Deal or No Deal (MATH) PDF

Deal or No Deal?
Decision Making under Risk in a Large-Payoff Game Show
Thierry Post, Martijn J. van den Assem, Guido Baltussen and Richard H. Thaler *
Published in the American Economic Review, March 2008 (98:1), 3871
Abstract
We examine the risky choices of contestants in the popular TV game show Deal or No Deal and
related classroom experiments. Contrary to the traditional view of expected utility theory, the
choices can be explained in large part by previous outcomes experienced during the game. Risk
aversion decreases after earlier expectations have been shattered by unfavorable outcomes or
surpassed by favorable outcomes. Our results point to reference-dependent choice theories such as
prospect theory, and suggest that path-dependence is relevant, even when the choice problems are
simple and well-defined, and when large real monetary amounts are at stake.
JEL: D81
A WIDE RANGE OF THEORIES OF RISKY CHOICE have been developed, including the
normative expected utility theory of John von Neumann and Oskar Morgenstern (1944) and
the descriptive prospect theory of Daniel Kahneman and Amos Tversky (1979). Although
risky choice is fundamental to virtually every branch of economics, empirical testing of these
theories has proven to be difficult.
Many of the earliest tests such as those by Maurice Allais (1953), Daniel Ellsberg
(1961), and the early work by Kahneman and Tversky were based on either thought
experiments or answers to hypothetical questions. With the rising popularity of experimental
economics, risky choice experiments with real monetary stakes have become more popular,
but because of limited budgets most experiments are limited to small stakes. Some
experimental studies try to circumvent this problem by using small nominal amounts in
developing countries, so that the subjects face large amounts in real terms; see, for example,
Hans P. Binswanger (1980, 1981) and Steven J. Kachelmeier and Mohamed Shehata (1992).
Still, the stakes in these experiments are typically not larger than one months income and
thus do not provide evidence about risk attitudes regarding prospects that are significant in
relation to lifetime wealth.
Nonexperimental empirical research is typically plagued by what amounts to joint
hypothesis problems. Researchers cannot directly observe risk preferences for most real-life
1
problems, because the true probability distribution is not known to the subjects and the
subjects beliefs are not known to the researcher. For example, to infer the risk attitudes of
investors from their investment portfolios, one needs to know what their beliefs are regarding
the joint return distribution of the relevant asset classes. Were investors really so risk averse
that they required an equity premium of 7 percent per year, or were they surprised by an
unexpected number of favorable events or worried about catastrophic events that never
occurred? An additional complication arises because of the possible difference between risk
and uncertainty: real-life choices rarely come with precise probabilities.
In order to circumvent these problems, some researchers analyze the behavior of
contestants in TV game shows, for example Card Sharks (Robert H. Gertner, 1993),
Jeopardy! (Andrew Metrick, 1995), Illinois Instant Riches (Philip L. Hersch and Gerald
S. McDougall, 1997), Lingo (Roel M. W. J. Beetsma and Peter C. Schotman, 2001),
Hoosier Millionaire (Connel R. Fullenkamp, Rafael A. Tenorio and Robert H. Battalio,
2003) and Who Wants to be a Millionaire? (Roger Hartley, Gauthier Lanot and Ian Walker,
2006). The advantage of game shows is that the amounts at stake are larger than in experi-
ments and that the decision problems are often simpler and better defined than in real life.
The game show we use in this study, Deal or No Deal, has such desirable features that
it almost appears to be designed to be an economics experiment rather than a TV show. Here
is the essence of the game. A contestant is shown 26 briefcases which each contain a hidden
amount of money, ranging from 0.01 to 5,000,000 (in the Dutch edition). The contestant
picks one of the briefcases and then owns its unknown contents. Next, she selects 6 of the
other 25 briefcases to open. Each opened briefcase reveals one of the 26 prizes that are not in
her own briefcase. The contestant is then presented a bank offer the opportunity to walk
away with a sure amount of money and asked the simple question: Deal or No Deal? If
she says No Deal, she has to open five more briefcases, followed by a new bank offer. The
game continues in this fashion until the contestant either accepts a bank offer, or rejects all
offers and receives the contents of her own briefcase. The bank offers depend on the value of
the unopened briefcases; if, for example, the contestant opens high-value briefcases, the bank
offer falls.
This game show seems well-suited for analyzing risky choice. The stakes are very high
and wide-ranging: contestants can go home as multimillionaires or practically empty-handed.
Unlike other game shows, Deal or No Deal involves only simple stop-go decisions (Deal
or No Deal) that require minimal skill, knowledge or strategy, and the probability
distribution is simple and known with near-certainty (the bank offers are highly predictable,
2
as discussed later). Finally, the game show involves multiple game rounds, and consequently
seems particularly interesting for analyzing path-dependence, or the role of earlier outcomes.
Richard H. Thaler and Eric J. Johnson (1990) conclude that risky choice is affected by prior
outcomes in addition to incremental outcomes due to decision makers incompletely adapting
to recent losses and gains. Although Deal or No Deal contestants never have to pay money
out of their own pockets, they can suffer significant paper losses if they open high-value
briefcases (causing the expected winnings to fall), and such losses may influence their
subsequent choices. (Throughout this study we will use the term outcomes to indicate not
only monetary pay-offs, but also new information or changed expectations.)
We examine the games of 151 contestants from the Netherlands, Germany and the
United States in 2002 2007. The game originated in the Netherlands and is now broadcast
around the world. Although the format of Deal or No Deal is generally similar across all
editions, there are some noteworthy differences. For example, in the daily versions from Italy,
France and Spain, the banker knows the amounts in the briefcases and may make informative
offers, leading to strategic interaction between the banker and the contestant. In the daily
edition from Australia, special game options known as Chance and Supercase are
sometimes offered at the discretion of the game-show producer after a contestant has made a
Deal. These options would complicate our analysis, because the associated probability
distribution is not known, introducing a layer of uncertainty in addition to the pure risk of the
game. For these reasons, we limit our analysis to the games played in the Netherlands,
Germany and the United States.
The three editions have a very similar game format, apart from substantial variation in
the amounts at stake. While the average prize that can be won in the Dutch edition is roughly
400,000, the averages in the German and US edition are roughly 25,000 and 100,000,
respectively. At first sight, this makes the pooled dataset useful for separating the effect of the
amounts at stake from the effect of prior outcomes. (Within one edition, the stakes are
strongly confounded with prior outcomes.) However, cross-country differences in culture,
wealth and contestant selection procedure could confound the effect of stakes across the three
editions. To isolate the effect of stakes on risky choice, we therefore conduct classroom
experiments with a homogeneous student population. In these experiments, we vary the prizes
with a factor of ten, so that we can determine if, for example, 100 has the same subjective
value when it lies below or above the initial expectations.
Our findings are difficult to reconcile with expected utility theory. The contestants
choices appear to be driven in large part by the previous outcomes experienced during the
3
game. Risk aversion seems to decrease after earlier expectations have been shattered by
opening high-value briefcases, consistent with a break-even effect. Similarly, risk aversion
seems to decrease after earlier expectations have been surpassed by opening low-value
briefcases, consistent with a house-money effect.
The orthodox interpretation of expected utility of wealth theory does not allow for these
effects, because subjects are assumed to have the same preferences for a given choice
problem, irrespective of the path traveled before arriving at this problem. Our results point in
the direction of reference-dependent choice theories, such as prospect theory, and indicate that
path-dependence is relevant, even when large real monetary amounts are at stake. We
therefore propose a version of prospect theory with a path-dependent reference point as an
alternative to expected utility theory.
Of course, we must be careful with rejecting expected utility theory and embracing
alternatives like prospect theory. Although the standard implementation of expected utility
theory is unable to explain the choices of losers and winners, a better fit could be achieved
with a nonstandard utility function that has convex segments (as proposed by, for example,
Milton Friedman and Leonard J. Savage, 1948, and Harry Markowitz, 1952), and depends on
prior outcomes. Therefore, this study does not reject or accept any theory. Rather, our main
finding is the important role of reference-dependence and path-dependence, phenomena that
are not standard in typical implementations of expected utility, but common in prospect
theory. Any plausible explanation of the choice behavior in the game will have to account for
these phenomena. A theory with static preferences cannot explain why variation of the stakes
due to the subjects fortune during the game has a much stronger effect than variation in the
initial stakes across different editions of the TV show and experiments.
The remainder of this paper is organized as follows. In Section I, we describe the game
show in greater detail. Section II discusses our data material. Section III provides a first
analysis of the risk attitudes in Deal or No Deal by examining the bank offers and the
contestants decisions to accept (Deal) or reject (No Deal) these offers. Section IV
analyzes the decisions using expected utility theory with a general, flexible-form expo-power
utility function. Section V analyzes the decisions using prospect theory with a simple
specification that allows for partial adjustment of the subjective reference point that separates
losses from gains. This implementation of prospect theory explains a material part of what
expected utility theory leaves unexplained. Section VI reports results from classroom
experiments in which students play Deal or No Deal. The experiments confirm the
important role of previous outcomes and suggest that the isolated effect of the amounts at
4
stake is limited compared to the isolated effect of previous outcomes. Section VII offers
concluding remarks and suggestions for future research. Finally, an epilogue gives a synopsis
of other Deal or No Deal studies that became available after our study was first submitted to
the American Economic Review in October 2005.
I. Description of the Game Show

