls could refer to reading or listening to conference calls (or possibly

watching them, which we do not test herein). Given that the standard does not supply explicit

guidance, it is unclear which media the auditor should use, and which of the two we examine is

more beneficial.

Reading transcripts stresses the linguistic component of the call whereas listening stresses

both the linguistic and vocal components. Interestingly, however, commercial providers of

conference call transcripts typically purge speech hesitations (ah’s, um’s) and other linguistic

features that likely provide important linguistic clues about cognitive dissonance, deception, or

both. Our experiment holds this constant as we manually reinsert these speech hesitations into

our written transcripts. This design choice likely biases against our analyses finding differences

between our two design mediums, but provides a cleaner manipulation of medium.

Prior research is mixed as to the benefits of adding audio on detection accuracy rates.

One the one hand, many psychology studies finding detection rates that fail to surpass chance

15
also include live or taped audio. It could be that audio distracts receivers from linguistic features

of interpersonal communication, as prior research suggests linguistic features are useful in

detecting deception in both general and business contexts (e.g., Zhou et al. [2004]; Larcker and

Zakolyukina [2012]). On the other hand, prior research suggests social information is contained

in sender’s voices (Burgoon et al. [2008]), and hearing the CEO’s voice could facilitate

comprehension of the linguistic cues themselves. Intuitively, this may be especially the case for

novices who likely are less familiar with industry-specific performance metrics or jargon

common in earnings conference calls, which would tax their working memory sources they

otherwise could devote to deception detection. Since we have no strong a priori basis for a

directional prediction we posit the following exploratory research question:

RQ1: Does the medium used to convey management’s responses during Q&A portions of
earnings conference calls—transcript alone or audio plus transcript—affect auditors
accuracy rates in fraud detection?

3. Method

3.1 PARTICIPANTS

To examine our hypotheses and provide a strong test (Kerlinger and Lee [2000], 459) of

the effect of experience on fraud detection, we gather data from very experienced audit

professionals and inexperienced audit students. We examine very experienced auditors because

research suggests that expertise often takes very extensive, deliberate, and practical experience,

in general and professional contexts (Ericsson et al. [1993]). We also recognize that our molar

experience variable almost certainly encompasses several molecular constructs known to

influence auditor task performance, such as technical knowledge, tacit knowledge, general

diagnostic reasoning, etc. We intentionally use a coarse proxy for this study, however, as it is a

relevant variable of interest in the natural audit ecology (many of these constructs are

16
confounded in the real world) and because our study is a first step in examining determinants of

auditors’ ability to detect the presence versus absence of fraud from the Q&A sessions of

earnings conference calls.7

Our thirty-one very experienced participants are current or retired audit professionals

from multiple large public accounting firms with an average of 24 years in the audit, assurance,

and/or fraud/forensic services, and 22 years in public company audit. Twenty-one of them are

partners or retired partners, four are managers, senior managers, or directors, two are seniors, and

four are staff.8 All but three are CPAs. In addition, 180 students at a large public university

enrolled in an auditing class as accounting majors provided usable judgments.

3.2 SPEECH CORPUS SELECTION

Each participant provided deception judgments to excerpts from the question and answer

portion of a quarterly earnings conference call for four companies. Due to time constraints and

our desire to solicit audit partners, participants could only reasonably evaluate four companies.

We draw each company from a population of 10 public company quarter-end earnings

conference calls containing five companies with deceptive discussions and five without. The

allocation of four companies to each participant is random with the stipulation that participants

have an 80 percent chance of evaluating two fraud companies, a nine percent chance of one fraud

company and three non-fraud companies, a nine percent chance of one non-fraud company and

three fraud companies, a one percent chance of all fraud companies, and a one percent chance of

7
We encourage future research that disentangles main and interactive effects of these various constructs. We do not
have a sufficiently specified theory to a priori predict specific ranks at which there are significant increases in
auditors’ ability to detect the absence of fraud or, when coupled with a prompt to consider manager’s cognitive
dissonance, to detect the presence of fraud. Nor are we able to predict how precipitously and gradually the increases
occur. We leave empirical tests of these matters to future research.
8
Though we requested responses from only auditors at only the partner level, we used all responses received from
auditors, whether from partners, staff, etc. However, our results are robust (all p < 0.06) to excluding all staff, all
staff and seniors, and any participant that identifies themselves as a forensic specialist.

17
no fraud companies.9 Additionally, adjustments were made to minimize any one company being

evaluated much more frequently than another, and to ensure that any one company was not

frequently presented in any one specific order position (e.g., always first).

All participants are made aware that the expected rate of fraud in the companies they will

evaluate is 50%. The 10 company quarters are a subset of the 1,572 company-quarter earnings

conference calls studied in Hobson et al. [2012], which originally were broadcast during calendar

year 2007. We use two selection criteria. First, the fraudulent companies do not systematically

differ from non-fraud companies in terms of financial statement predictors of fraud, and second,

the companies are not generally well known, as described in more detail below.

Each of the 10 conference call excerpts are the first five minutes of CEO responses to

analyst questions in the quarterly earnings conference call, including any analyst questions that

elicit those first five minutes of CEO dialogue. We characterize excerpts as deceptive/fraudulent

and code them as DECEP_COMPANY equals 1 if the company’s quarterly financial statements

being discussed in the conference call audio were restated (i.e. the fiscal quarter end falls

between RES_BEGIN_DATE and RES_END_DATE on Audit Analytics via WRDS), the

restatement adversely impacted the financial statements (RES_ADVERSE = 1 on Audit

Analytics via WRDS) and any of the following “irregularity conditions” hold: the restatement

was deemed fraudulent (RES_FRAUD = 1 on Audit Analytics via WRDS), a regulatory

investigation followed the restatement (RES_SEC_INVESTIGATION = 1 on Audit Analytics

9
We use this distribution and multiple observations per participant for two reasons. First, we use a high-risk
population because this is our population of interest. A natural-world equivalent is a client acceptance decision for a
pool of risky potential clients. Second, we maximize value from the scarce resource of audit partner participants. A
more realistic rate of detected fraud requiring restatement would dramatically increase the number of participants
needed to draw useful inferences for fraud companies. We hold ex ante fraud risk constant by matching companies
on F-Score, as discussed below. Also, see section 4.5 for a discussion of our selected fraud rate.

18
via WRDS), or a class action lawsuit followed it (DAYS_TO_SECURITIES_CLASS_ACTION

> 0 on Audit Analytics via Audit Analytics online feed).

To meet the two criteria we have for selection of our 10 companies, we first calculate F-

Scores (Dechow et al. [2011]) for all 1,572 observations in Hobson et al. [2012]. We then sort

all observations by F-Score and for each irregularity quarter, we select the observation from the

same two-digit industry with the closest F-Score, without replacement. If no such match is

available within 10 observations in either direction, we use one digit SIC code, and if that fails,

we take the closest F-Score without matching on industry. We then eliminate fraudulent

observations and their related pair based on survey responses from accounting doctoral students

that indicate the fraud was familiar and likely to be known to a general audience. Among the

remaining companies, we chose the 10 with the widest absolute difference in vocal cognitive

dissonance as measured in Hobson et al. [2012] such that fraud companies have higher levels of

cognitive dissonance. This helps ensure the potential markers of fraud are sufficiently salient to

be considered by participants.

Of the five pairs that comprise the 10 observations, three are matched on two-digit SIC

code, one is matched on 1 digit SIC code, and one is not matched on SIC code. The average

(median) F-Score is 1.67 (1.90) for the five fraud companies and 1.94 (2.15) for the five non-

fraud companies (t8 = 0.43, p = 0.68).

3.3 SPEECH CORPUS PREPARATION

We manually transcribe each conference call excerpt rather than rely on commercial

transcription as used in prior literature (Larcker and Zakolyukina [2012]). Manual transcription

is necessary to ensure a clean manipulation of medium given commercial transcripts are purged

of speech hesitations (ah’s, um’s etc.) that potentially provide useful information. Additionally,

19
while Hobson et al. [2012] isolated only the voice of the CEO in the conference call, thereby

purging any important context from the CEO-analyst exchange, we include the analyst

question(s) to which the CEO is responding so as to provide appropriate context. Analyst

questions are read by a generic, computerized male or female voice. We use excerpts instead of

the full question and answer period given practical limits on participant time.

3.4 PROCEDURE & VARIABLES

We focus on two key variables. First, we measure auditor experience at two levels

(inexperienced audit student or very experienced audit professional). Second, between

participants, we manipulate whether auditors are prompted to consider management’s cognitive

dissonance (COGDIS). This manipulation has three parts. First, initial instructions for all

participants preceding each company evaluation stated “Note: Research indicates that certain

cues in what a CEO says and how s/he says it can help in the detection of deception.” Next, half

of our participants were made aware of cognitive dissonance and additionally received the

following:

One cue found to be useful in detecting deception in these CEO responses is cognitive
dissonance. Cognitive dissonance is the negative, uncomfortable emotion a person
feels when they are saying something that they know is not true. Those experiencing
cognitive dissonance feel uncomfortable, uneasy, and bothered.

Next, after answering our principal dependent measure, participants prompted about

management’s cognitive dissonance are asked to assess “how much cognitive dissonance the

CEO felt during this excerpt of the conference call.” Overall, then, this prompt defines cognitive

dissonance (including its effect on individuals), links cognitive dissonance to deception, and

encourages participants to observe cognitive dissonance in the conference call.

