2013 Fraud Dinamics in Organizations-Davis

Accounting, Organizations and Society 38 (2013) 469–483
Contents lists available at SciVerse ScienceDirect
Accounting, Organizations and Society

journal homepage: www.elsevier.com/locate/aos
Fraud dynamics and controls in organizations

Jon S. Davis ⇑, Heather L. Pesch 1
Department of Accountancy, University of Illinois at Urbana–Champaign, 1206 S. Sixth Street, Champaign, IL 61821, USA
a b s t r a c t
This paper develops an agent-based model to examine the emergent dynamic characteris-
tics of fraud in organizations. In the model, individual heterogeneous agents, each of whom
can have motive and opportunity to commit fraud and a pro-fraud attitude, interact with
each other. This interaction provides a mechanism for cultural transmission through which
attitudes regarding fraud can spread. Our benchmark analysis identifies two classes of
organizations. In one class, we observe fraud tending toward a stable level. In the other
class, fraud dynamics are characterized by extreme behaviors; organizations with mostly
honest behavior suddenly change their state to mostly fraudulent behavior and vice versa.
These changes seem to occur randomly over time. We then modify our model to examine
the effects of various mechanisms thought to impact fraud in organizations. Each of these
mechanisms has different impacts on the two classes of organizations in our benchmark
model, with some mechanisms being more effective in organizations exhibiting stable
levels of fraud and other mechanisms being more effective in organizations exhibiting
unstable extreme behavior. Our analysis and results have general implications for design-
ing programs aimed at preventing fraud and for fraud risk assessment within the audit
context.
Ó 2012 Elsevier Ltd. All rights reserved.
Introduction of fraud with financial statement reporting choices and cor-

