Can we trust consumers survey answers when dealing with

insurance fraud? Evidence from an experiment

Kerstin Puchstein*
Jrg Schiller
Frauke von Bieberstein

Abstract
Consumer surveys (e.g. questionnaires, telephone surveys) are an important means to
measure the acceptability and willingness to commit insurance fraud as well as related
influencing factors. However, for such a sensitive issue, the data validity and reliability is
unclear. We use a two stage within-subject procedure which consists of an experiment and a
questionnaire. In the experiment participants are incentivized and have the opportunity to
commit fraud. We compare participants behavior in the experiment and their answers in the
questionnaire and do not find a strong correlation between self-stated attitude towards
insurance fraud and behavior in the experiment. Our results indicate that influencing factors
for insurance fraud differ in their direction and significance with subject to the data collection
method used.

30 October 2014

Financial support of the German Insurance Science Foundation (Deutscher Verein fr Versicherungswissenschaft e.V.) is gratefully acknowledged. We thank Christian Biener und Martin Nell for
helpful comments.
*
University of Hohenheim, Department of Accounting and Finance, Schloss Osthof, 70593 Stuttgart,
Germany, E-Mail: [email protected] (corresponding author).

University of Hohenheim, Chair in Insurance and Social Systems, Fruwirthstr. 48, 70593 Stuttgart,
Germany, E-Mail: [email protected].

University of Bern, Institute for Organization and HRM, Engehaldenstr. 4, 3012 Bern, Switzerland, EMail: [email protected].

1 Introduction
Insurance fraud is a serious problem in the insurance sector all over the world. For Germany,
fraudulent activities are estimated to cause a damage of 4 billion EUR (Debreband, 2011).
Surveys indicate that about 4% of all German households committed insurance fraud in the
last five years and another 7% of respondents have information about fraudulent behavior
among their acquaintances (John, 2011). In the United States the annual extent of insurance
fraud is appraised to a range between 89 to 300 billion USD. According to the Insurance
Research Council over one third of all bodily-injury claims of auto accidents include fraud.
This implies that about 17-20 USD cents of every dollar of these claims cover consequences
of insurance fraud. About 5.2 to 6.3 billion USD are added to automobile insurance premiums
in the American market due to dishonest behavior (Zalma, 2011).
Insurance fraud is only possible as the occurrence and/or the specific loss size is (partly)
private information of the policyholder. Current research focusses both on theoretical and
empirical issues of insurance fraud.1 The two main theoretical approaches that deal with
major forms of insurance fraud are costly state falsification and costly state verification. In
costly state falsification models (Crocker and Morgan, 1998) policyholders can exaggerate
claims by costly activities, which are unobservable for insurance companies. The extent of
falsification (claim build-up) can therefore only be limited by an efficient design of the
indemnity function, where small losses are overcompensated and high losses are
undercompensated. The second approach of costly state verification goes back to Townsend
(1979) and Picard (1996) and considers a situation, where policyholders are able to report
fictitious losses and insurance companies can only limit these activities by costly audits.
An important focus of empirical papers is on identifying different factors which may influence
the attitude towards committing insurance fraud. For example, Tennyson (2002) finds that
there are significant differences between the attitude towards insurance fraud of women and
men. Fullerton et al. (1996) identify that the tolerance of unethical behavior decreases in age.
Most of these studies are based on consumer surveys like questionnaire or telephone
surveys. The use of questionnaire data supposes that consumers attitude towards insurance
fraud is closely related with their real behavior in insurance transactions. By comparing
fraudulent death benefits in the life insurance market Colquitt and Hoyt (1997) found that (...)
the publics perception of the acceptability of fraudulent claiming behavior is an important
indicator of the extent of fraud committed in the state(Colquitt and Hoyt, 1997, p. 475).
These results are confirmed by Cummins and Tennyson (1996) who find that a higher
acceptability of fraudulent practices in automobile bodily injury liability is positively related to
1

For example, Picard (2013) and Dionne (2013) provide an excellent overview on the current state of
the art.

higher rates of liability insurance claiming in this state. Dionne et al. (2009) analyze 1,000
random claim reports of a European auto insurer which were not classified as fraudulent by
regular claim adjusters. They found that 55 out of the 1,000 claims contained some indication
of fraud.
There are only a few studies using actual claims when analyzing insurance fraud. For
example Artis et al. (2002) confirm the finding that the tendency to defraud declines in age
when analyzing actual claims. However, we are not aware of any study which is based on
actual claims data that focusses on socio-demographic factors of policyholders, like gender
and income. In the insurance industry, the use of computer-based expert systems for the
detection of insurance fraud is very common. These systems rely both on observable claims
and policyholder information. For the improvement of such systems a better understanding
about reliable factors which may affect the attitude for committing fraud is important.
Although many studies use consumer surveys to assess the impact of socio-demographic
factors, the validity and reliability of these surveys are unclear. Our main research question is:
do consumer surveys provide a reliable and valid instrument when they are used to
investigate the sensitive issue of insurance fraud? Validity is given if the design of the survey
is reasonable to investigate the stated research question. A high level of validity allows the
generalizability of the results.2 Validity can be categorized in different sub-categories. 3 Here,
the criterion validity is of relevance. A test possesses criterion validity if the relationship
between the measure of the test and the measured external criterion in real world is close.
Especially, for sensitive questions the use of self-reported measures can lead to difficulties of
criterion validity (Thornberry and Krohn, 2000, p. 52). Reliability is given if a measure is
robust over a repeated exercise and free of measurement errors. Reliability can be tested in
different ways.4 Here, we use a procedure based on a parallel test. In a parallel test, the
same question is investigated using two different test forms for the same set of individuals
within a short period of time. A direct comparison of peoples survey responses to their actual
behavior is almost impossible in an insurance fraud context. We therefore compare the actual
behavior in an incentivized insurance experiment, where participants can exaggerate actual
losses or can claim losses that never occurred, with their responses to standard survey
questions.

2
3
4

Some authors classify these characteristics of validity as internal and external validity (see e.g.
Campbell, 1957).
Such categories are e.g. content validity, face validity and construct validity (see e.g. Thornberry and
Krohn, 2000 or Moosbrugger and Kelava 2012).
Other methods to test for reliability are for example test-retest reliability or internal consistency (see
e.g. Thornberry and Krohn, 2000 or Moosbrugger and Kelava, 2012).

The advantage of an economic experiment is that real people earn real money for making
real decisions about abstract claims that are just as "real" as a share of General Motors
(Smith, 1976, p. 274). In a lab experiment, the researcher is able to control for the decision
environment. As a direct negative consequence, the generalizability is also unclear due to the
artificial decision environment (Richter et al., 2014). In field experiments, decision-making is
embedded in a more natural setting. Consequently, as noted by Harrison and List (2004)
there might be substantial differences between field and lab experiments. Apart from the
decision environment, one important difference between the two forms of experiments is
usually the nature of the subjects pool. Lab experiments are mainly conducted with students,
whereas field experiments often use pools of subjects with a greater variety of sociodemographic characteristics which allows controlling for biases (Blackburn et al., 1994).
Naturally, there is a tendency to favor the field to the lab. However, findings of Harrison et al.
(2007) with respect to risk attitudes for example do not lead to a clear superiority of field
experiments. Inferring consumer preferences from actual choices (reveled preferences) is an
integral part of economic theory. When it comes to illegal or sensitive activities, this general
approach is often not feasible, as the extent of activity is unobservable. In these areas data
from anonymous surveys might be the best source for analyzing consumers preferences.
But the validity and reliability of survey data is particularly important for sensitive topics like
the consumption of illegal drugs, sexual behavior, the personal health situation or criminal
offences like insurance fraud. Here, results from anonymous questionnaires may be
susceptible for biases.
According to Tourangeau et al. (2009) a sensitive question comprises three properties:
Firstly, it is intrusive, in the sense that the topic is considered as very private information.
Secondly, there are social norms pertaining to the question. Therefore, respondents might
opt for a socially desirable reply or give responses which they consider as moderate. Thirdly,
there is a threat of disclosure if the reply was known by a third party.
Insurance fraud is a criminal offence and socially not accepted. The question of supporting or
doing illegal activities could be classified as a private issue. Survey respondents who are
declaring in doing so might become scared of the consequences. Regarding the second
criterion, different fraud types have to be distinguished. While claim build-up is often
considered as common practice and a weak form of insurance fraud, reporting fictitious
claims is contemplated as a stronger form of fraud for which a higher level of criminal energy
is necessary (Psychonomics, 1996). Thus, social desirability might play a stronger role when
it comes to fictitious claims. Overall, since all three properties apply, we consider the
question of accepting or practicing insurance fraud in surveys as a sensitive one for which a

