Optimism, overconfidence, and insurance decisions

Jennifer Coatsa, Vickie Bajtelsmita,*

aDepartment of Finance and Real Estate, Colorado State University, 1272 Campus Delivery, Fort Collins,

CO 80523, USA

Abstract

We report experimental evidence regarding overconfidence, optimism, and insurance decisions.

Our design distinguishes between an individual’s optimism bias and overconfidence bias, a contribu-

tion particularly important for understanding insurance decisions related to risks beyond the purchas-

er’s control. Results show that optimistic participants incur a higher total cost of risk and are more

likely to underinsure than non-optimistic participants, even when purchasing insurance maximizes

expected payoffs. In contrast, we find that overconfidence does not significantly affect the decision

to insure. However, participants with higher overall overconfidence show larger differences in insur-

ance behavior when the risk of loss arises from their own mistakes. © 2021 Academy of Financial

Services. All rights reserved.

Keywords: C9 Design of experiments; C91 Laboratory experiments; D81 Decision-making under uncertainty;

Overconfidence; Insurance demand

1. Introduction

Optimism bias, or the tendency to assign higher subjective probabilities to favorable out-

comes, is well documented in the psychology and economics literature. In fact, after decades

of controlled studies in psychology, the only individuals identified as consistently free from

this bias are the clinically depressed (Pyszczynski, Holt, & Greenberg, 1987). Overconfidence,

which can be viewed as a special case of optimism, relates to having a biased perception of

one’s own skills, prospects, or knowledge. Both optimism and overconfidence theoretically

affect decision-making under conditions of risk and uncertainty (De Bondt & Thaler, 1985).

Although previous studies differ in how these factors are defined and operationalized, the

*Corresponding author. Tel.: +1-970-491-0610.

E-mail address: vickie.bajtelsmit@colostate.edu (V. Bajtelsmit)

1057-0810/21/$ – see front matter © 2021 Academy of Financial Services. All rights reserved.

Financial Services Review 29 (2021) 1–28



conclusions overwhelmingly suggest that underestimation of the risk of negative outcomes has

an economically significant effect on individual and societal well-being.1 General optimism

about events outside of one’s own control may cause individuals to underestimate their actual

risk, which may lead them to make suboptimal financial decisions. For example, underestima-

tion of expected losses may result in reduced demand for insurance (Kunreuther & Pauly,

2004) and optimism about market performance may affect portfolio allocations (Jacobsen, Lee,

Marquering, & Zhang, 2014). Similarly, overestimation of one’s financial skills may result in

excessive and costly financial market trading (Barber & Odean, 2001) or investment in subopti-

mal business projects (Malmendier & Tate, 2005).2 In this article, we focus on the degree to

which heterogeneous risk perceptions, which may be affected by both optimism and overconfi-

dence, influence insurance and risk management decisions. To the extent that these biases

result in underestimation of personal risks, individuals are hypothesized to have reduced

demand for insurance and greater total cost of risk.

Although overconfidence and optimism are widely discussed in the psychology and eco-

nomics literatures, the influence of these behavioral biases on insurance decision-making

has received less attention. Information asymmetry in insurance markets, in which appli-

cants for insurance know more about their own risk characteristics than do insurers, is shown

theoretically to create the potential for market failure. With heterogeneous risk types, equi-

librium solutions result in separating contracts that encourage individuals to select price and

coverage policies that are appropriate to their risk type.3 For example, a high risk individual

might prefer a contract that provides relatively full coverage for a higher premium rate,

whereas an individual with a lower risk of loss might select a partial-coverage policy for a

lower premium rate. Critical to the success of self-selection equilibria models is that individ-

uals can correctly self-identify their risk type. Bajtelsmit and Thistle (2015) develop a model

in which noisy or imperfect information about risk types could result in suboptimal insur-

ance and risk management decisions. In this article, we consider information imperfection

caused by individual psychological biases that influence an individual’s risk perceptions.

Our study uses a novel experimental design which connects, through large monetary

incentives, an earnings task, a frequency estimation task, and insurance decisions.

Participants develop subjective estimates of their own and others’ performance on an earn-

ings task and decide, in light of their subjective probabilities of loss, whether to fully insure

against loss of earnings. In this article, we first focus on the origins of subjective probability

estimation, as it depends on task performance, and then develop measures to distinguish

between optimism and overconfidence biases in the estimation task. Finally, we investigate

how the biases affect the risk management decision to fully insure against loss. Our analysis

of insurance decisions in this article continues a stream of research from a large experimen-

tal design which, as a whole, considers an extensive set of variables shown theoretically to

affect risk management and insurance decisions.4

The experiment design allows us to distinguish between an individual’s optimism, or the

general tendency to overestimate favorable outcomes (“wishful thinking”), and their over-

confidence, the tendency to overestimate their own performance or skills (“I think I’m better

than I am”) and the relationship between these outlooks.5 We assess optimism with a treat-

ment manipulation in which a participant’s expected payoff is independent of their own per-

formance on a task, but positively related to the performance of others on the same task.

2 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



Participants’ outlooks range from underconfident to highly overconfident in their own per-

formance and from pessimistic to optimistic about others’. Participants who are optimistic

about others’ performance are generally confident in their own as well. We find that over-

confidence bias does not reduce the likelihood of insurance purchase, but that the highly op-

timistic are less likely to purchase insurance and, controlling for risk preferences, tend to

underinsure.

The remainder of the article is organized as follows. Section 2 describes related literature

on insurance and overconfidence. Section 3 explains the experimental design and hypothe-

ses. We provide results in Section 4, and discuss conclusions in the final section.

2. Related literature

The empirical link between overconfidence and the purchase of insurance is an important

contribution in support of recent theoretical findings (Huang, Liu, & Tzeng, 2010;

Spinnewijn, 2013) and leads to policy implications applicable to insurance contract design

and government policy interventions. Huang et al. (2010) show theoretically that hidden

overconfidence can lead to insurance market equilibria that are consistent with observed

anomalies in insurance markets. They suggest that insurers could use an overconfidence

proxy to screen prospective policyholders so as to achieve an advantageous selection equi-

librium and that regulators might need to intervene in markets that are operating imperfectly

as a result of this type of asymmetric information.6

In recent years, researchers have distinguished several different categories of overconfi-

dence bias. However, overconfidence, as a subset of optimism bias, has often remained con-

founded with a general optimistic outlook. The optimism bias leads individuals to

underweight the probability of negative outcomes that are beyond their control, such as an

airline crash or a wildfire, and overweight the probability of positive outcomes that are

beyond their control, such as winning a lottery. Spinnewijn (2013) refers to this phenom-

enon as “baseline optimism.” For example, Landry and Jahan-Parvar (2011) suggest that

failure to purchase flood insurance may be related to residents’ reliance on community pro-

tection policies (seawalls, beach replenishment, etc.). If the households are overly optimistic

regarding the success of the communities’ risk management efforts, then they may underin-
sure against loss.

In addition to a baseline level of optimism, individuals may be overconfident with respect

to their influence and/or abilities, which then leads to unrealistic optimism about outcomes,

a quality Spinnewijn (2013) terms “control optimism.” Royal and Tasoff (2017) show theo-

retically that overconfident agents are likely to reduce their payoffs by investing too much in

capital that complements their ability and too little in capital that substitutes for their own

ability. Their experimental results support the theoretical predictions. In the insurance con-

text, this could lead an individual to underinsure if they unrealistically believe they can

reduce the frequency or severity of loss through their own actions. A large literature in be-

havioral economics distinguishes further between different types of control optimism.

“Absolute overconfidence” is a term applied to individuals who overestimate their own

knowledge, ability, or performance against a given benchmark, such as the prediction that

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 3



they will perform better on a test or run faster than they actually do (Moore & Healy, 2008).

Perhaps the most well-known type of overconfidence is the “better-than-average effect,”

also termed “over-placement,” which refers to overestimation of one’s relative performance

in a group. An important quality of over-placement is “reference group neglect” in which

individuals rank their placement equally high among groups of self-selected members and

groups of exogenously assigned members (Camerer & Lovallo, 1999; Moore & Healy,

2008). Other types of control optimism include the “illusion of control,” which refers to the

case in which individuals erroneously perceive their actions to influence independent events,

and “calibration-based overconfidence,” or overconfidence in the precision of knowledge.7

Fellner and Krugel (2012) find evidence that overconfidence assessed with different methods

actually reflects separate and distinct biases.

Several recent studies suggest that heterogeneity of risk perceptions (Spinnewijn, 2013)

or risk preferences (De Meza & Webb, 2001) can explain the negative correlation between

risk and insurance coverage found in some markets.8 For example, as modeled in Sandroni

and Squintani (2007), differences in risk perceptions may lead some high-risk types to

believe that they are low-risk types, which can decrease investment in precaution and/or

reduce demand for insurance at offered prices. Huang et al. (2010) model optimistic indi-

viduals who have subjective loss probabilities that are lower than their objective loss prob-

abilities and “rational” individuals who assess their loss probability correctly. In their

model, higher optimism leads to lower likelihood of insurance purchase. Arad (2014)

presents a study in which participants are aware of objective probabilities but may assign a

higher or lower likelihood due to their own personal motivations that are unrelated to the

random event. Arad labels this phenomenon “magical thinking” and notes that beliefs

about one’s own good or bad luck, regardless of probability distribution can also lead to

suboptimal insurance decisions. Honl, Meissner, and Wulf (2017) develop a model in

which individual risk-taking behavior depends on cognitive processing of outcomes and

probabilities, affect in judgment and decision-making, and upon contextual factors. To the

extent that insurance purchasers do not know the objective likelihood of a loss event, it is

important to consider biases in estimating risk, and insurance decisions in light of subjec-

tive probability estimation.

While our experiment is not motivated as a study of gender effects or risk preferences per

se, extant studies suggest that it is important to control for both factors in the analysis.