porate governance variables (e.g., Beasley, 1996; Jones,
Fraud2 has become a popular area of inquiry among Krishnan, & Melendrez, 2008; Sharma, 2004).
accounting academics because of the magnitude of losses In light of the high cost of corporate fraud, one might
(estimated by the Association of Certified Fraud Examiners expect considerable research activity investigating the effi-
in 2010 to be US$2.9 trillion worldwide) and requirements cacy of mechanisms designed to prevent or reduce fraud.
imposed on auditors to explicitly address the problem (AIC- However, despite its potential importance, a review of re-
PA, 2002). Research has addressed fraud risk assessment by search reveals a striking dearth of work examining the
auditors (e.g., Bell & Carcello, 2000; Carpenter, 2007; Wilks effectiveness of various prevention mechanisms, except
& Zimbelman, 2004), fraud detection (e.g., Cleary & Thibo- for the deterring role of audits (e.g., Finley, 1994; Schneider
deau, 2005; Hoffman & Zimbelman, 2009; Matsumura & & Wilner, 1990; Uecker, Brief, & Kinney, 1981). This lack of
Tucker, 1992), fraud incentives (e.g., Erickson, Hanlon, & work on prevention is likely due to fraud being a hidden
Maydew, 2006; Gillett & Uddin, 2005), and the correlation crime. Because the extent of fraud is usually unknown in
an organization, measuring the effectiveness of prevention
mechanisms is difficult using traditional empirical
⇑ Corresponding author. Tel.: +1 217 300 0489; fax: +1 217 244 0902.
methods.
E-mail addresses: [email protected] (J.S. Davis), [email protected]
(H.L. Pesch). Research on fraud in organizations tends to focus on
1
Tel.: +1 217 300 0514; fax: +1 217 244 0902. either the individual or the organization (Holtfreter,
2
For the purposes of our model, we define fraud as all forms of 2005; Pinto, Leana, & Pil, 2008). To date, very little work
occupational fraud, including asset misappropriation, corruption, and has attempted to explicitly link individual behaviors in
financial statement fraud.
0361-3682/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.aos.2012.07.005
470 J.S. Davis, H.L. Pesch / Accounting, Organizations and Society 38 (2013) 469–483
the organization to organizational outcomes within the observe a very different pattern in which the number of
fraud context.3 Understanding the individual–organization fraudsters in the organization vacillates over time between
link is important because a focus on either individual behav- extremes; either virtually no one in the organization is a
ior or the organization in isolation turns a blind eye to the fraudster or virtually everyone is.
social process through which individuals’ behaviors are When we consider mechanisms to prevent or eliminate
influenced by the organization as a whole and vice versa. fraud, we find that their impact is contingent on average
In other words, a narrow focus on individual behaviors or susceptibility to social influence within the organization.
on the organization ignores the organization’s sociology, A reduction in perceived opportunity or the introduction
which can have profound effects on fraud outcomes and of influential, honest managers (tone at the top) reduces
the efficacy of fraud prevention mechanisms. the number of fraudsters, but neither change to our model
We develop a model of fraud in organizations that al- is effective in eliminating outbreaks of fraud when suscep-
lows an evaluation of the relative efficacy of mechanisms tibility is moderate or high. Allowing honest employees to
designed to prevent fraud while explicitly recognizing be more influential than fraudsters has no qualitative ef-
the social processes underlying the formation of organiza- fect when susceptibility is low; however, it transforms
tional norms. To develop our model, we use a method that behavior when average susceptibility is moderate to high,
is relatively new in accounting research: agent-based mod- reducing the number of fraudsters to near zero and elimi-
eling (ABM). Designed to study the emergence of macro-le- nating fraud outbreaks. The contingent nature of this effect
vel phenomena from micro-level interactions, ABM is well may prove important in fraud risk assessments performed
suited to address questions involving organizational out- by auditors. We also find that efforts to remove fraudsters
comes (e.g., a culture of fraud) resulting from the interac- can effectively reduce the number of fraudsters to near
tions between individuals within an organization and zero regardless of the level of susceptibility, but such ef-
organizational variables. The use of ABM confers an addi- forts do not eliminate fraud outbreaks when susceptibility
tional advantage: It allows us to gain insights into fraud is moderate to high.
even when data in organizations are censored. This paper continues with a brief introduction to ABM.
Our model is comprised of an organization represented This methodological introduction is followed by the devel-
by 100 independent, heterogeneous agents (employees) opment of a benchmark model of an organization with no
and a set of simple interaction rules. Following Cressey’s interventions to prevent fraud. We then modify the bench-
(1953) characterization of occupational fraud (known as mark model to examine the effect of anti-fraud interven-
the fraud triangle hypothesis), any agent in our model pos- tions. This paper ends with our conclusions and a
sessing motive, opportunity, and an attitude that frames discussion of the implications of our analysis.
the fraudulent act as acceptable will commit fraud. We al-
low agents to repeatedly interact, with an eye toward
emergent aggregate fraud levels and the dynamics of fraud Agent-based models
over time. We begin with a benchmark model in which all
agents have opportunity and motive. We then modify our We use ABM to address our research question because
model to investigate the impact of mechanisms to prevent it confers several advantages. As noted by Epstein (2006),
or detect fraud. We first investigate the impact of modify- the method avoids several shortcomings in traditional the-
ing the likelihood that agents perceive the opportunity to oretical work in the social sciences. When aggregate
commit fraud. Next, we consider a hierarchy in which behavior is the research subject in traditional theory,
higher-level honest employees exert greater influence than heroic assumptions about individual behavior and the pop-
lower-level employees (i.e., ‘‘tone at the top’’). Then we ulation being modeled are typically required to maintain
consider the impact of asymmetric influence exerted by tractability (e.g., the perfect rationality and homogeneity
fraudsters relative to honest employees (which can arise of individual actors). While these assumptions bear little
as a result of ethical training, the implementation of a code resemblance to human behavior, it is often argued that
of ethics, or a variety of other interventions). Finally, we such simplifications are necessary because no rigorous
consider the impact of detection and termination efforts. method exists that would allow their relaxation. Similarly,
Two patterns emerge from the analysis of our bench- in traditional theory, the notion of equilibrium plays a pre-
mark model, depending upon how susceptible individual dominant role as a solution concept; models attempt to ex-
agents are to social influence. When average susceptibility plain social phenomenon by identifying the behavior of
is low, the number of fraudsters in the organization tends interest as an equilibrium. While this approach can yield
toward a specific level and remains relatively stable over valuable insights, there are limits. In many cases, phenom-
time. When average susceptibility is moderate to high we ena of interest may involve disequilibrium dynamics or the
identified equilibria may be unattainable either outright or
3 within acceptable time scales (Epstein, 2006, p. 72). Exper-
Pinto et al. (2008) discuss the ways in which corruption can exist in an
organization, with a particular focus on how individual corruption can imental research in the social sciences that attempts to test
spread to the point where it becomes an organizational phenomenon. hypotheses linking heterogeneous individual behaviors to
Chang and Lai (2002) use econometrics to model corrupt organizations as a aggregate behavioral outcomes is also challenging because
pandemic arising from individual interactions. Kim and Xiao (2008) of an exiguity of theoretical work linking realistic individ-
examine the link between individual behaviors and aggregate outcomes
in the context of fraud in a public health care delivery program. In a broader
ual behaviors to aggregate phenomena. Finally, the stove-
context, Davis, Hecht, and Perkins (2003) develop a model that links pipe structure of the social science disciplines (e.g.,
individual behaviors to societal outcomes related to tax evasion. sociology, economics, psychology, anthropology) tends to
J.S. Davis, H.L. Pesch / Accounting, Organizations and Society 38 (2013) 469–483 471
limit the perspective that each discipline develops (much Once the elements of the agent-based model have been
like the parable of the blind men and the elephant).4 developed (usually instantiated in a computer program6),
The approach taken by ABM promises a solution to agents are allowed to repeatedly interact with each other
many of the shortcomings identified in traditional social and with the environment. As interactions occur, the re-
science research. The method allows for the specification search focuses on the spontaneous emergence of aggregate
of heterogeneous actors who follow simple and realistic or group behaviors resulting from interactions at the micro
(boundedly rational) behavioral rules. It focuses on the (agent) level. Epstein and Axtell (1996) characterize this ap-
emergence of macro-level phenomena from micro-level proach as attempting to ‘‘grow’’ social phenomena from the
behavior and is therefore well suited for the study of social bottom up (modeling individual behavior) rather than from
phenomena (including the emergence of fraud outbreaks the top down (directly modeling aggregate phenomena).
in organizations).5 Because ABM is dynamic, the natural fo- Thus, ABM does not involve macro models but, rather, mod-
cus of the research is on changes in behavior over time els of individual behavior and social interaction that produce
rather than on equilibria that may never be achieved. Final- macro-level outcomes.
ly, ABMs naturally cross the boundaries of the various social Since agent-based models are simulations, they rely on
sciences because of the requisite focus on modeling individ- both deduction and induction. While each such model is a
ual behaviors (e.g., psychology), spatial and ecological con- strict deduction and constitutes a sufficiency theorem (that
siderations (e.g., anthropology), social interaction (e.g., proves existence but not uniqueness), multiple simulations
sociology), and a wide range of other potential cross-disci- can be thought of as a distribution of theorems that to-
plinary considerations. gether are used to inductively test a proposition (Epstein,
A fully specified agent-based model consists of agents 2006). Thus, the flavor of ABM research is more similar to
(e.g., individuals in an organization, traders in a market, laboratory experimentation than to deductive proofs.7 That
organizations in an economy, or trees in a forest), an envi- is, each instantiation of an agent-based model can be used as
ronment in which the agents ‘‘live,’’ and a set of rules that an empirical observation; theoretical support can be gar-
govern agent interactions with other agents and their envi- nered for a theory through induction, but no deductive proof
ronment. Agents are characterized by a collection of inter- obtained.
nal states, endowments (knowledge and assets), and In the context of scientific inquiry, ABM has been em-
behaviors that may recognize the existence of bounded ployed as a tool for both prediction (akin complex econo-
rationality. Some of these characteristics are static, while metric models) and understanding and explanation (akin
others can change in response to interactions with the envi- to analytical models). When models are used for predic-
ronment or with other agents. Similarly, environments can tion, researchers typically focus on incorporating as much
be static or can change on the basis of rules or as a result of realism as possible. When models are used as a tool for
interactions with agents. Environments can be represented improving understanding, researchers strive for simplicity,
in a variety of ways. When location or resource placement with the ultimate goal of incorporating only the features
is important in a model (an issue in anthropological or eco- necessary to generate the phenomenon of interest (Axel-
logical research), the environment can be represented spa- rod, 1997; Grimm et al., 2005; Miller & Page, 2007; Ragan,
tially, using a lattice or a landscape developed from a 2010; Tesfatsion & Judd, 2006). Because the goal of this pa-
geographic information system. In economics, the environ- per is to improve understanding and to provide insights
ment of choice is a market (e.g., Gode & Sunder, 1993). In into the effects of various interventions on fraud, we adopt
research examining sociological or sociopsychological phe- the latter approach.
nomena (the development of a culture of fraud in organiza-
tions, in our instance), a social network may be favored
Benchmark model
(e.g., Breiger & Carley, 2003). Finally, a set of institutional
rules is specified that defines how agents interact with each
Most (if not all) organizations face some level of fraud.
other and with their environment. As with agent and envi-
There is also anecdotal evidence of fraud outbreaks in orga-
ronmental characteristics, these rules can be static, dy-
nizations where fraud becomes institutionalized as accept-
namic, or adaptive. A simple rule for an agent might be to
able behavior. Consider the case of bribery at Siemens. A
sell to another agent if the offer price is less than the bid
newspaper article (Schubert & Miller, 2008) reports that
price or, in the context of our fraud model, to commit fraud
illegal bribe payments were so widely accepted at Siemens
if the agent believes that it is acceptable behavior and if
that they were treated as any other line item in the financial
both motive and opportunity are present.
statements. The authors note that ‘‘bribery was Siemens’s
6
Our model employs Mathematica 7.0 (Wolfram Research Inc., 2008) as
the programming platform.
7
Epstein (1999) observes that ABM represents a new approach to
4
In the parable of the blind men and the elephant, a group of blind men science that can be used as an instrument for either theory development or
touch an elephant to learn what it is like. Each one feels a different part, but theory testing. Rather than fitting neatly into inductive or deductive realms,
only one part (e.g., the elephant’s tail, tusk, leg, etc.). The blind men ABM is more appropriately thought of as a generative form of science. To
compare notes and find themselves in disagreement. explain a macroscopic social phenomenon, the generativist asks, ‘‘How
5
The micro-level focus of agent-based models confers another benefit: It could decentralized local interactions of heterogeneous autonomous agents
avoids the problematic representative agent assumption that is common in generate the given regularity?’’ (Epstein, 1999, p. 41), Put another way, the
macro-level models (for a discussion of the shortcomings of the represen- generativist’s motto is ‘‘If you didn’t grow it, you didn’t explain its
tative agent assumption, see Kirman, 1992). emergence’’ (Epstein, 1999, p. 46).
business model . . . Siemens had institutionalized corrup- that each of these variables is binary (either present or
tion.’’ Organizational level fraud (or fraud outbreaks) can absent).11
arise from the creation of a culture of fraud in the organiza- Each of the agents in our model may or may not
tion, wherein fraud grows from the bottom up, as individ- have a motive for fraud (determined randomly with
ual employees spread a culture of fraud to others through some predetermined agent-specific probability at each
socialization or other means (e.g., see the discussion of step of the simulation).12 Implicit in motive is some form
organizations of corrupt individuals by Pinto et al. of mental calculus performed by individual agents evalu-
(2008)). Alternatively, organization-level fraud can be engi- ating the perceived costs and benefits of committing
neered by the organization’s leaders and pushed from the fraud.13 Over time, an agent’s motive to commit fraud
top down, using powerful mechanisms such as obedience can change in response to her personal situation (e.g.,
to authority (Milgram, 1963).8 Consistent with the land- unexpected medical bills or loss of employment by a
scape of fraud observed in the natural ecology, our modeling family member).
goal is to identify a minimal set of plausible assumptions suf- Opportunity is an organization-level variable that
ficient to generate organizations that experience both fraud proxies for the overall strength of a system of internal
outbreaks and stable levels of fraud. This goal guides our controls. It is represented as the probability that each
modeling choices, described below. agent will perceive an opportunity to commit fraud. For
We base our model on Cressey’s (1953) fraud triangle example, a 20% opportunity implies that each agent inde-
hypothesis, which still prevails as a popular view of the pendently has a 20% chance of perceiving an opportunity
necessary and sufficient conditions for an individual’s deci- to commit fraud at each point in the simulation. An agent
sion to commit fraud.9 Cressey’s characterization identifies might perceive such an opportunity when a loophole is
three key determinants of fraud. When the conjunction of discovered in the control system (e.g., a fellow employee
these determinants exists, an individual will commit fraud. fails to log off his computer or a manager ‘‘turns her
The determinants are a perceived non-sharable financial back’’).
need (or, more generally, a motive), a perceived opportunity Finally, an individual’s ability to rationalize fraud in
to commit fraud, and an attitude or belief that frames the our model is a function of both influence by others within
fraudulent act as acceptable behavior. In our model, agents the organization and factors exogenous to this influence
become fraudsters if they face a conjunction of perceived (i.e., events not related to the work environment, for
opportunity (O), motive (M), and an attitude (A) that fraud example, shifts in the attitudes of other reference groups
is an acceptable behavior, or such as family and friends). The plasticity of attitude in
response to influence by others in the organization
O ^ M ^ A $ F: (whether through socialization among rank and file
employees or through the command of organizational
Conversely, over time, if a fraudster no longer has opportu-
leadership) is central to the potential for fraud outbreaks
nity, motive, or attitude, then he or she will reform to
in our model.
become an honest agent.10 For simplicity, we assume
Characterizing agents
We begin developing our model of fraud in organiza-