higher level of carefulness in analyzing and interpreting is given. Therefore, the truthfulness
of replies must at least be put into question.
Several studies indicate that misreporting in surveys is quite common for sensitive topics.
Tourangeau and Yan (2007) describe three possible distortions: The number of participants
of the whole survey could decrease if they expect sensitive questions to be asked. The
number of respondents of a certain, sensitive question could decline and the response
accuracy could be reduced. In our analysis, we will focus on the last category. Since the
respondent had no possibility to skip certain questions or to jump over the whole
questionnaire, this is the only relevant category in our experimental analysis. According to
Tourangeau and Yan (2007) there exists a systematic underreporting for socially unaccepted
forms of behavior and a systematic overreporting for socially accepted kinds of behavior. For
example this is confirmed by the results of Heeb and Gmel (2005), who are investigating selfreported consumption of alcoholic drinks or Fendrich and Vaughn (1994) who analyze the
effect of underreporting in the consumption of cocaine and marijuana in a survey.
Since insurance fraud is a criminal act and its practice is socially not respectable, it is straight
forward to presume that individuals self-expression in consumer surveys differs substantially
from their actual attitude and behavior. The fact that individuals might deliberately misreport
their attitude with respect to fraudulent activities could cause a serious distortion of survey
results. By combining common questionnaire answers with actual behavior in an incentivized
economic experiment, we try to evaluate how reliable and valid surveys about consumers
attitude towards insurance fraud are. Of course, our analysis can only offer an indirect test
and we cannot evaluate the reliability and validity of survey answers concerning insurance
fraud. However, linking survey responses of individuals with their behavior in an experiment
is in our view a promising first step.
In the experiment participants make decisions in an insurance setting. Losses are private
information. Therefore, participants can honestly report their claims, inflate their loss size
(claim build-up) or can report losses that have never occurred (fictitious claims). After the
experiment, participants answer typical survey questions querying individuals sociodemographic characteristics as well as attitudes and experiences towards insurance fraud.
We compare individuals responses in the questionnaire with their actual behavior in the
experiment.
We find that participants behavior in the experiment does not always correspond with their
stated attitude towards insurance fraud in the questionnaire. Conformity of individuals
behavior in the two data sources is only available for approximately half of participants. There
are also differences between the investigated fraud types. We therefore separate two
5

criteria when we identify influencing factors on insurance fraud: the type of fraud and the
method of data collection (experiment/questionnaire). Depending on the method of data
collection, different impact factors are significant. Our findings reinforce doubts on the
reliability and validity of survey responses, as there is no strong relationship between
individuals acceptance of insurance fraud in the questionnaire and their behavior in the
experiment.
Given that participants first make decisions in the experiment and then answer the
questionnaire, answers in the questionnaire could be influenced by actual behavior in the
experiment. In particular, the theory of cognitive dissonance (Festinger, 1962) suggests that
people try to avoid a discrepancy between their stated values and their actions. Note that if
this effect was present in our study, we would underestimate the difference between
questionnaire replies and experimental actions. In this respect, our findings can be seen as a
conservative estimate of these differences.
The paper is organized as follows: Following the introduction, Section 2 presents a literature
overview of recent empirical literature of influencing factors for insurance fraud. Section 3
provides a brief overview of the design and proceedings of the analyzed experiment. Section
4 is split in two parts. The first part descriptively matches the data generated out of the
experiment with the data out of the questionnaire. In the second part of the section the
results of our econometric model are presented. Section 5 concludes.

2 Literature Overview
2.1 Socio-demographic and Environmental Factors
The attitude towards insurance fraud is strongly influenced by the social and cultural
environment as well as other socio-demographic factors. Fullerton et al. (1996) investigate
consumers judgment of ethically disputable actions in several situations using a
questionnaire. Their analysis indicates that with an increase in age, the tolerance for
unethical behavior declines. By investigating automobile claims from accidents in Spain Artis
et al. (2002) find that younger individuals have a significantly higher fraud probability. Another
interesting finding is that a higher level of education and a higher income are related to a
greater acceptance of unethical behavior. Tennyson (2002) builds different income categories
and shows that individuals in the highest income category are less likely to accept fraud than
those of lower income categories. According to a study of Psychonomics (1996) the
satisfaction with the personal financial situation is more important than income. It turns

out that individuals who are less content with their status quo have a higher affinity to
insurance fraud.
Another common finding in the literature are gender differences with respect to social issues
and risk bearing. Males are often more willing to accept fraudulent behavior than females
(Tennyson, 2002). Weeks et al. (1999) find significant differences in ethical judgment
between women and men. In a broader sense, women are more risk averse than men (e.g.
Croson and Gneezy, 2009). Several other studies have shown that the neighborhood and the
living area have an effect on individual behavior (Elliot et al., 1996). This implies that people
also learn and adapt their behavior from their environment. If among their acquaintances or
in their neighborhood insurance fraud is approvable and common, it is more likely that
individuals follow this attitude (Tennyson, 1997). Tennyson (2002) also analyzes the impact of
residential areas and shows that suburban residents are less willing to agree with insurance
fraud than urban residents.
In our analysis we control for gender effects and risk aversion. However, as participants are
mainly students from the area of Munich, we do not control for age, income and residential
area.

2.2 Experience and Perception towards Insurance Companies

In general, a limited understanding of insurance mechanisms may result in mistrust and a
negative perception of the insurance industry. However, the insurance industry may also not
sufficiently support customers education for a better understanding of their products. Both
factors may lead to an unappealing environment encouraging fraudulent behavior (Hoyt,
1990). Using data of an US survey of 1,987 individuals, Tennyson (1997) explores the
relationship between the attitude towards claim build-up in automobile insurance claims and
the perception of insurance companies. Individuals who are less confident with the financial
stability of their insurer tend more towards a fraud supporting attitude. Those who consider
their premiums as a financial burden have a higher acceptability of recouping deductibles but
not of recouping their premiums paid.
Tennyson (2002) tests whether or not there is a systematic relationship between insurance
experience and the acceptability of deception. Experience is defined as the number and type
of policies held as well as the type of losses claimed in the past two years. She found that the
number of types of insurance contracts is not significantly associated with the attitude
towards insurance fraud. But, participants with claiming experience in recent years have a
lower level of acceptability of fraudulent behavior. One possible explanation is that through
7

the process of claims settlement and the communication between the two parties, the insurer
signals deception as not inacceptable. Using automobile claims data from a Swiss insurance
company, Mller (2013) examines policyholder characteristics, vehicle and policy
specifications as well as loss characteristics as potential fraud drivers. In the final regression
model, she finds that a lower number of previous records leads to a higher fraud probability.
Contrary, by analyzing Spanish automobile claims Artis et al. (1999) show that individuals
with more claiming experience have higher tendency to defraud. Experience is thereby also
measured by the number of previous claims.
Based on the findings in the literature, we include four variables into our analysis: the number
of insurance contracts, recent claiming experience, the individuals general attitude and
perception towards the insurance industry, as well as the individuals satisfaction with his or
her actual insurance contracts and the claims settlement process.

2.3 Different Contract Settings

Moreno et al. (2006) show that bonus-malus contracts may be an effective means to prevent
insurance fraud. Gabaldon et al. (2009) analyze the impact of bonus-malus contracts on
fraudulent claiming behavior in an experiment. They only find a mild tempering effect of
bonus-malus contracts in their experiments.
Deductible contracts are often perceived as unfair and may therefore trigger fraudulent
behavior. The fact that a higher deductible causes a lower premium is thereby often
neglected by the policyholders (Miyazaki, 2009). Dionne and Gagne (2001) investigate claim
build-up in deductible contracts using filed claims after automobile accidents. They find that
the amount of the deductible is a significant determinant for the reported loss, at least when
no other vehicle is involved in the accident. (Dionne and Gagne, 2001, p. 290). Using a
questionnaire, Miyazaki (2009) investigates whether higher deductibles cause a greater
sense of justification for fraudulent behavior and if this feeling ends in applying lower ethical
standards. The results show that in case of high deductibles claim build-up will be considered
as less unethical. These effects are stronger for individuals with lower ethical standards as
for respondents with stricter ethical standards. In their experiment that is also the basis for
the current paper, von Bieberstein and Schiller (2014) find that people are more likely to
commit fraud given a deductible contract compared to a bonus-malus contract.
As the perception of the different contract forms is varying in our analysis, we control for
differences in the contract settings. Table 1 provides a brief overview of important influencing
factors on insurance fraud which are discussed above.
8