Recent studies designed to investigate gender differences in decision-making find that opti-

mism about conditions or others’ performance varies across men and women. In an analysis

of survey data that includes several different indicators of optimism, Jacobsen et al. (2014)

find men to be more optimistic than women in their expectations about the general economic

outlook. In a study spanning three years, Foster and Frijters (2014) find that male university

students are more consistently overconfident than females about their future grades. In an

experiment task where payoffs depend on team performance, Kuhn and Villeval (2015) find

that both men and women expect to outperform others on their team, although women are

more optimistic than men about their team members’ performance.

Laboratory experiments designed to model insurance decisions offer the opportunity to

measure and control for beliefs, risk attitudes, and the set of risk management alternatives.

For example, Harrison and Ng (2016) find that, after measuring and controlling for risk

4 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



preferences, experimental participants tend to make welfare-reducing insurance decisions.

In another controlled laboratory experiment, Jaspersen and Aseervatham (2017) find that in-

surance demand decisions are often driven by biases and the use of heuristics. Laury and

McInnes (2003) find that information provided by actuarially fair insurance prices can

reduce experimental participants’ reliance on heuristics and improve decisions. When

experiment participants are allowed to insure against losses that depend on relative perform-

ance, Hales and Kachelmeier (2008) show that insurance decisions are affected by biases in

performance estimation. For a comprehensive survey on the experimental literature on insur-

ance demand, see Jaspersen (2016) who concludes that the decision context, decision task,

and the use of salient incentives all heavily influence experimental results. In this article, we

present the results of an experimental study in which we first elicit participants’ beliefs about

risk perceptions, and then investigate their subsequent decisions in laboratory insurance de-

cision tasks.

3. Experiment procedures and design and hypotheses

In this section, we first describe the design and procedures used in this laboratory experi-
ment, and then explain how this design can be used to test various hypotheses related to the
effects of overconfidence and optimism on risk management decisions. The experiment is
designed with incentive-compatible earnings and risk management tasks. The design also
includes an indirectly incentive-compatible frequency estimation task to elicit subjective
probabilities for losses that depend on the participants’ own ability and losses that are out-
side of their influence.9 The reported subjective probabilities allow us to measure partici-
pants’ overconfidence in their own performance as well as baseline optimism regarding
favorable outcomes. We use these to estimate the impact of overconfidence and optimism
biases on incentivized insurance purchase decisions under different risk conditions.

3.1. Procedures overview

Students were recruited from business classes at a large university to participate in a paid

experiment. All sessions were conducted with Z-tree (Fischbacher, 2007) in a networked

computer lab with partitioned stations. We conducted six sessions, each with ten partici-

pants, between June and October 2013. The experiment proceeded through two stages with

steps as summarized in Table 1, including participation payment, earnings task, instructions,

estimation task, and risk management task.

After participants were paid $15 up-front in cash (never at risk of loss) and learned the

experiment procedures, they earned $60 for correctly answering at least eight questions on a

quiz comprised of twenty questions drawn from previous driver licensing exams for the state

in which their university was located.10,11 For each question, participants were asked to indi-

cate whether they were sure they had answered it correctly. To incentivize participants to

carefully make these assessments, the instructions clearly explained the relationship between

their quiz performance, others’ quiz performance, the probability of loss, and their expected

earnings from the experiment. After participants demonstrated their understanding of these

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 5



T
ab
le

1
E
x
p
er
im

en
ta
l
p
ro
ce
d
u
re
s

F
ir
st
st
ag
e

S
te
p
1

•
$
1
5
in

ca
sh

im
m
ed
ia
te
ly

u
p
o
n
en
te
ri
n
g
la
b
.

P
ar
ti
ci
p
at
io
n
p
ay
m
en
t
an
d
p
ro
ce
d
u
re
s

o
v
er
v
ie
w

•
A
g
en
d
a
fo
r
ea
rn
in
g
s
an
d
ri
sk

m
an
ag
em

en
t
ta
sk
s.

S
te
p
2

•
T
w
en
ty

m
u
lt
ip
le
ch
o
ic
e
d
ri
v
in
g
q
u
iz
q
u
es
ti
o
n
s
d
is
tr
ib
u
te
d
o
n
p
ap
er
.

E
ar
n
in
g
s
ta
sk

•
M
o
n
et
ar
y
in
ce
n
ti
v
es

to
an
sw

er
co
rr
ec
tl
y
.

•
In
d
ic
at
e
w
h
et
h
er

su
re

o
f
an
sw

er
.

S
te
p
3

•
R
is
k
m
an
ag
em

en
t
ta
sk

in
st
ru
ct
io
n
s
d
is
tr
ib
u
te
d
an
d
re
ad

al
o
u
d
.

In
st
ru
ct
io
n
s

•
R
el
at
io
n
b
et
w
ee
n
ea
rn
in
g
s,
es
ti
m
at
io
n
,
an
d
ri
sk

m
an
ag
em

en
t
ta
sk
s
ex
p
la
in
ed

w
it
h
ex
am

p
le
s.

•
In
st
ru
ct
io
n
s
as
se
ss
m
en
t
an
d
re
v
ie
w
.

S
te
p
4

•
E
n
te
r
d
ri
v
in
g
q
u
iz
fi
n
al
an
sw

er
s
an
d
w
h
et
h
er

su
re

o
f
ea
ch
.

E
st
im

at
io
n
ta
sk

•
E
n
te
r
es
ti
m
at
e
o
f
o
w
n
sc
o
re
.

•
E
n
te
r
es
ti
m
at
e
o
f
av
er
ag
e
sc
o
re

ea
rn
ed

b
y
o
th
er

p
ar
ti
ci
p
an
ts
.

S
ec
o
n
d
st
ag
e

S
te
p
1

•
N
o
M
is
ta
k
es
,
O
w
n
M
is
ta
k
es
,
an
d
O
th
er
s’
M
is
ta
k
es

tr
ea
tm

en
ts
.

R
is
k
m
an
ag
em

en
t
ta
sk
s

•
P
re
ca
u
ti
o
n
,
in
su
ra
n
ce
,
an
d
in
it
ia
l
p
ro
b
ab
il
it
y
o
f
lo
ss

tr
ea
tm

en
t
m
an
ip
u
la
ti
o
n
s
p
re
se
n
te
d
in

ra
n
d
o
m

o
rd
er
.

S
te
p
2

•
C
o
m
p
le
te
al
l
d
ec
is
io
n
s,
th
en

re
v
ie
w
ea
ch
,
o
n
e
at
a
ti
m
e.

R
ev
ie
w
al
l
d
ec
is
io
n
s

•
M
u
st
ac
ti
v
el
y
co
n
fi
rm

o
r
re
v
is
e
ea
ch

in
d
iv
id
u
al
d
ec
is
io
n
.

S
te
p
3

•
R
an
d
o
m

d
ra
w
b
y
a
p
ar
ti
ci
p
an
t
d
et
er
m
in
es

tr
ea
tm

en
t
ap
p
li
ed

fo
r
se
ss
io
n
p
ay
m
en
t.

S
el
ec
ti
o
n
o
f
a
tr
ea
tm

en
t
fo
r
p
ay
o
ff

•
P
ar
ti
ci
p
an
ts
ar
e
ea
ch

p
ai
d
ac
co
rd
in
g
to

th
ei
r
d
ec
is
io
n
in

th
e
d
ra
w
n
tr
ea
tm

en
t.

6 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



relationships, we proceeded with the estimation task in which they recorded an estimate of

the number of questions they had answered correctly and an estimate of the average number

correct for the other participants in the session.

In the second stage, participants made risk management and insurance decisions in sev-

eral treatments. Participants finalized their choices in the program only after experiencing all

decisions for the experiment, and no losses or outcomes were realized until after confirma-

tion of all decisions. A random draw at the end of the experiment was used to select the

treatment used to determine payoff.

3.2. Treatment design

We use a within-subjects design with 60 participants each completing 8 treatments.12

This results in 480 participant-treatment observations, some of which are used in the primary

analyses, and some of which are used only to check participant rationality or as controls. To

minimize order effects, treatment manipulations are randomized across participants in each

session.

In each treatment, participants are exposed to a risk of losing $45 from their $60 earnings.

Treatment manipulations include the initial loss probability (10% or 32%), and the determi-

nants of overall loss probability. For each initial loss probability, three treatment manipula-

tions that differ in the way that quiz performance determines the overall probability of loss

as follows:

• No Mistakes: The risk of loss is implemented as a computer-generated random number—explained

with the analogy of a random draw from 100 white and orange ping-pong balls. The risk of loss is

expressed as a percentage (10% or 32% orange balls) and also described in terms of number of or-

ange and white balls, respectively. Participants are told that, if an orange ball is drawn, they lose

$45.
• Own Mistakes: As in the No Mistakes treatments, there is a draw from a known distribution of or-

ange and white ping-pong balls (10% or 32% orange). Participants are told that they lose $45 if an

orange ball is drawn but, if a white ball is drawn, there is a random draw from their own driving

quiz questions. If they answered the drawn question correctly, they do not lose any money, but if

they answered it incorrectly, they lose $45.
• Others’ Mistakes: As in the Own Mistakes treatments, there is a draw from a known distribution

of orange and white ping-pong balls (10% or 32% orange). Participants are told that they lose $45

if an orange ball is drawn but, if a white ball is drawn, there is a random draw from a different

participant’s driving quiz questions. If the other participant answered it incorrectly, they lose $45.

Before learning about whether they experienced a loss, participants made risk manage-

ment (insurance or precaution) decisions described as the option to pay a dollar cost from

their earnings to replace orange balls with white balls. In each case, they were presented

with a menu of incremental options as in the example in the Appendix. Consistent with

Jasperson and Aseervatham (2015), who emphasize the importance of a choice frame in lab-

oratory experiments of insurance decisions, the insurance and precaution decisions are

framed as choice tasks rather than elicitations of willingness to pay for insurance. The cost

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 7



of insurance is $14.50 and, if purchased in any treatment, it reduces their probability of loss

to zero. Each $1.50 spent on precaution reduces the probability of loss by one percentage

point in the low probability of loss treatments and by four percentage points in the high

probability of loss treatments.13 In the No Mistakes treatments, participants could reduce

their initial probability of loss to zero through buying the maximum level of precaution,

making this choice equivalent to full insurance. Buying full precaution is more expensive

than insuring in the low probability treatments, but less expensive in the high probability

treatments. Our within-participants design allows us to observe that participants who

wish to reduce risk to zero make rational choices between precaution and insurance. In

the Mistakes treatments where the probability of loss depends on quiz performance, in-

surance is the only option for reducing risk to zero, because even with the purchase of

full precaution (replacing all orange balls with white balls), the risk of loss from mistakes

remains.