8
tions by characterizing individual agents. In particular,
Social influence may also play a part in top-down fraud as more
each agent is a vector whose elements indicate the
employees participate in the activity.
9
Wolfe and Hermanson (2004) expand the fraud triangle to a fraud
agent’s attributes (endowments and individual behavioral
diamond, suggesting capability as a fourth condition required for fraud. We rules). Each attribute is a number or a list of numbers. In
elect to build on Cressey’s original approach for two reasons. First, the our benchmark model, agent attributes include the
original fraud triangle has the advantage of simplicity: Capability is already following:
included in the hypothesis as part of opportunity. Second, the fraud triangle
remains the dominant characterization in the professional literature. For
11
example, it is incorporated in a discussion of the characteristics of fraud in It is possible that the levels of these variables are a matter of degree in
the AICPA’s 2002 Statement on Auditing Standards (SAS) 99, which addresses organizations and that a high value of one variable (e.g., motivation)
the auditor’s responsibilities relating to fraud in an audit of financial compensates for the low value of another (opportunity or attitude). To
statements. model continuous values of O, M, and A, one would likely employ a
10
Our model allows for free flow between the honest and fraudster threshold rule (where fraud occurs when the combination of the three
classes. In the natural ecology, such a free flow is not always observed. In factors exceeds some level). We conjecture that the use of a threshold rule
some cases individuals commit one fraud or a series of frauds to meet a in our model could lead to unanticipated responses to changes in
particular goal and then reform. As an example, a sales manager may parameters such as the phase transition observed in the threshold model
fraudulently record a large sale as occurring before year-end to meet his in Davis et al. (2003).
12
forecast in one year but then report properly in future periods when he is At the outset, for simplicity, we assume that every agent has a motive
able to meet his forecast honestly. In other cases, the fraudster continues to for fraud but the model admits the possibility of heterogeneity in motives
commit fraud in perpetuity or until discovered. This can occur when the across individual agents.
13
fraud is committed to solve an ongoing problem (e.g., living beyond one’s Our model subsumes both fraud committed on behalf of individual
means or hitting unrealistic growth expectations year after year) or if the agents and fraud committed on behalf of the organization. Underlying the
decision to commit fraud in one period requires future fraud to keep the mental calculus noted here, the benefits could accrue to the individual
original fraud hidden. Because fraud is censored, we have no sense of the committing the fraud or to the organization. Similarly, the costs arising
relative frequencies of one-shot versus continued fraud. Our model permits from the fraud could be imposed on either the individual or the
either type. organization (see Pinto et al. (2008) for a discussion).
A unique integer that identifies the agent (a ‘‘name’’). for example, a situation in which an agent goes to lunch
Four independent probabilities, each drawn from a with co-workers and is told that one of them believes it
bounded uniform distribution. The first two represent is acceptable to claim a personal meal as a reimbursable
the likelihood of the agent taking on a particular business expense. We assume that the more often this oc-
characteristic (motive for fraud and perceived oppor- curs, the more likely the agent will begin to rationalize
tunity to commit fraud) and the second two repre- committing similar behavior. We also allow for the possi-
sent (i) the likelihood (q) of one’s attitude toward bility that one’s attitude toward fraud can change (with
fraud changing as a result of interacting with other probability p) independently of observing others inside
agents and (ii) the likelihood (p) of one’s attitude the organization.
toward fraud changing as a result of factors In the model, agents’ attitudes about the acceptability
unrelated to interactions with other agents. These of fraud change according to their particular experiences
probabilities are static and, with the exception of with other agents and their personal proclivities. At any
the opportunity likelihood, randomly determined for given time, agent attitudes may be heterogeneous be-
each agent. The opportunity likelihood is a static cause each agent has a unique set of characteristics and
organization-level variable. different experiences in the organization. One particular
Four binary variables indicating whether a motive for agent can be a fraudster at a given time while other
fraud is present, whether perceived opportunity is pres- agents may be honest. When interacting with an honest
ent, whether a pro-fraud attitude is present, and, finally, agent in the organization, the fraudster can trigger a
whether the agent is committing fraud. The presence of pro-fraud attitude in the honest agent (and possibly con-
motive, opportunity, and attitude is determined at each vert the honest agent into a fraudster, if both motive and
step in the simulation for each agent, based on the perceived opportunity exist) or the honest agent can con-
assigned probabilities above. Fraud exists if motive, vert the fraudster into an honest agent. In the case of two
opportunity, and attitude co-exist. honest agents interacting, one or both agents may begin
to view fraud as acceptable because of factors exogenous
to their interactions. These rules for interaction and
The organization
behavior define a social dynamic. The culture changes
dynamically in response to both agent interactions and
To create an organization,14 we develop a matrix. Each
their spontaneous behavior. The model is ‘‘bottom up,’’
row in the matrix is an agent vector. To allow for interac-
in that a culture of fraud can emerge spontaneously in
tions between agents, we represent a social network (a set
the population as a result of the interactions of individual
of co-workers) for each agent by adding a list to each
agents with each other.
agent’s vector, where each element in the list is an integer
referring to another agent’s name. In our social network,
agent relationships are symmetric (if agent 1 is included
An example
in agent 3’s social network, then the reverse is also true).
The list of co-workers assigned to each agent is randomly
Recall that our model begins with an initial organiza-
determined (subject to the symmetry constraint) and re-
tion represented by a matrix in which each row represents
mains static.
an agent and each element in a row represents an agent
attribute. For clarity, we present an example of a matrix
Observing the organization over time with a population size of five:
After establishing a starting population in the organi- 0 1

ID pðMÞ M p q A pðOÞ O F E S SN
zation, we allow agents to repeatedly interact and up- B 1 0:00449253 0 0:0242858 0:90035
B 0 0:95 1 0 0 1 f4g C C
date agent states. Interactions between agents are B C
B 2 0:426662 0 0:0292374 0:585986 0 0:95 1 0 0 0 f g C
B C
accomplished by repeatedly and randomly pairing each B
B 3 0:354574 0 0:0322799 0:511431
C:
B 0 0:95 1 0 0 2 f4;5g C
C
agent with a member of his or her social network.15 B C
@ 4 0:0897837 0 0:036227 0:525595 0 0:95 1 0 0 2 f1;3g A
The pairing allows each agent to observe another. As
noted previously, the observing agent will emulate the 5 0:496052 0 0:0205544 0:892721 0 0:95 1 0 0 1 f3g
attitude16 of the observed with probability q. Consider,
The elements in each column represent the following agent
14
In our model, the term ‘‘organization’’ can apply to an entire business attributes:
entity, a branch, a division, an office, or even a department within an office. ID = agent’s unique identifier,
15
We employ an algorithm developed by Gaylord and Davis (1999) to p(M) = likelihood of an agent’s motive to commit fraud
accomplish this pairing.
16 changing,
The model presented in this paper assumes that agent attitude is
influenced by other agents’ attitudes. We also evaluated an alternative M = binary variable indicating the presence of motive,
model where agent attitude is influenced only by other agents’ actions. The p = likelihood of an agent changing attitude as a result
only qualitative difference observed was in the conditions under which of factors exogenous to agent interactions,
perceived opportunity to commit fraud is reduced. In this instance, q = likelihood of an agent changing attitude as a result
organizations with moderate to high levels of social influence respond to
of interactions with other agents,
reduced opportunity in a fashion similar to that observed in our
asymmetric influence condition (i.e., fraud outbreaks are eliminated and A = binary variable indicating the presence of an atti-
the mean level of fraud is reduced). tude supporting fraud,
Table 1
Summary of model analysis.
Starting population attitude Random (n = 10) All pro-fraud (n = 5) All anti-fraud (n = 5)

Model variant
Benchmark q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95}
Opportunity q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95}
p(O) = {0.9, 0.8, . . ., 0.5} p(O) = {0.9, 0.8, . . ., 0.5} p(O) = {0.9, 0.8, . . ., 0.5}
Tone at top q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95}
Managers = {1, 2, 4, 6} Managers = {1, 2, 4, 6} Managers = {1, 2, 4, 6}
Asymmetric q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95}
Influence c = {99, 19, 9, 5.7, 4, 3} c = {99, 19, 9, 5.7, 4, 3} c = {99, 19, 9, 5.7, 4, 3}
Termination q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95} q = {0.05, 0.10, . . ., 0.95}
e = {0.05, 0.10, . . ., 0.20} e = {0.05, 0.10, . . ., 0.20} e = {0.05, 0.10, . . ., 0.20}
For all model variants, 10 starting organizations were used. Attitude regarding fraud was either randomly assigned to agents (for all 10 organizations) or set
to anti-fraud or pro-fraud for a subset of five organizations. For all model variants, average emulation likelihood (q) was varied from 0.05 to 0.95 in 5%
increments. Within each value of q, the model variant parameters (i.e., perceived opportunity, number of managers, influence scalar, and termination
likelihood) were evaluated at the levels indicated.
p(O) = likelihood of an agent perceiving an opportunity Benchmark model design