Socio-demographic factors
variable
author

Influence
Unethical behavior declines with an
Fullerton et al.(1996)
increase in age.
age
Younger individuals have a higher fraud
Artis et al. (2002)
probability.
The acceptance for unethical behavior
Fullerton et al. (1996)
increases with income.
High income individuals are less likely to
Tennyson (2002)
income
accept insurance fraud.
Income is not relevant. More relevant is
Psychonomics (1996)
the satisfaction with the individual financial
situation.
Croson and Gneezy (2009)
Women are more risk averse than men.
Men are more willing to accept fraudulent
Tennyson (2002)
gender
behavior.
There are differences in ethical judgment
Weeks et al. (1999)
between women and men.
The living area influences individuals
Elliot et al. (1996)
behavior.
residential
The surrounding field has an impact on
area
Tennyson (1997)
individuals attitude towards insurance
fraud.
Experience and perception towards the insurance industry
variable
author
influence
A negative perception increases the
perception
Tennyson (1997)
tendency to cheat in insurance contracts.
Recent claiming experience lowers the
Tennyson (2002)
acceptability of insurance fraud.
A higher number of records is associated
experience
Artis et al. (1999)
with a higher fraud probability.
The lower the number of previous records,
Mller (2013)
the higher is the tendency to defraud.
Contract settings
variable
author
influence
bonusPrevious experience with bonus-malus
Gabalon et al. (2009)
malus
contracts reduces the fraud probability.
The level of the deductible is a significant
Dionne and Gagne (2001)
factor for the submitted claim.
deductibles
For higher deductibles individuals consider
Miyazaki (2009)
claim build-up as more justified.
bonusIndividuals commit more fraud given a
von Bieberstein and Schiller
malus vs.
deductible contract compared to a payoff(2014)
deductibles
equivalent bonus-malus contract.
Table 1: Overview of influencing factors on insurance fraud

3 Experimental Design
The experiment was conducted at the MELESSA laboratory at the Ludwig-MaximiliansUniversity in Munich. Six treatments took place from March to July 2009. Two further
treatments were done in July 2012. Differences between the first and the second data
collection were group size and transaction costs for filing claims. In the first set of
experiments 4 participants belonged to one group and transaction costs were 40 percent. In
the second set of experiments the group size was 24 participants and transactions costs
were 20 percent. The second set of experiments was designed as a control with higher fraud
incentives due to a larger group size and lower transactions costs. In order to account for the
differences between treatments, we control for the specific contract type and the group size
in our regression analysis. The total database consists of 576 participants with 72
participants in each treatment. Almost all participants were students with various majors, but
there is no predominant major.
The experiment is based on the idea of collective risk sharing. It consists of two computerbased parts. In the experiment subjects are randomly and anonymously allocated into fixed
groups of four or twenty-four. For every group, a specific insurance account from which all
incoming and outgoing payments are made is introduced. All payments from and to the group
members are settled via this account. At the end of the game the insurance account is
automatically balanced and all profits and losses in this account are equally shared between
the group members. Each group plays five periods of the following insurance game: At the
.5 Participants

beginning of each period every participant obtains a period endowment

are informed that they have to insure against the risk to suffer no loss

( )

, a small loss

with a 0 probability of

or a large loss
member

0.7,= 0.2

= 0)

The specific losses are independently determined for each group 0

and

( = 10)

. The losses occur in each period

( . = 15)

in each period.

In surance is mandatory in the experiment. Thus, each group

= 0.1

member in each period must pay an insurance premium to the group-specific insurance
account that finances all indemnities paid to the group members. Hence, in our experiment,
we apply a mutual insurance setup. In order to have sufficient data, we used the strategy
method. By choosing their claiming strategy

with

in

decisions about how much

to claim from the

each period, participants make contingent

group-account depending on the loss realization. This means, participants make their
= ( ( ), (

), ( ))

( ,

decision about the level of the indemnity for all=possible0,=10losses before=15 the actual loss is known. For each loss only the indemnities 0 and could be
selected.

Hence, participants were able to inflate the size of the loss by claiming an indemnity that is
5

During the experiment the players are dealing in points, which will be converted by 1 point equal to
10 Euro-cents afterwards.

10

higher than the loss (claim build-up) or by claiming a positive indemnity when no loss
occurred (fictitious claims). In the next step the players obtain information about their actual
loss occurred in this period, the indemnity they reported and their preliminary payment,
without considering the group insurance account. In accordance with the chosen strategy,
the indemnity will be paid virtually. After the last period, they receive their preliminary final
result over all rounds. While indemnities are paid as claimed, the group insurance account is
debited with the amount of the indemnity plus an additional 40 % respectively 20 % for
transaction costs.
One important aspect of the experimental setup is that there is no monitoring and no
sanctions for fraudulent behavior. The reason for this is that we wanted to concentrate on the
preferences of the participants and not on strategic aspects or risk aversion related to costly
monitoring and sanctions. To analyze and compare different features, the whole experiment
consists of six different treatments. For the Base Treatment, as described above the
endowment is
period premium

is

for each period and the possible losses are

and
. The
the possible losses with
, which equals the expected value of

= 25

= 10

= 15
Additionally there are three modification s of the Base Treatment. The first

transaction costs. = 5
type does not use insurance specific wording e.g. the groups insurance account is named
group account (No Insurance Wording Treatment). In the second modified treatment again no
insurance specific wording is used and in addition no transaction costs are comprised (No
Transaction Costs Treatment). In the third modification the participants get information about
the amount of the insurance account and its changes at the end of each period (Feedback
Treatment). Since the selectable indemnity corresponds to the loss sizes, the Base
Treatment depicts full insurance.
Furthermore, two treatments without full insurance are applied. In the Deductible Treatment a
deductible of 5 is introduced by increasing the amount of the possible losses
, while the claimable indemnity remains unchanged. Hence, the
increased

to

endowment

and
is

= 15

, wherea s the premium is

. With the Bonu s-Malus Treatment an

= 20
rating is introduced. Thereby the premium is not a fixed amount, it depends on

=5
experience = 27
the indemnity of the previous periods. In the first period it starts with
period

the premium

+1

period

. In the following

indemnity

was received in

=5

and decreases by 1 point in case no indemnity was paid. The new starting point for
+1

the next period

Treatment have
6

increases by 2 points in case an

+2

is the premium in

. The Deductible Treatment and the Bonus-Malus

been applied in the first and second round of data collection. 6

+1

For more details, please see von Bieberstein and Schiller (2014) and Lammers and Schiller (2010).

11

The second part is based on a questionnaire about the individuals characteristics and
attitudes. The questionnaire provides information about participants attitudes concerning the
general compliance of rules or laws, experience with insurance contracts and companies, as
well as insurance fraud. The aim of the experiment is to match behavior in the experiment
with the participants survey responses. For their attendance the participants received a fixed
show-up fee of 4 EUR as well as additional payments which depended on their result in the
first part. Average earnings ranged from 12.85 to 15.01 EUR including the show-up fee.
Answers in the questionnaire did not affect the final payment. Each session took about 50-60
minutes.

4 Results
4.1 Behavior versus stated attitudes
In our analysis, we distinguish different categories of insurance fraud. In the experimental
part, participants have the possibility to cheat in two different ways: By making fictitious
claims when they do not face a loss, or by claim build-up, when they exaggerate a given loss
size. Participants report their claims for all possible loss situations before they know the real
loss realization (strategy method). All (actual and hypothetical) contingent decisions are
included in our analysis. Table 2 displays the frequencies of fraudulent behavior for the five
rounds of the experiment, separated by reporting fictitious claims and claim build-up. Since
the distributions of fictitious and inflated claims differ, we distinguish both types of fraudulent
behavior.

Number of defrauds

Fictitious claims

Claim build-up

22.40 %

33.33 %

12.85 %

11.28 %

10.42 %

7.47 %

12.33 %

7.29 %

9.55 %

7.29 %

32.47 %

33.33 %

Table 2: Number and share of defrauds in the experiment

12

Along the lines of the experiment, the questionnaire entails questions whether or not
individuals in general accept fraudulent behavior or could imagine reporting wrong
information to their insurance company.7 These questions separately focus on reporting
fictitious claims, claim build-up, redefinition of claims and insurance fraud in general. The
answers are given on a seven-point Likert-scale, whereby zero corresponds to no
agreement and six implies full agreement with the respective statement or fraud type. The
middle position three indicates a neutral position to the written statement. In our analysis, we
divide responses to these questions into three categories. Answers on the scale from zero to
two are classified into the first group as individuals who do not accept that behavior. Answers
on the scale from four to six are sorted into the second group containing participants who
accept this behavior. Individuals whose answer is three are categorized as neutral, as they
neither agree nor disagree with insurance fraud. Figure 1 shows the distribution of the
answers in the questionnaire.
Claims exaggeration
Fictitious claims
66.67%

60,00%

60,00%

40,00%

40,00%

20,00%

15.80%

16.67%

11.98 24.83%
%
14.58%
10.94%
5.21%

19.27%
20,00%

6.08% 4.17%

0,00%
Disagreement
92.02 %

2.43% 1.22% 0.17%

0,00%
Disagreement
43.41 %

Agreement
3.82%

In general, I can imagine to report a claim that never

occurred to receive an indemnity from the insurance
company.