The TV game show Deal or No Deal was developed by the Dutch production
company Endemol and was first aired in the Netherlands in December 2002. The game show
soon became very popular and was exported to dozens of other countries, including Germany
and the United States. The following description applies to the Dutch episodes of Deal or No
Deal. Except for the monetary amounts, the structure of the main game is similar in the
German and US versions used in this study.
Each episode consists of two parts: an elimination game based on quiz questions in
order to select one finalist from the audience, and a main game in which this finalist plays
Deal or No Deal. Audience members have not been subjected to an extensive selection
procedure: players in the national lottery sponsoring the show are invited to apply for a seat
and tickets are subsequently randomly distributed to applicants. Only the main game is the
subject of our study. Except for determining the identity of the finalist, the elimination game
does not influence the course of the main game. The selected contestant has not won any prize
before entering the main game.
The main game starts with a fixed and known set of 26 monetary amounts ranging from
0.01 to 5,000,000, which have been randomly allocated over 26 numbered and closed
briefcases. One of the briefcases is selected by the contestant and this briefcase is not to be
opened until the end of the game.
The game is played over a maximum of nine rounds. In each round, the finalist chooses
one or more of the other 25 briefcases to be opened, revealing the prizes inside. Next, a
banker tries to buy the briefcase from the contestant by making her an offer. Contestants
have a few minutes to evaluate the offer and to decide between Deal and No Deal, and
may consult a friend or relative who sits nearby.1 The remaining prizes and the current bank
offer are displayed on a scoreboard and need not be memorized by the contestant. If the
1
In the US version and in the second German series, three or four friends and/or relatives sit on stage nearby the
contestant. In the Dutch version and in the first German series, only one person accompanies the contestant.
5
26 briefcases open 6 cases
bank offer 1 20 briefcases open 5 cases
bank offer 5 6 briefcases open 1 case
open contestants briefcase
Figure 1: Flow Chart of the Main Game. In each round, the finalist chooses a number of
briefcases to be opened, each giving new information about the unknown prize in the contestants
own briefcase. After the prizes in the chosen briefcases are revealed, a bank offer is presented to
the finalist. If the contestant accepts the offer (Deal), she walks away with the amount offered
and the game ends; if the contestant rejects the offer (No Deal), play continues and she enters
the next round. If the contestant decides No Deal in the ninth round, she receives the prize in her
own briefcase. The flow chart applies to the Dutch and US editions and the second German series.
The first German series involves one fewer game round and starts with 20 briefcases.
contestant accepts the offer (Deal), she walks away with this sure amount and the game
ends; if the contestant rejects the offer (No Deal), the game continues and she enters the
next round.
In the first round, the finalist has to select six briefcases to be opened, and the first bank
offer is based on the remaining 20 prizes. The numbers of briefcases to be opened in the
6
13,000
0.01 7,500
0.20 10,000
0.50 25,000
1 50,000
5 -------------------- 75,000
10 close-up of the 100,000
20 contestant is 200,000
50 shown here 300,000
100 -------------------- 400,000
500 500,000
1,000 1,000,000
2,500 2,500,000
5,000 5,000,000
Figure 2: Example of the Main Game as Displayed on the TV Screen. A close-up of the
contestant is shown in the center of the screen. The possible prizes are listed in the columns to the
left and right of the contestant. Prizes eliminated in earlier rounds are shown in a dark color and
remaining prizes are in a bright color. The top bar above the contestant shows the bank offer. This
example demonstrates the two options open to the contestant after opening six briefcases in the
first round: accept a bank offer of 13,000 or continue to play with the remaining 20 briefcases,
one of which is the contestants own. This example reflects the prizes in the Dutch episodes.
maximum of eight subsequent rounds are 5, 4, 3, 2, 1, 1, 1, and 1. Accordingly, the number of

prizes left in the game decreases to 15, 11, 8, 6, 5, 4, 3, and 2. If the contestant rejects all nine
offers she receives the prize in her own briefcase. Figure 1 illustrates the basic structure of the
main game.
To provide further intuition for the game, Figure 2 shows a typical example of how the
main game is displayed on the TV screen. A close-up of the contestant is shown in the center
and the original prizes are listed to the left and the right of the contestant. Eliminated prizes
are shown in a dark color and remaining prizes are in a bright color. The bank offer is
displayed at the top of the screen.
As can be seen on the scoreboard, the initial prizes are highly dispersed and positively
skewed. During the course of the game, the dispersion and the skewness generally fall as
more and more briefcases are opened. In fact, in the ninth round, the distribution is perfectly
symmetric, because the contestant then faces a 50/50 gamble with two remaining briefcases.
7
Bank Behavior
Although the contestants do not know the exact bank offers in advance, the banker
behaves consistently according to a clear pattern. Four simple rules of thumb summarize this
pattern:
Rule 1. Bank offers depend on the value of the unopened briefcases: when the lower
(higher) prizes are eliminated, the average remaining prize increases
(decreases) and the banker makes a better (worse) offer.
Rule 2. The offer typically starts at a low percentage (usually less than 10 percent) of
the average remaining prize in the first round and gradually increases to 100
percent in the later rounds. This strategy obviously serves to encourage
contestants to continue playing the game and to gradually increase excitement.
Rule 3. The offers are not informative, that is, they cannot be used to determine which
of the remaining prizes is in the contestants briefcase. Only an independent
auditor knows the distribution of the prizes over the briefcases. Indeed, there is
no correlation between the percentage bank offer and the relative value of the
prize in the contestants own briefcase.
Rule 4. The banker is generous to losers by offering a relatively high percentage of the
average remaining prize. This pattern is consistent with path-dependent risk
attitudes. If the game-show producer understands that risk aversion falls after
large losses, he may understand that high offers are needed to avoid trivial
choices and to keep the game entertaining to watch. Using the same reasoning,
we may also expect a premium after large gains; this, however, does not occur,
perhaps because with large stakes, the game is already entertaining.
Section III gives descriptive statistics on the bank offers in our sample and Section IV
presents a simple model that captures the rules of thumb noted above. The key finding is that
the bank offers are highly predictable.
II. Data
We examine all Deal or No Deal decisions of 151 contestants appearing in episodes
aired in the Netherlands (51), Germany (47), and the United States (53).
The Dutch edition of Deal or No Deal is called Miljoenenjacht (or Chasing
Millions). The first Dutch episode was aired on December 22, 2002 and the last in our
sample dates from January 1, 2007. In this time span, the game show was aired 51 times,
8
divided over eight series of weekly episodes and four individual episodes aired on New
Years Day, with one contestant per episode. A distinguishing feature of the Dutch edition is
the high amounts at stake: the average prize equals roughly 400,000 (391,411 in episode
1 47 and 419,696 in episode 48 51). Contestants may even go home with 5,000,000.
The fact that the Dutch edition is sponsored by a national lottery probably explains why the
Dutch format has such large prizes. The large prizes may also have been preferred to
stimulate a successful launch of the show and to pave the way for exporting the formula
abroad. Part of the 51 shows were recorded on videotape by the authors and tapes of the
remaining shows were obtained from the Dutch broadcasting company TROS.
In Germany, a first series of Deal or No Deal Die Show der GlcksSpirale started
on June 23, 2005 and a second series began on June 28, 2006.2 Apart from the number of
prizes, the two series are very similar. The first series uses 20 prizes instead of 26 and is
played over a maximum of 8 game rounds instead of 9. Because these 8 rounds are exactly
equal to round 2 9 of the regular format in terms of the number of remaining prizes and in
terms of the number of briefcases that have to be opened, we can analyze this series as if the
first round has been skipped. Both series have the same maximum prize (250,000) and the
averages of the initial set of prizes are practically equal (26,347 versus 25,003
respectively). In the remainder of the paper we will consider the two German series as one
combined subsample. The first series was broadcast weekly and lasted for 10 episodes, each
with two contestants playing the game sequentially. The second series was aired either once
or twice a week and lasted for 27 episodes, with one contestant per episode, bringing the total
number of German contestants in our sample to 47. Copies of the first series were obtained
from TV station Sat.1 and from Endemols local production company Endemol Deutschland.
The second series was recorded by a friend of the authors.
In the United States, the game show debuted on December 19, 2005, for five
consecutive nights and returned on TV on February 27, 2006. This second series lasted for 34
episodes until early June 2006. The 39 episodes combined covered the games of 53
contestants, with some contestants starting in one episode and continuing their game in the
next. The regular US format has a maximum initial prize of $1,000,000 (roughly 800,000)
and an average of $131,478 (105,182). In the games of six contestants, however, the top
prizes and averages were larger to mark the launch and the finale of the second series. All US
2
An earlier edition called Der MillionenDeal started on May 1, 2004. The initial average prize was 237,565
and the largest prize was 2,000,000. This edition however lasted for only 6 episodes and is therefore not
included here.
9
shows were recorded by the authors. US Dollars are translated into Euros by using a single
fixed rate of 0.80 per $ (the actual exchange rate was within 5 percent of this rate for both
the 2005 and 2006 periods).
For each contestant, we collected data on the eliminated and remaining prizes, the bank
offers, and the Deal or No Deal decisions in every game round, leading to a panel dataset
with a time-series dimension (the game rounds) and a cross-section dimension (the
contestants).
We also collected data on each contestants gender, age and education. Age and
education are often revealed in an introductory talk or in other conversations during the game.
The level of education is coded as a dummy variable, with a value of 1 assigned to contestants
with a bachelor degree level or higher (including students) or equivalent work experience.
Although a contestants level of education is usually not explicitly mentioned, it is often clear
from the stated profession. We estimate the missing values for age based on the physical
appearance of the contestant and information revealed in the introductory talk, for example,
the age of children. However, age, gender and education do not have significant explanatory
power in our analysis. In part or in whole, this may reflect a lack of sampling variation. For
example, during the game, the contestant is permitted to consult with friends, family
members, or spouse, and therefore decisions in this game are in effect taken by a couple or a
group, mitigating the role of the individual contestants age, gender or education. For the sake
of brevity, we will pay no further attention to the role of contestant characteristics. Moreover,
prior outcomes are random and unrelated to characteristics and therefore the characteristics
probably would not affect our main conclusions about path-dependence, even if they would
affect the level of risk aversion.
Table 1 shows summary statistics for our sample. Compared to the German and US
contestants, the Dutch contestants on average accept lower percentage bank offers (76.3
percent versus 91.8 and 91.4 percent) and play roughly three fewer game rounds (5.2 versus
8.2 and 7.7 rounds). These differences may reflect unobserved differences in risk aversion due
to differences in wealth, culture or contestant selection procedure. In addition, increasing
relative risk aversion (IRRA) may help to explain the differences. As the Dutch edition
involves much larger stakes than the German and US editions, a modest increase in relative
risk aversion suffices to yield sizeable differences in the accepted percentages. Furthermore,
the observed differences in the number of rounds played are inflated by the behavior of the
banker. The percentage bank offer increases with relatively small steps in the later game
10
Table 1
Summary Statistics
The table shows descriptive statistics for our sample of 151 contestants from the Netherlands (51;
panel A), Germany (47; panel B) and the United States (53; panel C). The contestants
characteristics age and education are revealed in an introduction talk or in other conversations
between the host and the contestant. Age is measured in years. Gender is a dummy variable with a
value of one assigned to females. Education is a dummy variable that takes a value of one for
contestants with a bachelor degree or higher (including students) or equivalent work experience.
Stop Round is the round number in which the bank offer is accepted. The round numbers from the
first series of German episodes are adjusted by +1 to correct for the lower initial number of
briefcases and game rounds; for contestants who played the game to the end, the stop round is set
equal to 10. Best Offer Rejected is the highest percentage bank offer the contestant chose to reject
(No Deal). Offer Accepted is the percentage bank offer accepted by the contestant (Deal), or
100 percent for contestants who rejected all offers. Amount Won equals the accepted bank offer in
monetary terms, or the prize in the contestants own briefcase for contestants who rejected all
offers.
Mean Stdev Min Median Max