For exploratory purposes and robustness, we also manipulate conference call medium at

two levels. Some participants receive the conference call excerpt only as a transcript (AUDIO =

20
0) while others receive the transcript and accompanying audio (AUDIO = 1). We include the

transcript in both conditions to isolate the effect of audio and because transcripts often are

available long after audio feeds have been removed from company web sites.

Appendix A provides a timeline of the online experiment, which on average took about

1.75 hours.10 First, participants view a brief orientation video that instructs them to complete the

experiment in one sitting without accessing outside information. Next, participants complete a

guided tour and practice evaluating a fictional company. Written instructions and three narrated

videos aid in this process. For example, we tell participants that their task is to use the CEO

responses to analyst questions to (1) determine whether they think the results being discussed are

fraudulent and to (2) identify potential red flags in the audit of the company. We provide time

and encouragement for participants to practice the experimental tasks during this example.

The example and each of the four real companies participants evaluate consist of three

parts. Part 1 consists of background company information listing the company name, the quarter

being reported, a short business overview, and the four basic financial statements. Part 2 presents

the transcript or the transcript with the audio and an area to record red flags. Participants in the

transcript only (audio) condition are told that their task is to read the transcript (listen to the

audio and follow along on the transcript) and identify red flags. We tell them that a red flag, in

the context of the present experiment, exists any time they feel that the CEO's comments are

suspicious, give them pause, or require additional investigation. We tell them that they must read

the entire transcript in the transcript only condition and listen to the full audio in the audio only

condition.

10
IRB approval was provided for the experiment.

21
 In Part 3, we collect responses from participants. The first question posed to participants,

and our primary dependent measure is, “Next, provide an overall judgment of whether it is more

likely than not that fraud was being committed at this company during this quarter. That is, did

this company later restate this quarter’s financial results due to one or more of the following:

fraudulent financial results, a regulatory investigation, or a class action lawsuit?” Participants

respond, “Yes, fraud was likely being committed during this quarter” or “No, fraud was not

likely being committed during this quarter.” We then ask participants to state how confident they

are about this judgment.11 We present all four evaluated companies in this manner. After

evaluating the fourth company, participants complete a post-experiment questionnaire that asks

several questions about work experience and experience detecting deception.

4. Results

4.1 DESCRIPTIVE STATISTICS

4.1.1 General Descriptives Each of the 211 (31 very experienced and 180 inexperienced)

participants provided four judgments, one per conference call, yielding 844 responses. We

eliminate 17 observations because participants indicated they were familiar before participating

in the study with fraud at the company, leaving 827 total observations, 211 auditors.12 Table 1

provides descriptive statistics for the entire sample and for very experienced and inexperienced

auditors separately, as well as univariate tests of differences between very experienced and

inexperienced. Overall, average audit experience (AUD_EXP) is 3.48 years, with 121 (14.63%)

11
Additional questions ask the participant whether they thought the CEO was lying, what areas of the financial
statements appear problematic, and how familiar the participant was with the company before starting the
experiment.
12
We eliminated observations when the participant responded “Yes” to the following question: “This company may
or may not have had to restate this quarter’s financial results due to one or more of the following: fraudulent
financial results, a regulatory investigation, or a class action lawsuit. Before participating in this study, were you
aware of any financial improprieties for this company?” Seventeen responses (2.01% of all responses) were omitted:
14 from inexperienced auditors (1.94% of the inexperienced auditor responses) and three from experienced auditors
(2.42% of experienced responses).

22
of the responses from very experienced auditors (EXP). Average audit experience for very

experienced auditors is 23.62 years, which is statistically greater than the average of 0.02 years

for inexperienced auditing students (p < 0.01).13 The average length of the conference call

information provided, in audio time (AUDIO_TIME), was 7.24 minutes; the financial statement

based fraud score (FSCORE) was 1.82, and 51% of the conference calls participants reviewed

were deceptive (DECEP_COMPANY). There was no statistical difference in AUDIO_TIME

(7.26 vs. 7.24, p = 0.71), FSCORE (1.89 vs 1.81, p = 0.32), or DECEP_COMPANY (0.52 vs.

0.51, p = 0.79) between very experienced and inexperienced auditors.

4.1.2 Accuracy Rates We first empirically test whether auditors, overall, detect the

presence and absence of fraud from earnings conference calls at better than chance rates. This

test is motivated by prior meta-analysis that finds better-than-chance deception detection is rare

(Bond and DePaulo [2006]; [2008]; Vrij [2008]). We observe that the overall accuracy rate

(ACCURACY) of 55% (see Figure 1 and Table 1) is statistically greater than chance levels of

50% (p < 0.05), which is our population rate and the rate used in the vast majority of prior

deception research (Levine et al. 2014). This accuracy is in line with the average accuracy rate of

54% that meta-analysis reveals is obtained when subjects face a 50/50 population rate (Levine

[2010]). Next, as a preliminary test of H1, we examine whether very experienced auditors are

more accurate than inexperienced auditors in detecting the absence versus presence of fraud. We

find that experience improves overall accuracy as auditors detect fraud at better than chance rates

(67%, p < 0.01), while audit students do not (53%, p > 0.10). Additionally, the overall accuracy

rate is higher for very experienced than inexperienced auditors (ܺ ଶ (1) = 8.96, p < 0.01 one

tailed; see Figure 1 and Table 1). Experienced auditors’ accuracy rate (67%) is remarkably good,

13
A small fraction of inexperienced auditors had worked as audit interns, yielding a non-zero value of AUD_EXP.

23
given reported accuracy rates for deception detection by experts in prior literature (Bond and

DePaulo [2006]; [2008]).14

4.2 H1—EFFECT OF EXPERIENCE ON ACCURACY RATES FOR FRAUD & NO FRAUD

COMPANIES

H1 predicts a specific pattern of results that has two key parts, resulting in an ordinal

interaction between experience (EXP) and company type (fraud versus non-fraud,

DECEP_COMPANY). First, we predict a positive simple main effect for EXP given a non-fraud

company. Second, we predict a negative simple main effect for DECEP_COMPANY given very

experienced auditors, i.e., a lower accuracy rate for fraud versus non-fraud companies.

Univariate analysis supports these predictions, as indicated in Table 1. Participants

judged 45% of companies to be deceptive (DECEP_JUDG). However, experienced auditors

predict fraud only 40% (49/121) of the time, which is statistically less than the disclosed fraud

rate of 50% (p < 0.05, see Table 1). While directionally consistent with our theory, this is

somewhat surprising since we informed participants of the 50% fraud base rate.15 Inexperienced

auditors predict fraud 46% (326/706) of the time which differs significantly from the disclosed

50% rate (p < 0.01) but not from the experienced auditors’ judged deception rate (p > 0.10, see

Table 1).

14
In all of our analysis we estimate generalized linear models (GLIMMIX via SAS 9.4) with random effects by
subject to account for within subject correlation, use identity link functions for continuous outcome variables, and
use logit link function for dichotomous outcome variables.
15
During training participants are told that, “[The] four companies [you will evaluate] were taken from a larger set
of companies. In this larger set, 50% of the companies had to restate their earnings due to fraud. Specifically, in half
of the companies in this larger set, the quarterly and/or yearly financial results being discussed in the conference call
were later restated….Since you have a sample of only four companies, you will not know for sure how many of
these companies committed fraud. The most likely scenario is that you will evaluate two fraud companies and two
clean companies. However, you could have any combination of fraud and clean companies, including all fraud
companies or all clean companies. ” Also, see section 4.5 for a discussion of our selected fraud rate.

24
 Figure 2 and Panel A of Table 2 show that while experienced auditors are very accurate

for non-fraud companies (78%, specificity, true negative rate)16, they are considerably less

accurate for fraud companies (57%, sensitivity, true positive rate). In fact, their accuracy rate for

fraud companies is no better than chance (p > 0.10), and no better than that of inexperienced

auditors (p > 0.10). Put another way, the proportion of false positive judgments/Type I errors are

almost half as prevalent for experienced compared to inexperienced auditors (0.11 vs. 0.21, p <

0.01), yet false negatives/Type II errors are not statistically different between very experienced

and inexperienced auditors (p > 0.10).

We test H1 directly in a multivariate, 3-way, repeated measures logistic analysis, in

which ACCURACY is the dependent variable, EXP, DECEP_COMPANY, and COGDIS and

their interactions are the independent variables of interest, and AUDIO is a covariate. Figure 2

and Panel A of Table 2 present mean rates of accuracy, collapsed across COGDIS, and are

consistent with H1. Panels B and C of Table 2 present analysis results. We directly test the

pattern of results specified in H1 in a linear contrast. Specifically, collapsing across COGDIS,

we test that accuracy rates will be highest for very experienced auditors assessing non-fraud

companies, significantly lower for very experienced auditors assessing fraud companies and

similarly low for inexperienced auditors assessing both fraud and non-fraud companies., using

weights of 3 (Experienced/Non-Fraud Company), -1 (Experienced/Fraud Company), -1

(Inexperienced/Non-Fraud Company), and -1 (Inexperienced/Fraud Company). Panel C displays

the results of this contrast, which is significant, confirming H1 (ܺ ଶ (1) = 10.84, p < 0.01 one

tailed).