Agreement
44.62%

In general, I can imagine to slightly exaggerate claims

in case a loss occurred.

Insurance fraud in general

Redefinition of claims
60,00%
60,00%

40,00%

40,00% 27.26%
20,00%

32.99%

20,00%
21.18%

10.24 15.63%

11.28%

9.55% 4.86%

0,00%
0

33.33%

Disagreement
59.72 %

Agreement
30.04%

In general, in case a loss occurred, I can imagine to report

wrong information regarding the course of events towards
the insurance company to receive an indemnity at all.

14.06% 7.99% 8.51%

2.60%

0,00%
0

Disagreement
80.38 %

0.52%

Agreement
11.63%

In general, it is acceptable to report wrong information

to the insurer.

Figure 1: Fraud in the questionnaire

The written questions can be found in Appendix 1.

13

Figure 1 indicates that the acceptance of fictitious claims is very low. Only 3.82% could
imagine reporting claims that never happened, while 92.02% disagree with reporting fictitious
claims. 4.17% of the individuals are neutral in this question. The distribution for claim build-up
strongly differs from the previous fraud type. Here, the percentages of agreement and
disagreement are almost equal. While 43.41% could not imagine to build-up claims, 44.62%
accept claim build-up. The discrepancy between individuals who agree and individuals who
disagree with redefinition of claims is not as high as for reporting fictitious claims, but the
distribution is also not as equally spread as in case of claim build-up. While 30.04% approve
redefinition of claims, 59.72% do not accept this fraud type. For fraud in general 80.38%
disagree and only 11.63% accept this kind of behavior.
A comparison of the distributions shows that for fictitious claims and fraud in general the rate
of disagreement is very high, whereas the rate is much lower for claim build-up and
redefinition of claims. Thus, respondents seem to consider the latter two activities as less
severe than giving wrong information to the insurer in general. These distributions are in line
with findings in the literature, stating that claim build-up and redefinition of claims is often
considered as a more acceptable form of insurance fraud since an actual loss occurred (Nell,
1998). Since we cannot measure redefinition in the experimental part, we will not focus on
this type of fraud in the following.
To investigate whether we can trust consumers answers in questionnaires concerning
insurance fraud, we first compare participants behavior in the experiment and their stated
behavior in the questionnaire. Table 3 to Table 5 display the relation between answers in the
questionnaire and behavior in the experiment. For the questionnaire we use the
categorization described above: Answers from zero to two are assigned to disagreement with
fraud while answers from four to six are classified as agreement with fraud. Individuals whose
answer in the questionnaire is three are classified as neutral. They are not included in the
following comparison. The behavior in the experiment refers to all five rounds of the
experiment. Again, we build two categories: participants who cheat in zero to two out of the
five rounds of the experiment are considered as individuals who disagree with fraudulent
behavior. Participants who commit any fraudulent claim in more than two rounds of the
experiment are classified as individuals who agree with insurance fraud. We choose this
segmentation since it forms two equal groups (with each three scale items) and therefore
best represents the separation made in the questionnaire.8

An alternative segmentation for the frequency of defrauds in the experimental part would be to classify
individuals behavior as non-fraudulent if they do not report any fraudulent claim in all periods of the
experiment. If an individual cheats in one or more periods, its behavior is classified as fraudulent in the
alternative segmentation. Thereby, no separation of the number of periods in which the

14

Fictitious claims
Disagree with fraud in the
questionnaire
Agree with fraud in the
questionnaire

Disagree with fraud in the

experiment
44.92 %
(248)
1.09 %
(6)

Agree with fraud in the

experiment
51.09 %
(282)
2.90 %
(16)

Table 3: Match of fictitious claims in experiment and questionnaire (n=552)9

Focusing on fictitious claims (Table 3), 44.92% of the participants state to disagree with
reporting fictitious claims in the questionnaire and cheat in two rounds of the experiment at
maximum. Only 2.90% state that they could imagine reporting fictitious claims and confirm
their attitude in the experimental part. With 51.09% more than half of the individuals disagree
with reporting fictitious claims in the questionnaire but they do so in the experiment.
Therefore, it is straight forward to assume that there exists a difference in the expressed
behavior for a sensitive topic like insurance fraud in questionnaires and their real actions in
experiments. A McNemars chi-square statistic indicates a statistically significant difference in
the proportions at a 1 percent level.
The numbers change when claim build-up in Table 4 is considered. Thereby 26.04% disagree
with fraud in the questionnaire as well as in the experiment. 24.46% of the participants agree
with claim build-up in the questionnaire and do so in the experiment. Here, 23.27% state that
they do not agree with claim build-up in the questionnaire but exaggerate claims in more than
two rounds of the experiment. Contrary 26.23% accept fraud in the questionnaire but behave
honestly in the experiment. A McNemars chi-square statistic indicates no statistically
significant difference.
Claim build-up
Disagree with fraud in the
questionnaire
Agree with fraud in the
questionnaire

Disagree with fraud in the

experiment
26.04 %
(132)
26.23 %
(133)

Agree with fraud in the

experiment
23.27 %
(118)
24.46 %
(124)

Table 4: Match of claim build-up in experiment and questionnaire (n=507)

individual cheats is made. The tables in which the alternative segmentation is used are provided in
Appendix 2.
9
Table 3 includes observations of 552 individuals. 24 observations are not included because their
attitude concerning fictitious claims is categorized as neutral in the questionnaire. The number of
observations for each category is stated in parenthesis below the percentages.

15

Table 5 shows the results for fraud in general10. 34.34% of all participants disagree with fraud
in the questionnaire and defraud maximum twice in the experiment, while 9.25% agree with
fraud in the questionnaire and defraud in the experiment. 53.02% of the individuals cheat in
three or more periods of the experiment but state not to accept fraudulent behavior in general
in the questionnaire. The contrary case does only apply for 3.40% of the participants. A
McNemars chi-square statistic indicates a statistically significant difference in the proportions
at a 1 percent level.
Fraud in general
Disagree with fraud in the
questionnaire
Agree with fraud in the
questionnaire

Disagree with fraud in the

experiment
34.34 %
(182)
3.40 %
(18)

Agree with fraud in the

experiment
53.02 %
(281)
9.25 %
(49)

Table 5: Match of fraud in general in experiment and questionnaire (n=530)

For all three fraud types the Tables 3-5 show, that only approximately half of the participants
behave consistently in the questionnaire and in the experiment. In total there exists a positive
match for 47.82% of the individuals for fictitious claims, for 50.50% of the individuals for claim
build-up and for 43.59% for fraud in general. The aim of the experiment is to gain information
about individuals attitude towards insurance fraud and their moral concept of value.
Therefore no sanctions have been included into the experimental setup. Of course, the
absence of sanctions could impact individuals behavior. However, there are also no
sanctions for the agreement with insurance fraud in the questionnaire. Conformity between
the attitude towards insurance fraud in the questionnaire and in the experiment should
therefore be even stronger in absence of any sanctions. A comparison of the three fraud
types shows that when analyzing insurance fraud, the single types have to be considered
separately since they are perceived differently. The range for individuals who disagree with
fraud in the questionnaire and in the experiment lies between 26.04% - 44.92%. The fraction
of participants who agree with fraud in both parts is low for fictitious claims but much higher
for claim build-up.
To provide an assessment of the reliability and validity of questionnaire based surveys on
insurance fraud, the category disagree with fraud in the questionnaire / agree with fraud in
the experiment is a matter of particular interest. This category covers observations where a
negative answer is given in the questionnaire that does not coincide with the behavior in the
experimental part. A comparison of fictitious claims and claim build-up indicates that in case
10

Fraud in general is defined as either reporting fictitious claims or exaggerating claims or both types
of fraud in one round of the experiment.