In the No Mistakes treatments, participants know their risk of loss with certainty whereas,

in the Mistakes treatments, they must make subjective assessments over the risk of quiz mis-

takes to determine their probability of loss. Table 2 summarizes the way in which partici-

pants’ estimates of quiz scores impact estimates of the probability of loss prior to any

investments in risk mitigation.

Table 3 summarizes the optimal insurance decisions for a risk-neutral participant in each

of the treatments used in this study.14 The treatments that do not provide an insurance option

are used to categorize and control for participants’ risk preferences but are not otherwise

included in the primary analysis. The No Mistakes low probability treatment provides a

check of participant rationality but is not included in the primary analysis.

Several treatments allow for the possibility of underinsurance by risk-neutral (or risk-
averse) participants. In all of the High (32%) Initial Probability of Loss treatments, purchas-
ing insurance is optimal for risk-neutral (or risk-averse) participants in that it results in the
highest expected payoff. In the Low (10%) Initial Probability treatments with loss probabil-
ity independent of performance, the highest expected payoff results from not purchasing any
risk mitigation. However, in the treatments where loss probability depends on performance,
purchasing insurance is optimal for risk-neutral or risk-averse participants who performed
poorly on the quiz. Participants maximize expected payoffs by purchasing insurance if less

Table 2 Probability of losing $45 before risk management decisions in different treatments

Treatments Initial probability of loss

Low (10%) High (32%)

Own Mistakes: Dependent on own
performance

0:10 þ 0:90 � 20�OwnQuiz Scoreð Þ
20

0:32 þ 0:68 � 20�OwnQuiz Scoreð Þ
20

Others’ Mistakes: Dependent on
others’ performance

0:10 þ 0:90 � 20�Others’ Avg: Scoreð Þ
20

0:32 þ 0:68 � 20�Others’ Avg: Scoreð Þ
20

No Mistakes: Independent of
performance

0.10 0.32

8 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



than 76% of quiz questions are answered correctly. Therefore, overestimation of perform-
ance can lead to underinsurance.

3.3. The incentive for accurate estimation in this experiment design

In addition to the show-up fee paid at the beginning of Stage 1, participants received a

payment at the end of Stage 2 (in private and in cash) based upon their performance, risk

management decisions, and chance. Monetary incentives connect the earnings, estimation,

and risk management tasks across the two stages. A higher score on the driving quiz

decreases the probability of loss in the Own Mistakes treatments in the same way that loss

event probability estimation and insurance decisions correspond to expected wealth effects

in practice. That is, the estimation task facilitates comparison by the participants between

their expected payoff without insurance versus their payoff with insurance. More accurate

score estimates improve a participant’s ability to make an optimal insurance decision. In

summary, higher scores on the quiz reduce the risk of loss in the mistakes treatments, but

overestimation of scores could lead to suboptimal insurance decisions and lower expected

payoffs, and this correspondence rewards participants for accuracy in their estimation.15 For

example, a risk-neutral participant with an initial 10% probability of loss who estimates

scoring 90% correct on the driving quiz in the Own Mistakes treatment (expected loss of

$8.55) is better off not purchasing insurance for $14.50. However, if the participant’s actual

performance on the driving quiz is 60% (expected loss of $20.70), the expected payoff is

higher with insurance. Participants recorded their estimated scores only after they had

received all the instructions and passed the instructions assessment, demonstrating they

understood how both the probability of mistakes on the quiz and the cost of insurance

affected their expected payoffs. At that point, participants were aware of how their quiz

Table 3 Optimal risk management decisions to minimize expected loss in each treatment

Treatments Initial probability of loss

Low (10%) High (32%)

Own Mistakes: Risk depends on own performance Insure if quiz score < 76%
No risk mitigation if > 76%

Insure

Others’ Mistakes: Risk depends on others’ performance Insure if quiz score < 76%
No risk mitigation if > 76%

Insure

No Mistakes: Risk is independent
Risk mitigation-precaution and insurance No risk mitigation Full precautiona

Risk mitigation-precaution onlyb No risk mitigation Full precaution

a Participants pay to reduce the initial probability of loss before a mistake is drawn in increments of 10 percent-

age points. In the No Mistakes treatments, the purchase of full precaution reduces the probability of loss to zero

and is, therefore, the risk mitigation equivalent to buying insurance in this design. Full precaution is the more ef-

ficient means to reduce risk to zero in the high probability treatments, while insurance is the more efficient

means in low probability treatments.
b Precaution is the only risk management tool. These treatments are used only to categorize and control for

risk attitudes.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 9



scores and score estimation accuracy combined with their risk management and insurance

decisions to affect their earnings in the experiment.

As in actual insurance purchase decisions, a participant in the experiment who chooses

not to insure based on an optimistic or overconfident estimate of loss probability faces a

larger total cost of risk compared with insurance decisions based on more accurate assess-

ments. All responses were completely anonymous. There were no financial or risk manage-

ment incentives to report higher or lower scores than estimated, and optimal decisions

depended on using best estimates. Therefore, participants had incentives to accurately esti-

mate their risk of loss and faced no incentive to inaccurately report their estimates.16

Consequently, the earnings task is directly incentive-compatible, and the corresponding esti-

mation task is indirectly incentive-compatible, with respect to maximizing final payoffs in

the experiment.

Participants completed all decisions before receiving their earnings from the estimation

and risk management task. After all decisions were completed, reviewed, and confirmed, a

public random draw of a numbered ping-pong ball by a participant determined the treatment

used to pay out earnings. Each participant’s individual earnings for the chosen treatment

depended on a computer-generated random number representing either an orange ball or

white ball, and if applicable, random selection of quiz question.

3.4. Hypotheses

In this section, we draw on the existing literature to categorize participants as optimistic

or overconfident based on decisions made in the experiment and develop hypotheses about

the expected impact of optimism and overconfidence on participants’ insurance decisions.

Participants who overestimate their own quiz scores are classified as overconfident in
their own knowledge or abilities. We measure a participant’s overconfidence as the percent-
age by which they overestimate (or underestimate) their own score and we call this measure
the gross overconfidence bias.17 Gross overconfidence bias ranges from negative (under con-
fident) to positive (overconfident) so that results closer to zero reflect smaller biases.

We measure a participant’s general level of optimism (or pessimism) about an unknown
probability of loss outside of his or her own control, as the percentage by which they overesti-
mate (or underestimate) the average score of other participants. Our measure of optimism bias
also ranges from negative (pessimistic) to positive (optimistic), with measures closer to zero
reflecting smaller biases. Because an individual may be generally optimistic or pessimistic about
outcomes that beyond their control, the accuracy of the participant’s estimate of others’ average
score is used as a proxy for a general tendency to underestimate or overestimate the risk of loss
in the experiment, independent of their own knowledge or ability. Participants who overestimate
others’ scores are classified as optimistic because overestimation of scores corresponds to under-
estimating the probability of loss due to errors that are beyond their own control.

An optimistic outlook that causes participants to underestimate the chance of loss in general

may also influence participants’ estimation of their own scores. In other words, a participant’s

overestimation of their own score may be attributable to overconfidence in their own abilities, a

general optimistic outlook, or a combination of the two biases. Therefore, we also measure net

overconfidence bias as the difference between the errors in estimates of own and others’ scores.

10 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



Again, there is no compensation associated with relative performance, or reward for

above average performance in this experiment. The larger the positive bias in a partici-

pant’s own estimate compared with the bias in estimating others’ average score, the

higher that participant’s net overconfidence bias. Participants who overestimate their

own performance by the same or smaller percentage than they overestimate others’ do

not exhibit the net overconfidence bias.

Based on the theoretical literature on insurance demand, and the literature on optimism

and overconfidence, we develop the following hypotheses to be tested using the experiment

design described in the previous section:

Hypothesis 1: Individual biases

A. Gross Overconfidence Bias: Participants will exhibit gross overconfidence bias and

will, therefore, overestimate their own performance. This measure of bias includes

the possibly confounded effects of general optimism and overconfidence in their own

ability to minimize the risk of loss.

B. Optimism Bias: Participants will exhibit optimism bias and will, therefore, overesti-

mate others’ average performance.

C. Net Overconfidence Bias: Participants will exhibit net overconfidence bias and will

overestimate their own performance to a larger extent than they overestimate others’

performance. After adjusting for general optimism about overall performance, partic-

ipants’ estimates of their own performance will reflect a positive bias.

Hypothesis 2: Effect of biases on decision-making

By causing individuals to underestimate the risk of loss, optimism and overconfidence

biases will lead to underinsurance against losses. We define underinsurance as declining to

insure when the expected payoff is higher with insurance than without it.18

A. Gross overconfidence bias will increase the likelihood of underinsurance. Our mea-

sure of gross overconfidence bias directly reflects an underestimate of the probability

of loss in the Own Mistakes treatments. Therefore, the gross overconfidence bias will

be more likely to increase underinsurance against a loss depending on a participant’s

own performance, than a loss depending on others’ mistakes.

B. Optimism bias will increase the likelihood of underinsurance. Errors in assessing risk

due to general optimism do not depend on the source of loss. However, our measure of

optimism bias directly reflects an underestimate of the probability of loss in the Others’

Mistakes treatments. Therefore, the optimism bias will be more likely to increase under-

insurance against a loss resulting from others’ mistakes than a participant’s own

mistakes.