16
This is very accurate in the sense of O’Sullivan and Ekman’s “wizard” or genius level of deception detection that
they expect only one to two percent of individuals in expert populations to achieve (O'Sullivan and Ekman [2004],
271).

25
 This table also tests the two necessary parts of the prediction in H1. First, Panel C shows

a significant simple main effect for EXP given Non-Fraud Company, indicating that very

experienced auditors have higher accuracy rates than those of inexperienced auditors when

evaluating companies that did not commit fraud (t = 3.04, p < 0.01 one tailed). Second, Panel C

shows that the simple effect of DECEP_COMPANY given Very Experienced is significant (t =

2.34, p = 0.01, one tailed), indicating that very experienced auditors’ accuracy significantly

decreases when they evaluate fraud companies relative to non-fraud companies. Additional,

untabulated examination reveals that, collapsing across prompt conditions, the accuracy rate of

very experienced auditors evaluating fraud companies is not statistically different from that of

inexperienced auditors when they are evaluating fraud companies (t206 = 1.23, p = 0.22) nor that

of inexperienced auditors evaluating non-fraud companies (t206 = 0.13, p = 0.90). Finally,

inexperienced auditors’ accuracy rates also decrease significantly when evaluating fraud

companies relative to non-fraud companies (t206 = 2.02, p = 0.04). In summary, very experienced

auditors’ superior accuracy, relative to that of inexperienced auditors, predominantly stems from

their ability to correctly classify clean companies, thus avoiding false positive (Type I) errors.

When evaluating fraud companies, highly experienced auditors do not outperform audit students.

4.3 H2—COGNITIVE DISSONANCE PROMPT: DOES IT HELP?

In H2, we propose that a remedy that nudges auditors to reinterpret cues that are

ambiguously suggestive of fraud will help prevent false negatives. We predict that accuracy rates

for fraud companies will be highest for very experienced auditors that are prompted about

management cognitive dissonance, significantly lower for very experienced auditors without a

cognitive dissonance prompt and similarly low for inexperienced auditors, regardless of whether

or not they receive a cognitive dissonance prompt.

26
 Univariate analysis supports these predictions. Very experienced participants judged just

30% of companies to be deceptive (DECEP_JUDG) when they did not receive a cognitive

dissonance prompt, which is significantly less than 50% (p = < 0.01, untabulated). However,

when experienced auditors received the cognitive dissonance prompt, they predicted fraud 50%

of the time, which is significantly higher than the rate when no prompt is given (p = 0.02). Figure

3 and Panel A of Table 2 present means that show a significant improvement in accuracy for

fraud companies when very experienced auditors are (70%, which is greater than 50% or chance,

p < 0.05) versus are not (43% which is not less than 50% or chance, p > 0.10) given a cognitive

dissonance prompt.

We test H2 directly using the same model discussed above. Figure 3 and Panel A of

Table 2 present mean rates of accuracy that are consistent with H2. Panels B and C of Table 2

present analysis results. We directly test the pattern of results specified in H2 using two linear

contrasts. First, we fit linear contrast weights to each mean in the model, such that, very

experienced auditors receive a 2.5 when evaluating a non-fraud company whether or not a

prompt is provided and when evaluating a fraud company when a cognitive dissonance prompt is

given. All five other means are weighted with -1.50. This contrast is significant, confirming H2

(ܺ ଶ (1) = 17.46, p < 0.01 one tailed). Next, we use a linear contrast to test only judgments of

fraud companies. Specifically, very experienced auditors are given a weight of 3 when they

receive a cognitive dissonance prompt (and judge a fraud company). All other judgments of

fraud companies are given a weight of -1. This contrast is significant, confirming H2 (ܺ ଶ (1) =

5.70, p = 0.01 one tailed). Additionally, we find that the accuracy rate of 0.70 for very

experienced auditors given a cognitive dissonance prompt and evaluating a fraud company, is

27
greater than each of the other three accuracy rates evaluating a fraudulent company (all p < 0.03,

one tailed).

[Figure 3 about here]

4.4 RESEARCH QUESTION —AUDIO VERSUS TRANSCRIPT

Collapsing across COGDIS, we examine our research question, RQ1, through an

untabulated, repeated measures logistic analysis, in which ACCURACY is the dependent

variable and EXP and AUDIO are the independent variables. The pattern of means in this

analysis—0.49 (Inexperienced / Text), 0.55 (Inexperienced / Audio + Text), 0.69 (Experienced /

Text), and 0.65 (Experienced / Audio + Text)—appears consistent with an ordinal interaction

such that accuracy rates are highest for very experienced auditors when audio is and is not

available, equal or lower for inexperienced auditors when audio is available, and lowest for

inexperienced auditors when only text is available. Indeed, we find a significant simple main

effect for EXP given text (t = 3.11, p < 0.01 one tailed), and a significant simple main effect of

AUDIO given Inexperienced (t = 2.17, p = 0.02 one tailed). Further, the simple effect of AUDIO

given Very Experienced is not significant (t = 0.48, p = 0.63). Additional analysis reveals that

when audio is provided, accuracy rates of very experienced auditors (65%) and inexperienced

auditors (55%) are indistinguishable at conventional levels (t = 1.14, p = 0.26). In summary,

providing audio to inexperienced auditors improves their accuracy more so than it improves that

of experienced auditors. Interestingly, having audio improves inexperienced auditors’ accuracy

rates such that they are statistically indistinguishable from those of very experienced auditors.

4.5 ROBUSTNESS TESTS

4.5.1 Time Participants spent an average (median) of 19 (11) minutes consuming each

conference call, with inexperienced participants (16 / 11) using significantly less time per call

28
than that of experienced participants (39 / 15, t = 2.21, p = 0.03). To assess whether time spent

plays a significant role in accuracy, we rerun our analyses for H1 and H2 with time as a covariate

(control variable, untabulated). We estimate the natural logarithm of time spent as the variable is

right skewed. The coefficient on time spent is positive but insignificant (p = 0.39), and our

inferences remained unchanged.

4.5.2 Familiarity with the Company We ask the following question to assess

participants’ familiarity with each company they are evaluating: “What general level of

familiarity did you have with this company before doing this study?”, where 0=Very Low and

10=Very High. Overall, familiarity is quite low, at 0.05 and does not significantly differ between

inexperienced and very experienced auditors (p = 0.28). Our inferences are not changed when we

add familiarity as a covariate (control variable) in our analysis.

4.5.3 50-50 Fraud Rates One potential alternative explanation of our results is that

auditors were not accustomed to our experiment’s 50% rate of fraud. Even though participants

were told that the most likely scenario was receiving two fraud companies and two clean

companies, very experienced auditors’ may have difficulty overcoming a lifetime of

experiencing low rates of fraud while inexperienced auditors, on the other hand, would be less

subject to this issue given that fraud is frequently discussed in audit lectures and audit text and

case books. However, this is an unlikely explanation of our findings for at least two reasons.

First, very experienced auditors did not select fraud less often than did inexperienced

participants, as discussed above. Second, our results that follow in section 4.6 examining red flag

accuracy show that very experienced auditors, particularly those that are given a cognitive

dissonance prompt, are more accurate at identifying flags for fraud. It is difficult to see how our

selected fraud rates may have caused these effects. However, as a robustness test we rerun

29
analysis without the very experienced participants who said zero frauds, since these participants

may have been calibrated incorrectly. Results are at least marginally consistent with our

hypotheses (all p < 0.10).

4.6 SUPPLEMENTAL PROCESS ANALYSIS

We examine process data related to auditor detection of red flags. As mentioned above,

for each of the four companies they evaluated, we ask auditors to read (listen to) the transcript

(audio) and write down the line number from the transcript (audio timestamp) of each red flag

and as much detail about the red flag as possible. We examine this process data in two ways—(1)

coding auditors’ notations about red flags, and (2) red flag detection accuracy.

4.6.1 Auditors’ Red Flagged Issues Two separate raters with knowledge of accounting

but with no knowledge of the hypotheses being tested nor the experience level of the auditor

providing the response, independently coded the open-ended notes auditors made when

identifying red flags. The coders counted the number of issues for 14 categories. The two coders’

counts in each category were highly, positively correlated (all p < 0.01).17 We average their

responses together and then arrange the 14 categories into three broad groups: financial

statement items (FIN), speech (SPEECH), and suspect knowledge (SKNOW). The FIN group

contains red flagged issues pertaining to financial statements, financial results, company strategy,

and topics regarding the appropriate application of judgment when applying accounting

standards given the operating environment. The SPEECH group captures speech hesitations,

filled pauses (ah’s and um’s), language use that avoids answering questions or deflects blame,

and speech that suggests the speaker was nervous. The SKNOW group contains red flagged

17
All but six of the 14 categories have Cronbach's alphas greater than 0.70. Inferences are identical if these six
categories are omitted with the exception that RF_ISSUES significantly increase for fraud companies relative to
non-fraud companies for inexperienced auditors.

30
issues identifying instances where the executive appeared to lack knowledge of the subject

matter by giving answers that seem incomplete, inaccurate, internally inconsistent, or lacking

appropriate level of confidence. The sum of FIN, SPEECH and SKNOW represents the total

number of red flagged issues identified (RF_ISSUES).