16

of fictitious claims the share of people who fall in this category is nearly twice as high as for
claim build-up. Of all individuals who disagree with fictitious claims in the questionnaire,
46.79% disagree in the experiment, too and 53.21% agree. Likewise of all participants who
disagree with claim build-up in the questionnaire, 52.80% accept this fraud type in the
experiment while 47.20% do not. Given our findings, trusting results out of the questionnaire
without further considerations may lead to misleading conclusions.
Differences between fraud types may be due to differences in the level of sensitivity of
fictitious claims and claim build-up. As defined above, a sensitive question is characterized by
the level of intrusiveness, of social desirability and of threat of disclosure. Since questions
about both types of fraudulent behavior are classified as very private information, there is no
difference in the level of intrusiveness. About the threat of disclosure, no concrete statement
could be made since the punishment also depends on the amount of fraud and not only on
the type of insurance fraud. Considering the criterion of social desirability, prior research
indicates that fictitious claims could be more socially undesirable than claim build-up
(Psychonomics, 1996). The difference in social desirability could be one possible explanation
for the diverging results for both fraud types. Another possible explanation is that filing
fictitious claims in the experiment was relatively easy, while it takes considerable criminal
effort in the real world where a loss situation has to be fabricated. Therefore, the experiment
may not represent differences in both fraud types in the same magnitude than they exist in
the questionnaire. The correlation matrix in Table 6 shows that there exists a significant
correlation of 0.6002 between fictitious claims and claim build-up within the experimental part.
Contrary the correlation of the attitude towards both fraud types within the questionnaire is
0.1992. The discrepancies in these two correlation coefficients suggest that the fraud types
are perceived differently in both measurement methods. Based on the correlation coefficients
the perceived differences between fictitious claims and claim build-up are weaker when fraud
is generated out of experimental data than in case fraud is generated out of the
questionnaire.

17

Fictitious
claims
(questionnaire)
Fictitious
claims
(questionnaire)
Claim build-up
(questionnaire)
Fraud in
general
(questionnaire)
Fictitious
claims
(experiment)
Claim build-up
(experiment)
Fraud in
general
(experiment)

Claim buildup
(questionnaire)

Fraud in
general
(questionnaire)

Fictitious
claims
(experiment)

Claim buildup (experiment)

Fraud in
general
(experiment)

1.0000
0.1992 ***

1.0000

0.1873 ***

0.3646 ***

1.0000

0.0766 *

0.0444

0.1079 **

1.0000

0.0089

0.0105

0.0856 **

0.6002 ***

1.0000

0.0452

0.0189

0.0853 **

0.8482 ***

0.7457 ***

1.0000

Note: The table shows the pairwise correlation coefficients. The variables used are
categorized in line with the categorization above in this chapter. Significance levels p: ***
p < 0.01; ** p < 0.05; and * p < 0.1.
Table 6: Correlation matrix of fraud in the experiment and the questionnaire
Furthermore, Table 6 shows the relationship between behavior in the experiment and stated
attitude towards insurance fraud in the questionnaire. The correlation between the answers in
the questionnaire and the behavior in the experiment is quite low. With a weak significant
correlation coefficient of 0.0766 for fictitious claims and an insignificant coefficient of 0.0105
for claim build-up in the questionnaire and in the experiment, the results are nearly
uncorrelated. For fraud in general the coefficient shows a significant but low correlation of
0.0853.
The above results show that stating not to accept insurance fraud in a questionnaire does not
necessarily imply an according action in an experiment. There is no clear favorability of one
data set. One could argue that the experimental data may lead to more realistic results, as
real actions are practiced and monetary incentives are set. The disadvantages of
questionnaires, due to the particularities of sensitive topics, may increase the favorability of
the use of experimental data to investigate insurance fraud. However, different perception of
the distinction between claim build-up and fictitious claims may also be explained by potential
problems to adequately represent the different fraud types in the experiment.

18

4.2 Influencing factors

The results above show that the expressed behavior in the questionnaire does not always
coincide with the behavior in the experimental part. Fraud detection systems and prevention
activities of insurance companies are partly based on policyholders characteristics, as they
might be correlated with the tendency to defraud (see e.g. Schiller, 2006). The relevant
policyholders characteristics are often adapted from questionnaire-based surveys. However,
if they do not coincide with real behavior, detection systems may lead to wrong conclusions.
Based on prior research, we therefore compare the impact of well-established influencing
factors on insurance fraud by applying two different methods to measure insurance fraud. On
one hand, fraud is generated out of questionnaire responses. On the other hand, the
behavior in the experiment is used to investigate influencing factors on insurance fraud.
Accordingly, two regression models are developed. In the first model fraud is concerned with
the stated attitudes in the questionnaire and the second model measures the fraudulent
behavior in the experiment. Both models use the same set of influencing factors as
independent variables.
To further explore the relation between stated attitudes in the questionnaire and behavior in
the experiment, we consider a third model. Here, we use fraud out of the experiment as a
dependent variable and treat the stated attitude towards insurance fraud as an independent
variable in addition to the influencing factors. To account for differences between fraud types,
we estimate three different regressions for fictitious claims and claim build-up.

/=

1) = (

10

+ )

In the following we present results for six different logistic regressions. The models 1a, 1b
and 1c consider fictitious claims. Model 1a is related to the stated attitude in the
questionnaire, whereas Model 1b measures the behavior in the experiment. In Model 1c, we

further incorporate the impact of the stated attitude in the questionnaire on the behavior in
the experiment. The models 2a, 2b and 2c measure claim build-up in the same manner as
the first three models. We report the results for fraud in general in Appendix 4. The estimated
models can be summarized as follows:

(= 1) = (

10

11

+ )

19

The explanatory variables are chosen in line with results of existing literature described in
Section 3. The variable measures the impact of an individuals satisfaction

on the tendency to defraud. The variable is calculated out of two questions concerning the
satisfaction with insurance contracts in general if existing, and the satisfaction with prior claim
settlement processes if experienced. Therefore the variable is generated as a binary variable
which is equal to 1 in case the participant is not satisfied with its insurances
contracts in general or with the claims settlement process in case she already claimed a loss.
The variable equals 0 if no negative experience with existing insurance contracts was
reported.11 The variable is related to the fact whether participants have a
negative attitude towards the insurance industry in general. The variable is the arithmetic
mean of answers to two questions concerning the attitude towards insurance companies in
general. The scale of the coefficient remains on a seven point Likert-scale. The variables
and

measure the participants experience with insurance relationships.

The variable

corresponds to the number of participants insurance contract s.

coefficient

is equal to 1 if the respondent ha s reported a claim in the last two

The
in case of ab sence of any claims. The explanatory dummy variable

years and is 0

accounts for

measures gender differences (1 = female). Furthermore, the variable

measured on a seven

differences in the participants stated risk attitude. The variable is also

point Likert-scale.
Finally, we included four additional explanatory variables due to different modes in the
experimental part. Existing studies (e.g. Frey and Meier, 2004, and Frank and Schulze, 2000)
provide evidence that in experimental settings the behavior of individuals with a major in
economics or business differs from the behavior of individuals with other majors. The dummy
variable
is 1 for participants with a major in economics or business and 0 otherwise.

The dummy variable

is equal to 1 for all individuals belonging to experiment

treatments with the larger group size of 24 participants (instead of 4) and reduced transaction
costs. As the specific contract form in the experiment may influence behavior, we also control
for two contract types. The dummy variable

equals 1 for all participants of treatments

the dummy variable

with bonus-malus contracts. Respectively,

is equal to 1 for all

reference group in this category is a

participants of with deductible insurance contracts. The

full coverage contract. All used questions can be found in detail in Appendix 3. Table 7
provides a summary statistics of all used variables.

11

The detailed questions used can be found in Appendix 3. The answers of these questions are given
on a seven score Likert-scale. In line with the categorization of the previous chapters, we assign
answers from zero to two as not satisfied with insurance contracts or with the claims settlement
process. For all other answers we assume that no negative experience with insurance contracts
exists.

