Can we trust consumers survey answers when dealing with
insurance fraud? Evidence from an experiment#
Kerstin Puchstein*Jrg SchillerFrauke 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 Versiche-rungswissenschaft 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, E-Mail:
[email protected].
1
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.
2
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 Some authors classify these characteristics of validity as
internal and external validity (see e.g. Campbell, 1957).
3 Such categories are e.g. content validity, face validity and
construct validity (see e.g. Thornberry and Krohn, 2000 or
Moosbrugger and Kelava 2012). 4 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).
3
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 socio-demographic 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
4
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 self-reported 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 socio-demographic 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
6
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
variableauthorInfluence
Fullerton et al.(1996)Unethical behavior declines with an
ageincrease in age.
Artis et al. (2002)Younger individuals have a higher fraud
probability.
Fullerton et al. (1996)The acceptance for unethical behavior
increases with income.
Tennyson (2002)High income individuals are less likely to
incomeaccept 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.
Tennyson (2002)Men are more willing to accept fraudulent
genderbehavior.
Weeks et al. (1999)There are differences in ethical judgment
between women and men.
Elliot et al. (1996)The living area influences individuals
residentialbehavior.
The surrounding field has an impact on
area
Tennyson (1997)individuals attitude towards insurance
fraud.
Experienceand perception towards the insurance industry
variableauthorinfluence
perceptionTennyson (1997)A negative perception increases the
tendency to cheat in insurance contracts.
Tennyson (2002)Recent claiming experience lowers the
acceptability of insurance fraud.
experienceArtis et al. (1999)A higher number of records is
associated
with a higher fraud probability.
Mller (2013)The lower the number of previous records,
the higher is the tendency to defraud.
Contract settings
variableauthorinfluence
bonus-Gabalon et al. (2009)Previous experience with
bonus-malus
maluscontracts reduces the fraud probability.
Dionne and Gagne (2001)The level of the deductible is a
significant
deductiblesfactor for the submitted claim.
Miyazaki (2009)For higher deductibles individuals consider
claim build-up as more justified.
bonus-von Bieberstein and SchillerIndividuals commit more fraud
given a
malus vs.deductible contract compared to a payoff-
(2014)
deductiblesequivalent bonus-malus contract.
Table 1: Overview of influencing factors on insurance fraud
9
3 Experimental Design
The experiment was conducted at the MELESSA laboratory at the
Ludwig-Maximilians-University 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 computer-based 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 thebeginning of each
period every participant obtains a period endowment.5
Participants
are informed that they have to insure against the risk to suffer
no loss( ), a small loss
or a large loss. The losses occur in each periodwith a0
probability of
(= 0)
andThe specific losses are independently determined for each
group0
( = 10)( . = 15)=
memberin each period.Insurance is mandatory in the experiment.
Thus, each group
0.7,= 0.2= 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 strategy0with0in
decisions about how muchto 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 isfor each period and the possible losses aretheand.
The
is, which equals the expected value ofpossible losseswith
period premium = 25= 10= 15
transaction costs.Additionally there are three modifications of
the Base Treatment. The first
= 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 lossesand
,while theclaimable indemnityremains unchanged. Hence,
theendowmentis
= 15
increasedto, whereas the premium is. With the Bonus-Malus
Treatment an
= 20
rating is introduced. Thereby the premium is not a fixed amount,
it depends on
experience= 27= 5
the indemnity of the previous periods. In the first period it
starts with. In the following
periodthe premiumincreases by 2 points in case anindemnitywas
received in
period+1= 5
and decreases by
1 point in case no indemnity was paid. The newstarting point
for
+ 1
the nextperiodis the premium in. The Deductible Treatment and
the Bonus-Malus
been applied in the first and second round of data collection.
6
Treatment have + 2+ 1
6 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 defraudsFictitious claimsClaim build-up
022.40 %33.33 %
112.85 %11.28 %
210.42 %7.47 %
312.33 %7.29 %
49.55 %7.29 %
532.47 %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.
Fictitious claims
66.67%
60,00%
40,00%
19.27%20,00%6.08% 4.17%2.43% 1.22% 0.17%0,00%
DisagreementAgreement
92.02 %3.82%
In general, I can imagine to report a claim that never occurred
to receive an indemnity from the insurance company.