A. Netherlands (N = 51)
Age (years) 45.31 11.51 21.00 43.00 70.00
Gender (female = 1) 0.27 0.45 0.00 0.00 1.00
Education (high = 1) 0.55 0.50 0.00 1.00 1.00
Stop Round 5.22 1.75 3.00 5.00 10.00
Best Offer Rejected (%) 55.89 32.73 10.17 55.32 119.88
Offer Accepted (%) 76.27 30.99 20.77 79.29 165.50
Amount Won () 227,264.90 270,443.20 10.00 148,000.00 1,495,000.00
B. Germany (N = 47)
Age (years) 36.47 8.17 20.00 35.00 55.00
Gender (female = 1) 0.34 0.48 0.00 0.00 1.00
Education (high = 1) 0.47 0.50 0.00 0.00 1.00
Stop Round 8.21 1.53 5.00 8.00 10.00
Offer Accepted (%) 91.79 19.15 52.78 95.99 149.97
Amount Won () 20,602.56 25,946.69 0.01 14,700.00 150,000.00
C. United States (N = 53)
Age (years) 34.98 10.03 22.00 33.00 76.00
Gender (female = 1) 0.57 0.50 0.00 1.00 1.00
Education (high = 1) 0.49 0.50 0.00 0.00 1.00
Stop Round 7.70 1.29 5.00 8.00 10.00
Offer Accepted (%) 91.43 15.31 49.16 97.83 112.50
Amount Won ($) 122,544.58 119,446.18 5.00 94,000.00 464,000.00
rounds and consequently a modest increase in relative risk aversion can yield a large
reduction in the number of game rounds played. Thus, the differences between the Dutch
contestants on the one hand and the German and US contestants on the other hand are
consistent with moderate IRRA.
11
Cross-Country Analysis
Apart from the amounts at stake, the game show format is very similar in the three
countries. Still, there are some differences in how contestants are chosen to play that may
create differences in the contestant pool. In the Dutch and German episodes in our sample
there is a preliminary game in which contestants answer quiz questions, the winner of which
gets to play the main game we study. One special feature of the Dutch edition is the existence
of a bail-out offer at the end of the elimination game: just before a last, decisive question,
the two remaining contestants can avoid losing and leaving empty-handed by accepting an
unknown prize that is announced to be worth at least 20,000 (approximately 5 percent of the
average prize in the main game) and typically turns out to be a prize such as a world trip or a
car. If the more risk-averse pre-finalists are more likely to exit the game at this stage, the
Dutch finalists might be expected to be less risk averse on average. In the United States,
contestants are not selected based on an elimination game but rather the producer selects each
contestant individually, and the selection process appears to be based at least in part on the
appearance and personalities of the contestants. (The Web site for the show tells prospective
contestants to send a video of themselves and their proposed accompanying friends and
relatives. The show also conducts open casting calls.) Contestants (and their friends) thus
tend to be attractive and lively. Another concern is that richer and more risk-seeking people
may be more willing to spend time attempting to get onto large-stake editions than onto
small-stake editions. To circumvent these problems, Section VI complements the analysis of
the TV shows with classroom experiments that use a homogeneous student population.
III. Preliminary Analysis

To get a first glimpse of the risk preferences in Deal or No Deal, we analyze the
offers made by the banker and the contestants decisions to accept or reject these offers in the
various game rounds.
Several notable features of the game can be seen in Table 2. First, the banker becomes
more generous by offering higher percentages as the game progresses (Rule 2). The offers
typically start at a small fraction of the average prize and approach 100 percent in the later
rounds. The strong similarity between the percentages in the Dutch edition (panel A), the
German edition (panel B) and the US edition (panel C) suggest that the banker behaves in a
12
Table 2
Bank Offers and Contestants Decisions
The table shows summary statistics for the percentage bank offers and contestants decisions in
our sample of 151 contestants from the Netherlands (51; panel A), Germany (47; panel B) and the
United States (53; panel C). The average bank offer as a percentage of the average remaining prize
(%BO), the average remaining prize in Euros (Stakes) and the number of contestants (No.) are
reported for each game round (r = 1,,9). The statistics are also shown separately for contestants
accepting the bank offer (Deal) and for contestants rejecting the bank offer (No Deal). The
round numbers from the first series of German episodes are adjusted by +1 to correct for the lower
initial number of briefcases and game rounds.
Unconditional Deal No Deal