We first observe that while the number of characters provided in the text box

(ISSUE_CHARS) from which the red flagged issues were coded are almost identical across

experience level (242 versus 243 for inexperienced versus very experienced auditors), very

experienced auditors provide 3.07 issues per company, on average, which is statistically larger

than the 2.32 inexperienced auditors provide (t = 2.72, p = 0.01, see Table 3). Thus, very

experienced auditors are more efficient at detecting potential problems in conference calls. Next,

when we compare the number of red flagged issues found based upon whether a company did or

did not commit fraud, we find that very experienced auditors find significantly more issues

overall and in each of the three categories (all p < 0.10) for fraud versus non-fraud companies.

However, inexperienced auditors do not find more issues overall for fraud versus non-fraud

companies (p = 0.19), and only find more issues for fraud companies in the FIN category. This

appears to indicate that experienced auditors are more attuned to accurate red flags for fraud.

Additionally, experienced auditors are better able to use “softer” issues (e.g., the issues

categorized into the SPEECH and SKNOW categories).

[Table 3 about here]

4.6.2 Red Flag Detection Accuracy Next, we conduct an alternative analysis to more

closely examine accuracy in red flag detection. This enables us to examine two key aspects of

our results. First, does greater audit experience lead to more accurate red flag detection of

fraudulent sentences? That is, does the superior accuracy of very experienced auditors relative to

31
inexperienced auditors noted above result because experienced auditors are better able to detect

actual fraudulent statements? Second, does the cognitive dissonance prompt help experienced

auditors better identify fraudulent statements?

Examining only our five fraud companies, we construct a measure of red flag detection

accuracy (RF_ACCURACY) in three steps. First, for each CEO sentence, participants are given

a 1 (0) if they do (do not) identify a CEO’s sentence as a red flag. Second, each CEO sentence is

classified as fraudulent or not fraudulent. This subjective classification is done by a coauthor and

several teams of research assistants with no knowledge of the red flag results, using hand

collected data about each of the five frauds.18 Finally, RF_ACCURACY is 1 if the auditor

correctly flags a fraudulent sentence or does not flag a non-fraudulent sentence, and 0 otherwise.

To answer the above questions , we examine RF_ACCURACY for only the fraudulent

CEO sentences of our five fraudulent companies, using EXP, COGDIS, and their interaction.

Thus, we are examining whether audit experience and the cognitive dissonance prompt help

auditors accurately identify fraudulent CEO statements; we are testing auditors’ sensitivity to

true red flags. Using a multivariate, repeated measures logistic analysis, we find that very

experienced auditors more accurately identify fraudulent statements (0.27) than do inexperienced

auditors (0.17, ܺ ଶ (1) = 12.30, p < 0.01, two tailed, Figure 4 and Table 4).

Next, we find that the cognitive dissonance prompt marginally significantly increases

very experienced auditors’ red flag accuracy (0.21 to 0.32, t = 1.78, p = 0.08, two tailed), but not

the accuracy of inexperienced auditors (0.17 to 0.17, t = 0.43, p = 0.67, two tailed). A linear

contrast testing the ordinal interaction in which the accuracy of very experienced auditors

18
Just over 100 second-year audit students were divided into small groups and asked to discover and examine the
background information behind the fraud for each fraud company, and then to categorize each CEO sentence based
upon how directly the fraud was discussed. Six groups were assigned to each fraud company in our sample.

32
receiving the cognitive dissonance prompt is given a weight of 3 while the other three means are

given a weight of -1 is significant (p < 0.01, two tailed).

5. CONCLUSION AND IMPLICATIONS

We use an experiment to examine how extensive audit experience and a prompt to attend

to management’s cognitive dissonance individually and jointly influence auditors’ ability to

detect deception from question and answer (Q&A) portions of earnings conference calls. We also

explore which medium of the call, audio or transcript, is most helpful to the auditor in that effort.

We ask very experienced auditors (average experience of 24 years) and audit students to evaluate

the conference calls of four companies, and to predict whether the companies were committing

fraud. We identify fraud by whether the financial results released at the time of the conference

call are eventually restated due to an irregularity.

We find both good and bad news. Examining overall accuracy, we find that very

experienced auditors predict accurately 67% of the time, and significantly outperform both

inexperienced auditors and chance. This is good news as meta-analyses document that experts

rarely outperform chance in detecting deception (e.g., Bond and DePaulo [2006]; [2008]). This

accuracy rate increases even more, to 78%, when experienced auditors judge non-fraud

companies. This rate is approaches O’Sullivan and Ekman’s “wizard” or genius level of

deception detection that they expect only one to two percent of individuals in expert populations

to achieve (O'Sullivan and Ekman [2004], 271).

The bad news, however, is that this accuracy stems principally from fewer false positives.

That is, very experienced auditors outperform chance and inexperienced auditors only for

companies that do not commit fraud, contrary to the likely preference of financial statement

users. Nevertheless, this pattern of findings is consistent with psychology-based theory. We

33
theorize that auditors’ experiential learning with respect to fraud does not primarily make it

adaptive for them to avoid or minimize false negatives but rather to avoid or minimize false

positives (Friedrich [1993]).

We also develop a psychology-based remedy to help experienced auditors overcome their

experientially learned avoidance of false positives: a prompt to attend to management’s cognitive

dissonance. With this prompt, experienced auditors are significantly more accurate in detecting

fraud (from 43% to 70%) and pick fraud more often (31% to 50%). In process analysis, we find

that very experienced auditors more accurately identify fraudulent statements in conference calls

as red flags. Further, our cognitive dissonance prompt increases this accuracy for very

experienced auditors. Finally, providing the audio of conference calls improves the accuracy

rates of inexperienced auditors to match the rates of unprompted, very experienced auditors.

Our research is subject to limitations. First, our results generalize most readily to

situations in which the auditor is unfamiliar with the client. Future research could examine the

robustness of our effects to situations in which the auditor assesses their own client. We expect

familiarity would exacerbate experienced auditors’ avoidance of false positives and so present a

stronger challenge for our cognitive dissonance prompt remedy (Bowlin et al. [2015]). Second,

we do not examine interactions between auditor judgments and other systems audit firms may

have in place for detecting fraud. The extent to which auditors view such systems as threats to

their own expertise, in turn influencing their detection accuracy, remains unexplored (Elkins et

al. [2013]).

This research contributes to both audit practice and research. From a practice perspective,

we show that conference calls are a potentially useful source for auditors to find cues to potential

fraud. Conference calls are plentiful and low-cost avenues of information about a current or

34
potential client. For example, while the prospect of growth in audit revenues is high in BRIC

(Brazil, Russia, India, and China) and other developing countries, auditors have expressed to us

hesitancy accepting prospective clients in some parts of the world with higher corruption indices

(etc.) due to unknown and potentially elevated levels of fraud risk. To the degree that our theory

and experimental findings helps auditors to better discriminate which clients to accept/retain

versus reject because of heightened risk of fraud, it will be useful to the profession and to

investors.

Further, we begin to address the question of who at the audit firm should consume the

conference call. While highly experienced audit personnel are seemingly ideal for this task,

inexperienced auditors provided with the audio version of the call can predict fraud from

conference calls at better than chance rates.

From a research perspective, we demonstrate a setting in which individuals successfully

detect deception from dialogue. In particular, we show a setting in which very experienced

auditors (experts) perform well in avoiding false positives. We also highlight the need for

additional research examining experienced auditors’ reluctance to predict fraud or “truth bias.”

Experienced auditors’ elevated Type II error rates should be of concern to both academics,

regulators, investors, and other financial statement users. Future research could examine the

particular dispositional and institutional factors that are most influential in this bias. Finally, we

identify a potentially useful remedy—prompting auditors to consider the cognitive dissonance of

the speaker—to aid experienced auditors in mitigating these Type II errors.

35
Appendix A: Time Line

Video guided tour through

the mechanics of the
Evaluate four companies, in
experiment, including Exit survey
turn
practicing with a hypothetical
company.

For each company,

participants receive
the following:

Part 1: background company

Part 3: response variables,
information, the quarter being
Part 2: transcript (transcript and including, first, our main
reported, a short business
audio) if Audio = 0 (Audio = 1) dependent measure: "...was
overview, balance sheet, income
and an area to enter red flags fraud being committed during
statement, and statement of cash
this quarter?"
flows

36
Appendix B: Variable Definitions

ACCURACY Indicator variable that identifies whether the participant correctly

classifies the conference call: equals 1 if (DECEP_COMPANY =
1 and DECEP_JUDG = 1 or DECEP_COMPANY = 0 and
DECEP_JUDG = 0) and 0 otherwise.
AUD_EXP Number of years of experience as an assurance professional
reported by the participant.
AUDIO Indicator variable that equals 1 if the conference call excerpt was
provided both as a transcript and as audio, 0 if as a transcript only.
AUDIO_TIME Duration in minutes of the audio version of the conference call
assigned to the participant.
COGDIS Cognitive dissonance prompt is a between subjects manipulated
variable that has two parts. First, participants receiving this
manipulation are given additional initial instructions that said the
following: “One cue found to be useful in detecting deception in
these CEO responses is cognitive dissonance. Cognitive dissonance
is the negative, uncomfortable emotion a person feels when they
are saying something that they know is not true. Those
experiencing cognitive dissonance feel uncomfortable, uneasy, and
bothered.” Second, after answering our principal dependent
measure, these participants are asked “how much cognitive
dissonance the CEO felt during this excerpt of the conference call.”
COGDIS = 1 when both of these components are present and 0
otherwise.
DECEP_COMPANY Indicator variable that equals 1 if conference call pertains to a
quarter that was eventually restated due to an irregularity (Fraud
Companies), and 0 otherwise (Non-Fraud Companies).
DECEP_JUDG Indicator variable that equals 1 if the participant judges the
conference call to be related to a company quarter that will
eventually be restated due to an irregularity, and 0 otherwise.
EXP Indicator variable that equals 1 if the response is from a very
experienced audit professional, and 0 if the response is from an
inexperienced accounting major enrolled in an auditing course.
FSCORE Fraud score of the company for the fiscal quarter being discussed
on the earnings conference call.
FIN For each of the four companies they evaluated, we ask auditors to
read (listen to) the transcript and write down the time stamp (line
number from the transcript) of each red flag and as much detail
about the red flag as possible. Two separate raters with knowledge
of accounting but with no knowledge of the hypotheses being
tested nor the experience level of the auditor providing the
response, independently coded the open-ended notes auditors make
when noting red flags. The coders counted the number of issues
and sorted them into 14 categories. We average their responses
together and then arrange the 14 categories into three broad groups:

37
 financial statement items (FIN), speech (SPEECH), and suspect
knowledge (SKNOW). The FIN group contains red flagged issues
pertaining to financial statements, financial results, company
strategy, and topics regarding the appropriate application of
judgment when applying accounting standards given the operating
environment.
SPEECH The SPEECH group captures speech hesitations, filled pauses (ah’s
and um’s), language use that avoids answering questions or
deflects blame, and speech that suggests the speaker was nervous.
SKNOW The SKNOW group contains red flagged issues identifying
instances where the executive appeared to lack knowledge of the
subject matter by giving answers that seem incomplete, inaccurate,
internally inconsistent, or lacking appropriate level of confidence.
RF_ISSUES The sum of FIN, SPEECH, and SKNOW represents the total
number of red flagged issues identified (RF_ISSUES).
ISSUE_CHARS ISSUE_CHARS is the number of characters provided in the text
box from which the red flagged issues were coded.
RF_ACCURACY Red flag detection accuracy, analyzed for only the five fraud
companies, is constructed in three steps. First, each participant is
given a 1 (0) if they identify (do not identify) a CEO’s sentence as
a red flag. This is done for each CEO sentence. Second, each CEO
sentence is classified as fraudulent or not fraudulent. This
subjective classification is done by a coauthor and several teams of
research assistants with no knowledge of the red flag results, using
hand collected data about each of the five frauds. Finally,
RF_ACCURACY receives a 1 if the auditor correctly flags a
fraudulent sentence or does not flag a non-fraudulent sentence, and
a 0 otherwise.

38
References

ALM, E. C. O. Affect in Text and Speech. Doctoral Dissertation. Universeity of Illinois at

Urbana-Champaign, 2008. http://cogcomp.cs.illinois.edu/papers/Alm%20thesis(1).pdf
AMBADY, N. and M. WEISBACH. “Nonverbal behavior” in Handbook of Social Psychology
Volume 1 (5th Edition), Edited by S. T. Fiske, D. T. Gilbert and G. Lindsey. John Wiley
& Sons, Inc. Hobeken, New Jersey, 2010.
ASARE, S. K.; G. M. TROMPETER and A. M. WRIGHT. “The Effect of Accountability and
Time Budgets on Auditors' Testing Strategies.” Contemporary Accounting Research 17
(2000): 539-560.
AUDIT ANALYTICS STAFF 'Analysis of Audit Fees by Industry Sector,' in Book Analysis of
Audit Fees by Industry Sector, edited by Editor. City, 2014.
BERNARDI, R. A. “Fraud Detection: The Effect of Client Integrity and Competence and
Auditor Cognitive Style.” Auditing 13 (1994): 68-84.
BLAU, B. M.; J. R. DELISLE and S. M. PRICE. “Do sophisticated investors interpret earnings
conference call tone differently than investors at large? Evidence from short sales.”
Journal of Corporate Finance 31 (2015): 203-219.
BOND, C. F. and B. M. DEPAULO. “Accuracy of deception judgments.” Personality and Social
Psychology Review 10 (2006): 214-234.
BOND, C. F. and B. M. DEPAULO. “Individual differences in judging deception: Accuracy and
bias.” Psychological Bulletin 134 (2008): 477-492.
BONNER, S. E. and B. L. LEWIS. “DETERMINANTS OF AUDITOR EXPERTISE.” Journal
of Accounting Research 28 (1990): 1-20.
BOWLIN, K. O.; J. L. HOBSON and M. D. PIERCEY. “The Effects of Auditor Rotation,
Professional Skepticism, and Interactions with Managers on Audit Quality.” The
Accounting Review 90 (2015).
BRAUN, R. L. “The effect of time pressure on auditor attention to qualitative aspects of
misstatements indicative of potential fraudulent financial reporting.” Accounting,
Organizations and Society 25 (2000): 243-259.
BRAZEL, J. F.; T. D. CARPENTER and J. G. JENKINS. “Auditors' Use of Brainstorming in the
Consideration of Fraud: Reports from the Field.” Accounting Review 85 (2010): 1273-
1301.
BURGOON, J. K.; J. P. BLAIR and R. E. STROM. “Cognitive Biases and Nonverbal Cue
Availability in Detecting Deception.” Human Communication Research 34 (2008): 572-
599.
CARCELLO, J. V. and A. L. NAGY. 'Auditor industry specialization and fraudulent financial
reporting.' Symposium on Auditing Problems, 2002.
CENTER FOR AUDIT QUALITY 'Deterring and Detecting Financial Reporting Fraud,' in Book
Deterring and Detecting Financial Reporting Fraud, edited by Editor. City, 2010.
CHAN, E. W. and D. V. MOSER. “Client Management’s Perceptions of Audit Offshoring and
Their Willingness to Trade Off Audit Quality.” Working Paper, U. o. Pittsburg, (2014).
CHRISTENSEN, B. E.; S. M. GLOVER, T. C. OMER, and M. K. SHELLEY. "Understanding
Audit Quality: Insights from Audit Partners and Investors" forthcoming in Contemporary
Accounting Research. (2015).
DECHOW, P. M.; W. GE, C.R. LARSON, and R. G. SLOAN. "Predicting Material Accounting
Misstatements" Contemporary Accounting Research 28 (2011): 17-82.

39
DEPAULO, B. M.; J. J. LINDSAY; B. E. MALONE; L. MUHLENBRUCK; K. CHARLTON
and H. COOPER. “Cues to deception.” Psychological Bulletin 129 (2003): 74-118.
DIPIAZZA, S. A.; D. MCDONNELL; W. G. PARRETT; M. D. RAKE; F. SAMYN and J. S.
TRULEY 'Global Capital Markets and the Global Economy: A Vision From the CEOs of
the International Audit Networks,' in Book Global Capital Markets and the Global
Economy: A Vision From the CEOs of the International Audit Networks, edited by Editor.
City, 2006.
DITTO, P. H. and D. F. LOPEZ. “Motivated Skepticism: Use of Differential Decision Criteria
for Preferred and Nonpreferred Conclusions.” Journal of Personality & Social
Psychology 63 (1992): 568-584.
DOTY, J. R. 'Keynote Address: Baruch College 13th Annual Financial Reporting Conference,' in
Book Keynote Address: Baruch College 13th Annual Financial Reporting Conference,
edited by Editor. City, 2014.
DYCK, I. J. A.; A. MORSE and L. ZINGALES. “Who Blows the Whistle on Corporate Fraud?”
Journal of Finance 65 (2010): 2213-2253.
DYCK, I. J. A.; A. MORSE and L. ZINGALES. “How pervasive is corporate fraud.” Working
Paper (2013). http://faculty.haas.berkeley.edu/morse/research/papers/DyckMorseZingales2013Apr9.pdf.
ELFENBEIN, H. A. and N. AMBADY. "On the universality and cultural specificity of emotion
recognition: A meta-analysis." Psychological Bulletin 128 (2002): 203-235.
ERICSSON, K. A. “Deliberate Practice and the Acquisition and Maintenance of Expert
Performance in Medicine and Related Domains.” Academic Medicine 79 (2004): S70-
S81.
ELKINS, A.C; N.E. DUNBAR, B. ADAMS and J.F. NUNAMAKER. "Are Users Threatened by
Creditibliy Assessment Systems?" Journal of Management Information Systems 29
(2013): 249-262.
ERICSSON, K. A.; R. T. KRAMPE and C. TESCH-RÖMER. “The role of deliberate practice in
the acquisition of expert performance.” Psychological Review 100 (1993): 363-406.
FRIEDRICH, J. “Primary error detection and minimization (PEDMIN) strategies in social
cognition: A reinterpretation of confirmation bias phenomena.” Psychological Review
100 (1993): 298-319.
FRIEDRICH, J.; G. LUCAS and E. HODELL. “Proportional reasoning, framing effects, and
affirmative action: Is six of one really half a dozen of another in university admissions?”
Organizational Behavior and Human Decision Processes 98 (2005): 195-215.
HARTSHORNE, J. K. and L. T. GERMINE. “When Does Cognitive Functioning Peak? The
Asynchronous Rise and Fall of Different Cognitive Abilities Across the Life Span.”
Psychological Science 26 (2015): 433-443.
HELLER, M. 'Accounting-related Securities Class Actions Climb 47%,' CFO.com, 2015.
HOBSON, J.; W. J. MAYEW and M. VENKATACHALAM. “Analyzing Speech to Detect
Financial Misreporting.” Journal of Accounting Research 50 (2012): 349-392.
HOGAN, C. E. Z. REZAEE, R.A. RILEY JR., and U.K. VELURY. "Financial statement fraud:
Insights from the academic literature." Auditing: A Journal of Practice and Theory 27
(2008): 231-252.