to commit fraud,
O = binary variable indicating the perceived opportu- We use 10 unique starting organizations to investigate
nity to commit fraud, our model.17 To establish a benchmark, we initially assume
F = binary variable indicating whether an agent is com- that every agent has a motive for fraud and perceives oppor-
mitting fraud, tunity. We set the size of the population to 100 agents and
E = binary variable to be used later when agent removal observe 15,000 interaction periods. Each agent is initially as-
is considered, signed (with a 50% probability) to either a pro-fraud or an
S = social network size, and anti-fraud attitude.18 Finally, we begin by allowing agents
SN = list of an agent’s social network. to have a very small but heterogeneous likelihood of chang-
ing their attitude toward fraud as a result of factors exoge-
In the sample matrix above, agent 3’s likelihood of a nous to agent interaction (probability p drawn
change in motive is 0.354574, opportunity is currently independently for each agent from a uniform distribution
perceived as present, no fraud is being committed, from 0 to 0.005).19 The design of our benchmark analysis
the termination variable is set to zero, the agent’s so- (and subsequent analysis of anti-fraud interventions) is de-
cial network size is two, and the social network con- scribed in Table 1.
sists of agents 4 and 5. The matrix represents the
organization’s and agents’ characteristics for one period
in time.
As noted previously, agents are iteratively paired and
allowed to interact over time. For example, moving for- 17
Our design and analysis were constrained by computational power and
ward one interaction period from the initial matrix above, practicality (approximately 5000 hour of computing time were expended).
if agent 1 begins to rationalize fraud as a result of factors We chose fewer organizations and more interaction periods because
additional interaction periods were low cost relative to examining
exogenous to agent interaction, his attitude variable will
additional organizations (due to the extensive analysis required for each
change from zero to one. Further, if we assume agent 4 is organization across the range of parameters that we examine). Similarly,
subsequently paired with agent 1 during this period and additional agents (beyond 100) placed excessive demands on our compu-
is influenced by agent 1, agent 4 will also adopt an attitude tational resources and we felt that the additional insight provided was
likely to be limited.
supportive of fraud, with a consequent change in his atti- 18
For a subset of five of the organizations, we examined alternative
tude variable (from zero to one). The following matrix starting conditions, where all agents start with a pro-fraud attitude or all
illustrates the result: agents start with an anti-fraud attitude. As these starting populations
evolved, the patterns of behavior converged to match those observed in
0 1 populations where agents were initially assigned to either pro-fraud or
ID pðMÞ M p q A pðOÞ O F E S SN
B 1 0:00449253 0 0:0242858 0:90035 anti-fraud attitudes with equal likelihood. We consequently conclude that
B 1 0:95 1 0 0 1 f4g C C
B C assignment of attitudes in the starting population has no effect on
B 2 0:426662 0 0:0292374 0:585986 0 0:95 1 0 0 0 f g C outcomes over time.
B C:
B 3 0:354574 0 0:0322799 0:511431 0 0:95 1 0 0 2 f4;5g C 19
B C Our choice of a small probability here is consistent with a belief that
B C
@ 4 0:0897837 0 0:036227 0:525595 1 0:95 1 0 0 2 f1;3g A individuals’ attitudes regarding the acceptability of fraud are relatively
5 0:496052 0 0:0205544 0:892721 0 0:95 1 0 0 1 f3g stable in response to non-organizational factors. As noted by the AICPA
(2002) in SAS 99, the ability to rationalize can be affected, in part, by one’s
attitude, character, and set of ethical values (which we believe to be stable
In the matrix we can see that although agents 1 and 4 now with only a small likelihood of change over time). This is the aspect of
rationalization captured by our small exogenous probability p. In contrast,
hold a pro-fraud attitude (A = 1) and both perceive the the likelihood of being influenced via interactions in the organization, q,
opportunity to commit fraud (O = 1), neither has motive captures a second source of rationalization noted in SAS 99, the organiza-
(M = 0) and therefore neither commits fraud. tional environment.
Emergence: Two states which few agents are fraudsters to a period in which virtu-
ally everyone is a fraudster. We find that the characteristic
The benchmark analysis studies the effect of a system- U shape and sudden shifts between extreme states persist
atic change in individuals’ tendencies to be impacted by across conditions ranging from very high average emula-
the behavior of others in the organization (emulation like- tion probabilities down to approximately 30%.22 The dra-
lihood, q). We explore the entire parameter range by vary- matic shifts in behavior arise from the overwhelming
ing the average emulation likelihood in increments of 5%, effect of emulation relative to the random effects introduced
from q = 0.05 to 0.95.20 Each agent is initially assigned an by other parameters in the model.
independent probability drawn from a uniform distribution At low average emulation likelihoods (30% and below),
between a minimum and a maximum value for q. The as- a very different state distribution emerges in every organi-
signed probability is static for each agent. Since the likeli- zation that we observe. Instead of tending toward ex-
hood is bounded, as we approach the extremes we tremes, the population tends toward a specific level of
manipulate either the minimum or the maximum to reach fraud. The shape of the associated histogram is an inverted
the desired average.21 U, as illustrated in Fig. 1, Panel B. Notice in this example,
Our analysis focuses on both state distributions in the roughly half of the population in the majority of periods
form of histograms illustrating the number of fraudsters are fraudsters. When we consider the change in fraud
in the population that emerge and on time series graphs behavior over time in Fig. 1, Panel D, we see that the rapid
illustrating the change in the number of fraudsters over extreme shifts in the number of fraudsters found in the U-
time, with an eye toward the periodic existence of fraud shaped distribution disappear for the most part. Instead,
outbreaks. Each histogram shows the number of periods we see smaller (noisy) movements around a specific num-
during which n agents are fraudsters, where n ranges from ber of fraudsters.
0 to 100 (the entire organization). In Fig. 1, Panels A and B, We perform additional analysis at parameter values
for example, the y-axis represents the number of periods in around the transition from a U-shaped organization to an
which a specific number of fraudsters is observed, while inverted U-shaped organization to evaluate whether the
the x-axis represents the number of fraudsters. Panels C change in state is gradual or a bifurcation point exists.
and D of Fig. 1 illustrate changes in the number of fraud- Our analysis indicates that the change is gradual. As q de-
sters over time. In each time series, the y-axis represents creases, the population spends less time at the extremes
the number of fraudsters in a given period and ranges from and more time at moderate levels of fraud; the height of
0 to 100 and the x-axis represents time (or interaction peri- the upper and lower supports of the U shape in the histo-
ods) and ranges from 1 to 15,000. gram decrease and the middle range becomes thicker until
Two state distributions emerge in our analysis. At mod- organizations reach a point where the state distribution is
erate and high emulation likelihoods, the majority of peri- flat (all fraud outcomes are equally likely over time). As q
ods exhibit extreme behavior in all 10 organizations, continues to decrease, the population tends toward a spe-
where either virtually every agent is a fraudster or virtually cific number of fraudsters, leading to an inverted U shape
no agents are fraudsters. We call this distribution U shaped in the histogram. The location of the transition from U to
after the appearance of its associated histogram, shown in inverted U differs slightly across different organizations
Fig. 1, Panel A. A typical time series associated with U- and depends on the relative values of p (likelihood of
shaped distributions is illustrated in Fig. 1, Panel C. The changing attitude in response to exogenous factors) and
time series is characterized by rapid swings between ex- q (likelihood of emulating the attitude of others in the
tremes. For example, in Panel C, at approximately the organization). As p increases, increasingly large values of
11,000th pairing, a rapid shift occurs from a period in q are required to generate a U-shaped state distribution.
High emulation likelihoods within the organization rela-
tive to external influences on attitude facilitate cascades
20
While we have no direct evidence regarding the level of q in real world from mostly honest to mostly fraudster populations and
organizations, research by Milgram (1965) suggests that the level of q
vice versa.
across organizations may vary. In the study, the extent of obedience to
authority (one aspect of emulation likelihood) was investigated at Yale and In Table 2 we provide descriptive statistics (mean and
at an office building in Bridgeport, Connecticut. Subjects were more standard deviation) to illustrate the impact of interven-
obedient at Yale; 65% were fully obedient versus 48% at the office in tions on the statistical properties of fraud in our models.
Bridgeport. The study (and others conducted by Milgram) suggests that the
For the benchmark analysis, the table shows a reduction
effectiveness of efforts to recruit (q) in our model may depend in important
ways on the organizational setting (e.g., background authority inherent in
in the average standard deviation across organizations as
the organization, etc.). q is reduced, while the mean remains relatively constant.
21
For example, if the maximum likelihood is set at 100% and the This pattern is consistent with the change in state distribu-
minimum at zero, the average likelihood is 50%. Holding the floor constant tions noted previously.
at zero as the ceiling is reduced from 100%, the average will drop below
The possibility of two states of fraud dynamics observed
50%. To examine the parameter space where averages exceed 50%, we
increase the floor and hold the ceiling constant at 100%. Since this approach in our benchmark model raises an additional question: Do
to manipulating the mean value of q also varies the range from which the efforts to prevent or detect fraud have differential effects
value of q is drawn, we performed additional analysis using a variety of on the two fraud dynamics observed? We now turn our
mean-preserving ranges. Our results are qualitatively similar (both time
series patterns and state distributions of fraud over time are consistent) in
22
both the mean-preserving and mean-shifting analyses, with the exception Unless noted otherwise, the results are consistent and clearly
that a U-shaped distribution is observed more frequently when mean- observable across all organizations and all initial fraud distributions in
preserving ranges are reduced. the populations.
Paneel A: Hiistoggramm whhen aaverrage emuulatiion likellihoood (q) Paanel B: H Histoograam w