C. Net overconfidence bias should not affect the probability of underinsurance in this

design because the estimated probability of loss does not depend on relative perform-

ance in any way.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 11



Hypothesis 3: Effect of biases on the total cost of risk

In this experiment, participants can spend money on partial risk mitigation and insurance.
Participants who do not either insure or purchase precaution face expected losses that
depend on their quiz performance, while those who insure incur the known cost of the insur-
ance premium. Therefore, our measure of the total cost of risk is the sum of the expected
loss resulting from the risk event and the known amount spent on precaution or insurance. If
the overconfident or optimistic choose to pay for precaution rather than fully insuring, they
may actually increase their total cost of risk. We hypothesize that:

A. Higher gross overconfidence bias will be associated with greater total cost of risk.
The increase in the total cost of risk will be higher when the probability of a loss
depends on one’s own performance (because the overconfidence bias directly meas-
ures errors in assessing risk in the Own Mistakes treatment).

B. Higher optimism bias will be associated with greater total cost of risk. The increase

in the total cost of risk will be higher when the probability of loss depends on others’

performance (because the optimism bias directly measures errors in assessing risk in

the Others’ Mistakes treatment).

C. Net overconfidence, the difference between a participant’s gross overconfidence and

optimism, does not inform the estimation of risk in either Mistakes treatment.

Therefore, it will not affect participants’ total cost of risk.

The next section presents and discusses the outcomes of participant decisions. Then it

introduces controls used in the analysis and presents tests of these hypotheses.

4. Results and analysis

4.1. Summary statistics

Table 4 presents summary statistics for the participants’ performance on the earnings and

estimation tasks, and summarizes the definitions used for participants’ optimism, gross and net

overconfidence biases. The summary statistics show that participants overestimate their own

performance to a greater extent (9.26%) than they overestimate others’ performance (2.10%)

and that there is a great deal of within-sample variation in these bias measures.

Given our experiment parameters ($45 loss, $14.50 insurance premium, and precaution
alternatives), risk-neutral and risk-averse participants are predicted to purchase insurance (or
the full precaution equivalent) for all high initial probability treatments. If they do not pur-
chase insurance, their expected payoff decreases with lower quiz scores in the Mistakes
treatments. However, risk-seeking participants in the 32% initial probability No Mistakes
treatments may prefer not to purchase insurance in the Mistakes treatments as well. Under
the low initial probability of loss, the expected loss is $4.50 in the No Mistakes treatments,
and a risk-neutral participant should not purchase insurance. However, in the low initial
probability Mistakes treatments, any participant who answers 76% or fewer quiz questions

12 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



correctly has higher expected payoffs from purchasing insurance, making insurance the opti-
mal choice for risk averse and risk neutral participants. As discussed above, overconfidence
and/or optimism regarding quiz scores could lead participants to underestimate their risk of
loss, which could lead to underinsurance in light of their risk preferences.

While Table 4 shows the average levels of the biases, Table 5 categorizes participants in

terms of which biases they exhibit. The table illustrates that participants are relatively

unlikely to exhibit one of these biases and not the other. They are more likely to have gross

overconfidence and optimism bias or show neither bias. Results in Tables 4 and 5 reveal that

while 50% of participants exhibit some degree of optimism, the average level of optimism

does not differ significantly from zero. However, 60% of participants exhibit gross overcon-

fidence, and the average level of gross overconfidence is significantly above zero. It follows

that their average net overconfidence bias is also significant. In fact, 60% of participants ex-

hibit the net overconfidence bias. In summary, we find descriptive evidence in support of

parts A (gross overconfidence bias) and C (net overconfidence bias) of Hypothesis 1, but not

part B (optimism bias).

We summarize the insurance and precaution purchase decisions in Table 6. In the No

Mistakes treatments, most participants insure against loss, all purchase some form of risk

mitigation in the initial 32% probability of loss treatments, and most purchase some risk mit-

igation in the initial 10% probability of loss treatments. Participants insure against loss more

frequently under the Others’ Mistakes treatment than the Own Mistakes treatment. Some

participants purchase precaution to reduce their risk of loss. For example, in the No

Mistakes 32% initial probability treatment, those participants who do not insure, purchase

Table 4 Summary statistics: driving quiz, estimation task, and biases, N = 60 participants

Mean Minimum Maximum Standard deviation

Estimated own quiz score (out of 20)
Percent correct

16.4
82%

11
55%

19
95%

2.06

Estimated others’ quiz score (out of 20)
Percent correct

15.4
77%

10
50%

18
90%

1.94

Actual quiz score (out of 20)
Percent correct

15.1
75%

12
60%

18
90%

1.50

GOCbias ¼ Own estimate� own score
Own score 9.26%*** �18.75% 41.67% 15.80%

Optimism bias ¼ Estimated others’score� others’ score
Others’score 2.10% �33.86% 19.46% 12.89%

NOC bias = Overconfidence Bias – Optimism Bias 7.16%*** �23.83% 39.65% 14.72%

***Significantly different from zero at the 99% confidence level, according to the t test and nonparametric

signed rank test. GOC and NOC denote gross and net overconfidence, respectively.

Table 5 Classification of participants by optimism and gross overconfidence

Not overconfident Overconfident (gross overconfidence)

Not optimistic 33% 17%
Optimistic 7% 43%

Differences in proportions are significant at the 99% level in a x2 test.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 13



T
ab
le

6
P
er
ce
n
ta
g
es

ch
o
o
si
n
g
ea
ch

ri
sk

m
it
ig
at
io
n
al
te
rn
at
iv
e,
b
y
tr
ea
tm

en
t

M
is
ta
k
es

tr
ea
tm

en
t

N
o
M
is
ta
k
es

O
w
n
M
is
ta
k
es

O
th
er
s’
M
is
ta
k
es

In
it
ia
l
p
ro
b
ab
il
it
y
o
f
lo
ss

3
2
%

1
0
%

3
2
%

1
0
%

3
2
%

1
0
%

B
u
y
in
su
ra
n
ce

o
r
fu
ll
p
re
ca
u
ti
o
n
a

6
5
%

4
3
%

5
7
%

5
5
%

6
8
%

6
7
%

B
u
y
p
ar
ti
al
p
re
ca
u
ti
o
n

(a
v
er
ag
e
re
d
u
ct
io
n
in

in
it
ia
l
p
ro
b
ab
il
it
y
)b

3
5
%

(�
2
0
%
)

3
8
%

(�
5
%
)

4
3
%

(�
2
2
%
)

2
3
%

(�
5
%
)

3
2
%

(�
2
1
%
)

1
7
%

(�
3
%
)

N
o
ri
sk

m
it
ig
at
io
n

0
%

1
9
%

0
%

2
2
%

0
%

1
6
%

a
T
h
e
p
u
rc
h
as
e
o
f
fu
ll
p
re
ca
u
ti
o
n
re
d
u
ce
s
th
e
p
ro
b
ab
il
it
y
o
f
lo
ss

to
ze
ro

an
d
is
,
th
er
ef
o
re
,
th
e
eq
u
iv
al
en
t
to

b
u
y
in
g
in
su
ra
n
ce

in
th
is
d
es
ig
n
(a
lt
h
o
u
g
h
n
o
t

th
e
sa
m
e
co
st
).

b
P
ar
ti
ci
p
an
ts

p
ay

to
re
d
u
ce

th
e
in
it
ia
l
p
ro
b
ab
il
it
y
o
f
lo
ss

b
ef
o
re

a
m
is
ta
k
e
is

d
ra
w
n
in

in
cr
em

en
ts

o
f
1
0
p
er
ce
n
ta
g
e
p
o
in
ts
.
T
h
e
n
u
m
b
er
s
in

p
ar
en
th
es
es

sh
o
w

th
e
av
er
ag
e
p
er
ce
n
ta
g
e
p
o
in
t
re
d
u
ct
io
n
in

in
it
ia
l
p
ro
b
ab
il
it
y
.
F
o
r
ex
am

p
le
,
in

th
e
N
o
M
is
ta
k
es

3
2
%

in
it
ia
l
p
ro
b
ab
il
it
y
tr
ea
tm

en
t,
3
5
%

o
f
p
ar
ti
ci
p
an
ts

re
d
u
ce

th
ei
r
p
ro
b
ab
il
it
y
o
f
lo
ss

b
y
an

av
er
ag
e
o
f
2
0
%
.

14 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



sufficient precaution to reduce the initial risk from 32% to 12% on average. In the No

Mistakes 10% initial probability treatments, participants who buy precaution instead of in-

surance reduce the risk of loss from 10% to 5% on average.

4.2. Risk attitudes and gender controls

The primary purpose of this research is to analyze the relationship between overconfi-

dence, optimism, and insurance purchase. However, participants make insurance and precau-

tion decisions in light of their estimation of the risk of loss and their attitudes about

accepting different levels of risk. Previous literature suggests that overconfidence bias (that

affects estimation of risk) may vary systematically with gender. At the same time, optimal

insurance purchase decisions will differ based on risk attitudes. This subsection discusses

summary statistics particular to the risk attitude and gender control variables included in the

analysis.

Risk-neutral or risk-averse participants should purchase insurance whenever the expected

loss exceeds the insurance premium. All else equal, risk-seeking participants would be less

likely to purchase insurance. We use the Precaution Only No Mistakes treatments to provide

information about participants’ risk attitudes. The within-subject design allows us to observe

individual participants’ choices across each different treatment. In the Precaution Only No

Mistakes treatments, full precaution provides equivalent risk mitigation to insurance. Under

an initial probability of loss of 10%, expected payoff is decreasing in precaution, and full

precaution (the equivalent of insurance) costs roughly three times the expected loss. Under

an initial probability of loss of 32%, expected payoff is increasing in precaution, and the

cost of full precaution is only 83% of the expected loss. Therefore, we can identify partici-

pants whose behavior is consistent with risk-seeking preferences in the 32% probability of

loss treatments, and those who make choices consistent with risk aversion in the 10% proba-

bility of loss treatments.

We analyze each participant’s choices across No Mistakes Precaution Only treatments to

broadly classify risk attitudes. We classify participants as risk averse if they purchase any

precaution in the 10% initial probability treatment and also purchase full precaution in the

32% initial probability treatment. Participants who exhibit risk-averse behavior under the

10% initial probability treatment, but risk-seeking behavior under the 32% initial probability

treatment are considered to be “reflexive.”19

We classify participants as risk-seeking if they purchase less than full precaution in the

32% initial probability treatment and also do not purchase any precaution in the 10% initial

probability treatment. The risk-seeking classification also controls for an interpretation of

optimism in which some participants are optimistic about their “luck” in the outcome of a

random draw with a known distribution, even if they are realistic about the distribution itself.