HOLLANDER, S.; M. PRONK and E. ROELOFSEN. "Does Silence Speak? An Empirical

Analysis of Disclosure Choices During Conference Calls." Journal of Accounting
Research 48 (2010): 531-563.

40
HOUSTON, R. W. “The Effects of Fee Pressure and Client Risk on Audit Seniors' Time Budget
Decisions.” AUDITING: A Journal of Practice & Theory 18 (1999): 70-86.
JAMAL, K.; P. E. JOHNSON and R. G. BERRYMAN. “Detecting Framing Effects in Financial
Statements*.” Contemporary Accounting Research 12 (1995): 85-105.
JOHNSON, P. E.; S. GRAZIOLI; K. JAMAL and R. GLEN BERRYMAN. “Detecting
deception: adversarial problem solving in a low base-rate world.” Cognitive Science 25
(2001): 355-392.
KADOUS, K.; S. J. KENNEDY and M. E. PEECHER. “The effect of quality assessment and
directional goal commitment on auditors' acceptance of client-preferred accounting
methods.” The Accounting Review 78 (2003): 759-778.
KELLEY, T. and L. MARGHEIM. “Survey On The Differential Effects Of Time Deadline
Pressure Versus Time Budget Pressure.” Journal of Applied Business Research 15
(1999).
KERLINGER, F. N. and H. B. LEE. Foundations of behavioral research. New York: Harcourt
College Publishers, 2000.
KNAPP, C. A. and M. C. KNAPP. “The effects of experience and explicit fraud risk assessment
in detecting fraud with analytical procedures.” Accounting, Organizations and Society 26
(2001): 25-37.
KUNDA, Z. “The case for motivated reasoning.” Psychological Bulletin 108 (1990): 480-498.
LARCKER, D. F. and A. A. ZAKOLYUKINA. “Detecting Deceptive Discussions in Conference
Calls.” Journal of Accounting Research 50 (2012): 495-540.
LEE, C.-C. and R. B. WELKER. “The Effect of Audit Inquiries on the Ability to Detect
Financial Misrepresentations.” Behavioral Research in Accounting 19 (2007): 161.
LEE, C.-C. and R. B. WELKER. “Identification of Perceived Interviewee Behaviors that
Influence Auditors' Assessment of Deception.” International Journal of Auditing 12
(2008): 205-220.
LEVINE, T.R. "A few transparent liars: Explaining 54% accuracy in deception detection
experiments." Communication Yearbook 34 (2010): 40-61.
LEVINE, T.R.; D.D. CLARE, T. GREEN, K.B. SEROTA and H.S. PARK. "The effects of
Truth-Lie Base Rate in Interactive Deception Detection Accuracy." Human
Communication Research 40 (2014): 350-372.
MATSUMOTO, D.; M. PRONK and E. ROELOFSEN. “What Makes Conference Calls Useful?
The Information Content of Managers' Presentations and Analysts' Discussion Sessions.”
The Accounting Review 86 (2011): 1383-1414.
MAYEW, W. J. and M. VENKATACHALAM. “The Power of Voice: Managerial Affective
States and Future Firm Performance.” The Journal of Finance 67 (2012): 1-43.
NELSON, M. W. “A Model and Literature Review of Professional Skepticism in Auditing.”
Auditing-a Journal of Practice & Theory 28 (2009): 1-34.
NUNAMAKER, J. F., JR.; J. K. BURGOON; N. W. TWYMAN; J. G. PROUDFOOT; R.
SCHUETZLER and J. S. GIBONEY. 'Establishing a foundation for automated human
credibility screening.' Intelligence and Security Informatics (ISI), 2012 IEEE
International Conference, 2012.
O'SULLIVAN, M. and P. EKMAN 'The wizards of deception detection,' in Deception detection
in forensic contexts, edited by P. A. Granhag and C. A. Stromwall. Cambridge:
Cambridge University Press, 2004: 269-286.

41
PEECHER, M. E.; I. SOLOMON and K. T. TROTMAN. “An accountability framework for
financial statement auditors and related research questions.” Accounting, Organizations
and Society 38 (2013): 596-620.
PRICE, S. M.; J. S. DORAN; D. R. PETERSON and B. A. BLISS. “Earnings conference calls
and stock returns: The incremental informativeness of textual tone.” Journal of Banking
& Finance 36 (2012): 992-1011.
PUBLIC COMPANY ACCOUNTING OVERSIGHT BOARD. Observations on auditors'
implementation of PCAOB standards relating to auditors' responsibilities with respect to
fraud. 2007.
PUBLIC COMPANY ACCOUNTING OVERSIGHT BOARD. Identifying and Assessing Risks
of Material Misstatement. Auditing Standard No. 12. Washington, D.C.: PCAOB, 2010.
PUBLIC COMPANY ACCOUNTING OVERSIGHT BOARD. PCAOB Announces Standing
Advisory Group Meeting on Nov. 13-14, 2013 and SAG Meeting Dates for 2014.
Washington, DC, 2013a.
PUBLIC COMPANY ACCOUNTING OVERSIGHT BOARD. PCAOB Reproposes Auditing
Standard, Related Parties, and Related Amendments, Including Amendments Regarding
Significant Unusual Transactions. Washington, DC, 2013b.
SHAUB, M. K. and J. E. LAWRENCE. “Differences in auditors’ professional skepticism across
career levels in the firm.” Advances in Accounting Behavioral Research 2 (1999): 61-83.
SWETS, J. A.; R. M. DAWES and J. MONAHAN. “Psychological Science Can Improve
Diagnostic Decisions.” Psychological Science in the Public Interest 1 (2000): 1-26.
VRIJ, A. Detecting Lies and Deceit: Pitfalls and Opportunities. West Sussex: Wiley-
Interscience, 2008.
VRIJ, A.; S. MANN; E. ROBBINS and M. ROBINSON. “Police officers ability to detect
deception in high stakes situations and in repeated lie detection tests.” Applied Cognitive
Psychology 20 (2006): 741-755.
WILLETT, C. and M. PAGE. “A SURVEY OF TIME BUDGET PRESSURE AND
IRREGULAR AUDITING PRACTICES AMONG NEWLY QUALIFIED UK
CHARTERED ACCOUNTANTS.” The British Accounting Review 28 (1996): 101-120.
ZHOU, L.; J. BURGOON; J. NUNAMAKER and D. TWITCHELL. “Automating Linguistics-
Based Cues for Detecting Deception in Text-Based Asynchronous Computer-Mediated
Communications.” Group Decision and Negotiation 13 (2004): 81-106.
ZUCKERMAN, M.; B. M. DEPAULO and R. ROSENTHAL 'Verbal and nonverbal
communication of deception,' in Advances in experimental social psychology, edited by
L. Berkowitz. New York: Academic Press, 1981: 1-59.

42
 FIGURE 1
Descriptive Statistics

Overall Accuracy Rates

100.00%
90.00%
Accuracy Rate (ACCURACY)
80.00%
70.00% 66.94%

60.00% 52.69% 54.78%

50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
Very Experienced Inexperienced (EXP=0) All Responses N=827
(EXP=1) N=121 N=706

EXP is an indicator variable that equals 1 if the response was from a very experienced audit professional (24 years
in audit on average), and 0 if the response was from an inexperienced accounting major enrolled in an auditing
course. Fraud and non-fraud companies are identified using DECEP_COMPANY, which is an indicator variable
that equals 1 if conference call pertains to a quarter that was eventually restated due to an irregularity (Fraud
Company), and 0 otherwise (Non-Fraud Company). DECEP_JUDG is an indicator variable that equals 1 if the
participant judges the conference call to be related to a company quarter that will eventually be restated due to an
irregularity, and 0 otherwise. The dependent measure is ACCURACY, which is an indicator variable that identifies
whether the participant correctly classifies the conference call: equals 1 if (DECEP_COMPANY = 1 and
DECEP_JUDG = 1 or DECEP_COMPANY = 0 and DECEP_JUDG = 0) and 0 otherwise.

43
 FIGURE 2
Accuracy of Very Experienced and Inexperienced Auditors for Fraud and Non-Fraud
Companies

0.80
0.78
0.75
0.70
0.65
Accuracy

0.60
0.57 0.57
0.55
0.50 0.49
0.45
0.40
Non-Fraud Company Fraud Company
Inexperienced Very Experienced

Very experienced and inexperienced auditors are identified using EXP, which is an indicator variable that equals 1 if
the response was from a very experienced audit professional (24 years in audit on average), and 0 if the response
was from an inexperienced accounting major enrolled in an auditing course. Fraud and non-fraud companies are
identified using DECEP_COMPANY, which is an indicator variable that equals 1 if conference call pertains to a
quarter that was eventually restated due to an irregularity (Fraud Company), and 0 otherwise (Non-Fraud Company).
DECEP_JUDG is an indicator variable that equals 1 if the participant judges the conference call to be related to a
company quarter that will eventually be restated due to an irregularity, and 0 otherwise. The dependent measure is
ACCURACY, which is an indicator variable that identifies whether the participant correctly classifies the
conference call: equals 1 if (DECEP_COMPANY = 1 and DECEP_JUDG = 1 or DECEP_COMPANY = 0 and
DECEP_JUDG = 0) and 0 otherwise.