whenn aveeragge emmulaationn likeelihoood (q)
is 0.995 (U
U-shhapeed orrgannization)). U-shhapeed orrgannizatiion).
is 0.055 (innvertted U
PPaneel C:: Timme sseriees off agggregaate ffraudd whhen aaverrage Paanel D: T Timee serries of aaggreegatte fraaud wheen avveragge eemullation
eemuulatioon liikelihhoodd (q) is 00.955 (U--shaaped organizzatioon). U-shaapedd orgganizzatioon).
likkelihhoodd (q)) is 00.05 (invverteed U
Fig. 1. Illustrative results from analysis of a single organization in the benchmark model, including histogram of number of fraudsters and associated time
series. Panel A: Histogram when average emulation likelihood (q) is 0.95 (U-shaped organization). Panel B: Histogram when average emulation likelihood
(q) is 0.05 (inverted U-shaped organization). Panel C: Time series of aggregate fraud when average emulation likelihood (q) is 0.95 (U-shaped organization).
Panel D: Time series of aggregate fraud when average emulation likelihood (q) is 0.05 (inverted U-shaped organization).
attention to the impact of attempts to prevent or eliminate trated by comparing the benchmark time series in Fig. 1,
fraud in the organization. Panel C to the time series in Fig. 2, Panel A. The associated
state distribution histogram is illustrated in Fig. 2, Panel B.
Perceived opportunity As opportunity decreases, the upper support of the U shape
is lowered in the histogram.
We begin our investigation into the effectiveness of The same result emerges when we consider low emula-
fraud prevention mechanisms by systematically manipu- tion likelihood organizations. Relative to the benchmark
lating perceived opportunity. Our analysis examines analysis, the qualitative characteristics of each organiza-
opportunity likelihood parameter values from 90% to 50% tion’s time series (the frequency, duration and location of
in 10% increments. For each value of p(O), we investigate peaks and valleys) are unchanged, but the maximum fraud
the entire emulation likelihood (q) range, focusing our level drops as p(O) is reduced (in Table 3, for organizations
attention on the two state outcomes (U-shaped and in- where q = 0.05, the maximum level of fraud drops from
verted U-shaped distributions) identified during the 91% in the benchmark model to 58% when p(O) = 0.5). A
benchmark analysis. time series for a representative organization is illustrated
When we evaluate the effect of reducing opportunity in in Fig. 2, Panel C. In Table 2, both the mean and the stan-
U-shaped organizations (with moderate and high average dard deviation of fraud are reduced as perceived opportu-
emulation levels), most of the time series characteristics nity is reduced. The inverted U state distribution retains its
are unchanged. We observe the same extreme changes be- basic shape, but it is shifted to the left and becomes more
tween honesty and fraud at the same points in time within peaked (consistent with the reduced mean and standard
an organization. The frequency of the peaks and valleys are deviation). An illustrative histogram is presented in
the same relative to the benchmark time series for each Fig. 2, Panel D.
organization, as is the duration spent in each state. How- Thus, overall, we find that reducing opportunity in the
ever, as indicated in Table 3, when perceived opportunity organization shifts the average fraud level closer to zero
is reduced, we observe a reduction in the maximum level and reduces the standard deviation of fraud, but it does
of fraud (e.g., in the q = 0.95 condition, the maximum level not otherwise affect the shape of the distribution of fraud
of fraud changes from 100% when p(O) = 1.0 to 68 percent over time or the nature of fraud dynamics. The two emer-
when p(O) = 0.50). The effect on the time series is illus- gent states observed in our benchmark model persist and
Table 2
Mean percentage (standard deviation) of agents who commit fraud within 10 organizations, where agent attitude (pro-fraud versus anti-fraud) is randomly
assigned in the initial time period.
Mean emulation likelihood (q) 0.05 0.25 0.50 0.75 0.95

Benchmark 50.4 (13.5) 48.1 (22.9) 52.4 (26.6) 47.8 (34.8) 51.0 (35.3)
Opportunity (p(O))
0.90 45.4 (12.4) 43.3 (20.1) 47.2 (24.1) 43.0 (31.4) 45.9 (31.8)
0.80 40.3 (12.4) 38.5 (15.9) 41.0 (20.4) 38.3 (28.0) 40.8 (28.4)
0.70 40.3 (11.2) 33.7 (15.9) 36.7 (19.0) 33.5 (24.6) 35.7 (24.9)
0.60 30.2 (8.8) 28.9 (13.8) 31.5 (16.4) 28.7 (21.0) 30.6 (21.5)
0.50 25.2 (7.6) 24.1 (11.6) 26.2 (13.8) 23.9 (17.7) 25.5 (18.0)
Tone at top
1 manager 42.5 (12.8) 33.6 (17.8) 30.1 (20.2) 15.2 (17.5) 15.4 (18.2)
2 managers 38.5 (11.6) 25.8 (14.8) 21.3 (15.8) 8.8 (11.2) 9.3 (12.1)
4 managers 32.3 (10.4) 17.2 (10.9) 13.2 (11.0) 5.0 (7.0) 4.9 (7.0)
6 managers 27.1 (9.4) 13.7 (8.9) 9.9 (8.7) 3.4 (5.0) 3.4 (5.2)
Asymmetric influence (c/(c + 1))
0.99 47.0 (13.0) 43.9 (22.3) 44.3 (26.6) 33.7 (32.4) 27.9 (30.1)
0.95 44.4 (12.3) 33.0 (19.5) 23.5 (19.0) 7.5 (10.7) 5.5 (8.2)
0.90 37.9 (12.0) 20.4 (13.2) 12.7 (11.3) 3.3 (5.2) 2.4 (4.0)
0.85 31.6 (10.5) 14.7 (10.4) 8.6 (8.0) 2.1 (3.5) 1.5 (2.8)
0.80 27.0 (9.6) 10.7 (8.0) 6.5 (6.1) 1.5 (2.7) 1.1 (2.1)
0.75 24.6 (8.9) 8.6 (6.3) 5.3 (4.9) 1.2 (2.1) 0.8 (1.7)
Termination likelihood (e)
0.01 28.4 (11.0) 31.9 (19.8) 33.2 (25.4) 39.2 (34.0) 36.6 (33.8)
0.05 7.4 (4.7) 8.7 (7.9) 9.2 (10.3) 11.7 (17.0) 12.4 (17.4)
0.10 3.7 (3.0) 4.2 (4.5) 4.5 (5.7) 5.3 (7.8) 5.3 (8.2)
0.15 2.3 (2.4) 2.8 (3.3) 2.9 (4.1) 3.5 (5.5) 3.7 (5.8)
0.20 1.7 (1.9) 2.1 (2.7) 2.2 (3.2) 2.6 (4.2) 2.7 (4.4)
Table 3
Maximum (minimum) fraud level observed across all organizations in the final 10,000 time steps by level of emulation likelihood for the benchmark and model
variants most likely to impede fraud.
Mean emulation likelihood (q) 0.05 0.25 0.50 0.75 0.95