Because risk-seekers could optimally choose to remain uninsured in cases where purchasing

insurance is optimal for risk-neutral or risk-averse individuals, we control for evidence of

risk-seeking behavior in our analysis of the relationship between optimism, overconfidence,

and insurance purchase. To avoid multicollinearity problems between this control variable

and others, we use the participants’ decisions in the Precaution Only No Mistakes treatments

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 15



(that are not included in the main regressions) only to estimate risk attitudes of the

participants.

Table 7 presents the classification of participant risk attitudes, optimism, and overconfi-

dence. Risk-seeking behavior is evident in 17% of the sample. As discussed above, partici-

pants who overestimate others’ average score are classified as optimistic and Table 4

showed no statistically significant optimism bias in the sample (though this is conservative

given the small sample size). However, Table 7 reveals that half of the participants do not

overestimate others’ scores. (The participants who overestimate others’ scores have an aver-

age error of 11.74%, while those who underestimate others’ scores have an average error of

only 7.5%.) We classify participants who overestimate their own score as exhibiting gross

overconfidence in their own knowledge or abilities; but some or all of that overconfidence

may be due to general optimism. Therefore, we also report the participants classified as

exhibiting net overconfidence bias. Comparing these metrics by gender, we find that a larger

percentage of men are overconfident, both before and after adjusting for general optimism.

Results in Tables 5, 6, and 7 confirm the presence of optimism and overconfidence. Table

8 summarizes the respective sample correlations of these measures, along with the categori-

cal variables Risk-Seeking and Male. Correlations with the categorical variable Male in

Table 8 show that there are no significant gender differences on average in gross overconfi-

dence or optimism biases. However, women’s average optimism is slightly higher and gross

overconfidence slightly lower than men’s. When these effects are combined, the difference-

in-differences between men and women is significant for net overconfidence bias. Men ex-

hibit a larger difference between gross overconfidence in their own quiz performance and

optimism about others’ quiz performance, compared with women. These results are consist-

ent with those reported in Kuhn and Villeval (2015), although the earnings tasks in the two

experiments are very different.

Table 7 Classification of risk attitudes, optimism, and overconfidence

Percent of participants Percent of females Percent of males

Risk seeking 17% 17% 17%
Optimism 50% 50% 50%
Gross Overconfidence 60% 54% 64%
Net overconfidence 67% 63% 69%

Table 8 Biases, risk attitude, and gender correlations

Gross overconfidence Optimism Net overconfidence Risk seeking Male

Gross overconfidence 1
Optimism 0.49*** 1
Net overconfidence 0.65*** �0.35*** 1
Risk seeking 0.08 0.02 0.06 1
Male 0.19 �0.14 0.33*** 0 1

***Significant at the 99% level.

16 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



4.3. The effect of overconfidence and optimism on the decision to insure

We now turn our attention to the second set of hypotheses, which considers the effects of

the biases on the decision to insure and on the total cost of risk. Because full precaution is

equivalent to insurance in the No Mistakes treatments, participants who purchase either of

those options in the No Mistakes treatments are counted as buying insurance.

Although some participants purchase partial precaution rather than insurance in the

Mistakes treatments, this decision results in their being underinsured when the probability of

loss is 32% or higher (unless they are risk-seeking). Participants who are overconfident in

their own quiz performance or precaution decisions, or optimistic with respect to others’

quiz performance, will underestimate this probability and therefore could make suboptimal

insurance decisions in the Mistakes treatments. Hypothesis 2 predicts that the likelihood of

underinsurance in the Others’ Mistakes treatment will increase with higher levels of opti-

mism bias. It also predicts that insurance decisions in the Own Mistakes treatments will

depend on the gross overconfidence bias, which as discussed above, may also include gen-

eral optimism. We expect both the optimism bias and the gross overconfidence bias to

increase the likelihood of underinsurance in the Own Mistakes treatment.

Fig. 1 presents the incidence of underinsurance (from a risk-neutral perspective) for par-

ticipants in each treatment, according to whether they exhibit optimism (Panel A), gross

overconfidence (Panel B), or net overconfidence (Panel C). The optimism bias is associated

with underinsurance in the Others’ Mistakes low probability treatment, but also in the No

Mistakes and Own Mistakes high probability treatments. We find that, as expected, the gross

overconfidence bias is associated with underinsurance when payoffs depend on one’s own

performance, but not others’ performance, and the net overconfidence bias has an insignifi-

cant impact on underinsurance in this experiment.

We further analyze the effect of the biases on underinsurance through logit regressions

presented in Table 9, in which the dependent variable is a dummy variable where

Underinsurance =1 indicates underinsurance from a risk-neutral perspective. Categorical in-

dependent variables include gender (female is the omitted category), risk treatment type

(10% initial probability is the omitted category), and risk-seeking attitude (no evidence of

risk seeking is the omitted category).20 Regression coefficients are presented in terms of log

odds. Due to the strong relationship between the biases, we run separate regression models,

including different bias measures as independent variables in each.

The top panel of Table 9 presents analysis of the Mistakes treatments (N = 240 decisions,

with 60 standard error clusters). Conceptually, we would expect optimism and gross over-

confidence (that encompasses optimism) to influence insurance decisions in both Mistakes

treatments. At the same time, our measure of optimism directly informs subjective probabil-

ity of loss in the Others’ Mistakes treatment and our measure of gross overconfidence

directly informs subjective probability of loss in the Own Mistakes treatment. Therefore, we

are especially interested in the significance of interaction terms for the biases and Mistakes

treatment. In particular, we would expect the gross overconfidence measure to have a more

significant impact on underinsurance in the Own Mistakes treatments and general optimism

to be more influential in the Others’ Mistakes treatments.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 17



Fig. 1. Panel A: Percent underinsuring by treatment and optimism. Panel B: Percent underinsuring by treatment

and gross overconfidence (GOC). Panel C: Percent underinsuring by treatment and net overconfidence (NOC)

Notes: Differences in proportions are significant at the 90% level (*) based on a x2 test. P(loss) represents the

initial probability 32% and 10% treatments; No Mistakes, Own Mistakes, and Others’ Mistakes denote No

Mistakes, Own Mistakes and Others’ Mistakes treatments, respectively.

18 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



T
ab
le

9
D
et
er
m
in
an
ts
o
f
u
n
d
er
in
su
ra
n
ce
,
lo
g
it
re
g
re
ss
io
n
co
ef
fi
ci
en
t
es
ti
m
at
es

an
d
(s
ta
n
d
ar
d
er
ro
rs
)

M
is
ta
k
es

tr
ea
tm

en
ts
,
N
=
2
4
0
,
st
an
d
ar
d
er
ro
rs
cl
u
st
er
ed

b
y
p
ar
ti
ci
p
an
t

M
o
d
el
1
:
G
ro
ss

o
v
er
co
n
fi
d
en
ce

M
o
d
el
2
:
O
p
ti
m
is
m

M
o
d
el
3
:
N
et
o
v
er
co
n
fi
d
en
ce

M
ai
n
ef
fe
ct
s

In
te
ra
ct
io
n
ef
fe
ct
s

M
ai
n
ef
fe
ct
s

In
te
ra
ct
io
n
ef
fe
ct
s

M
ai
n
ef
fe
ct
s

In
te
ra
ct
io
n
ef
fe
ct
s

In
te
rc
ep
t

�1
.5
0
5
0
*
*
*
(0
.3
9
7
2
)

�1
.6
0
8
9
*
*
*
(0
.4
1
0
4
)

�1
.7
8
9
4
*
*
*
(0
.4
0
9
5
)

�1
.8
0
6
5
*
*
*
(0
.4
1
2
1
)

�1
.4
9
8
6
*
*
*
(0
.4
2
4
8
)

�1
.5
7
0
6
*
*
*
(0
.4
3
5
7
)

O
th
er
s’
M
is
ta
k
es

�0
.7
1
2
2
*
*
*
(0
.2
3
9
3
)

�0
.4
9
6
5
*
(0
.2
5
0
7
)

�0
.7
6
0
7
*
*
*
(0
.2
5
3
2
)

�0
.7
2
0
2
*
*
*
(0
.2
3
6
9
)

�0
.7
3
3
2
*
*
*
(0
.2
4
1
9
)

�0
.6
0
5
5
*
*
(0
.2
6
0
0
)

G
O
C
b
ia
s

0
.7
3
8
8
(1
.5
1
4
0
)

1
.7
0
1
3
(1
.6
0
5
4
)

G
O
C
�

O
th
er
s’
M
is
ta
k
es

�2
.1
8
0
4
*
(1
.2
4
8
1
)

O
p
ti
m
is
m

b
ia
s

5
.0
2
8
1
*
*
*
(1
.5
4
2
7
)

5
.4
4
0
7
*
*
*
(1
.6
5
0
8
)

O
p
ti
m
is
m
�

O
th
er
s’
M
is
ta
k
es

�0
.9
3
3
2
(1
.7
8
6
0
)

N
O
C
b
ia
s

�3
.1
2
8
8
*
(1
.7
9
8
8
)

�2
.1
2
9
8
(1
.7
5
3
2
)

N
O
C
�

O
th
er
s’
M
is
ta
k
es

�2
.4
3
0
6
(1
.5
3
1
1
)

M
al
e

0
.5
6
0
0
(0
.4
6
4
4
)

0
.5
6
4
7
(0
.4
6
7
2
)

0
.8
2
7
1
*
(0
.4
4
3
9
)

0
.8
2
8
7
*
(0
.4
4
6
0
)

0
.9
2
8
7
*
(0
.4
9
1
5
)

0
.9
3
6
1
*
(0
.4
9
2
4
)

In
it
ia
l
p
ro
b
ab
il
it
y
=
3
2
%

0
.6
2
4
3
*
(0
.2
7
1
6
)

0
.6
2
8
5
*
*
(0
.2
7
4
3
)

0
.6
6
6
9
*
*
(0
.2
9
0
8
)

0
.6
6
7
7
*
*
(0
.2
9
1
9
)