44
 FIGURE 3
The Effect of Audit Experience and Cognitive Dissonance Prompt on Accuracy for Fraud
and Non-Fraud Companies

0.90 Inexperienced
0.80
Accuracy

0.70

0.60 0.57
0.51
0.50 0.56

0.47
0.40
Non-Fraud Company Fraud Company
Cog. Diss. Not Prompted Cog. Diss. Prompted

0.90 Very Experienced

0.83
0.80
0.72
Accuracy

0.70 0.70

0.60

0.50
0.43
0.40
Non-Fraud Company Fraud Company
Cog. Diss. Not Prompted Cog. Diss. Prompted

Very experienced and inexperienced auditors are identified using EXP, which is an indicator variable that equals 1 if
the response was from a very experienced audit professional (24 years in audit on average), and 0 if the response
was from an inexperienced accounting major enrolled in an auditing course. Fraud and non-fraud companies are
identified using DECEP_COMPANY, which is an indicator variable that equals 1 if conference call pertains to a
quarter that was eventually restated due to an irregularity (Fraud Company), and 0 otherwise (Non-Fraud Company).
AUDIO is an indicator variable that equals 1 if the conference call excerpt was provided both as a transcript and as
audio, 0 if as a transcript only. DECEP_JUDG is an indicator variable that equals 1 if the participant judges the
conference call to be related to a company quarter that will eventually be restated due to an irregularity, and 0
otherwise. The dependent measure is ACCURACY, which is an indicator variable that identifies whether the
participant correctly classifies the conference call: equals 1 if (DECEP_COMPANY = 1 and DECEP_JUDG = 1 or
DECEP_COMPANY = 0 and DECEP_JUDG = 0) and 0 otherwise. Cognitive dissonance prompt is a between

45
subjects manipulated variable that has two parts. First, participants receiving this manipulation are given additional
initial instructions that said the following: “One cue found to be useful in detecting deception in these CEO
responses is cognitive dissonance. Cognitive dissonance is the negative, uncomfortable emotion a person feels when
they are saying something that they know is not true. Those experiencing cognitive dissonance feel uncomfortable,
uneasy, and bothered.” Second, after answering our principal dependent measure, these participants are asked “how
much cognitive dissonance the CEO felt during this excerpt of the conference call.” COGDIS = 1 when both of
these components are present and 0 otherwise.

46
 FIGURE 4
The Effect of Audit Experience and Cognitive Dissonance Prompt on Accuracy of Red Flag
Detection for Fraudulent Statements

0.33
0.31 0.32
Red Flag Accuracy

0.29
0.27
0.25
0.23 0.21
0.21
0.19 0.17
0.17
0.15 0.17
Cog. Diss. Not Prompted Cog. Diss. Prompted
Inexperienced Very Experienced

Very experienced and inexperienced auditors are identified using EXP, which is an indicator variable that equals 1 if
the response was from a very experienced audit professional (24 years in audit on average), and 0 if the response
was from an inexperienced accounting major enrolled in an auditing course. Cognitive dissonance prompt is a
between subjects manipulated variable that has two parts. First, participants receiving this manipulation are given
additional initial instructions that said the following: “One cue found to be useful in detecting deception in these
CEO responses is cognitive dissonance. Cognitive dissonance is the negative, uncomfortable emotion a person feels
when they are saying something that they know is not true. Those experiencing cognitive dissonance feel
uncomfortable, uneasy, and bothered.” Second, after answering our principal dependent measure, these participants
are asked “how much cognitive dissonance the CEO felt during this excerpt of the conference call.” COGDIS = 1
when both of these components are present and 0 otherwise. The dependent measure, red flag detection accuracy
(RF_ACCURACY), analyzed for only the five fraud companies, is constructed in three steps. First, each
participant is given a 1 (0) if they identify (do not identify) a CEO’s sentence as a red flag. This is done for each
CEO sentence. Second, each CEO sentence is classified as fraudulent or not fraudulent. This subjective
classification is done by a coauthor and several teams of research assistants with no knowledge of the red flag
results, using hand collected data about each of the five frauds. Finally, RF_ACCURACY receives a 1 if the auditor
correctly flags a fraudulent sentence or does not flag a non-fraudulent sentence, and a 0 otherwise.

47
 TABLE 1
Descriptive Statistics

Descriptive Statistics
Test of
Inexperienced Very Experienced All Responses Experienced vs.
(N = 706) (N = 121) (N = 827) Inexperienced
Mean Mean Mean
Test Statistic
(std.) (std.) (std.)
(p-value)
Variable Max. | Min. Max. | Min. Max. | Min.
0.00 1.00 0.14
EXP (0.00) (0.00) (0.35) NA
0.00 | 0.00 1.00 | 1.00 1.00 | 0.00
0.02 23.62 3.48
AUD_EXP (0.13) (13.08) (9.72) N/A
1.00 | 0.00 38.00 | 1.00 38.00 | 0.00
7.24 7.26 7.24
0.14
AUDIO_TIME (0.60) (0.61) (0.60)
(0.71)
8.25 | 6.30 8.25 | 6.30 8.25 | 6.30
1.81 1.89 1.82
0.99
FSCORE (0.89) (0.88) (0.89)
(0.32)
2.97 | 0.41 2.97 | 0.42 2.97 | 0.42
0.51 0.52 0.51
0.07
DECEP_COMPANY (0.50) (0.50) (0.50)
(0.79)
1.00 | 0.00 1.00 | 0.00 1.00 | 0.00
0.46 0.40 0.45
1.71
DECEP_JUDG (0.50) (0.49) (0.50)
(0.19)
1.00 | 0.00 1.00 | 0.00 1.00 | 0.00
0.53 0.67 0.55
8.96
ACCURACY (0.50) (0.47) (0.50)
(<0.01)
1.00 | 0.00 1.00 | 0.00 1.00 | 0.00

Very experienced and inexperienced auditors are identified using EXP, which is an indicator variable that equals 1 if
the response was from a very experienced audit professional (24 years in audit on average), and 0 if the response
was from an inexperienced accounting major enrolled in an auditing course. AUD_EXP is the number of years of
experience as an assurance professional reported by the participant. AUDIO_TIME is the duration in minutes of the
audio version of the conference call assigned to the participant. FSCORE is the fraud score of the company for the
fiscal quarter being discussed on the earnings conference call. Fraud and non-fraud companies are identified using
DECEP_COMPANY, which is an indicator variable that equals 1 if conference call pertains to a quarter that was
eventually restated due to an irregularity (Fraud Company), and 0 otherwise (Non-Fraud Company). DECEP_JUDG
is an indicator variable that equals 1 if the participant judges the conference call to be related to a company quarter
that will eventually be restated due to an irregularity, and 0 otherwise. The dependent measure is ACCURACY,
which is an indicator variable that identifies whether the participant correctly classifies the conference call: it equals
1 if (DECEP_COMPANY = 1 and DECEP_JUDG = 1 or DECEP_COMPANY = 0 and DECEP_JUDG = 0) and 0
otherwise.

Test statistics are χ2 or F-statistics for dichotomous or continuous variables, respectively that control for repeated
measures. All p-values are two tailed. Results are excluded when the participant indicated that they were aware of
financial improprieties for the company before participating in this experiment.

48
 TABLE 2
Accuracy of Very Experienced and Inexperienced Auditors for Fraud and Non-Fraud
Companies

Panel A: Descriptive Statistics – ACCURACY by EXP, DECEP_COMPANY, and

COGDIS
Inexperienced Very Experienced
Non-Fraud Fraud Non-Fraud Fraud
Company Company Total Company Company Total
Not Prompted 0.56 0.51 0.53 Not Prompted 0.83 0.43 0.63
(Std. Dev.) (0.50) (0.50) (0.50) (Std. Dev.) (0.38) (0.50) (0.49)
N 170 184 354 N 29 30 59
Prompted 0.57 0.47 0.52 Prompted 0.72 0.70 0.71
(Std. Dev.) (0.50) (0.50) (0.50) (Std. Dev.) (0.45) (0.47) (0.46)
N 178 174 352 N 29 33 62
Total 0.56 0.49 0.53 Total 0.78 0.57 0.67
(Std. Dev.) (0.50) (0.50) (0.50) (Std. Dev.) (0.42) (0.50) (0.47)
N 348 358 706 N 58 63 121

Panel B: Analysis Results

Source DF ܺଶ p-value
EXP 1 10.13 < 0.01
DECEP_COMPANY 1 8.34 < 0.01
COGDIS 1 0.16 0.69
EXP*DECEP_COMPANY 1 2.27 0.13
EXP*COGDIS 1 0.51 0.48
DECEP_COMPANY*COGDIS 1 3.03 0.08
EXP* DECEP_COMPANY*COGDIS 1 4.15 0.04
AUDIO 3.32 0.07