Benchmark 91 (9) 100 (0) 100 (0) 100 (0) 100 (0)
Opportunity (p(O))
0.50 58 (1) 68 (0) 67 (0) 68 (0) 68 (0)
Tone at top
6 managers 64 (2) 60 (0) 64 (0) 52 (0) 70 (0)
Asymmetric influence (c/(c + 1))
0.75 68 (0) 48 (0) 44 (0) 25 (0) 20 (0)
Termination likelihood (e)
0.20 13 (0) 33 (0) 38 (0) 50 (0) 65 (0)
the extreme swings between honesty and widespread managers and staff. We assume that managers are more
fraud in U-shaped organizations remain. The results of likely than staff to influence the attitudes of others.
our analysis suggest that, while reducing opportunity can We introduce a hierarchy by adding a status identifier
limit the extent of fraud in an organization, firm culture to each agent vector in the benchmark model. Agents are
and individual susceptibility to influence continue to play randomly identified as managers in the organization prior
an important role in the nature of fraud dynamics. to the first period of agent interaction. Agents identified as
managers are assigned a hierarchical status identifier equal
to two (resulting in twice the impact on emulation likeli-
Tone at the top
hood),23 while the remaining agents are considered staff
and assigned a hierarchical status identifier equal to one.
Tone at the top is often noted as an important charac-
We use this identifier to scale emulation likelihoods during
teristic in organizations for controlling fraud (Association
of Certified Fraud Examiners, 2007). To better understand
the impact of tone at the top on dynamics in our model, 23
We tested the sensitivity of the emulation likelihood scalar introduced
we begin by extending our benchmark analysis to create when interacting with a manager by increasing it from two to four. No
a hierarchical organization with two levels of employees: qualitative effect was observed on aggregate outcomes.
Fig. 2. Representative histograms and time series of aggregate fraud levels when perceived opportunity, p(O), equals 0.50. Panel A: Time series when
average emulation likelihood (q) is 0.95 (U-shaped organization). Panel B: Histogram when average emulation likelihood (q) is 0.95 (U-shaped
organization). Panel C: Time series when average emulation likelihood (q) is 0.05 (inverted U-shaped organization). Panel D: Histogram when average
emulation likelihood (q) is 0.05 (inverted U-shaped organization).
agent interactions. For example, assume agent 1 is assigned the organization (span of control) and investigate the impact
an emulation likelihood of 0.15 when the organization is of this manipulation over the range of emulation likelihood
formed. In our benchmark model, when agent 1 observes an- (q) values (as in the benchmark model).
other agent, agent 1 will change her attitude to match the We find that a monolithic, honest management team in
observed agent’s attitude with a 15% probability. However, a U-shaped organization changes the characteristic shape
in our hierarchical model the likelihood of emulating an- of the state distribution. With even one honest manager,
other agent is doubled, to 30%, when the agent observed is we see the tendency to achieve maximum levels of fraud
a manager.24 with high frequency disappear (the upper tail in the U
The notion of tone at the top incorporates more than a shape is gone, as illustrated in Fig. 3, Panel A).26 A clearer
hierarchical organization with influential management. It picture of the qualitative effect of the manipulation is pro-
is usually characterized as a management team with a con- vided by an examination of the dynamics of fraud over time.
sistent message regarding ethical values and appropriate While there is an overall tendency toward honesty in orga-
behavior. To capture this idea, all of the managers in the nizations with a positive tone at the top, the time series
model are identical. Each is assigned an anti-fraud attitude exhibits short-lived outbreaks of fraud in the organization
in the initial period of the simulation. To ensure their atti- (see Fig. 3, Panel B). This remains the case regardless of
tudes do not change, the managers’ likelihood of emulating the number of managers we examined. The large influence
another agent (q) and likelihood of spontaneous attitude exerted by management in U-shaped organizations can be
change (p) are both set to zero.25 For the purposes of our attributed to both the consistent behavior exhibited by man-
analysis, we make no other modifications relative our agement and the high level of susceptibility to influence
benchmark. Motive and opportunity remain constant, at exhibited in the agent population.
100%. We systematically vary the number of managers in To complete our analysis of a monolithic, honest man-
agement team, we consider its impact on the aggregate
fraud outcomes and dynamics when average emulation
24
Since q is bounded (q 6 1), managers can only increase another agent’s (q) levels are low (i.e., inverted U-shaped organizations).
emulation likelihood parameter to 1, but no higher.
25
We also evaluated a hierarchical model using non-identical managers
26
with attitudes that can change to investigate the effect of introducing a Notably, when a monolithic, dishonest management team is introduced
hierarchy. The introduction of a hierarchy by itself had no qualitative into the model, the mirror image of the behavior exhibited here is observed.
impact relative to the benchmark model. The lower support is eliminated and fraud becomes the prevalent behavior.
Panel A: Histogram for U-shaped organizations illustrating Panel B: Time series for U-shaped organizations illustrating
loss of upper support in tone at the top, asymmetric fraud outbreaks under tone at the top and termination
influence, and termination model variants. model variants.
Panel C: Time series of aggregate fraud levels exhibiting no fraud outbreaks in U-shaped organizations in the asymmetric influence
model variant and in inverted U-shaped organizations in the termination model variant.
Fig. 3. Other histograms and time series of aggregate fraud levels observed in the analysis of model variants. Panel A: Histogram for U-shaped organizations
illustrating loss of upper support in tone at the top, asymmetric influence, and termination model variants. Panel B: Time series for U-shaped organizations
illustrating fraud outbreaks under tone at the top and termination model variants. Panel C: Time series of aggregate fraud levels exhibiting no fraud
outbreaks in U-shaped organizations in the asymmetric influence model variant and in inverted U-shaped organizations in the termination model variant.
We find that, while introducing a monolithic, honest man- vention would be applied across the entire population.
agement team reduces mean fraud level and its associated During agent interaction, the scalar changes the likelihood
standard deviation (as can be seen in Table 2), it has no of emulating a fraudster relative to an honest agent in the
other effect on either the inverted-U shape of the histo- following manner:
gram or characteristics of the related time series. Further-
more, increasing the relative size of the management team Q ¼ qc=ðc þ 1Þ when the agent observes another agent
does not alter this general finding. with a pro-fraud attitude and
Q ¼ q otherwise;
Asymmetric influence
where Q is the emulation likelihood in the asymmetric
Some interventions are aimed at enhancing the ethical influence model and q is the emulation likelihood in the
tendencies of employees. For example, organizations may benchmark model. To illustrate the application of this rule
engage in ethical training or encourage personal adherence set, assume agent A’s emulation likelihood (q) is 0.50 and
to a code of ethics. A clearly communicated emphasis on the scalar (c) is 9 in the organization. When agent A ob-
ethical behavior could have effects on both fraudsters serves another agent with a pro-fraud attitude, he will
and others in the organization. Non-fraudsters would be have a 45% likelihood of emulating the attitude. However,
better informed about what actions constitute fraud and if agent A observes an anti-fraud attitude in another agent,
engage in more careful thought before conforming with he will have a 50% likelihood of emulating the attitude.
undesirable norms (rather than thoughtless conformity to Thus, the smaller the value of c, the higher the likelihood
norms). Fraudsters would feel less comfortable in their that an agent will emulate an honest agent relative to a
recruitment efforts and less likely to reveal their attitudes fraudster, and the more effective the ethics intervention.
to others. To model these effects, we allow asymmetric To evaluate this version of our model, we systematically
influence, where honest employees exert more influence vary c at six levels (99, 19, 9, 5.7, 4 and 3)27 for every
on their peers than fraudsters. To achieve asymmetry, we
add a scalar (c) to our benchmark model. The scalar is con- 27
The levels of c were chosen to scale the influence of pro-fraud relative
stant across all agents because we assume that the inter- to anti-fraud attitudes to 99%, 95%, 90%, 85%, 80% and 75%, respectively.
emulation likelihood level used in the benchmark model so for the number of fraudsters in the organization, N is the
that the effect of asymmetric influence can be assessed number of agents in the organization (N = 100), and f is
across both U-shaped (moderate to high emulation likeli- the number of fraudsters in the organization at the start
hood) organizations and inverted U-shaped (low emulation of the simulation period.29
likelihood) organizations. We compare the results of these When a fraudster is discovered and replaced, our termi-
parameter sweeps to matched (identical) starting organiza- nation model assigns new values for all agent attributes
tions examined in our benchmark model. (using the process described in the benchmark model),
For U-shaped organizations, asymmetric influence dra- with the exception of the agent identifier (ID), social net-
matically affects both the shape of the state distribution work (SN), and social network size (S).30 These three attri-
and the dynamics of fraud behavior over time. As the asym- butes remain unchanged under the assumption that when
metry in influence is increased (by decreasing the scalar c), an agent is replaced, the new agent will interact with the
the upper support of the U shape rapidly disappears. The same individuals as the previous one. Our analysis of the ter-
resulting state distribution resembles that observed in the mination model evaluates the effect of a change in detection
tone at the top condition, in Fig. 3, Panel A.28 However, and termination likelihood (e) over the entire range of emu-
the periodic fraud outbreaks observed in the tone at the lation likelihood values examined in the benchmark model
top condition are completely eliminated. As can be seen in to investigate the impact of termination on our two classes
Table 3, the maximum level of fraud observed when of organizations (U shaped and inverted U shaped). In our
q = 0.95 is 20 at the highest level of asymmetric influence, analysis, e is examined at five levels: 0.01, 0.05, 0.10, 0.15
compared to 100 in the benchmark model. The resulting time and 0.20.
series is illustrated in Fig. 3, Panel C. Intuitively, when emu- When we consider organizations with moderate to high
lation likelihoods are high and asymmetric, the number of average emulation likelihoods (U-shaped organizations),
honest agents in the population will continue to grow until we find that the introduction of a termination regime leads
there are a sufficient number of honest agents to prevent to results similar to those observed in the tone at the top
an outbreak. A very different result emerges when we ob- model. The upper support of the U-shaped state distribu-
serve the impact on the inverted-U distribution found in tion is gone (a tendency toward low levels of fraud), while
organizations with low average emulation likelihoods. As periodic fraud outbreaks persist in the time series (see Pan-
asymmetry in influence increases, the mean level and stan- els A and B of Fig. 3). For organizations with low emulation
dard deviation of fraud decreases, but the inverted-U shape likelihoods, we find the first instance of an intervention
of the state distribution and the qualitative characteristics triggering a dramatic change in the inverted-U shape of
of the time series in the benchmark model are preserved. the state distribution. As the likelihood of detection and
The results suggest that the role of ethical interventions in termination increases, the level of fraud approaches zero
fraud prevention is contingent on the nature of the organiza- (the state distribution resembles Fig. 3, Panel A) and fraud
tion, a relation not previously recognized in the literature outbreaks are not observed in the time series (see Fig. 3,
that could prove important in fraud risk assessment. Panel C). Our analysis suggests that detection and termina-
tion efforts can be effective at reducing average fraud lev-
Removing fraudsters from the organization els to near zero in both types of organizations (U and
inverted U shaped). However, the effectiveness of such ef-
Our investigation so far has focused on organizational forts at preventing widespread fraud outbreaks appears
interventions associated with fraud prevention. We now contingent on the type of organization and related individ-
turn our attention to the effect of detecting and eliminat- ual susceptibilities to fraud.
ing fraud in the organization through the termination of
fraudsters. To model termination, we assume that fraud- Discussion and conclusions
sters will be detected with some probability t(e). Once de-
tected, agents are removed from the organization We use an ABM to investigate the dynamics of fraud
(‘‘terminated’’) and replaced with a new agent. within organizations and the impact of fraud prevention
We expect the existence of budget constraints with re- and detection mechanisms. Our model consists of an orga-
gard to termination efforts. As a result, we assume fraud nization, agents (employees), and a set of simple social
detection and termination programs will be more success- interaction rules. In accordance with Cressey’s (1953)
ful in organizations with lower levels of fraud. At higher occupational fraud model, any agent within our model
fraud levels, we anticipate that collusion among employees possessing the conjunction of motive, opportunity, and a
will impede such efforts. Therefore, in each period, we pro-fraud attitude commits fraud. We allow agents to
determine the likelihood of detection and termination as repeatedly interact, resulting in changes in agent attitudes
tðeÞ ¼ eððN f Þ=NÞ; 29