0
.6
4
2
8
*
*
(0
.4
9
1
5
)

0
.6
4
4
9
*
*
(0
.2
8
4
9
)

R
is
k
-s
ee
k
in
g

1
.6
7
4
6
*
*
*
(0
.3
8
7
6
)

1
.6
8
9
0
*
*
*
(0
.3
9
7
7
)

1
.8
0
3
9
*
*
*
(0
.4
4
3
9
)

1
.8
0
5
7
*
*
*
(0
.3
6
8
2
)

1
.8
3
0
5
*
*
*
(1
.7
9
8
8
)

1
.8
5
9
9
*
*
*
(0
.3
9
5
9
)

N
o
M
is
ta
k
es

p
(l
o
ss
)
=
3
2
%

tr
ea
tm

en
t,
N
=
6
0

M
o
d
el
1
:
G
ro
ss

o
v
er
co
n
fi
d
en
ce

M
o
d
el
2
:
O
p
ti
m
is
m

M
o
d
el
3
:
N
et
o
v
er
co
n
fi
d
en
ce

In
te
rc
ep
t

�1
.1
6
1
5
(0
.5
2
0
5
)

�1
.2
1
9
7
*
*
(0
.4
9
5
3
)

�1
.0
9
1
2
*
*
(0
.4
8
1
8
)

G
O
C
b
ia
s

1
.0
1
5
2
(1
.8
6
1
8
)

O
p
ti
m
is
m

b
ia
s

2
.4
2
9
8
(2
.3
5
7
4
)

N
O
C
b
ia
s

�0
.7
2
5
2
(0
.6
0
3
7
)

M
al
e

0
.7
0
4
6
(0
.5
9
4
6
)

0
.8
5
6
3
(0
.6
0
1
5
)

0
.8
3
3
8
(0
.6
0
3
7
)

N
o
te
s:
*
,
*
*
,
*
*
*
d
es
ig
n
at
e
si
g
n
ifi
ca
n
ce

at
th
e
9
0
%
,
9
5
%
,
an
d
9
9
%

co
n
fi
d
en
ce

in
te
rv
al
s,
re
sp
ec
ti
v
el
y
.
T
h
e
d
ep
en
d
en
t
v
ar
ia
b
le

is
ze
ro

if
th
e
p
ar
ti
ci
p
an
t
d
id

n
o
t
u
n
d
er
in
su
re
,
an
d
1
if
th
ey

d
id
.
C
o
ef
fi
ci
en
ts
ar
e
p
re
se
n
te
d
in

lo
g
o
d
d
s
te
rm

s.
G
O
C
an
d
N
O
C
d
en
o
te
g
ro
ss

an
d
n
et
o
v
er
co
n
fi
d
en
ce
,
re
sp
ec
ti
v
el
y
.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 19



Controlling for the biases, initial probability of loss, and risk-seeking preferences, we find

that participants are significantly less likely to underinsure in the Others’ Mistakes treat-

ments than in the Own Mistakes treatments. Model 1 shows that gross overconfidence itself

is not a significant predictor of underinsurance. However, the interaction effect in the second

column reveals that the difference in the probability of underinsuring between the Own and

Others’ Mistakes treatment does depend on gross overconfidence. As gross overconfidence

increases, there is a greater increase in the likelihood of underinsuring in the Own Mistakes

compared with the Others’ Mistakes treatments. In other words, participants with higher

gross overconfidence are even more likely (compared with those who are less overconfident)

to make suboptimal insurance decisions when the risk of loss depends on their own perform-

ance than when it depends on others’ performance. This is somewhat consistent with the

Hypothesis 2 predictions about biases.

In Model 2, we find that participants who exhibit higher levels of optimism bias are sig-

nificantly more likely to underinsure. However, the interaction between Optimism and the

Others’ Mistakes treatment is not significant, which suggests that the effect of optimism is

no more influential in the Others’ Mistakes treatment than in the Own Mistakes treatment.

Finally, In Model 3, we find that net overconfidence bias has a negative relationship with

underinsurance that is weakly significant and independent of the Mistakes treatment. We at-

tribute this result to the fact that net overconfidence is decreasing in optimism, which as dis-

cussed above, has a strong positive and significant relationship with the probability of

underinsuring across both Mistakes treatments.

The second panel of Table 9 displays results for the No Mistakes condition. In this treat-

ment, participants can only underinsure in the 32% initial probability of loss treatment,

because purchasing no risk mitigation is the payoff-maximizing choice in the 10% initial

probability treatment. Therefore, we examine underinsurance decisions separately for the

No Mistakes, 32% probability of loss treatment (N = 60), and find that, as expected, none of

the biases have a significant effect.21 This suggests that participants’ bias stems from their

beliefs about the probability of loss rather than optimism or pessimism about luck when con-

fronting a known distribution. However, the insignificant effect on the insurance decision

may also be attributable to the smaller sample size using only one treatment for this statisti-

cal test.

4.4. The effect of overconfidence and optimism on the total cost of risk

As described in Hypothesis 3, the relationship between the total cost of risk (the sum of

the expected loss resulting from the risk event and the known amount spent on precaution or

insurance) and the biases provides a way to examine the cost of underinsurance, in particular

when participants have access to alternative risk mitigation measures instead of insuring.

For example, given the experiment parameters in the 10% initial loss probability Own

Mistakes treatments, a participant who answers 70% of the quiz questions correctly and

insures faces a total cost of $14.50 (out of their $60 earnings). If they underinsure by neither

purchasing insurance nor precaution, they face an expected loss of $16.70 in the initial 10%

probability treatment. However, if that participant pays for precaution to reduce the initial

20 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



probability of loss to 5%, then the total cost of risk, including the cost of precaution, is

actually even higher at $22.60. Participants who are overconfident about the effects of pre-

caution increase their total cost of risk when they choose to purchase partial precaution

instead of either insuring or doing nothing.

Table 10 presents results of a generalized least squares regression estimating the impact

of the biases and control variables on the total cost of risk. As in the previous regressions,

we separately analyze the effects of the three types of biases. The results for Model 1 show

that gross overconfidence bias is a statistically and economically significant factor increasing

the cost of loss in the Own Mistakes treatment. Furthermore, the effect is significantly lower

for participants in the Others’ Mistakes treatments as compared with the Own Mistakes

treatments. Overconfidence in one’s own ability is more costly when the expected loss

depends on own performance.

In Model 2, we find that optimism bias also increases the total cost of risk, but there is no

significant difference in this relationship across mistakes treatments. This too is expected,

since errors in assessing risk due to general optimism do not depend on the source of loss.

Finally, Model 3 confirms that net overconfidence bias does not significantly increase the
total cost of risk. However, as the combined results for gross overconfidence and optimism

imply, the impact of net overconfidence on the total cost of risk is lower in the Others’

Mistakes treatment than in the Own Mistakes treatment. As in the previous table, the smaller

sample size when using only the data from the No Mistakes treatment (N = 60) reduces the

power of the test but, in this case, we still find the effect of net overconfidence to be margin-

ally significant.

5. Conclusions

This article contributes to the literature by providing experimental evidence regarding

the effects of overconfidence and optimism on insurance decisions. In the existing litera-

ture, optimism is sometimes confounded with overconfidence, and we contribute an inno-

vative design that distinguishes between overconfidence regarding the likelihood of a

favorable outcome resulting from one’s own performance versus optimism regarding the

likelihood of a favorable outcome outside of one’s own influence. We find that these psy-

chological biases have important implications for insurance in that they can cause indi-

viduals to underestimate their risk and, therefore, underinsure, resulting in a higher total

cost of risk. This effect is stronger for insurance over risks that depend on one’s own per-

formance as compared with exogenous risks. Our results contribute to the growing body

of literature on the effect of psychological biases on financial decisions and reinforce the

importance of careful measurement of overconfidence bias in laboratory experiments.

This distinction is particularly important for understanding insurance decisions related to

risks outside of the purchaser’s control. The relationship between overconfidence and op-

timism may also help explain the perceived tradeoffs between risk mitigation and insur-

ance decisions.

Our results show that, after controlling for risk-seeking behavior, participants are more

likely to make suboptimal insurance decisions when the risk of loss depends on their own

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 21



T
ab
le

1
0

D
et
er
m
in
an
ts
o
f
to
ta
l
co
st
o
f
ri
sk
,
g
en
er
al
iz
ed

le
as
t
sq
u
ar
es

re
g
re
ss
io
n
co
ef
fi
ci
en
t
es
ti
m
at
es

an
d
(s
ta
n
d
ar
d
er
ro
rs
)

D
et
er
m
in
an
ts
o
f
u
n
d
er
in
su
ra
n
ce

in
M
is
ta
k
es

tr
ea
tm

en
ts
,
N
=
2
4
0
,
st
an
d
ar
d
er
ro
rs
cl
u
st
er
ed

b
y
p
ar
ti
ci
p
an
t

M
o
d
el
1
:
G
ro
ss

o
v
er
co
n
fi
d
en
ce

M
o
d
el
2
:
O
p
ti
m
is
m

M
o
d
el
3
:
N
et
o
v
er
co
n
fi
d
en
ce

M
ai
n
ef
fe
ct
s

In
te
ra
ct
io
n
ef
fe
ct
s

M
ai
n
ef
fe
ct
s

In
te
ra
ct
io
n
ef
fe
ct
s

M
ai
n
ef
fe
ct
s

In
te
ra
ct
io
n
ef
fe
ct
s

In
te
rc
ep
t

1
4
.6
3
1
1
(0
.3
9
8
8
)

1
4
.8
3
7
3
*
*
*
(0
.4
0
3
4
)

1
4
.5
5
8
3
*
*
*
(0
.3
9
3
1
)

1
4
.5
2
9
7
*
*
*
(0
.3
9
2
2
)

1
4
.8
0
5
6
*
*
*
(0
.4
3
7
3
)

1
4
.6
5
0
4
*
*
*
(0
.4
3
4
0
)

O
th
er
s’
M
is
ta
k
es

�0
.4
8
4
7
(0
.3
3
0
2
)

0
.0
3
2
0
(0
.2
9
0
0
)