Round %BO Stakes No. %BO Stakes No. %BO Stakes No.
1 6% 387,867 51 - - 0 6% 387,867 51
2 14% 376,664 51 - - 0 14% 376,664 51
3 34% 369,070 51 36% 409,802 10 33% 359,135 41
4 61% 348,820 41 69% 394,860 11 58% 331,939 30
5 77% 317,618 30 82% 557,680 7 76% 244,555 23
6 88% 234,877 23 90% 237,416 12 87% 232,107 11
7 98% 243,868 11 104% 414,106 6 91% 39,582 5
8 96% 50,376 5 100% 78,401 3 90% 8,338 2
9 106% 11,253 2 91% 17,500 1 120% 5,005 1
B. Germany (N = 47)
1 8% 24,277 27 - - 0 8% 24,277 27
2 15% 24,915 47 - - 0 15% 24,915 47
3 34% 23,642 47 - - 0 34% 23,642 47
4 46% 21,218 47 - - 0 46% 21,218 47
5 59% 22,304 47 59% 29,976 2 59% 21,963 45
6 72% 20,557 45 67% 48,038 7 73% 15,494 38
7 88% 15,231 38 85% 21,216 5 88% 14,324 33
8 98% 15,545 33 91% 28,813 10 101% 9,776 23
9 103% 14,017 23 109% 13,925 11 99% 14,101 12
1 11% 152,551 53 - - 0 11% 152,551 53
2 21% 151,885 53 - - 0 21% 151,885 53
3 36% 147,103 53 - - 0 36% 147,103 53
4 50% 148,299 53 - - 0 50% 148,299 53
5 62% 148,832 53 79% 118,517 1 61% 150,434 52
6 73% 150,549 52 74% 139,421 9 73% 152,879 43
7 88% 154,875 43 91% 204,263 15 86% 128,416 28
8 92% 114,281 28 96% 183,917 14 88% 44,644 14
9 98% 39,922 14 99% 53,825 8 97% 21,384 6
similar way across the three editions.3 The number of remaining contestants in every round
clearly shows that the Dutch contestants tend to stop earlier and accept relatively lower bank
offers than the German and US contestants do. Again, this may reflect the substantially larger
3
A spokesman from Endemol, the production company, confirmed that the guidelines for bank offers are the
same for all three editions included in our sample.
13
stakes in the Dutch edition, or, alternatively, unobserved differences in risk aversion due to
differences in wealth, culture or contestant selection procedure. Third, the contestants
generally exhibit what might be called only moderate risk aversion. In the US and German
sample, all contestants keep playing until the bank offer is at least half the expected value of
the prizes in the unopened briefcases. In round 3 in the Netherlands, 20 percent of the
contestants (10 out of 51) do accept deals that average only 36 percent of the expected value
of the unopened briefcases, albeit at stakes that exceed 400,000. Many contestants turn down
offers of 70 percent or more of amounts exceeding 100,000. Fourth, there can be wide
discrepancies, even within a country, in the stakes that contestants face. In the Dutch show,
contestants can be playing for many hundreds of thousands of Euros, down to a thousand or
less. In the later rounds, the contestant is likely to face relatively small stakes, as a
consequence of the skewness of the initial set of prizes.
It is not apparent from this table what effect the particular path a player takes can have
on the choices she makes. To give an example of the decisions faced by an unlucky player,
consider poor Frank, who appeared in the Dutch episode of January 1, 2005 (see Table 3). In
round 7, after several unlucky picks, Frank opened the briefcase with the last remaining large
prize (500,000) and saw the expected prize tumble from 102,006 to 2,508. The banker
then offered him 2,400, or 96 percent of the average remaining prize. Frank rejected this
offer and play continued. In the subsequent rounds, Frank deliberately chose to enter unfair
gambles, to finally end up with a briefcase worth only 10. Specifically, in round 8, he
rejected an offer of 105 percent of the average remaining prize; in round 9, he even rejected a
certain 6,000 in favor of a 50/50 gamble of 10 or 10,000. We feel confident to classify this
last decision as risk-seeking behavior, because it involves a single, simple, symmetric gamble
with thousands of Euros at stake. Also, unless we are willing to assume that Frank would
always accept unfair gambles of this magnitude, the only reasonable explanation for his
choice behavior seems to be a reaction to his misfortune experienced earlier in the game.
In contrast, consider the exhilarating ride of Susanne, an extremely fortunate contestant
who appeared in the German episode of August 23, 2006 (see Table 4). After a series of very
lucky picks, she eliminated the last small prize of 1,000 in round 8. In round 9, she then
faced a 50/50 gamble of 100,000 or 150,000, two of the three largest prizes in the German
edition. While she was concerned and hesitant in the earlier game rounds, she decidedly
rejected the bank offer of 125,000, the expected value of the gamble; a clear display of risk-
seeking behavior and one that proved fortuitous in this case as she finally ended up winning
150,000.
14
Table 3
Example Frank
The table shows the gambles presented to a Dutch contestant named Frank and the Deal or No
Deal decisions made by him in game rounds 1 9. This particular episode was broadcast on
January 1, 2005. For each game round, the table shows the remaining prizes, the average
remaining prize, the bank offer, the percentage bank offer and the Deal or No Deal decision.
Frank ended up with a prize of 10.
Game Round (r)

Prize () 1 2 3 4 5 6 7 8 9
0.01 X X
0.20 X X
0.50 X X X X X X X
1 X X X X X
5
10 X X X X X X X X X
20 X X X X X X X X
50
100
500
1,000 X
2,500 X X X
5,000 X X
7,500
10,000 X X X X X X X X X
25,000 X X
50,000 X X X X
75,000 X X X
100,000 X X X
200,000 X X X X
300,000 X
400,000 X
500,000 X X X X X X
1,000,000 X
2,500,000
5,000,000 X
Average () 383,427 64,502 85,230 95,004 85,005 102,006 2,508 3,343 5,005
Offer () 17,000 8,000 23,000 44,000 52,000 75,000 2,400 3,500 6,000
Offer (%) 4% 12% 27% 46% 61% 74% 96% 105% 120%
Decision No Deal No Deal No Deal No Deal No Deal No Deal No Deal No Deal No Deal
Thus both unlucky Frank and lucky Susanne exhibit very low levels of risk aversion,
even risk-seeking, whereas most of the contestants in the shows are at least moderately risk
averse. Franks behavior is consistent with a break-even effect, a willingness to gamble in
order to get back to some perceived reference point. Susannes behavior is consistent with a
house-money effect, an increased willingness to gamble when someone thinks she is
playing with someone elses money.
15
Table 4
Example Susanne
The table shows the gambles presented to a German contestant named Susanne and the Deal or
No Deal decisions made by her in game rounds 1 9. This particular episode was broadcast on
August 23, 2006. For each game round, the table shows the remaining prizes, the average
remaining prize, the bank offer, the percentage bank offer, and the Deal or No Deal decision.
Susanne ended up with a prize of 150,000.
Game Round (r)

Prize () 1 2 3 4 5 6 7 8 9
0.01 X X X X
0.20 X X X
0.50 X X X X X X X
1
5
10
20 X X
50 X X
100 X X X X
200
300 X X X
400 X
500
1,000 X X X X X X X X
2,500 X X X X X X
5,000 X
7,500
10,000 X X
12,500 X X X
15,000 X
20,000 X X
25,000 X X X X X
50,000 X
100,000 X X X X X X X X X
150,000 X X X X X X X X X
250,000 X
Average () 32,094 21,431 26,491 34,825 46,417 50,700 62,750 83,667 125,000
Offer () 3,400 4,350 10,000 15,600 25,000 31,400 46,000 75,300 125,000
Offer (%) 11% 20% 38% 45% 54% 62% 73% 90% 100%
Decision No Deal No Deal No Deal No Deal No Deal No Deal No Deal No Deal No Deal
To systematically analyze the effect of prior outcomes such as the extreme ones
experienced by Frank and Suzanne, we first develop a rough classification of game situations
in which the contestant is classified as a loser or a winner and analyze the decisions of
contestants in these categories separately.
Our classification takes into account the downside risk and upside potential of rejecting
the current bank offer. A contestant is a loser if her average remaining prize after opening one
additional briefcase is low, even if the best-case scenario of eliminating the lowest remaining
16
prize would occur. Using x r for the current average, the average remaining prize in the best-
case scenario is:
n r x r x rmin
(1) BC r =
nr 1
where nr stands for the number of remaining briefcases in game round r = 1, L ,9 and x rmin
for the smallest remaining prize. Similarly, winners are classified by the average remaining
prize in the worst-case scenario of eliminating the largest remaining prize, x rmax :
n r x r x rmax
(2) WC r =
nr 1
More specifically, we classify a contestant in a given game round as a loser if BC r

belongs to the worst one-third for all contestants in that game round and as a winner if WC r
belongs to the best one-third.4 Game situations that satisfy neither of the two conditions (or
both in rare occasions) are classified as neutral.
Of course, there are numerous ways one could allocate players into winner and loser
categories. The results we show are robust to other classification schemes, provided that the
classification of winners accounts for the downside risk of continuing play: the house-money
effect a decreased risk aversion after prior gains is weak if incremental losses can exceed
prior gains. For example, partitioning on just the current average ( x r ) does not distinguish
between situations with different dispersion around that average, and therefore takes no
account of the downside risk of continuing play.
Table 5 illustrates the effect of previous outcomes on the contestants choice behavior.
We see that, compared to contestants who are in the neutral category, both winners and losers
have a stronger tendency to continue play. While 31 percent of all Deal or No Deal choices
in the neutral group are Deal in the Dutch sample, the Deal percentage is only 14 percent
for losers despite the generous offers they are presented (Rule 4). The low Deal
percentage for losers suggests that risk aversion decreases when contestants have been
4
To account for the variation in the initial set of prizes within an edition (see Section II), BCr and BWr are scaled
by the initial average prize.
17
Table 5
Deal or No Deal Decisions after Bad and Good Fortune
The table summarizes the Deal or No Deal decisions for our sample of 151 contestants from the
Netherlands (51; panel A), Germany (47; panel B) and the United States (53; panel C). The
samples are split based on the fortune experienced by contestants during the game. A contestant is
classified as a loser if her average remaining prize after eliminating the lowest remaining prize is
among the worst one-third for all contestants in the same game round; she is a winner if the
average after eliminating the largest remaining prize is among the best one-third. For each
category and game round, the table displays the percentage bank offer (%BO), the number of
contestants (No.) and the percentage of contestants choosing Deal (%D). The round numbers
from the first series of German episodes are adjusted by +1 to correct for the lower initial number
of briefcases and game rounds.
Loser Neutral Winner