Claims exaggeration
60,00%
40,00%24.83%
20,00%15.80%16.67%11.98
%14.58%
10.94%5.21%
0,00%
DisagreementAgreement
43.41 %44.62%
In general, I can imagine to slightly exaggerate claims in case
a loss occurred.
Redefinition of claims
60,00%
40,00%27.26%21.18%10.24
20,00%15.63%
11.28%%9.55%4.86%
0,00%
0123456
DisagreementAgreement
59.72 %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.
Insurance fraud in general
60,00%
40,00% 33.33% 32.99%20,00%14.06% 7.99%8.51%
0,00%2.60% 0.52%
0123456
DisagreementAgreement
80.38 %11.63%
In general, it is acceptable to report wrong information to the
insurer.
Figure 1: Fraud in the questionnaire
7 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
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 theAgree with fraud in the
experimentexperiment
Disagree with fraud in the44.92 %51.09 %
questionnaire(248)(282)
Agree with fraud in the1.09 %2.90 %
questionnaire(6)(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 theAgree with fraud in the
experimentexperiment
Disagree with fraud in the26.04 %23.27 %
questionnaire(132)(118)
Agree with fraud in the26.23 %24.46 %
questionnaire(133)(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 theAgree with fraud in the
experimentexperiment
Disagree with fraud in the34.34 %53.02 %
questionnaire(182)(281)
Agree with fraud in the3.40 %9.25 %
questionnaire(18)(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
FictitiousClaim build-Fraud inFictitiousClaim build-Fraud in
claimsupgeneralclaimsgeneral
up (experi-
(question-(question-(question-(experi-(experi-
ment)
naire)naire)naire)ment)ment)
Fictitious
claims1.0000
(questionnaire)
Claim build-up0.1992 ***1.0000
(questionnaire)
Fraud in
general0.1873 ***0.3646 ***1.0000
(questionnaire)
Fictitious
claims0.0766 *0.04440.1079 **1.0000
(experiment)
Claim build-up0.00890.01050.0856 **0.6002 ***1.0000
(experiment)
Fraud in
general0.04520.01890.0853 **0.8482 ***0.7457 ***1.0000
(experiment)
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) = ( 0 + 1+ 2 +
3 + 4 + 5 + 6 +
7 + 8 + 9 + 10 + )
In the following we present results for six different logistic
regressions. The models 1a, 1b
and 1c considerfictitious claims. Model1ais relatedto
thestatedattitude 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) = ( 0 + 1 + 2 +
3+ 4 + 5 + 6 +
7 + 8 + 9 + 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
andmeasure the participants experience with insurance
relationships.
The variablecorresponds to the number of participants insurance
contracts.
coefficientis equal to 1 if the respondent has reported a claim
in the last two
The
in case of absence of any claims. The explanatory dummy
variable
years and is 0accounts for
measures gender differences (1 = female). Furthermore, the
variable
differences in the participants stated risk attitude. The
variable is alsomeasured on a seven
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
variableis 1 for participants with a major in economics or
business and 0 otherwise.
The dummy variableis 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 variableequals 1 for all
participants of treatments
with bonus-malus contracts. Respectively,the dummy variableis
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
VariableObservationsMeanStandardMinMax
Deviation
Fictitious Claims (Questionnaire)5520.0400.19601
Claim build-up (Questionnaire)5070.5070.50001
Fraud in general (Questionnaire)5300.1260.33301
Fictitious Claims (Experiment)5760.5430.49901
Claim build-up (Experiment)5760.4790.50001
Fraud in general (Experiment)5760.6230.48501
Satisfaction5760.1680.37501
Satisfaction general3533.8471.33506
Satisfaction claims2693.9111.75106
Negative view5763.0271.38006
Complicated contracts5763.8301.69606
Earnings5762.2241.67606
Claims5760.4430.49701
Recent claims5760.7080.86002
Contracts5761.0991.21506
Female5760.5990.49101
Risk5762.7851.47806
Major5760.1420.35001
Group5760.2500.43301
Boma5760.2500.43301
Deduct5760.2500.43301
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
fullregression tables can be found in Appendix 5. For the
independent variables
major difference in
andthere is no,
the direction and significance of the investigated effect. We
now find a weak significant and
,,,
positive impactofrecent claimingexperienceandinthe
questionnaire. Additionally
there are minordifferences in thedirectionlevel ofsignificance
for some control
()
variables.