Presumably, the introduction of a termination regime might also
where t(e) is the likelihood of detection and termination, e decrease agents’ perceived opportunity or motivation to commit fraud.
is the likelihood of detection and termination unadjusted However, for the purposes of our analysis, we investigate effects indepen-
dently as an initial step toward understanding our model.
30
While we assume newly hired agents are randomly selected from the
28
If our asymmetric influence model were re-conceptualized to examine general population, one could also model an organization where hiring is
a setting where fraudsters have more influence that honest agents, the biased toward a particular type of agent. For example, an organization of
lower support in U-shaped organizations would be eliminated, with fraud fraudsters might be more likely to hire another fraudster (or someone who
becoming the dominant behavior. believes fraud to be acceptable behavior).
Table 4
Summary of findings.
Intervention Effect on U-shaped organization Effect on inverted U-shaped organization

Perceived opportunity Upper support lowered Average fraud level reduced
Dynamics preserved Dynamics preserved
Spontaneous outbreaks persist
Tone at the top Upper support removed Average fraud level reduced
Spontaneous outbreaks persist Dynamics preserved
Asymmetric influence Highly effective Average fraud level reduced
Upper support removed Dynamics preserved
Stops spontaneous outbreaks
Termination Upper support removed Highly effective
Spontaneous outbreaks persist Fraud reduced to near zero
No spontaneous outbreaks
toward fraud (via cultural transmission). We find that two and fraud dynamics are not affected under the interven-
types of organizations emerge, depending on how suscep- tion.31 Every other intervention affects the dynamics of
tible individual agents are to social influence. When aver- fraud. In every case, the effects on dynamics can be traced
age susceptibility to influence is low, a state distribution to the interaction of each intervention with q. Ethical inter-
emerges where aggregate fraud levels tend toward a spe- ventions (asymmetric influence) directly impact the value of
cific level of fraud. When average susceptibility is moder- q. Tone at the top affects q for some interactions (those be-
ate to high, we observe an outcome characterized by tween a manager and staff) and termination introduces new
extreme swings in behavior from an overall honest organi- agents (each with a new q) into the organization. Since each
zation to an overall fraudulent one. intervention affects q differently, they each have a different
After identifying these two types of organizations, we impact on fraud dynamics.
turn our attention to the effect of mechanisms to prevent, Our findings have important implications for auditors
detect, and eliminate fraud. We examine the effects of and other individuals responsible for assessing fraud risk
decreasing perceived opportunity to commit fraud, tone and detecting and preventing fraud. First, for certain types
at the top, asymmetric influence (resulting from ethical of organizations aggregate fraud levels can vary tremen-
interventions), and the removal of fraudsters from the dously over time. Furthermore, the effectiveness of mech-
organization. We find the effectiveness of these mecha- anisms to prevent and detect fraud can be contingent on
nisms is contingent on the type of organization. Our find- the type of organization and related individual susceptibil-
ings are summarized in Table 4. ities to social influence. Therefore, it may be inappropriate
The model suggests that ethical interventions aimed at for auditors to evaluate fraud prevention and detection
increasing the influence of honest employees relative to mechanisms in a uniform manner. Our results suggest that
fraudsters are particularly effective for moderate to high the same fraud prevention and detection mechanisms
average emulation likelihood organizations. However, eth- implemented in a similar manner in two different organi-
ical interventions are observed to be much less effective for zations cannot be expected to be equally effective without
organizations with low emulation likelihoods. A different considering the average susceptibilities to social influence
result emerges when we consider the impact of efforts to of the individuals therein.32 Similarly, some mechanisms
detect and remove fraudsters from the organization. Fraud can appear effective most of the time but still be ineffective
is reduced to near zero in both types of organizations when at preventing outbreaks of fraud. In general, there is no one-
the likelihood of terminating fraudsters is increased to size-fits-all fraud prevention (and/or detection) mechanism
0.20. However, increasing the likelihood of termination and fraud risk may be contingent on individual susceptibil-
does not prevent occasional outbreaks of widespread fraud ities to social influence in the organization.
in U-shaped organizations. The introduction of a mono-
lithic, honest management team (tone at the top) effec-
tively reduces the amount of fraud in both types of 31
The behavioral response to varying motivation in agents is the same as
organizations. Furthermore, while it eliminates the ten- the behavioral response to changes in opportunity because, like opportu-
dency to achieve maximum levels of fraud with high fre- nity, motivation is independent of q. Consequently, decreasing motivation
quency in U-shaped organizations, it does not eliminate decreases the average level of fraud. In the case of U-shaped organizations,
sporadic, short-lived fraud outbreaks. Reducing opportu- the decrease is exhibited through a reduction in the maximum level of the
upper support in the histogram. In the case of inverted U-shaped
nity (e.g., through stronger internal controls) merely re-
organizations, the distribution is simply shifted toward the origin. Because
duces the average mean level of fraud and associated both motivation and opportunity are independent of q, we do not expect
standard deviation within an organization. In our model, interaction effects between these variables and our model variants that do
reduced opportunity does not affect other aspects of fraud affect q.
32
dynamics, regardless of emulation likelihood levels. We conjecture that it might be possible to develop an instrument to
measure differences in q across organizations (e.g., see Bearden, Netemeyer,
The intuition behind these results lies in the relation be- & Teel, 1989) and that such an instrument might be useful to auditors in
tween each intervention and the emulation likelihood, q. their assessment of fraud risk and the relative effectiveness of anti-fraud
Reduced opportunity has no impact on q in our model interventions.
While the ABM method used in our investigation con- adding a ‘‘dollars available’’ variable to the model. Perhaps
fers a number of advantages over traditional models in this variable could be linked to one’s position in the orga-
the social sciences, it presents limitations as well. First, nization (manager versus staff). Finally, while our model
seemingly innocuous assumptions have sometimes been allows for selection in the termination intervention, it does
shown to have unexpected consequences that are not yet not select on the basis of relative fitness (allowing those
fully understood (e.g., Huberman & Glance, 1993). In the more skilled at fraud to avoid detection). Our model could
context of our model, additional research might examine be extended to investigate fraud in this evolutionary
alternative updating rules (e.g., one alternative model context.
might randomly select an agent and then randomly deter-
mine whether the agent will interact with another agent or Acknowledgements
make a decision about committing fraud). The results we
report may also be sensitive to other modeling choices. The authors acknowledge financial support provided
For example, we use Cressey’s (1953) fraud triangle to rep- the American Institute of Certified Public Accountants
resent the agent’s decision to commit fraud. Other formu- (AICPA) Center for Audit Quality. The paper benefited from
lations are possible, such as the fraud diamond (Wolfe & comments provided by David Piercey and participants at
Hermanson, 2004) or perhaps a utility-maximizing ap- the Queens University Fraud Conference, the University
proach. Our characterization of a social network assumes of Illinois Audit Symposium, and the Fraud in Accounting,
a static organization with uniform mixing for the sake of Organizations and Society Conference. We also acknowl-
simplicity. An alternative model might assume a dynamic edge comments from two anonymous reviewers.
network with local mixing or local mixing together with
random connections among agents. Other characteriza- References
tions of social influence (e.g., normative social influence
in lieu of paired recruitment) could also have an impact. Adams, J. (1965). Inequity in social exchange. Advances in Experimental
Social Psychology, 62, 335–343.
The second limitation associated with ABM is that the solu-
American Institute of Certified Public Accountants Auditing Standards
tion concepts employed tend to be relatively weak and, to Board (2002). Consideration of fraud in a financial statement audit.
date, there has been little formal work on the replicability Statement on auditing standards. New York: American Institute of
of results using different models (but see Axtell, Axelrod, Certified Public Accountants.
Association of Certified Fraud Examiners (2007). Fraud examiners manual.