20

Variable

Observations

Mean

Fictitious Claims (Questionnaire)

Claim build-up (Questionnaire)
Fraud in general (Questionnaire)
Fictitious Claims (Experiment)
Claim build-up (Experiment)
Fraud in general (Experiment)
Satisfaction
Satisfaction general
Satisfaction claims
Negative view
Complicated contracts
Earnings
Claims
Recent claims
Contracts
Female
Risk
Major
Group
Boma
Deduct

552
507
530
576
576
576
576
353
269
576
576
576
576
576
576
576
576
576
576
576
576

0.040
0.507
0.126
0.543
0.479
0.623
0.168
3.847
3.911
3.027
3.830
2.224
0.443
0.708
1.099
0.599
2.785
0.142
0.250
0.250
0.250

Standard
Deviation
0.196
0.500
0.333
0.499
0.500
0.485
0.375
1.335
1.751
1.380
1.696
1.676
0.497
0.860
1.215
0.491
1.478
0.350
0.433
0.433
0.433

Min

Max

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

1
1
1
1
1
1
1
6
6
6
6
6
1
2
6
1
6
1
1
1
1

Table 7: Summary Statistics

Table 8 summarizes the regression results for fictitious claims. The only consistent and
significant impact over all three models is related to the stated risk. Therefore, participants
who stated to be less risk-averse both state a higher acceptance of filing fictitious claims in
the questionnaire and actually have a higher tendency to commit this type of fraud in the
experiment. A more negative view on the insurance industry in general only positively affects
the stated acceptance of fictitious claims in the questionnaire but did not have any significant
impact in the experiment. Surprisingly, we could not find any significant gender effect for the
questionnaire. Only in the experiment, we find a weakly significant effect of women filing
fewer fictitious claims than men. We also find significant effects for the control variables in the
experiment. Business and economics majors defraud more often in the experiment. In this
part, in the larger groups with the lower level of transaction costs fraud is more prevalent.
Bonus-malus contracts lead to a lower defraud rate, as a fictitious claim increases future
premiums which decreases the incentive to defraud.

21

As a robustness check, we run Model 1a and Model 1b a second time with a reduced
number of participants. According to our descriptive analysis we thereby excluded individuals
who state to agree with fraud in the questionnaire, but do not cheat in the experimental part.
Due to their divergent behavior in both parts there might be a methodological issue that
these participants have problems in understanding and competing the experiment. The full
regression tables can be found in Appendix 5. For the
independent variables
a
n there is
d no
the direction and significance of the investigated effect. We now find a weak significant and
,

positive
impact
there are
minor
variables.

major difference in

,,

o recent
f claiming
differences in
the

experien
an
ce
d
direction

in

level of

the questionnaire.
Additionally

significance for some control

()

Finally, when treating the stated attitude toward fictitious claims as an independent variable
(Model 1c), the latter is insignificant. Therefore, the stated attitude does not significantly
affect the behavior in the experiment. This finding is in line with our analysis in the previous
section. For this model, we do not provide the same robustness check as for Model 1a and
Model 1c. In the check described above we exclude all participants who agree with fraud in
the questionnaire but disagree in the experiment. After the exclusion, all individuals who do
not accept fraud in the questionnaire do also not accept fraud in the experiment. Therefore
the predicted fraud probability in the experiment for has to be one for this category (one-way
causation).

22
Model 1a
Questionnaire
FICTITOUS CLAIMS
(QUESTIONNAIRE)
SATISFACTION
NEGATIVE VIEW

dependent variable / model

Model 1b
Experiment
-

-0.6011028
(0.6312042)
0.6406695 ***
(0.185677)

Model 1c
Experiment
0.5446891
(0.5258146)

0.1274511
(0.25801)

0.1379444
(0.258641)

0.0085749
(0.0681258)

-0.0020064
(0.0690032)

CLAIMS

0.5137144
(0.4922188)

-0.2237958
(0.1948811)

-0.2354098
(0.1954144)

CONTRACTS

0.1839095
(0.1816373)

-0.0984059
(0.0755897)

-0.1020047
(0.0757705)

FEMALE

-0.7981881
(0.4938245)

-0.3599954 *
(0.1924543)

-0.3490582 *
(0.1927819)

RISK

0.4028744 **
(0.1690424)

0.2589551 ***
(0.0642794)

0.2525515 ***
(0.0645149)

MAJOR

0.7492009
(0.5792261)

0.9759309 ***
(0.2909073)

0.967264 ***
(0.2913237)

GROUP

-0.0358143
(0.5980268)

0.4395276 *
(0.2616858)

0.4433886 *
(0.2622353)

0.1905198
(0.680697)

-0.5933166 **
(0.2586237)

-0.5951297 **
(0.2593034)

0.3185786
(0.2631442)

0.3153274
(0.2635156)

-0.3318901
(0.3377757)

-0.2996076
(0.3390073)

BOMA
DEDUCT
Constant

0.3250499
(0.6133695)
-6.972874 ***
(1.110147)

Number of Observations
552
552
Log Likelihood
-78.03
-352.11
Significance levels p: *** p < 0.01; ** p < 0.05; * p < 0.1.
The standard errors are stated in parenthesis below the coefficients.

552
-351.54

Table 8: Logistic Regression for fictitious claims

Table 9 displays the results concerning claim build-up. Here, the influencing factors do not
have a consistent impact for all models. However, a negative perception of the insurance
industry in general only has a positive effect on agreement to fraud in the questionnaire.
Somehow surprisingly, participants with real-world claim experience have a higher
acceptance of claim build-up in the questionnaire. But a higher number of insurance
contracts has only in the questionnaire the opposite effect. As before, we find significant
effects for , and in the experiment. Participants in treatments with
deductible contracts have a higher tendency to defraud in the experiment and reply in the
questionnaire that claim inflation is acceptable. This finding is intuitive, since deductible 23

contracts may be perceived as unfair. As before, participants in larger groups with less
transaction costs do defraud more often in the experiment. But in the questionnaire (after the
experiment) these participants state a lower acceptance of claim build-up. Participants in
bonus-malus treatment state a higher acceptance in the questionnaire with respect to claim
build-up. This is very surprising, because premium increases after losses are not affected by
any claim inflation activities. Again, the stated attitude towards claim build-up does not
significantly affect the behavior in the experiment. The regression results show that for both
fraud types the impact of potential influencing factors on insurance fraud differs for fraud in
the questionnaire and fraud in the experiment.

Model 2a
Questionnaire
CLAIM BUILD-UP
(QUESTIONNAIRE)

dependent variable / model

Model 2b
Experiment
-

Model 2c
Experiment
-0.1130159
(0.2062041)

SATISFACTION

-0.117751
(0.2741898)

0.0591365
(0.2634079)

0.0565623
(0.2634022)

NEGATIVE VIEW

0.6352588 ***
(0.0807376)

0.0762532
(0.0687932)

0.0913426
(0.0741779)

0.3747613 *
(0.209563)

0.1265239
(0.2016172)

0.1358467
(0.2023874)

CONTRACTS

-0.1639345 *
(0.0846644)

0.0450852
(0.079509)

0.0415258
(0.0797409)

FEMALE

-0.2854212
(0.2094693)

-0.4239537 **
(0.1994057)

-0.4308138 **
(0.1998664)

RISK

0.1034024
(0.0676409)

0.1138266 *
(0.0644989)

0.1159491 *
(0.0646079)

MAJOR

0.1667317
(0.3008636)

0.8038195 ***
(0.2926293)

0.8064087 ***
(0.2923967)

GROUP

-0.840668 ***
(0.2847806)

0.7294585 ***
(0.2672125)

0.7094687 ***
(0.2695875)

BOMA

0.4953708 *
(0.2810854)

0.1911555
(0.2665228)

0.2037659
(0.2674406)

DEDUCT

0.5471097 *
(0.2849863)

0.5648161 **
(0.2683106)

0.5788966 **
(0.2698744)

Constant

-2.041513 ***
(0.3906996)

-0.9778661 ***
(0.3579997)

-0.9703043 ***
(0.3582354)

CLAIMS

Number of Observations
507
507
Log Likelihood
-303.29
-324.94
Significance levels p: *** p < 0.01; ** p < 0.05; * p < 0.1.
The standard errors are stated in parenthesis below the coefficients.

507
-324.79

Table 9: Logistic regressions for claim build-up

24

Robustness checks with reduced number of participants for Model 2a and Model 2b are
provided in Appendix 6. Again, we excluded participants who agreed with claim build-up in
the questionnaire but do not agree with this fraud type in the experiment. We now find
conformity between their answers and their behavior for the variables
as well as
still find only weak

We
,

and.

significant effects for recent claiming experience

in the

,,,,
of the experiment.

questionnaire. But the impact is insignificant when fraud is generated out ()

Participants belonging to the bonus-malus treatment have significantly higher acceptance of
claim build-up in the questionnaire. But this effect is still not observable when fraud comes
out of the experiment.

5 Conclusion
When analyzing the extent and influencing factors for insurance fraud, consumer surveys are
an important data source. Given that insurance fraud is a sensitive topic, an important
question is how reliable this kind of data is. We match individuals expressed attitude towards
insurance fraud in a questionnaire with their behavior in an experiment. For the fraud types
fictitious claims and claim build-up, we only find a weak link between the stated attitude
toward fraud in the questionnaire and the behavior in the experiment. For almost half of the
participants the self-stated attitude in the questionnaire does not coincide with their behavior
in the experiment. Hence, individuals do not accept insurance fraud in the questionnaire, but
cheat in the experiment. When closer analyzing factors that might influence the attitude
towards insurance fraud and respective behavior, we do not find many common effects of
influencing factors.
Of course a lab experiment is far from being realistic. Therefore, the generalizability of our
experiment is definitely limited. However, in an experiment, talk is not cheap. Behavior has
monetary consequences and participants may not feel a direct pressure to behave or state
attitudes according to social standards. Given our findings, we would be cautious to take
survey results on the self-stated attitude with respect to insurance fraud at face value. But
due to the fact that fraud is a major problem for the insurance industry, it seems worthwhile to
further explore.