Round %BO No. %D %BO No. %D %BO No. %D
1 6% 17 0% 6% 17 0% 6% 17 0%
2 15% 17 0% 12% 17 0% 15% 17 0%
3 40% 17 12% 29% 17 41% 31% 17 6%
4 69% 14 14% 58% 13 46% 54% 14 21%
5 82% 10 10% 71% 10 20% 78% 10 40%
6 94% 8 50% 85% 7 43% 86% 8 63%
7 99% 4 25% 97% 3 67% 99% 4 75%
8 105% 1 0% 91% 3 67% 100% 1 100%
9 120% 1 0% - 0 - 91% 1 100%
2-9 72 14% 70 31% 72 25%
B. Germany (N = 47)
1 7% 9 0% 7% 9 0% 8% 9 0%
2 16% 16 0% 13% 15 0% 14% 16 0%
3 35% 16 0% 33% 15 0% 33% 16 0%
4 46% 16 0% 44% 15 0% 47% 16 0%
5 65% 16 0% 54% 15 13% 57% 16 0%
6 83% 15 0% 67% 15 20% 66% 15 27%
7 107% 13 0% 80% 12 25% 76% 13 15%
8 117% 11 0% 89% 11 55% 86% 11 36%
9 107% 8 38% 106% 7 57% 98% 8 50%
2-9 111 3% 105 17% 111 13%
1 9% 18 0% 10% 17 0% 13% 18 0%
2 19% 18 0% 19% 17 0% 25% 18 0%
3 41% 18 0% 29% 17 0% 39% 18 0%
4 57% 18 0% 42% 17 0% 51% 18 0%
5 69% 18 0% 55% 17 6% 62% 18 0%
6 78% 18 11% 68% 16 31% 73% 18 11%
7 92% 15 27% 87% 13 23% 84% 15 53%
8 94% 9 22% 95% 10 70% 87% 9 56%
9 92% 4 50% 101% 6 67% 99% 4 50%
2-9 118 8% 113 18% 118 14%
18
unlucky in selecting which briefcases to open. In fact, the strong losers in our sample
generally exhibit risk-seeking behavior by rejecting bank offers in excess of the average
remaining prize.
The low Deal percentage could be explained in part by the smaller stakes faced by
losers and a lower risk aversion for small stakes, or increasing relative risk aversion (IRRA).
However, the losers generally still have at least thousands or tens of thousands of Euros at
stake and gambles of this magnitude are typically associated with risk aversion in other
empirical studies (including other game show studies and experimental studies). Also, if the
stakes explained the low risk aversion of losers, we would expect a higher risk aversion for
winners. However, risk aversion seems to decrease when contestants are lucky and have
eliminated low-value briefcases. The Deal percentage for winners is 25 percent, below the
31 percent for the neutral group.
Interestingly, the same pattern arises in all three countries. The overall Deal
percentages in the German and US editions are lower than in the Dutch edition, consistent
with moderate IRRA and the substantially smaller stakes. Within every edition, however, the
losers and winners have relatively low Deal percentages.
These results suggest that prior outcomes are an important determinant of risky choice.
This is inconsistent with the traditional interpretation of expected utility theory in which the
preferences for a given choice problem do not depend on the path traveled before arriving at
the choice problem. By contrast, path-dependence can be incorporated quite naturally in
prospect theory. The lower risk aversion after misfortune is reminiscent of the break-even
effect, or decision makers being more willing to take risks due to incomplete adaptation to
previous losses. Similarly, the relatively low Deal percentage for winners is consistent with
the house-money effect, or a lower risk aversion after earlier gains.
Obviously, this preliminary analysis of Deal percentages is rather crude. It does not
specify an explicit model of risky choice and it does not account for the precise choices (bank
offers and remaining prizes) the contestants face. Furthermore, there is no attempt at statistical
inference or controlling for confounding effects at this stage of our analysis. The next two
sections use a structural choice model and a maximum-likelihood methodology to analyze the
Deal or No Deal choices in greater detail.
19
IV. Expected Utility Theory
This section analyzes the observed Deal or No Deal choices with the standard
expected utility of wealth theory. The choice of the appropriate class of utility functions is
important, because preferences are evaluated on an interval from cents to millions. We do not
want to restrict our analysis to a classical power or exponential utility function, because it
seems too restrictive to assume constant relative risk aversion (CRRA) or constant absolute
risk aversion (CARA) for this interval. To allow for the plausible combination of increasing
relative risk aversion (IRRA) and decreasing absolute risk aversion (DARA), we employ a
variant of the flexible expo-power family of Atanu Saha (1993) that was used by Mohammed
Abdellaoui, Carolina Barrios and Peter P. Wakker (2007) and by Charles A. Holt and Susan
K. Laury (2002):
1 exp( (W + x)1 )
(3) u ( x) =

In this function, three parameters are unknown: the risk aversion coefficients and ,
and the initial wealth parameter W. The classical CRRA power function arises as the limiting
case where 0 and the CARA exponential function arises as the special case where = 0.
Theoretically, the correct measure of wealth should be lifetime wealth, including the present
value of future income. However, lifetime wealth is not observable and it is possible that
contestants do not integrate their existing wealth with the payoffs of the game. Therefore, we
include initial wealth as a free parameter in our model.
We will estimate the three unknown parameters using a maximum likelihood procedure
that measures the likelihood of the observed Deal or No Deal decisions based on the stop
value, or the utility of the current bank offer, and the continuation value, or the expected
utility of the unknown winnings when rejecting the offer. In a given round r, B( x r ) denotes
the bank offer as a function of the set of remaining prizes x r . The stop value is simply:
(4) sv( xr ) = u ( B( xr ))
Analyzing the continuation value is more complicated. We elaborate on the continuation

value, the bank offer model and the estimation procedure below.
20
Continuation Value
The game involves multiple rounds and the continuation value has to account for the
bank offers and optimal decisions in all later rounds. In theory, we can solve the entire
dynamic optimization problem by means of backward induction, using Richard E. Bellmans
principle of optimality. Starting with the ninth round, we can determine the optimal Deal or
No Deal decision in each preceding game round, accounting for the possible scenarios and
the optimal decisions in subsequent rounds. This approach assumes, however, that the
contestant takes into account all possible outcomes and decisions in all subsequent game
rounds. Studies on backward induction in simple alternating-offers bargaining experiments
suggest that subjects generally do only one or two steps of strategic reasoning and ignore
further steps of the backward induction process; see, for example, Johnson et al. (2002) and
Ken Binmore et al. (2002). This pleads for assuming that the contestants adopt a simplified
mental frame of the game.
Our video material indeed suggests that contestants generally look only one round
ahead. The game-show host tends to stress what will happen to the bank offer in the next
round should particular briefcases be eliminated and the contestants themselves often
comment that they will play just one more round (although they often change their minds
and continue to play later on). We therefore assume a simple myopic frame. Using this
frame, the contestant compares the current bank offer with the unknown offer in the next
round, and ignores the option to continue play thereafter.
Given the current set of prizes ( x r ), the statistical distribution of the set of prizes in the
next round ( x r +1 ) is known:
1
nr
(5) Pr[ x r +1 = y | x r ] = = pr
n r +1
for any given subset y of nr +1 elements from x r . In words, the probability is simply one
divided by the number of possible combinations of nr +1 out of nr . Thus, using ( x r ) for all
such subsets, the continuation value for a myopic contestant is given by:
(6) cv( x r ) = u ( B( y )) p
y ( x r )
r
21
Given the cognitive burden of multi-stage induction, this frame seems the appropriate
choice for this game. However, as a robustness check, we have also replicated our estimates
using the rational model of full backward induction and have found that our parameter
estimates and the empirical fit did not change materially. In the early game rounds, when
backward induction appears most relevant, the myopic model underestimates the continuation
value. Still, the myopic model generally correctly predicts No Deal, because the expected
bank offers usually increase substantially during the early rounds, so even the myopic
continuation value is generally greater than the stop value. In the later game rounds, backward
induction is of less importance, because fewer game rounds remain to be played and because
the rate of increase in the expected bank offers slows down. For contestants who reach round
nine, such as Frank and Susanne, the decision problem involves just one stage and the myopic
model coincides with the rational model. The low propensity of losers and winners in later
game rounds to Deal is therefore equally puzzling under the assumption of full backward
induction.
Bank Offers
To apply the myopic model, we need to quantify the behavior of the banker. Section I
discussed the bank offers in a qualitative manner. For a contestant who currently faces
remaining prizes x r and percentage bank offer br in game round r = 1,L ,9 , we quantify this
behavior using the following simple model:
(7) B( x r +1 ) = br +1 x r +1
(8) br +1 = br + (1 br ) (9 r )
where , 0 1 , measures the speed at which the percentage offer goes to 100 percent.
Since myopic contestants are assumed to look just one round ahead, the model predicts the
offer in the next round only. The bank offer in the first round needs not be predicted, because
it is shown on the scoreboard when the first Deal or No Deal choice has to be made.
B ( x10 ) = x10 and b10 = 1 refer to the prize in the contestants own briefcase.
The model does not include an explicit premium for losers. However, before misfortune
arises, the continuation value is driven mostly by the favorable scenarios and the precise
22
percentage offers for unfavorable scenarios do not materially affect the results. After bad luck,
the premium is included in the current percentage and extrapolated to the next game round.
For each edition, we estimate the value of by fitting the model to the sample of
percentage offers made to all contestants in all relevant game rounds using least squares
regression analysis. The resulting estimates are very similar for each edition: 0.832 for the
Dutch edition, 0.815 for the first German series, 0.735 for the second German series and
0.777 for the US shows. The model gives a remarkably good fit. Figure 3 illustrates the
goodness-of-fit by plotting the predicted bank offers against the actual offers. The results are
highly comparable for the three editions in our study and therefore the figure shows the
pooled results. For each individual sample, the model explains well over 70 percent of the
total variation in the individual percentage offers. The explanatory power is even higher for
monetary offers, with an R-squared of roughly 95 percent for each sample. Arguably, accurate
monetary offers are more relevant for accurate risk aversion estimates than accurate
percentage offers, because the favorable scenarios with high monetary offers weigh heavily
on expected utility. On the other hand, to analyze risk behavior following the elimination of
the largest prizes, accurate estimates for low monetary offers are also needed. It is therefore
comforting that the fit is good in terms of both percentages and monetary amounts. In
addition, if is used as a free parameter in our structural choice models, the optimal values
are approximately the same as our estimates, further confirming the goodness.
Since the principle behind the bank offers becomes clear after seeing a few shows, the
bank offer model (7) (8) is treated as deterministic and known to the contestants. Using a
stochastic bank offer model would introduce an extra layer of uncertainty, yielding lower
continuation values. For losers, the bank offers are hardest to predict, making it even more
difficult to rationalize why these contestants continue play.
Maximum Likelihood Estimation