�0
.4
8
4
7
(0
.3
3
0
2
)

�0
.4
2
7
5
(0
.3
1
0
2
)

�0
.4
8
4
7
(0
.3
3
0
2
)

�0
.1
7
4
3
(2
.6
1
5
4
)

G
O
C
b
ia
s

3
.2
6
3
7
(2
.0
0
6
4
)

6
.0
5
4
1
*
*
*
(2
.7
7
0
3
)

G
O
C
�

O
th
er
s’
M
is
ta
k
es

�5
.5
8
0
8
*
*
*
(0
.4
3
5
2
)

O
p
ti
m
is
m

b
ia
s

5
.6
8
8
5
*
*
*
(1
.6
2
5
4
)

7
.0
4
9
4
*
*
*
(1
.9
7
4
4
)

O
p
ti
m
is
m
�

O
th
er
s’
M
is
ta
k
es

�2
.7
2
1
8
(1
.9
3
7
3
)

N
O
C
b
ia
s

�0
.4
8
4
7
(0
.3
3
0
2
)

1
.4
1
7
9
(0
.5
8
9
8
)

N
O
C
�

O
th
er
s’
M
is
ta
k
es

�4
.3
3
6
9
*
(2
.3
5
5
5
)

M
al
e

0
.4
0
9
9
(0
.5
3
6
9
)

0
.4
0
9
9
(0
.5
3
8
1
)

0
.8
1
7
0
(0
.5
1
5
9
)

0
.8
1
7
0
(0
.5
1
7
0
)

0
.6
7
8
0
(0
.5
4
0
1
)

0
.6
7
8
0
(0
.5
4
1
3
)

In
it
ia
l
p
ro
b
ab
il
it
y
=
3
2
%

2
.1
9
5
9
*
*
*

(0
.4
3
0
6
)

2
.1
9
5
9
*
*
*
(0
.4
3
1
5
)

1
.8
2
7
2
*
*
*
(0
.4
9
3
7
)

2
.1
9
5
9
*
*
*
(0
.4
3
1
5
)

2
.1
9
5
9
(0
.4
3
0
6
)

2
.1
9
5
9
*
*
*
(0
.4
3
1
5
)

R
is
k
-s
ee
k
in
g

1
.7
5
9
6
*
*
*

(0
.5
1
8
6
)

1
.7
5
9
6
*
*
*
(0
.5
3
8
1
)

1
.8
2
7
2
*
*
*
(0
.4
9
4
7
)

1
.8
8
3
1
*
*
*
(0
.5
4
0
8
)

1
.8
8
3
1
*
*
*
(0
.5
4
1
9
)

D
et
er
m
in
an
ts
o
f
to
ta
l
co
st
o
f
ri
sk

in
N
o
M
is
ta
k
es

p
(l
o
ss
)
=
3
2
%

tr
ea
tm

en
t,
N
=
6
0

M
o
d
el
1
:
G
ro
ss

o
v
er
co
n
fi
d
en
ce

M
o
d
el
2
:
O
p
ti
m
is
m

M
o
d
el
3
:
N
et
o
v
er
co
n
fi
d
en
ce

In
te
rc
ep
t

1
2
.5
6
5
0
*
*
*

(0
.2
3
6
7
)

1
2
.6
4
4
6
*
*
*
(0
.2
2
2
2
)

1
2
.5
8
*
*
*
(0
.2
1
9
4
)

G
O
C
b
ia
s

0
.6
9
1
0
(0
.8
1
3
3
)

O
p
ti
m
is
m

b
ia
s

�1
.0
4
3
9
(1
.0
3
9
6
)

N
O
C
b
ia
s

1
.7
3
8
5
*
(0
.9
7
8
0
)

M
al
e

0
.0
2
1
2
(0
.9
2
9
9
)

0
.0
2
3
6
(0
.2
5
6
8
)

�0
.1
0
6
2
(0
.2
5
8
8
)

N
o
te
s:
*
,
*
*
,
*
*
*
d
es
ig
n
at
e
si
g
n
ifi
ca
n
ce

at
th
e
9
0
%
,
9
5
%
,
an
d
9
9
%

co
n
fi
d
en
ce

in
te
rv
al
s,
re
sp
ec
ti
v
el
y
.
G
O
C
an
d
N
O
C
d
en
o
te

g
ro
ss

an
d
n
et

o
v
er
co
n
fi
d
en
ce
,

re
sp
ec
ti
v
el
y
.

22 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



performance than others’ performance. Optimistic participants are more likely to underin-

sure than non-optimistic participants. Overconfidence does not have a significant effect on

the decision to purchase insurance, although participants with higher overall overconfidence

show larger differences in behavior when they are responsible for the risk of loss than when

it is beyond their control.

Analysis of the total cost of risk, including both the expected loss and the cost of precau-

tion or insurance, under different treatments provides similar evidence about the influence of

these psychological biases. Overconfidence and optimism both significantly increase the

total cost of risk. However, overconfidence has a significantly lower effect on total cost

when the loss event is triggered by someone else’s error. Optimism bias increases the cost

by about the same amount regardless of whether the risk depends on one’s own actions.

These results suggest that general optimism extends to outcomes that depend on one’s own

ability, but overconfidence in one’s own performance does not affect the decision to mitigate

risks due to factors outside of one’s own control, such as those resulting from nature or from

others’ errors.

The laboratory evidence reported in this article offers a potential explanation for the under-

insurance against catastrophe that has been observed in the market. Beliefs, together with risk

tolerance, preferences, or general probability misperceptions may provide alternative explana-

tions for some of the observed insurance decision puzzles. For example, Jacobsen et al. (2015),

who study asset allocation decisions under uncertainty, find that optimism about outcomes and

optimism about the level of risk are as important as risk aversion in explaining asset allocation.

Spinnewijn (2013) notes that, while heterogeneity in beliefs informs insurance policy design, it

is very difficult to obtain direct evidence about beliefs. Laboratory results such as ours provide

a first step in connecting beliefs to insurance decisions, with more control than surveys or be-

havioral proxies for perceptions. While convenience samples of students provide for a strong

degree of laboratory control, it would interesting to further explore these issues with a more

diverse participant population. Future research should also consider in greater detail the influ-

ence of these biases on individual perceptions about the effectiveness of different risk mitiga-

tion alternatives.

Notes

1 Theoretical explanations in economics and evolutionary biology have also illustrated

conditions under which optimism or overconfidence bias can be individually wel-

fare-improving (compared to rational expectations), a second best solution in the

presence of other biases, and even a necessary adaptation for species survival. See,

for instance Brunnermeier and Parker (2005), Besharov (2004), and Johnson and

Fowler (2011).

2 This definition of overconfidence assumes misperceptions rather than a well-cali-

brated assessment of one’s own relative abilities. Well-calibrated confidence

does not produce lower results. For example, Fielder (2011) finds that virtual

traders who self-report as better than average, in fact earn above average virtual

profits.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 23



3 The adverse selection literature in insurance originated with the seminal work of

Rothschild and Stiglitz (1976). See Dionne, Fombaron, and Doherty (2013) for a

more complete summary of this extensive literature.

4 In Bajtelsmit, Coats, and Thistle (2015), the authors focus on the other considera-

tions addressed by the experiment design, including the effect of ambiguity on risk

management decisions and the tradeoff between taking precaution and purchasing

insurance, as well as replicability of other researchers’ results.

5 We acknowledge that there is a great deal of inconsistency in terminology and trait

measurement across the overconfidence literature, both theoretical and empirical.

See Clark and Friesen (2009) and Spinnewijn (2013) for more complete reviews of

this literature.

6 De Meza and Webb (2001) show that insurers can design contracts that will result in

an equilibrium which they term “advantageous selection” in which the risk-averse

agent buys insurance and also invests in some precaution. The risk-neutral agent

does not take precaution or buy insurance.

7 However, recent research suggests that calibration-based overconfidence observed

in confidence interval reporting may be overstated because of the measurement

instrument (Blavatskky, 2009; Cesarini, Sandewall, & Johannesson, 2006; Glaser,

Langer, & Weber, 2013; Soll & Klayman, 2004).

8 See Chiappori and Salanie (2013) for a review of this literature.

9 Experiments by Cesarini et al (2006), Blavatskyy (2009), and Clark and Friesen

(2009) suggest that frequency estimation tasks provide better measures of over-

confidence relative to confidence interval estimation tasks because of the

improvement in incentive-compatibility, better alignment of accuracy and infor-

mation, and because framing a forecast as a frequency is a much more natural

cognitive task.

10 The earnings, probability estimation, and risk management tasks were explained in a

Power Point presentation at the front of the room, with the instructions read aloud.

Participants took an instructions assessment to confirm they understood how their

earnings would be determined and were able to ask questions.

11 Participants were required to have a valid state driver’s license as a condition of par-

ticipation in the experiment. The driving quiz was designed to include a sufficient

number of easy questions such that all participants were expected to be able to

achieve the minimum score, but also some more difficult questions to minimize the

number who could achieve a perfect 20 out of 20 correct.

12 The within-participants design provides for much greater statistical power than a

between-participants design of the same size (see Bellamare, Bissonnette, and

Kroger, 2014 and Charness, Gneezy, and Kuhn, 2012). We obtain more partici-

pant observations and cluster standard errors at the participant level in our

analysis.

13 The return to taking precaution is higher for the high-risk treatments than the

low-risk treatments due to the assumption of greater productivity of precaution

24 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



under high initial risk. For a theoretical justification, see Bajtelsmit and Thistle

(2015).

14 The full design also included four additional treatments, in which the probability of

loss depended on own and others’ performance under high and low initial probabil-

ities of loss. However, in those treatments, participants could neither reduce loss

probability to zero, nor insure against loss and, therefore, we do not include or ana-

lyze them in this article. See Bajtelsmit et al. (2015).

15 Although there are other methods of incentivizing participants, such as scoring

rules that reward estimates but penalize errors, participant risk preferences have

been shown to affect their choices. See, for example, Andersen, Fountain,

Harrison, and Rutström (2014) and Harrison, Martinex-Correa and Swarthout,

(2014). Other probability elicitation mechanisms depend on independence

between agents’ actions and the probability of the risky event (Armentier and

Treich, 2013) and on the agent having no stake in the risky event (Karni, 2009).