25

References
Artis, Manuel, Ayuso, Mercedes and Guillen, Montserrat (1999): Modelling different types of
automobile insurance fraud behavior in the Spanish market, Insurance: Mathematics
and Economics, 24(1-2): 67-81.
Artis, Manuel, Ayuso, Mercedes and Guillen, Montserrat (2002): Detection of Automobile
Insurance Fraud With Discrete Choice Models and Misclassified Claims, Journal of
Risk and Insurance, 69(3): 325-340.
von Bieberstein, Frauke and Schiller, Jrg (2014): Contract Design and Insurance fraud: an
Experimental Investigation, Working Paper, University of Hohenheim.
Blackburn, McKinley, Harrison, Glenn W., and Rutstrm, E. Elisabeth (1994): Statistical Bias
Functions and Informative Hypothetical Surveys, American Journal of Agricultural
Economics, 76(5): 1084-1088.
Campbell, Donald T. (1957): Factors relevant to the Validity of Experiments in Social Settings,
Psychological Bulletin 54(4): 297-312.
Colquitt, Lee L. and Hoyt, Robert E. (1997): An Empirical Analysis of the Nature and Cost of
Fraudulent Life Insurance Claims: Evidence From Resisted Claims Data, Journal of
Insurance Regulation 15(4): 451-479.
Crocker, Keith J. and Morgan, John (1998): Is Honesty the Best Policy? Curtailing Insurance
Fraud through Optimal Incentive Contracts, Journal of Political Economy 106(2): 355375.
Croson, Rachel and Gneezy, Uri (2009): Gender Differences in Preferences, Journal of
Economic Literature 47(2): 1-27.
Cummins, David J. and Tennyson, Sharon (1996): Moral Hazard in Insurance Claiming:
Evidence from Automobile Insurance, Journal of Risk and Uncertainty 12(1): 29-50.
Dionne, Georges (2013): The Empirical Measure of Information Problems with Emphasis on
Insurance Fraud and Dynamic Data, in: in: Dionne, Georges (Ed.): Handbook of
Insurance, Second Edition, New York, 423-448.
Debreband, Serge (2011): Versicherungsbetrug - Ist mir leider aus der Hand gefallen, GDV
press release of 09.06.2011, http://www.gdv.de/2011/06/hintergrund-ist-mir-leider-ausder-hand-gefallen/
Dionne, Georges and Gagne, Robert (2001): Deductible Contracts against Fraudulent
Claims: Evidence from Automobile Insurance, Review of Economics and Statistics
83(2): 290-301.
Dionne, Georges, Giuliano, Florence, and Picard, Pierre (2009): Optimal Auditing with
Scoring: Theory and Application to Insurance Fraud, Management Science 55(1): 5870.
Elliott, Delbert S., Wilson, William Julius, Huizinga, David, Sampson, Robert J., Elliott,
Amanda, and Rankin, Bruce (1996): The Effects of Neighborhood Disadvantage on
26

Adolescent Development, Journal of Research in Crime and Delinquency 33(4): 389426.

Fendrich, Michael and Vaughn Connie M. (1994): Diminished Lifetime Substance use Over
Time: An Inquiry into Differential Underreporting, Public Opinion Quarterly 58(1): 96123.
Festinger, Leon (1962): A theory of cognitive dissonance. Vol. 2. Stanford University Press.
Frank, Bjrn and Schulze, Gnther G. (2000): Does economics male citizens corrupt?,
Journal of Economic Behavior & Organization 43(1): 101-113.
Frey, Bruno S. and Meier, Stephan (2004): Pro-Social behavior in a natural setting, Journal of
Economic Behavior & Organization 54(1): 65-88.
Fullerton, Sam, Kerch, Kathleen B., and Dedge, H. Robert (1996): Consumer Ethics: An
Assessment of Individual Behavior in Market Place, Journal of Business Ethics 15(7),
805-815.
Gabaldon, Ignacio Moreno, Hernandez, Francisco Jose Vazquez, and Watt, Richard (2011):
The Effect of Contract Type on Insurance Fraud, Working Paper, Universidad de
Malaga.
Harrison, Glenn W. and List, John A. (2004): Field Experiments, Journal of Economic
Literature 42(4): 1009-1055.
Harrison, Glenn W., List, John A., and Towe, Charles (2007): Naturally Occurring Preferences
and Exogenous Laboratory Experiments: a Case Study of Risk Aversion,
Econometrica 75(2): 433-458.
Heeb, Jean-Luc and Gmel, Gerhard (2005): Measuring alcohol consumption: A comparison
of graduated frequency, quantity frequency, and weekly recall diary methods in a
general population survey, Addictive Behaviors 30(3): 403-413.
Hoyt, Robert E. (1990): The Effect of Insurance Fraud on the Economic System, Journal of
Insurance Regulation 8(3): 304-315.
John, Karsten (2011): Versicherungsbetrug: aktuelle Entwicklungen, Muster und ihre Abwehr,
GDV press release of 12.07.2011, http://www.gdv.de/wp-content/uploads/2011 /
11/PK_Versicherungsbetrug_2011_Praes1_Versicherungsbetrug_in_Deutschland_G
fK_KarstenJohn_n4.pdf.
Lammers, Frauke and Jrg Schiller (2010): Einflussfaktoren fr betrgerisches Verhalten von
Versicherungsnehmern: erste experimentelle Befunde, Zeitschrift fr die gesamte
Versicherungswissenschaft 99 (5): 649-663.
Miyazaki, Anthony D. (2009): Perceived Ethicality of Insurance Claim Fraud: Do Higher
Deductibles Lead to Lower Ethical Standards?, Journal of Business Ethics 87(4): 589598.
Moreno, Ignacio, Vazquez, Francisco J., and Watt, Richard (2006): Can Bonus-Malus
alleviate Insurance Fraud?, Journal of Risk and Insurance 73(1): 123-151.
27

Moosbrugger,
Helfried
and
Kelava,
Augustin
Fragebogenkonstruktion, Second Edition, Berlin.

(2012):

Testtheorie

und

Mller, Katja (2013): The Identification of Insurance Fraud - An Empirical Analysis -, Working
Papers on Risk Management and Insurance No. 137, University of St. Gallen.
Nell, Martin (1998): Das moralische Risiko und seine Erscheinungsformen, in: Frankfurter
Vortrge zum Versicherungswesen 29, Karlsruhe.
Picard, P. (1996): Auditing claims in the insurance market with fraud: The credibility issue,
Journal of Public Economics, 63(1): 27-56.
Picard, Pierre (2013): Economic Analysis of Insurance Fraud, in: Dionne, Georges (Ed.):
Handbook of Insurance, Second Edition, New York, 349-395.
Psychonomics (1996): Versicherungsbetrug: Erklrung und Prvention Die erste Studie zu
subjektiven Faktoren des Versicherungsbetrugs, Kln.
Richter, Andreas, Schiller, Jrg, and Schlesinger, Harris (2014): Behavioral Insurance:
Theory and Experiments, forthcoming: Journal of Risk and Uncertainty.
Schiller, Jrg (2006): The Impact of Insurance Fraud Detection Systems, Journal of Risk and
Insurance 73(3): 421-438.
Smith, Vernon (1976): Experimental Economics: Induced Value Theory, American Economic
Review 66(2): 274-279.
Thornberry, Terence P. and Krohn, Marvin D. (2000): The Self-Report Method for Measuring
Delinquency and Crime, in: Duffee, D. (Ed.): Measurement and Analysis of Crime and
Justice: Criminal Justice 2000, Volume 4, 33-83.
Tennyson, Sharon (1997): Economic institutions and individual ethics: A study of consumer
attitudes towards insurance fraud, Journal of Economic Behavior & Organization
32(2): 247-265.
Tennyson, Sharon (2002): Insurance Experience and Consumers Attitude toward Insurance
Fraud, Journal of Insurance Regulation 21(2): 35-55.
Tourangeau, Roger and Yan, Ting (2007): Sensitive Questions in Surveys, Psychological
Bulletin 133(5): 859-883.
Tourangeau, Roger, Rips, Lance J., and Rasinski, Kenneth A. (2009): The Psychology of
Survey Response, Cambridge.
Townsend, Robert M. (1979): Optimal Contracts and Competitive Markets with Costly State
Verification, Journal of Economic Theory, 21(2): 265-293.
Weeks, William A., Moore, Carlos W., McKinney, Joseph A., and Longenecker, Justin G.
(1999): The Effects of Gender and Career Stage on Ethical Judgment, Journal of
Business Ethics 20(4): 301-313.
Zalma, Barry (2010): Insurance Fraud, Coalition against insurance fraud press release of
25.02.2010, http://www.insurancefraud.org/article.lasso?RecID=2988.
28

Appendix 1: Questionnaire Data Fraudulent behavior

Fictitious claims:

In general, I can imagine to report a claim that never occurred

to receive an indemnity from the insurance company.