Epstein, and Cohen (1996) for an exception). Finally, ABM Austin, TX: Association of Certified Fraud Examiners.
still lacks a set of standard, generally accepted practices. Association of Certified Fraud Examiners (2010). Report to the nations on
Our initial investigation of fraud leads to a variety of occupational fraud and abuse. Austin, TX: Association of Certified
Fraud Examiners.
additional opportunities for future research. Our analysis Axelrod, R. (1997). Advancing the art of simulation in the social sciences.
examines specific changes to our model in isolation. Our In R. Conte, R. Hegselmann, & P. Terna (Eds.), Simulating social
design choice was driven by a desire to develop a deeper phenomena (pp. 21–40). Berlin: Springer.
Axtell, R., Axelrod, R., Epstein, J., & Cohen, M. (1996). Aligning simulation
understanding of the effects of various approaches taken models: A case study and results. Computational and Mathematical
to control fraud; however, in organizations, specific inter- Organization Theory, 1(2), 123–141.
ventions are seldom implemented alone. The relative effi- Bearden, W., Netemeyer, R., & Teel, J. (1989). Measurement of consumer
susceptibility to interpersonal influence. The Journal of Consumer
cacy of combining different interventions could be Research, 15(4), 473–481.
investigated. Beasley, M. (1996). An empirical analysis of the relation between the
A more detailed examination of each leg of the fraud tri- board of director composition and financial statement fraud. The
Accounting Review, 71(4), 443–465.
angle could be undertaken. Future research might refine
Bell, T., & Carcello, J. (2000). A decision aid for assessing the likelihood of
our characterization of motivation by explicitly consider- fraudulent financial reporting. Auditing: A Journal of Practice and
ing social psychological factors in the organization. Equity Theory, 19(1), 169–184.
theory (Adams, 1965) suggests that an honest individual in Breiger, R., & Carley, K. (2003). Dynamic social network modeling and
analysis: Workshop summary and papers. Washington, DC: National
an organization replete with fraudsters would be more Academies Press.
likely to have a motivation to commit fraud in order to Carpenter, T. (2007). Audit team brainstorming, fraud risk identification,
reestablish fairness in their relationships with co-workers and fraud risk assessment: Implications of Statement on Auditing
Standards No. 99. The Accounting Review, 82(5), 1119–1140.
and the organization. Our model could be modified to Chang, J., & Lai, C. (2002). Is the efficiency wage efficient? The social norm
incorporate this effect by making motivation a function and organizational corruption. Scandinavian Journal of Economics, 104,
of the number of fraudsters in the agent’s social network. 27–47.
Cleary, R., & Thibodeau, J. (2005). Applying digital analysis using Benford’s
Our model ignores ancillary changes in motive, oppor- law to detect fraud: The dangers of type I errors. Auditing: A Journal of
tunity, and attitude that might result from the interven- Practice and Theory, 24(1), 77–82.
tions that we investigate (or from social changes and Cressey, D. (1953). Other people’s money: A study in the social psychology of
embezzlement. Belmont, CA: Wadsworth Publishing.
other forms of regulation). For example, our model’s repre-
Davis, J., Hecht, G., & Perkins, J. (2003). Social behaviors, enforcement, and
sentation of the detection and removal (termination) of tax compliance dynamics. The Accounting Review, 78(1), 39–69.
fraudsters ignores the deterrent effect that can be gener- Epstein, J. (1999). Agent-based computational models and generative
social science. Complexity, 4, 41–60.
ated by an active termination program and concomitant ef-
Epstein, J. (2006). Generative social science: Studies in agent-based
fects on motivation. Future research could extend our computational modeling. Princeton, NJ: Princeton University Press.
study to incorporate these correlated effects. Our model Epstein, J., & Axtell, R. (1996). Growing artificial societies: Social science
also ignores the magnitude of fraud (dollars lost). Instead, from the bottom up. Washington, DC: The Brookings Institution.
Erickson, M., Hanlon, M., & Maydew, E. (2006). Is there a link between
we focus on the number of fraudsters in the organization. executive equity incentives and accounting fraud? Journal of
Insights on the magnitude of fraud might be gained by Accounting Research, 44(1), 113–143.
Finley, D. (1994). Game theoretic analysis of discovery sampling for Milgram, S. (1963). Behavioral study of obedience. The Journal of Abnormal
internal fraud control auditing. Contemporary Accounting Research, and Social Psychology, 67(4), 371–378.
11(1), 91–114. Milgram, S. (1965). Some conditions of obedience and disobedience to
Gaylord, R., & Davis, J. (1999). Modeling nonspatial socioeconomic authority. Human Relations, 18(1), 57–76.
interactions: An agent-based approach. Mathematica in Education Miller, J., & Page, S. (2007). Complex adaptive systems: An introduction to
and Research, 8(2), 30–33. computational models of social life. Princeton, NJ: Princeton University
Gillett, P., & Uddin, N. (2005). CFO intentions of fraudulent financial Press.
reporting. Auditing: A Journal of Practice and Theory, 24(1), 55–75. Pinto, J., Leana, C., & Pil, F. (2008). Corrupt organizations or organizations
Gode, D., & Sunder, S. (1993). Allocative efficiency of markets with zero- of corrupt individuals? Two types of organization-level corruption.
intelligence traders: Market as a partial substitute for individual Academy of Management Review, 33(3), 685–709.
rationality. Journal of Political Economy, 101(1), 119–137. Ragan, R. (2010). An embarrassment of riches: Parameter choice in agent-
Grimm, V., Revilla, E., Berger, U., Jeltsch, F., Mooij, W., Railsback, S., et al. based models. Unpublished manuscript. Department of Political
(2005). Pattern-oriented modeling of agent-based complex systems: Science, Duke University.
Lessons from ecology. Science, 310, 987–991. Schneider, A., & Wilner, N. (1990). A test of audit deterrent to financial
Hoffman, V., & Zimbelman, M. (2009). Do strategic reasoning and reporting irregularities using the randomized response technique. The
brainstorming help auditors change their standard audit procedures Accounting Review, 65(3), 668–681.
in response to fraud risk? The Accounting Review, 84(3), 811–837. Schubert, S., & Miller, T. (2008). Where bribery was just a line item. The
Holtfreter, K. (2005). Is occupational fraud ‘‘typical’’ white-collar crime? A New York Times (December 21).
comparison of individual and organizational characteristics. Journal of Sharma, V. (2004). Board of director characteristics, institutional
Criminal Justice, 33(4), 353–365. ownership, and fraud: Evidence from Australia. Auditing: A Journal of
Huberman, B., & Glance, N. (1993). Evolutionary games and computer Practice and Theory, 23(2), 105–117.
simulations. Proceedings of the National Academy of Sciences of the Tesfatsion, L., & Judd, K. (2006). Handbook of computational economics,
United States, 90, 7716–7718. volume 2: Agent-based computational economics. Amsterdam: North-
Jones, K., Krishnan, G., & Melendrez, K. (2008). Do models of discretionary Holland.
accruals detect actual cases of fraudulent and restated earnings? Uecker, W., Brief, A., & Kinney, W. (1981). Perception of the internal and
Contemporary Accounting Research, 25(2), 499–531. external auditor as a deterrent to corporate irregularities. The
Kim, Y., & Xiao, N. (2008). Fraudsim: Simulating fraud in a public delivery Accounting Review, 56(3), 465–478.
program. In L. Liu, & J. Eck (Eds.), Artificial crime analysis systems: Using Wilks, T., & Zimbelman, M. (2004). Decomposition of fraud-risk
computer simulations and geographic information systems (pp. 319– assessments and auditors’ sensitivity to fraud cues. Contemporary
338). Hershey, PA: Information Science Reference. Accounting Research, 21(3), 719–745.
Kirman, A. (1992). Whom or what does the representative agent Wolfe, D., & Hermanson, D. (2004). The fraud diamond: Considering the
represent? Journal of Economic Perspectives, 6(2), 117–136. four elements of fraud. The CPA Journal, 74(12), 38–42.
Matsumura, E., & Tucker, R. (1992). Fraud detection: A theoretical Wolfram Research, Inc. (2008). Mathematica version 7.0. Champaign, IL:
foundation. The Accounting Review, 67(4), 753–782. Wolfram Research, Inc.

2013 Fraud Dinamics in Organizations-Davis

Uploaded by

Copyright:

Available Formats

2013 Fraud Dinamics in Organizations-Davis

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2013 Fraud Dinamics in Organizations-Davis

Uploaded by

Copyright:

Available Formats

Accounting, Organizations and Society 38 (2013) 469–483

Contents lists available at SciVerse ScienceDirect

Accounting, Organizations and Society

Fraud dynamics and controls in organizations

Introduction of fraud with ﬁnancial statement reporting choices and cor-

We begin developing our model of fraud in organiza-

After establishing a starting population in the organi- 0 1

Starting population attitude Random (n = 10) All pro-fraud (n = 5) All anti-fraud (n = 5)

p(O) = likelihood of an agent perceiving an opportunity Benchmark model design

Paneel A: Hiistoggramm whhen aaverrage emuulatiion likellihoood (q) Paanel B: H Histoograam w

Mean emulation likelihood (q) 0.05 0.25 0.50 0.75 0.95

Mean emulation likelihood (q) 0.05 0.25 0.50 0.75 0.95

tðeÞ ¼ eððN f Þ=NÞ; 29

Intervention Effect on U-shaped organization Effect on inverted U-shaped organization

You might also like