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
dependent variable / model
Model 1aModel 1bModel 1c
QuestionnaireExperimentExperiment
FICTITOUS CLAIMS--0.5446891
(QUESTIONNAIRE)(0.5258146)
SATISFACTION-0.60110280.12745110.1379444
(0.6312042)(0.25801)(0.258641)
NEGATIVE VIEW0.6406695***0.0085749-0.0020064
(0.185677)(0.0681258)(0.0690032)
CLAIMS0.5137144-0.2237958-0.2354098
(0.4922188)(0.1948811)(0.1954144)
CONTRACTS0.1839095-0.0984059-0.1020047
(0.1816373)(0.0755897)(0.0757705)
FEMALE-0.7981881-0.3599954*-0.3490582*
(0.4938245)(0.1924543)(0.1927819)
RISK0.4028744**0.2589551***0.2525515***
(0.1690424)(0.0642794)(0.0645149)
MAJOR0.74920090.9759309***0.967264***
(0.5792261)(0.2909073)(0.2913237)
GROUP-0.03581430.4395276*0.4433886*
(0.5980268)(0.2616858)(0.2622353)
BOMA0.1905198-0.5933166**-0.5951297**
(0.680697)(0.2586237)(0.2593034)
DEDUCT0.32504990.31857860.3153274
(0.6133695)(0.2631442)(0.2635156)
Constant-6.972874***-0.3318901-0.2996076
(1.110147)(0.3377757)(0.3390073)
Number of Observations552552552
Log Likelihood-78.03-352.11-351.54
Significance levels p: *** p < 0.01; ** p < 0.05; * p <
0.1.The standard errors are stated in parenthesis below the
coefficients.
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.
dependent variable / model
Model 2aModel 2bModel 2c
QuestionnaireExperimentExperiment
CLAIM BUILD-UP---0.1130159
(QUESTIONNAIRE)(0.2062041)
SATISFACTION-0.1177510.05913650.0565623
(0.2741898)(0.2634079)(0.2634022)
NEGATIVE VIEW0.6352588***0.07625320.0913426
(0.0807376)(0.0687932)(0.0741779)
CLAIMS0.3747613*0.12652390.1358467
(0.209563)(0.2016172)(0.2023874)
CONTRACTS-0.1639345*0.04508520.0415258
(0.0846644)(0.079509)(0.0797409)
FEMALE-0.2854212-0.4239537**-0.4308138**
(0.2094693)(0.1994057)(0.1998664)
RISK0.10340240.1138266*0.1159491*
(0.0676409)(0.0644989)(0.0646079)
MAJOR0.16673170.8038195***0.8064087***
(0.3008636)(0.2926293)(0.2923967)
GROUP-0.840668***0.7294585***0.7094687***
(0.2847806)(0.2672125)(0.2695875)
BOMA0.4953708*0.19115550.2037659
(0.2810854)(0.2665228)(0.2674406)
DEDUCT0.5471097*0.5648161**0.5788966**
(0.2849863)(0.2683106)(0.2698744)
Constant-2.041513***-0.9778661***-0.9703043***
(0.3906996)(0.3579997)(0.3582354)
Number of Observations507507507
Log Likelihood-303.29-324.94-324.79
Significance levels p: *** p < 0.01; ** p < 0.05; * p <
0.1.The standard errors are stated in parenthesis below the
coefficients.
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
findconformity betweentheir answers and their behavior for the
variablesWe
as well asand. ,
still find only weaksignificant effects for recent claiming
experiencein 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): 355-375.
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-aus-der-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): 58-70.
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): 389-426.
Fendrich, Michael and Vaughn Connie M. (1994): Diminished
Lifetime Substance use Over Time: An Inquiry into Differential
Underreporting, Public Opinion Quarterly 58(1): 96-123.
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): 589-598.
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 (2012): Testtheorie
und Fragebogenkonstruktion, Second Edition, Berlin.