In the spirit of Gordon M. Becker, Morris H. DeGroot and Jacob Marschak (1963) and
John D. Hey and Chris Orme (1994), we assume that the Deal or No Deal decision of a
given contestant i = 1, L , N in a given game round r = 1, L ,9 is based on the difference
between the continuation value and the stop value, or cv( xi , r ) sv( xi , r ) , plus some error. The
errors are treated as independent, normally distributed random variables with zero mean and
standard deviation i, r . Arguably, the error standard deviation should be higher for difficult
23
A. Percentage bank offers
Actual
150% R2 = 0.737
100%
50%
Predicted
0%
0% 50% 100% 150%
B. Monetary bank offers
Actual ()
1,500,000
R2 = 0.962
1,000,000
500,000
Predicted ()
0
0 500,000 1,000,000 1,500,000
Figure 3: Predicted Bank Offers versus Actual Bank Offers. The figure displays the goodness
of our bank offer model by plotting the predicted bank offers versus the actual bank offers for all
relevant game rounds in our pooled sample of 151 contestants from the Netherlands, Germany and
the United States. Panel A shows the fit for the percentage bank offers and panel B shows the fit
for the monetary bank offers (in Euros). A 45-degree line (perfect fit) is added for ease of
interpretation.
24
choices than for simple choices. A natural indicator of the difficulty of a decision is the
standard deviation of the utility of the outcomes used to compute the continuation value:
(9) ( xi ,r ) = (u( B( y)) cv( x

y ( xi , r )
i ,r )) 2 p r
We assume that the error standard deviation is proportional to this indicator, that is,
i ,r = ( xi ,r ) , where is a constant noise parameter. As a result of this assumption, the
simple choices effectively receive a larger weight in the analysis than the difficult ones. We
also investigated the data without weighting. The (unreported) results show that the overall fit
in the three samples deteriorates. In addition, without weighting, the estimated noise
parameters in the three editions strongly diverge, with the Dutch edition having a substantially
higher noise level than the German and US editions. The increase in the noise level seems to
reflect the higher difficulty of the decisions in the Dutch edition relative to the German and
US editions; contestants in the Dutch edition typically face (i) larger stakes because of the
large initial prizes and (ii) more remaining prizes because they exit the game at an earlier
stage. The standard deviation of the outcomes (9) picks up these two factors. The
deterioration of the fit and the divergence of the estimated noise levels provide additional,
empirical arguments for our weighting scheme.
Given these assumptions, we may compute the likelihood of the Deal or No Deal
decision as:
cv( xi ,r ) sv( xi ,r )
if No Deal

( xi ,r )
(10) l ( xi , r ) =
sv ( x ) cv ( x i ,r )
i ,r
if Deal
( x )
i ,r
where () is the cumulative standard normal distribution function.5
5
This error model allows for violations of first-order stochastic dominance (FSD). The probability of Deal is
predicted to be larger than zero and smaller than unity, even when the bank offer is smaller than the smallest
outcome (No Deal dominates Deal) or larger than the largest outcome (Deal dominates No Deal). As
pointed out by an anonymous referee, a truncated error model can avoid such violations of FSD. In our dataset,
however, the bank offer is always substantially larger than the smallest and substantially smaller than the largest
outcome, and violations of FSD do not occur.
25
Aggregating the likelihood across contestants, the overall log-likelihood function of the
Deal or No Deal decisions is given by:
N Ri
(11) ln( L) = ln(l ( xi , r ))
i =1 r = 2
where Ri is the last game round played by contestant i.

To allow for the possibility that the errors of individual contestants are correlated, we
perform a cluster correction on the standard errors (see, for example, Jeffrey M. Wooldridge,
2003). Note that the summation starts in the second game round (r = 2). The early German
episodes with only eight game rounds effectively start in this game round and in order to align
these episodes with the rest of the sample, we exclude the first round (r = 1) of the editions
with nine game rounds. Due to the very conservative bank offers, the choices in the first
round are always trivial (no contestant in our sample ever said Deal); including these
choices does not affect the results, but it would falsely make the early German episodes look
more noisy than the rest of the sample.
The unknown parameters in our model (, , W, and ) are selected to maximize the
overall log-likelihood. To determine if the model works significantly better than a nave
model of risk neutrality, we perform a likelihood ratio test.
Results
Table 6 summarizes our estimation results. Apart from coefficient estimates and p-
values, we have also computed the implied certainty equivalent as a fraction of the expected
value, or certainty coefficient (CC), for 50/50 gambles of 0 or 10z, z = 1,,6. These values
help to interpret the coefficient estimates by illustrating the shape of the utility function.
Notably, the CC can be interpreted as the critical bank offer (as a fraction of the expected
value of the 50/50 gamble) that would make the contestant indifferent between Deal and
No Deal. If CC = 1, the contestant is risk neutral. When CC > 1, the contestant is risk
seeking, and as CC approaches zero, the contestant becomes extremely risk averse. To help
interpret the goodness of the model, we have added the hit percentage, or the percentage of
correctly predicted Deal or No Deal decisions.
In the Dutch sample, the risk aversion parameters and are both significantly
different from zero, suggesting that IRRA and DARA are relevant and the classical CRRA
26
Table 6
Expected Utility Theory Results
The table displays the estimation results of expected utility theory for our sample of 151
contestants from The Netherlands (51), Germany (47) and the United States (53). Shown are
maximum likelihood estimators for the and parameters and the wealth level (W, in Euros) of
the utility function (3), and the noise parameter . The table also shows the overall mean log-
likelihood (MLL), the likelihood ratio (LR) relative to the nave model of risk neutrality, the
percentage of correctly predicted Deal or No Deal decisions (Hits), and the total number of
Deal or No Deal decisions in the sample (No.). Finally, the implied certainty coefficient (CC;
certainty equivalent as a fraction of the expected value) is shown for 50/50 gambles of 0 or 10z,
z = 1,,6. p-values are shown in parentheses.
Netherlands Germany United States