49
Panel C: Planned Contrasts, Simple Effects, and Comparisons
Source DF ܺଶ / t p-value
H1: Accuracy rates will be highest for very experienced
auditors assessing non-fraud companies (contrast weight=3),
significantly lower for very experienced auditors assessing
fraud companies (-1) and similarly low for inexperienced
auditors assessing both fraud (-1) and non-fraud companies (-
1). 1 10.84 < 0.01*
EXP given Non-Fraud Company 1 3.04 < 0.01*
DECEP_COMPANY given Very Experienced 1 2.34 0.01*
H2: -1.5 (Inexperienced / Non-Fraud / No Prompt), -1.5
(Inexperienced / Non-Fraud / Prompt), -1.5 (Inexperienced /
Fraud / No Prompt), -1.5 (Inexperienced / Fraud / No Prompt),
2.5 (Experienced / Non-Fraud / No Prompt), 2.5 (Experienced
/ Non-Fraud / Prompt), -1.5 (Experienced / Fraud / No
Prompt), 2.5 (Experienced / Fraud / Prompt) 1 17.46 < 0.01*
H2: -1 (Inexperienced / Fraud / No Prompt), -1 (Inexperienced
/ Fraud / No Prompt), -1 (Experienced / Fraud / No Prompt), 3
(Experienced / Fraud / Prompt) 1 5.70 0.01*
Accuracy rate for very experienced auditor given a cognitive
dissonance prompt and evaluating a fraud company, is greater
than each of the other three accuracy rates evaluating a all p <
fraudulent company 0.03*

* These p-values are for effects that occur in the expected direction suggested by our theory, and are therefore the
one-tailed test of the signed t-statistic. Other reported p-values are two-tailed.

Very experienced and inexperienced auditors are identified using EXP, which is an indicator variable that equals 1 if
the response was from a very experienced audit professional (24 years in audit on average), and 0 if the response
was from an inexperienced accounting major enrolled in an auditing course. Fraud and non-fraud companies are
identified using DECEP_COMPANY, which is an indicator variable that equals 1 if conference call pertains to a
quarter that was eventually restated due to an irregularity (Fraud Company), and 0 otherwise (Non-Fraud Company).
DECEP_JUDG is an indicator variable that equals 1 if the participant judges the conference call to be related to a
company quarter that will eventually be restated due to an irregularity, and 0 otherwise. The dependent measure is
ACCURACY, which is an indicator variable that identifies whether the participant correctly classifies the
conference call: equals 1 if (DECEP_COMPANY = 1 and DECEP_JUDG = 1 or DECEP_COMPANY = 0 and
DECEP_JUDG = 0) and 0 otherwise. Cognitive dissonance prompt is a between subjects manipulated variable that
has two parts. First, participants receiving this manipulation are given additional initial instructions that said the
following: “One cue found to be useful in detecting deception in these CEO responses is cognitive dissonance.
Cognitive dissonance is the negative, uncomfortable emotion a person feels when they are saying something that
they know is not true. Those experiencing cognitive dissonance feel uncomfortable, uneasy, and bothered.” Second,
after answering our principal dependent measure, these participants are asked “how much cognitive dissonance the
CEO felt during this excerpt of the conference call.” COGDIS = 1 when both of these components are present and 0
otherwise. AUDIO is an indicator variable that equals 1 if the conference call excerpt was provided both as a
transcript and as audio, 0 if as a transcript only.

50
 TABLE 3
Very Experienced and Inexperienced Auditors’ Red Flagged Issues for Fraud and Non-
Fraud Companies

Descriptive Statistics – Auditors’ Red Flagged Issues by EXP and DECEP_COMPANY

Non-Fraud Company Fraud Company
Mean (std. dev.) Mean (std. dev.)
Variable % RF_ISSUES % RF_ISSUES t p-value
Inexperienced Auditors
N 348 358
ISSUE_CHARS 240.39 (251.82) 243.34 (271.50) 0.21 0.83
2.23 (1.81) 2.40 (2.12)
RF_ISSUES 1.31 0.19
100% 100%
0.54 (0.78) 0.87 (1.06)
FIN 5.42 < 0.01
24% 36%
0.98 (1.19) 0.86 (1.17)
SPEECH 1.55 0.12
44% 36%
0.71 (0.89) 0.67 (0.91)
SKNOW 0.55 0.58
32% 28%
Very Experienced Auditors
N 58 63
ISSUE_CHARS 169.41 (188.18) 310.68 (325.41) 3.61 < 0.01
2.15 (2.17) 3.91 (3.00)
RF_ISSUES 4.08 < 0.01
100% 100%
0.63 (0.98) 1.24 (1.17)
FIN 3.20 < 0.01
29% 32%
0.82 (1.26) 1.27 (1.39)
SPEECH 2.10 0.04
38% 32%
0.70 (1.04) 1.40 (1.59)
SKNOW 3.13 < 0.01
33% 36%
For each of the four companies they evaluated, we ask auditors to read (listen to) the transcript and write down the
time stamp (line number from the transcript) of each red flag and as much detail about the red flag as possible. Two
separate raters with knowledge of accounting but with no knowledge of the hypotheses being tested nor the
experience level of the auditor providing the response, independently coded the open-ended notes auditors make
when noting red flags. The coders counted the number of issues and sorted them into 14 categories. We average
their responses together and then arrange the 14 categories into three broad groups: financial statement items (FIN),
speech (SPEECH), and suspect knowledge (SKNOW). The FIN group contains red flagged issues pertaining to
financial statements, financial results, company strategy, and topics regarding the appropriate application of
judgment when applying accounting standards given the operating environment. The SPEECH group captures
speech hesitations, filled pauses (ah’s and um’s), language use that avoids answering questions or deflects blame,
and speech that suggests the speaker was nervous. The SKNOW group contains red flagged issues identifying
instances where the executive appeared to lack knowledge of the subject matter by giving answers that seem
incomplete, inaccurate, internally inconsistent, or lacking appropriate level of confidence. The sum of FIN,
SPEECH, and SKNOW represents the total number of red flagged issues identified (RF_ISSUES). ISSUE_CHARS
is the number of characters provided in the text box from which the red flagged issues were coded. Very
experienced and inexperienced auditors are identified using EXP, an indicator variable that equals 1 if the response
was from a very experienced audit professional (24 years in audit on average), and 0 if the response was from an
inexperienced accounting major enrolled in an auditing course. Fraud and non-fraud companies are identified using

51
DECEP_COMPANY, an indicator variable that equals 1 if conference call pertains to a quarter that was eventually
restated due to an irregularity (Fraud Company), and 0 otherwise (Non-Fraud Company). DECEP_JUDG is an
indicator variable that equals 1 if the participant judges the conference call to be related to a company quarter that
will eventually be restated due to an irregularity, and 0 otherwise.

52
 TABLE 4
The Effect of Audit Experience and Cognitive Dissonance Prompt on Accuracy of Red Flag
Detection for Fraudulent Statements

Panel A: Descriptive Statistics – RF_ACCURACY by EXP and COGDIS

Inexperienced Experienced Total

Not Prompted 0.17 0.21 0.18
(Std. Dev.) (0.38) (0.41) (0.38)
N 1362 224 1586
Prompted 0.17 0.32 0.20
(Std. Dev.) (0.38) (0.47) (0.40)
N 1290 241 1531
Total 0.17 0.27 0.19
(Std. Dev.) (0.38) (0.44) (0.39)
N 2652 465 3117

Panel B: Analysis Results

Source DF ܺଶ p-value
EXP 1 12.30 < 0.01
COGDIS 1 3.20 0.08
EXP*COGDIS 1 1.99 0.16

Panel C: Planned Contrasts, Simple Effects, and Comparisons

Source DF ܺଶ / t p-value
The accuracy of very experienced auditors receiving the
cognitive dissonance prompt is highest (contrast weight=3),
significantly lower for very experienced auditors that are not
prompted (-1), and similarly low for inexperienced auditors
who are both prompted (-1) and not prompted (-1). 1 11.97 < 0.01
COGDIS given Very Experienced 1 1.78 0.08
COGDIS given Inexperienced 1 0.43 0.67
Very experienced and inexperienced auditors are identified using EXP, which is an indicator variable that equals 1 if
the response was from a very experienced audit professional (24 years in audit on average), and 0 if the response
was from an inexperienced accounting major enrolled in an auditing course. Cognitive dissonance prompt is a
between subjects manipulated variable that has two parts. First, participants receiving this manipulation are given
additional initial instructions that said the following: “One cue found to be useful in detecting deception in these
CEO responses is cognitive dissonance. Cognitive dissonance is the negative, uncomfortable emotion a person feels
when they are saying something that they know is not true. Those experiencing cognitive dissonance feel
uncomfortable, uneasy, and bothered.” Second, after answering our principal dependent measure, these participants
are asked “how much cognitive dissonance the CEO felt during this excerpt of the conference call.” COGDIS = 1
when both of these components are present and 0 otherwise. The dependent measure, red flag detection accuracy
(RF_ACCURACY), analyzed for only the five fraud companies, is constructed in three steps. First, each
participant is given a 1 (0) if they identify (do not identify) a CEO’s sentence as a red flag. This is done for each
CEO sentence. Second, each CEO sentence is classified as fraudulent or not fraudulent. This subjective
classification is done by a coauthor and several teams of research assistants with no knowledge of the red flag
results, using hand collected data about each of the five frauds. Finally, RF_ACCURACY receives a 1 if the auditor
correctly flags a fraudulent sentence or does not flag a non-fraudulent sentence, and a 0 otherwise.

53