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 theAgree with fraud in the
experimentexperiment
Disagree with fraud in the22.28 %73.73 %
questionnaire(123)(407)
Agree with fraud in the0.18 %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 theAgree with fraud in the
experimentexperiment
Disagree with fraud in the17.95 %31.36 %
questionnaire(91)(159)
Agree with fraud in the16.17 %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 theAgree with fraud in the
experimentexperiment
Disagree with fraud in the17.74 %69.62 %
questionnaire(94)(369)
Agree with fraud in the1.70 %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
dependent variable / model
Model 3aModel 3bModel 3c
FRAUD IN GENERAL--0.5092978
(QUESTIONNAIRE)(0.3151518)
SATISFACTION0.2777229-0.0472918-0.0589142
(0.3447836)(0.2665708)(0.2679711)
NEGATIVE VIEW0.5440002***0.10289640.0769129
(0.1103533)(0.0711045)(0.072878)
CLAIMS0.6133291**-0.075904-0.1056978
(0.2959884)(0.2038155)(0.2051018)
CONTRACTS-0.1902664-0.0736677-0.0653826
(0.1304612)(0.0786408)(0.0790023)
FEMALE-0.4274328-0.2829882-0.2657588
(0.2880471)(0.2040918)(0.2047709)
RISK0.11348350.2487395***0.2456196***
(0.0936771)(0.066885)(0.0670047)
MAJOR-0.67578930.7900403**0.8147962**
(0.4863195)(0.3148028)(0.3155086)
GROUP-0.4879310.9805861***1.011056***
(0.3938714)(0.2880899)(0.2895946)
BOMA0.2774118-0.2950543-0.3070216
(0.3786757)(0.2614914)(0.2625096)
DEDUCT0.32859570.30960840.2924325
(0.3934913)(0.2791766)(0.2800345)
Constant-3.98013***-0.4802794-0.4624968
(0.5850051)(0.3610004)(0.3616341)
Number of Observations530530530
Log Likelihood-177.62-324.23-322.87
Significance levels p: *** p < 0.01; ** p < 0.05; * p <
0.1.The standard errors are stated in parenthesis below the
coefficients.
Table 13: Logistic regressions for fraud in general
32
Appendix 5: Regression table Model 1 with reduced number of
participants
dependent variable / model
Model 1aModel 1b
QuestionnaireExperiment
SATISFACTION-0.1372120.1713012
(0.843709)(0.263439)
NEGATIVE VIEW0.762166***0.0173268
(0.2285536)(0.0693252)
CLAIMS1.083246*-0.2745966
(0.5975732)(0.1976486)
CONTRACTS0.0449196-0.0901808
(0.2328893)(0.0764041)
FEMALE-0.6761494-0.4076423**
(0.5795864)(0.1949906)
RISK0.4688845**0.27348***
(0.2091866)(0.0652968)
MAJOR0.65082211.076541***
(0.7193627)(0.3021016)
GROUP0.39824280.3858759
(0.7533969)(0.2658366)
BOMA-0.1365368-0.5608003**
(0.8381087)(0.2610179)
DEDUCT-0.16942670.412277
(0.7858404)(0.2691268)
Constant-8.028374***-0.3563582
(1.356513)(0.3405476)
Number of Observations546546
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 2aModel 2b
QuestionnaireExperiment
SATISFACTION-0.109019-0.1707937
(0.3342281)(0.3364634)
NEGATIVE VIEW0.6371252***0.3926565***
(0.0998702)(0.0919389)
CLAIMS0.4700271*0.1788879
(0.2607505)(0.252064)
CONTRACTS-0.0730495-0.000414
(0.1072196)(0.0930073)
FEMALE-0.6248819**-0.5704957**
(0.2590952)(0.2525971)
RISK0.1623418*0.1474814*
(0.0838444)(0.0809779)
MAJOR0.6287427*0.7741426**
(0.341885)(0.3669833)
GROUP-0.41467230.2128278
(0.3349029)(0.3475813)
BOMA0.686782*0.4362015
(0.3589225)(0.3377403)
DEDUCT0.7971227**0.9377498**
(0.3553193)(0.3630112)
Constant-3.268936***-1.075555**
(0.5014541)(0.435562)
Number of Observations374374
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