0.424 (0.000) 1.58e-5 (0.049) 4.18e-5 (0.000)
0.791 (0.000) 0.000 (1.000) 0.171 (0.000)
W 75,203 (0.034) 544 (0.481) 101,898 (0.782)
0.428 (0.000) 0.467 (0.000) 0.277 (0.000)
MLL -0.365 -0.340 -0.260
LR 24.29 (0.000) 3.95 (0.267) 15.10 (0.002)
Hits 76% 85% 89%
No. 214 327 349
CC (0/101) 1.000 1.000 1.000
CC (0/102) 0.999 1.000 1.000
CC (0/103) 0.994 0.996 0.998
CC (0/104) 0.946 0.960 0.984
CC (0/105) 0.637 0.640 0.859
CC (0/106) 0.141 0.088 0.302
power function and CARA exponential function are too restrictive to explain the choices in
this game show. The estimated wealth level of 75,203 significantly exceeds zero. Still, given
that the median Dutch household income is roughly 25,000 per annum, the initial wealth
level seems substantially lower than lifetime wealth and integration seems incomplete. This
deviates from the classical approach of defining utility over wealth and is more in line with
utility of income or the type of narrow framing that is typically assumed in prospect theory. A
low wealth estimate is also consistent with Matthew Rabins (2000) observation that plausible
risk aversion for small and medium outcomes implies implausibly strong risk aversion for
large outcomes if the outcomes are integrated with lifetime wealth. Indeed, the estimates
imply near risk neutrality for small stakes, witness the CC of 0.994 for a 50/50 gamble of 0
or 1,000, and increasing the wealth level would imply near risk neutrality for even larger
gambles.
Rabins point is reinforced by comparing our results for large stakes with the laboratory
experiments conducted by Holt and Laury (2002) using the lower stakes typical in the lab.
Holt and Laurys subjects display significant risk aversion for modest stakes, which, as Rabin
27
notes, implies extreme risk aversion for much larger stakes behavior our contestants do not
display. Indeed, contestants with Holt and Laurys parameter estimates for the utility function
would generally accept a Deal in the first game round, in contrast to the actual behavior we
observe. We conclude, agreeing with Rabin, that expected utility of wealth models have
difficulty explaining behavior for both small and large stakes.
The model also does not seem flexible enough to explain the choices for losers and
winners simultaneously. The estimated utility function exhibits very strong IRRA, leading to
an implausibly low CC of 0.141 for a 50/50 gamble of 0 or 1,000,000. Indeed, the model
errs by predicting that winners would stop earlier than they actually do. If risk aversion
increases with stakes, winners are predicted to have a stronger propensity to accept a bank
offer, the opposite of what we observe; witness for example the Deal percentages in Table
5. However, strong IRRA is needed in order to explain the behavior of losers, who reject
generous bank offers and continue play even with tens of thousands of Euros at stake. Still,
the model does not predict risk seeking at small stakes; witness the CC of 0.946 for a 50/50
gamble of 0 or 10,000 roughly Franks risky choice in round 9. Thus, the model also errs
by predicting that losers would stop earlier than they actually do.
Interestingly, the estimated coefficients for the German edition are quite different from
the Dutch values. The optimal utility function reduces to the CARA exponential function
( = 0) and the estimated initial wealth level becomes insignificantly different from zero. Still,
on the observed domain of prizes, the two utility functions exhibit a similar pattern of
unreasonably strong IRRA and high risk aversion for winners. Again, the model errs by
predicting that losers and winners would stop earlier than they actually do. These errors are so
substantial in this edition that the fit of the expected utility model is not significantly better
than the fit of a naive model that assumes that all contestants are risk neutral and simply
Deal whenever the bank offer exceeds the average remaining prize.
Contrary to the Dutch and German utility functions, the US utility function
approximates the limiting case of the CRRA power function ( 0). The CC is again very
high for small stakes. For larger stakes, the coefficient decreases but at a slower pace than in
the other two countries, reflecting the relatively low propensity to Deal for US contestants
with relatively large amounts at stake. The decreasing pattern stems from the estimated initial
wealth level of 101,898, which yields near risk neutrality for small stakes. Still, initial wealth
is not significantly different from zero, because a similar pattern can be obtained if we lower
the value of beta relative to alpha and move in the direction of the CARA exponential
function.
28
To further illustrate the effect of prior outcomes, Table 7 shows separate results for
losers and winners (as defined in Section III). Confirming the low Deal percentages found
earlier, the losers and winners are less risk averse and have higher CCs than the neutral group.
The losers are in fact best described by a model of risk seeking, which is not surprising given
that the losers in our sample often reject bank offers in excess of the average remaining prize.
The same pattern arises in each of the three editions, despite sizeable differences in the set of
prizes. For example, the Dutch losers on average face larger stakes than the contestants in the
US and German neutral groups. Still, risk seeking (CC > 1) arises only in the loser group.
Overall, these results suggest that the expected utility model fails to capture the strong effect
of previous outcomes.
V. Prospect Theory
In this section, we use prospect theory to analyze the observed Deal or No Deal
choices. Contestants are assumed to have a narrow focus and evaluate the outcomes in the
game without integrating their initial wealth a typical assumption in prospect theory.
Furthermore, we will again use the myopic frame that compares the current bank offer with
the unknown offer in the next round. Although myopia is commonly assumed in prospect
theory, the choice of the relevant frame in this game is actually more important than for
expected utility theory. As discussed in Section IV, the myopic frame seems appropriate for
expected utility theory. For prospect theory, however, it can be rather restrictive. Prospect
theory allows for risk-seeking behavior when in the domain of losses and risk seekers have a
strong incentive to look ahead multiple game rounds to allow for the possibility of winning
the largest remaining prize. Indeed, contestants who reject high bank offers often explicitly
state that they are playing for the largest remaining prize (rather than a large amount offered
by the banker offer in the next round). Preliminary computations revealed that prospect theory
generally performs better if we allow contestants to look ahead multiple game rounds. The
improvements are limited, however, because risk seeking typically arises at the end of the
game. At that stage, only a few or no further game rounds remain and the myopic model then
gives a good approximation. Thus, we report only the results with the myopic model in order
to be consistent with the previous analysis using expected utility theory.
The stop value and continuation value for prospect theory are defined in the same way
as for expected utility theory, with the only difference that the expo-power utility function (3)
is replaced by the prospect theory value function, which is defined on changes relative to
some reference point:
29
Table 7
Path Dependence
The table shows the maximum likelihood estimation results of expected utility theory for our
sample of 151 contestants from the Netherlands (51; panel A), Germany (47; panel B) and the
United States (53; panel C). The samples are split based on the fortune experienced during the
game. A contestant is classified as a loser (winner) if her average remaining prize after
eliminating the lowest (highest) remaining prize is among the worst (best) one-third for all
contestants in the same game round. The results are presented in a format similar to the full-sample
results in Table 6.
Loser Neutral Winner

A. Netherlands
-244.904 (0.022) 0.044 (0.204) 0.125 (0.831)
0.993 (0.000) 0.687 (0.000) 0.736 (0.011)
W 0 (1.000) 304 (0.671) 3,061 (0.824)
0.627 (0.000) 0.323 (0.000) 0.309 (0.000)
MLL -0.300 -0.383 -0.325
Hits 89% 81% 83%
No. 72 70 72
CC (0/101) 1.330 0.994 0.999
CC (0/102) 1.338 0.945 0.992
CC (0/103) 1.347 0.723 0.928
CC (0/104) 1.355 0.392 0.630
CC (0/105) 1.363 0.150 0.216
CC (0/106) 1.371 0.032 0.035
B. Germany
-7.914 (0.117) 0.364 (0.000) 0.087 (0.000)
0.814 (0.000) 0.759 (0.000) 0.651 (0.000)
W 930 (0.825) 50,926 (0.481) 113,582 (0.180)
0.659 (0.000) 0.241 (0.000) 0.454 (0.000)
MLL -0.276 -0.257 -0.278
Hits 90% 87% 88%
No. 111 105 111
CC (0/101) 1.012 1.000 1.000
CC (0/102) 1.113 0.999 0.999
CC (0/103) 1.584 0.990 0.995
CC (0/104) 1.823 0.911 0.949
CC (0/105) 1.891 0.485 0.614
CC (0/106) 1.929 0.072 0.101
C. United States
-203.512 (0.006) 1.96e-5 (0.000) 0.938 (0.000)
0.995 (0.000) 0.086 (0.000) 0.998 (0.000)
W 54 (0.691) 934,904 (0.331) 29,468 (0.107)
0.193 (0.000) 0.308 (0.000) 0.326 (0.000)
MLL -0.194 -0.275 -0.253
Hits 92% 86% 91%
No. 118 113 118
CC (0/101) 1.004 1.000 1.000
CC (0/102) 1.023 1.000 0.999
CC (0/103) 1.054 0.999 0.992
CC (0/104) 1.071 0.986 0.927
CC (0/105) 1.081 0.863 0.646
CC (0/106) 1.089 0.252 0.289
30
( RP x) x RP
(12) v( x | RP) =
( x RP) x > RP
where > 0 is the loss-aversion parameter, RP is the reference point that separates losses
from gains, and > 0 measures the curvature of the value function. The original formulation
of prospect theory allows for different curvature parameters for the domain of losses
( x RP ) and the domain of gains ( x > RP ). To reduce the number of free parameters, we
assume here that the curvature is equal for both domains.6
Reference Point Specification