Claims exaggeration:

In general, I can imagine to slightly exaggerate claims in case a

loss occurred.

Redefinition of claims:

In general, in case a loss occurred, I can imagine to report

wrong information regarding the course of events towards the
insurance company to receive an indemnity at all.

Insurance fraud in general: In general, it is acceptable to report wrong information to the

insurer.

29

Appendix 2: Alternative Segmentation

Fictitious claims
Disagree with fraud in the
Agree with fraud in the
experiment
experiment
Disagree with fraud in the
22.28 %
73.73 %
questionnaire
(123)
(407)
Agree with fraud in the
0.18 %
3.80 %
questionnaire
(1)
(21)
Note: The table includes the observations of 552 individuals. 24 observations are not
included because their attitude concerning fictitious claims is categorized as neutral in the
questionnaire. The number of observations for each category is stated in parenthesis
below the percentages. A McNemars chi-square statistic indicates a statistically significant
difference at a 1 percent level.
Table 10: Match of fictitious claims in experiment and questionnaire
Claim build-up
Disagree with fraud in the
Agree with fraud in the
experiment
experiment
Disagree with fraud in the
17.95 %
31.36 %
questionnaire
(91)
(159)
Agree with fraud in the
16.17 %
34.52 %
questionnaire
(82)
(175)
Note: The table includes the observations of 507 individuals. 69 observations are not
included because their attitude concerning claim build- up is categorized as neutral in the
questionnaire. The number of observations for each category is stated in parenthesis
below the percentages. A McNemars chi-square statistic indicates a statistically significant
difference at a 1 percent level.
Table 11: Match of claim build-up in experiment and questionnaire
Fraud in general
Disagree with fraud in the
Agree with fraud in the
experiment
experiment
Disagree with fraud in the
17.74 %
69.62 %
questionnaire
(94)
(369)
Agree with fraud in the
1.70 %
10.94 %
questionnaire
(9)
(58)
Note: The table includes the observations of 530 individuals. 46 observations are not
included because their attitude concerning fraud in general is categorized as neutral in the
questionnaire. The number of observations for each category is stated in parenthesis
below the percentages. A McNemars chi-square statistic indicates a statistically significant
difference at a 1 percent level.
Table 12: Match of fraud in general in experiment and questionnaire

30

Appendix 3: Questionnaire Data Independent variables

SATISFACTION:
SATISFACTION GENERAL: How satisfied are you in general with your insurance? (Please
skip the question if you do not have any insurance contracts)
SATISFACTION CONTRACTS: If you already have claimed a loss, how satisfied have you
been with the claims settlement process? (Please skip the
question if you did not claim a loss)

NEGATIVE VIEW:
COMPLICATED CONTRACTS: To make a lot of money, insurance companies construct their
contracts very complicated.
EARNINGS:

Insurance companies earn so much money, that it is no

problem if customers retrieve their paid premiums.

RECENT CLAIMS:

When did you lastly claim a loss?

CONTRACTS:

How many insurance contracts do you have (not concluded by

your parents)?

FEMALE:

What is your gender?

RISK:

How do you assess yourself: Are you in general more risk

seeking or do you try to avoid risks?

MAJOR:

What are you studying?

31

Appendix 4: Regression table Model 3

Model 3a
FRAUD IN GENERAL
(QUESTIONNAIRE)

dependent variable / model

Model 3b
-

Model 3c
0.5092978
(0.3151518)

SATISFACTION

0.2777229
(0.3447836)

-0.0472918
(0.2665708)

-0.0589142
(0.2679711)

NEGATIVE VIEW

0.5440002 ***
(0.1103533)

0.1028964
(0.0711045)

0.0769129
(0.072878)

CLAIMS

0.6133291 **
(0.2959884)

-0.075904
(0.2038155)

-0.1056978
(0.2051018)

CONTRACTS

-0.1902664
(0.1304612)

-0.0736677
(0.0786408)

-0.0653826
(0.0790023)

FEMALE

-0.4274328
(0.2880471)

-0.2829882
(0.2040918)

-0.2657588
(0.2047709)

RISK

0.1134835
(0.0936771)

MAJOR

-0.6757893
(0.4863195)

0.7900403 **
(0.3148028)

0.8147962 **
(0.3155086)

GROUP

-0.487931
(0.3938714)

0.9805861 ***
(0.2880899)

1.011056 ***
(0.2895946)

BOMA

0.2774118
(0.3786757)

-0.2950543
(0.2614914)

-0.3070216
(0.2625096)

DEDUCT

0.3285957
(0.3934913)

0.3096084
(0.2791766)

0.2924325
(0.2800345)

Constant

-3.98013 ***
(0.5850051)

-0.4802794
(0.3610004)

-0.4624968
(0.3616341)

0.2487395 ***
(0.066885)

Number of Observations
530
530
Log Likelihood
-177.62
-324.23
Significance levels p: *** p < 0.01; ** p < 0.05; * p < 0.1.
The standard errors are stated in parenthesis below the coefficients.

0.2456196 ***
(0.0670047)

530
-322.87

Table 13: Logistic regressions for fraud in general

32

Appendix 5: Regression table Model 1 with reduced number of participants

dependent variable / model
Model 1a
Model 1b
Questionnaire
Experiment
SATISFACTION

-0.137212
(0.843709)

0.1713012
(0.263439)

NEGATIVE VIEW

0.762166 ***
(0.2285536)

0.0173268
(0.0693252)

CLAIMS

1.083246 *
(0.5975732)

-0.2745966
(0.1976486)

CONTRACTS

0.0449196
(0.2328893)

-0.0901808
(0.0764041)

FEMALE

-0.6761494
(0.5795864)

-0.4076423 **
(0.1949906)

RISK

0.4688845 **
(0.2091866)

0.27348 ***
(0.0652968)

MAJOR

0.6508221
(0.7193627)

1.076541 ***
(0.3021016)

GROUP

0.3982428
(0.7533969)

0.3858759
(0.2658366)

BOMA

-0.1365368
(0.8381087)

-0.5608003 **
(0.2610179)

DEDUCT

-0.1694267
(0.7858404)

0.412277
(0.2691268)

Constant

-8.028374 ***
(1.356513)

-0.3563582
(0.3405476)

Number of Observations
546
546
Log Likelihood
-57.85
-344.69
Significance levels p: *** p < 0.01; ** p < 0.05; * p < 0.1.
The standard errors are stated in parenthesis below the coefficients.

Table 14: Logistic Regression for fictitious claims

33

Appendix 6: Regression table Model 2 with reduced number of participants

dependent variable / model
Model 2a
Model 2b
Questionnaire
Experiment
SATISFACTION

-0.109019
(0.3342281)

-0.1707937
(0.3364634)

NEGATIVE VIEW

0.6371252 ***
(0.0998702)

0.3926565 ***
(0.0919389)

CLAIMS

0.4700271 *
(0.2607505)

CONTRACTS

-0.0730495
(0.1072196)

-0.000414
(0.0930073)

FEMALE

-0.6248819 **
(0.2590952)

-0.5704957 **
(0.2525971)

RISK

0.1623418 *
(0.0838444)

0.1474814 *
(0.0809779)

0.6287427 *
(0.341885)

0.7741426 **
(0.3669833)

MAJOR

0.1788879
(0.252064)

GROUP

-0.4146723
(0.3349029)

0.2128278
(0.3475813)

BOMA

0.686782 *
(0.3589225)

0.4362015
(0.3377403)

DEDUCT

0.7971227 **
(0.3553193)

0.9377498 **
(0.3630112)

Constant

-3.268936 ***
(0.5014541)

-1.075555 **
(0.435562)

Number of Observations
374
374
Log Likelihood
-196.65
-214.70
Significance levels p: *** p < 0.01; ** p < 0.05; * p < 0.1.
The standard errors are stated in parenthesis below the coefficients.

Table 15: Logistic Regression for claim build-up

34