16 Estimating one score to use in risk management decision making, but recording a

different score, while technically possible, would be inconsistent with payoff maxi-

mization efforts because it would increase the cognitive difficulty of the risk man-

agement task and potentially increase the chance of making a costly risk

management error and, therefore, would not be incentive compatible with maximiz-

ing experiment payoffs.

17 We thank an anonymous referee for suggesting the “gross overconfidence” and “net

overconfidence” labels.

18 Because this definition only applies to risk-averse and risk-neutral participants, we

control for risk-seeking behavior in our analysis.

19 This classification is discussed in detail in an earlier article by the authors where the

No Mistakes treatments is compared to the alternative to buy insurance against the

No Mistakes Precaution Only treatments and show that (1) participants purchase the

more efficient means of risk mitigation, and (2) that participants are consistent in

their risk mitigation decisions across treatments. Comparison of the Precaution Only

Mistakes treatments to the Precaution Only No Mistakes treatments shows that par-

ticipants respond predictably to the lower effectiveness of precaution by purchasing

less precaution in the Mistakes treatments.

20 Identified through behavior in the Precaution Only No Mistakes treatments.

21 The risk-seeking dummy variable is not included because, in the treatments used to

determine risk attitudes (Precaution Only No Mistakes), a risk-seeking identification

is indistinguishable from underinsurance in the high probability of loss insurance

treatments.

Appendix: Examples of choices available in high and low initial probability treatments

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 25



References

Andersen, S., Fountain, J., Harrison, G. W., & Rutström, E. E. (2014). Estimating subjective probabilities.

Journal of Risk and Uncertainty, 48, 207–229.

Arad, A. (2014). Avoiding greedy behavior in situations of uncertainty: The role of magical thinking. Journal of

Behavioral and Experimental Economics, 53, 17–23.

Armentier, O., & Treich, N. (2013). Eliciting beliefs: Proper scoring rules, incentives, stakes, and hedging.

European Economic Review, 62, 17–40.

Bajtelsmit, V., Coats, J., & Thistle, P. (2015). The effect of risk ambiguity on risk management choices: An ex-

perimental study. Journal of Risk and Uncertainty, 50, 249–280.

Bajtelsmit, V., & Thistle, P. (2015). Liability, insurance and the incentive to obtain information about risk. The

Geneva Risk and Insurance Review, 40, 171–193.

Initial high probability of loss example: Choose one of the following options below.

Decision
Up-front cost to replace
orange balls

New number of
orange balls

New number of
white balls

Probability orange
ball is drawn

A $0 32 68 32%
B $1.50 28 72 28%
C $3.00 24 76 24%
D $4.50 20 80 20%
E $6.00 16 84 16%
F $7.50 12 88 12%
G $9.00 8 92 8%
H $10.50 4 96 4%
I $12.00 0 100 0
J (Insurance) $14.50 32 68 N/A

Your decision in Scenario 2: ____________________

Initial low probability of loss example: Choose one of the following options below.

Decision
Up-front cost to replace
orange balls

New number of
orange balls

New number of
white balls

Probability an orange
ball is drawn

A $0 10 90 10%
B $1.50 9 91 9%
C $3.00 8 92 8%
D $4.50 7 93 7%
E $6.00 6 94 6%
F $7.50 5 95 5%
G $9.00 4 96 5%
H $10.50 3 97 3%
I $12.00 2 98 2%
J $13.50 1 99 1%
K $15.00 0 100 0
L (Insurance) $14.50 10 90 N/A

Your decision in Scenario 8: ___________________

26 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28



Barber, B. M., & Odean, T. (2001). Boys will be boys: Gender, overconfidence, and common stock investment.

The Quarterly Journal of Economics, 116, 261–292.

Bellemare, C., Bissonnette, L., & Kroger, S. (2014). Statistical Power of Within and Between-Subjects Designs

in Economic Experiments. Discussion Paper No. 8583. Germany: Institute for the Study of Labor.

Besharov, G. (2004). Second-best considerations in correcting cognitive biases. Southern Economic Journal, 71,

12–20.

Blavatskyy, P. R. (2009). Betting on own knowledge: Experimental test of overconfidence. Journal of Risk and

Uncertainty, 38, 39–49.

Brunnermeier, M., & Parker, J. (2005). Optimal expectations. American Economic Review, 95, 1092–1118.

Camerer, C. F., & Lovallo, D. (1999). Overconfidence and excess entry: an experimental approach. American

Economic Review, 89, 306–318.

Cesarini, D., Sandewall, O., & Johannesson, M. (2006). Confidence interval estimation tasks and the economics

of overconfidence. Journal of Economic Behavior & Organization, 61, 453–470 .

Charness, G., Gneezy, U., & Kuhn, M. (2012). Experimental Methods: Between-subject and Within-subject

Design. Journal of Economic Behavior and Organization, 91, 1–8.

Chiappori, P., & Salanie, B. (2013).Asymmetric information in insurance markets: Predictions and Tests. In G.

Dionne (Ed.), Handbook of Insurance (pp. 397–422). New York, NY: Springer.

Clark, J., & Friesen, L. (2009). Overconfidence in forecasts of own performance: An experimental study. The

Economic Journal, 119, 229–251.

De Bondt, F., & Thaler, R. (1985). Does the stock market overreact? The Journal of Finance, 40, 793–805.

De Meza, D., & Webb, D. C. (2001). Advantageous selection in insurance markets. The Rand Journal of

Economics, 32, 249–262.

Dionne, G., Fombaron, N., & Doherty, N. (2013).Adverse selection in insurance contracting. In G. Dionne (Ed.),

Handbook of Insurance (pp. 231–280). New York, NY: Springer

Fellner, G., & Krugel, S. (2012). Judgmental overconfidence: Three measures, one bias? Journal of Economic

Psychology, 33, 142–154.

Fiedler, M. (2011). Experience and confidence in an internet based asset market experiment. Southern Economic

Journal, 78, 30–52.

Fischbacher, U. (2007). z-Tree: Zurich toolbox for readymade economic experiments. Experimental Economics,

10, 171–178.

Foster, G., & Frijters, P. (2014). The formation of expectations: competing theories and new evidence. Journal of

Behavioral and Experimental Economics, 53, 66–81.

Glaser, M., Langer, T., & Weber, M. (2013). True overconfidence in interval estimates: Evidence based on a new

measure of miscalibration. Journal of Behavioral Decision Making, 2, 405–417.

Hales, J., & Kachelmeier, S. (2008). Predicting relative performance in economic competition. Journal of

Behavioral Finance, 9, 187–192.

Harrison, G., Martinez-Correa, J., & Swarthout, T. (2014). Eliciting subjective probabilities with binary lotteries.

Journal of Economic Behavior & Organization, 101, 128–140.

Harrison, G., & Ng, J. (2016). Evaluating the expected welfare gain from insurance. Journal of Risk and

Insurance, 83, 91–120.

Honl, H., Meissner, P., & Wulf, T. (2017). Risk attribution theory: An exploratory conceptualization of individ-

ual choice under uncertainty. The Journal of Behavioral and Experimental Economics, 67, 20–27.

Huang, R., Liu, Y., & Tzeng, L. (2010). Hidden overconfidence and advantageous selection. The Geneva Risk

and Insurance Review, 35, 93–107.

Jacobsen, B., Lee, J., Marquering, W., & Zhang, C. (2014). Gender differences in optimism and asset allocation.

Journal of Economic Behavior & Organization, 107, 630–651 .

Jaspersen, J. (2016). Hypothetical surveys and experimental studies of insurance demand: A review. Journal of

Risk and Insurance, 83, 217–255.

Jaspersen, J., & Aseervatham, V. (2017). The influence of affect on heuristic thinking in insurance demand.

Journal of Risk and Insurance, 84, 239–266.

Johnson, D., & Fowler, J. (2011). The evolution of overconfidence. Nature, 477, 317–320.

J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28 27



Karni, E. (2009). A mechanism for eliciting probabilities. Econometrica, 77, 603–606.

Kuhn, P., & Villeval, M. (2015). Are women more attracted to co-operation than men? The Economic Journal,

125, 115–140.

Kunreuther, H., & Pauly, M. (2004). Neglecting disaster: Why don’t people insure against large losses? Journal

of Risk and Uncertainty, 28, 5–21.

Landry, C., & Jahan-Parvar, M. (2011). Flood insurance coverage in the coastal zone. Journal of Risk and

Insurance, 78, 361–388.

Laury, S., & McInnes, M. (2003). The impact of insurance prices on decision making biases: An experimental

analysis. Journal of Risk & Insurance, 70, 219–233.

Malmendier, U., & Tate, G. (2005). Does overconfidence affect corporate investment?: CEO overconfidence

measures revisited. The Journal of Finance, 11, 649–659.

Moore, D., & Healy, P. (2008). The trouble with overconfidence. Psychological Review, 115, 502–517.

Pyszczynski, T., Holt, K., & Greenberg, J. (1987). Depression, self-focused attention, and expectancies for posi-

tive and negative future life events for self and others. Journal of Personality and Social Psychology, 52,

994–1001.

Rothschild, M., & Stiglitz, J. (1976). Equilibrium in competitive insurance markets: An essay on the economics

of imperfect information. The Quarterly Journal of Economics, 90, 629–649.

Royal, A., & Tasoff, J. (2017). When higher productivity hurts: The interaction between overconfidence and cap-

ital. Journal of Behavioral and Experimental Economics, 67, 131–142.

Sandroni, A., & Squintani, F. (2007). Overconfidence, insurance, and paternalism. American Economic Review,

97, 1994–2004.

Soll, J., & Klayman, J. (2004). Overconfidence in interval estimates. Journal of Experimental Psychology:

Learning, Memory, and Cognition, 30, 299–314.

Spinnewijn, J. (2013). Insurance and perceptions: How to screen optimists and pessimists. The Economic

Journal, 123, 606–633.

28 J. Coats, V. Bajtelsmit / Financial Services Review 29 (2021) 1–28