Kahneman and Tverskys (1979) original treatment of prospect theory equates the
reference point with the status quo. Since Deal or No Deal contestants never have to pay
money out of their own pockets, the reference point would then equal zero and contestants
would never experience any losses. The authors recognize, however, that there are situations
in which gains and losses are coded relative to an expectation or aspiration level that differs
from the status quo (286). They point out that a person who has not made peace with his
losses is likely to accept gambles that would be unacceptable to him otherwise (287). This
point is elaborated by Thaler and Johnson (1990), though neither team offers a formal model
of how the reference point changes over time. One recent effort along these lines is by Botond
Kszegi and Rabin (2006, 2007).
The specification of the subjective reference point (or the underlying set of
expectations) and how it varies during the game is crucial for our analysis, as it determines
whether outcomes enter as gain or loss in the value function and with what magnitude. Slow
adjustment or stickiness of the reference point can yield break-even and house-money effects,
or a lower risk aversion after losses and after gains. If the reference point adjusts slowly after
losses, relatively many remaining outcomes are placed in the domain of losses, where risk
seeking applies. Similarly, if the reference point sticks to an earlier, less favorable value after
gains, relatively many remaining prizes are placed in the domain of gains, reducing the role of
loss aversion.
6
Empirical curvature estimates are often very similar for gains and losses. Tversky and Kahneman (1992), for
example, find a median value of 0.88 for both domains. Furthermore, the curvature needs to be the same for both
domains in order to be consistent with the definition of loss aversion; see Veronica Kbberling and Wakker
(2005).
31
v (x ) v W (x| 25,000)
v N (x |50,000)
v L (x |75,000)
25,000 75,000 x
Figure 4: Break-Even and House-Money Effects in Prospect Theory. The figure displays the
prospect value function (12) for three different levels of the reference point (RP) and the
associated certainty equivalents (CEs) for a 50/50 gamble of 25,000 or 75,000. Value function
vN(x|50,000) refers to a neutral situation with RPN = 50,000 and CEN = 44,169, vW(x|25,000) to a
winner with RPW = 25,000 and CEW = 47,745, and vL(x|75,000) to a loser with RPL = 75,000
and CEL = 52,255. All three value functions are based on the parameter estimates of Tversky and
Kahneman (1992), or = 0.88 and = 2.25. The crosses indicate the certainty equivalents for the
50/50 gamble.
Figure 4 illustrates these two effects using a 50/50 gamble of 25,000 or 75,000.
Contestants in Deal or No Deal face this type of gamble in round 9. The figure shows the
value function using the parameter estimates of Tversky and Kahneman (1992), or = 0.88
and = 2.25, and three alternative specifications for the reference point. In a neutral situation
without prior outcomes, the reference point may equal the expected value (RPN = 50,000). In
this case, the contestant frames the gamble as losing 25,000 (50,000 25,000) or winning
25,000 (75,000 50,000). The certainty equivalent of the gamble is CEN = 44,169,
meaning that bank offers below this level would be rejected and higher offers would be
accepted. The risk premium of 5,831 is caused by loss aversion, which assigns a larger
weight to losses than to gains.
Now consider contestant L, who initially faced much larger stakes than 50,000 and
incurred large losses before arriving at the 50/50 gamble in round 9. Suppose that L slowly
32
adjusts to these earlier losses and places his reference point at the largest remaining prize
(RPL = 75,000). In this case, L does not frame the gamble as losing 25,000 or winning
25,000 but rather as losing 50,000 (75,000 25,000) or breaking even (75,000
75,000). Both prizes are placed in the domain of losses where risk seeking applies. Indeed,
L would reject all bank offers below the certainty equivalent of the gamble, CEL = 52,255,
which implies a negative risk premium of 2,255.
Finally, consider contestant W, who initially faced much smaller stakes than 50,000
and incurred large gains before arriving at the 50/50 gamble. Due to slow adjustment, W
employs a reference point equal to the smallest remaining prize (RPW = 25,000) and places
both remaining prizes in the domain of gains. In this case, W frames the gamble as one of
either breaking even (25,000 25,000) or gaining 50,000 (75,000 25,000). Since loss
aversion does not apply in the domain of gains, the risk aversion of W is lower than in the
neutral case and W would reject all bank offers below CEW = 47,745, implying a risk
premium of 2,255, less than the value of 5,831 in the neutral case.
It should be clear from the examples above that a proper specification of the reference
point and its dynamics is essential for our analysis. In fact, without slow adjustment, prospect
theory does not yield the path-dependence found in this study. Unfortunately, the reference
point is not directly observable and prospect theory alone provides minimal guidance for
selecting the relevant specification. We therefore need to give the model some freedom and
rely on the data to inform us about the relevant specification. To reduce the risk of data
mining and to simplify the interpretation of the results, we develop a simple structural model
based on elementary assumptions and restrictions for the reference point.
If contestants were confronted with the isolated problem of choosing between the
current bank offer and the risky bank offer in the next round, it would seem natural to link the
reference point to the current bank offer. The bank offer represents the sure alternative and the
opportunity cost of the risky alternative. Furthermore, the bank offer is linked to the average
remaining prize and therefore to current expectations regarding future outcomes. A simple
specification would be RPr = 1 B(xr). If 1 = 0, then the reference point equals the status quo
(RPr = 0) and all possible outcomes are evaluated as gains; if 1 > 0, the reference point is
strictly positive and contestant may experience (paper) losses, even though they never have to
pay money out of their own pockets. A reference point below the current bank offer, or 1 < 1,
is conservative (pessimistic) in the sense that relatively few possible bank offers in the next
round are classified as losses and relatively many possible outcomes are classified as gains.
33
By contrast, an optimistic reference point, or 1 > 1, involves relatively many possible
losses and few possible gains.
The actual game is dynamic and the bank offer changes in every round, introducing the
need to update the reference point. Due to slow adjustment, however, the reference point may
be affected by earlier game situations. We may measure the effect of outcomes after earlier
round j, 0 j < r, by the relative increase in the average remaining prize, or
d r( j ) = ( x r x j ) / x r . For j = 0, d r( j ) measures the change relative to the initial average, or x 0 .
Ideally, our model would include this measure for all earlier game rounds (and possibly
also interaction terms). However, due to the strong correlation between the lagged terms and
the limited number of observations, we have to limit the number of free parameters. We
restrict ourselves to just two terms: d r( r 2 ) and d r( 0 ) . The term d r( r 2 ) is the longest fixed lag
that can be included for all observations (our analysis starts in the second round) and
measures recent changes; d r( 0 ) , or the longest variable lag, captures all changes relative to the
initial game situation. Adding these two lagged terms to the static model, our dynamic model
for the reference point is:
(13) RPr = ( 1 + 2 d r( r 2 ) + 3 d r( 0 ) ) B ( x r )
In this model, 2 < 0 or 3 < 0 implies that the reference points sticks to earlier values
and that it is higher than the neutral value 1B(xr) after decreases in the average remaining
prize and lower after increases.
It is not immediately clear how strong the adjustment would be, or if the adjustment
parameters would be constant, but it seems realistic to assume that the adjustment is always
sufficiently strong to ensure that the reference point is feasible in the next round, i.e., not
lower than the smallest possible bank offer and not higher than the largest possible bank offer.
We therefore truncate the reference point at the minimum and maximum bank offer, i.e.
min B ( y ) RPr max B ( y ) . This truncation improves the empirical fit of our model and
y ( x r ) y ( x r )
the robustness to the specification of the reference point and its dynamics.
Our complete prospect theory model involves five free parameters: loss aversion ,
curvature , and the three parameters of the reference point model 1, 2 and 3. We estimate
these parameters and the noise parameter with the same maximum likelihood procedure
used for the expected utility analysis. We also apply the same bank offer model.
34
Our analysis ignores subjective probability transformation and uses the true
probabilities as decision weights. The fit of prospect theory could improve if we allow for
probability transformation. If losers have a sticky reference point and treat all possible
outcomes as losses, they will overweight the probability of the smallest possible loss,
strengthening the risk seeking that stems from the convexity of the value function in the
domain of losses. For example, applying the Tversky and Kahneman (1992) weighting
function and parameter estimates to a gamble with two equally likely losses, the decision
weight of the smallest loss is 55 percent rather than 50 percent. Still, we prefer to focus on the
effect of the reference point in this study and we ignore probability weighting for the sake of
parsimony. This simplification is unlikely to be material, especially in the most important
later rounds, when the relevant probabilities are medium to large and the decision weights
would be relatively close to the actual probabilities (as illustrated by the 50/50 gamble).
Results
Table 8 summarizes our results. For the Dutch edition, the curvature and loss aversion
parameters are significantly different from unity. The curvature of the value function is
needed to explain why some contestants reject bank offers in excess of the average remaining
prize; loss aversion explains why the average contestant accepts a bank offer below the
average prize. Both parameters take values that are comparable with the typical results in
experimental studies. Indeed, setting these parameters equal to the Tversky and Kahneman
(1992) parameter values does not change our conclusions.
The parameter 1 is significantly larger than zero, implying that contestants do
experience (paper) losses, consistent with the idea that the reference point is based on
expectations and that diminished expectations represent losses. The parameter is also
significantly smaller than unity, indicating that the reference point generally takes a
conservative value below the current bank offer.
The adjustment parameters 2 and 3 are significantly smaller than zero, meaning that
the reference point tends to stick to earlier values and is higher than the neutral value after
losses and lower after gains. In magnitude, 2 is much larger than 3, suggesting that the effect
of recent outcomes is much stronger than the effect of initial expectations. However, the
changes in the average remaining prize during the last two game rounds are generally much
smaller than the changes during the entire game, limiting the effect of the parameter value. In
35
Table 8
Prospect Theory Results
The table shows the estimation results of prospect theory for our sample of 151 contestants from
The Netherlands (51), Germany (47) and the United States (53). Shown are maximum likelihood
estimators for the loss aversion () and curvature () of the value function, the three parameters of
the reference point model 1, 2 and 3, and the noise parameter . The table also shows the
overall mean log-likelihood (MLL), the likelihood ratio (LR) relative to the nave model of risk
neutrality, the percentage of correctly predicted Deal or No Deal decisions (Hits), and the total
number of Deal or No Deal decisions in the sample (No.). p-values are shown in parentheses.
Netherlands Germany United States

2.375 (0.013) 4.501 (0.008) 4.528 (0.001)
0.516 (0.000) 0.486 (0.000) 0.836 (0.000)
1 0.474 (0.000) 1.096 (0.000) 1.163 (0.000)
2 -0.285 (0.000) -0.026 (0.000) 0.031 (0.329)
3 -0.028 (0.000) -0.052 (0.000) -0.093 (0.023)
0.345 (0.000) 0.533 (0.000) 0.193 (0.000)
MLL -0.309 -0.303 -0.228
LR 48.41 (0.000) 27.44 (0.000) 37.28 (0.000)
Hits 85% 89% 91%
No. 214 327 349
addition, in case of large changes, the reference point often falls outside the range of feasible
outcomes. In these cases, the reference point is set equal to the smallest or largest possible
bank offer (see above), further limiting the effect of the parameter value.
The slow adjustment of the reference point lowers the propensity of losers and winners
to Deal. Not surprisingly, the prospect theory model yields substantially smaller errors for
losers and winners and the overall log-likelihood is significantly higher than for the expected
utility model. While the expected utility model correctly predicted 76 percent of the Deal or
No Deal decisions, the hit percentage of the prospect theory model is 85 percent.
The results for the German and US samples are somewhat different from the results for
the Dutch sample, but still confirm the important role of slow adjustment. The difference
seems related to the relatively large stakes and the associated high propensity to Deal in the
Dutch edition. In the German and US samples, the reference point is substantially higher in
relative terms than in the Dutch sample. The relatively high reference point helps explain why
the German and US contestants stop in later rounds and demand higher percentage bank
offers than the Dutch contestants. Relatively many outcomes are placed in the domain of
losses, where risk seeking applies. In such a situation, a relatively strong loss aversion is
needed to explain Deals. Indeed, the loss aversion estimates are substantially higher than for
the Dutch sample. Again, stickiness is highly significant. However, the most recent outcomes
seem less important and the reference point now sticks primarily to the initial situation. This
36
seems related to the German and US contestants on average playing more game rounds than
the Dutch contestants. In later rounds, many briefcases have already been opened, but
relatively few briefcases have been opened in the last few rounds. The last two game rounds
played in the German and US edition therefore generally reveal less information than in the
Dutch edition. The model again materially reduces the errors for losers and winners and fits
the data significantly better than the expected utility model in these two samples.
These results are consistent with our earlier finding that the losers and winners have a
low propensity to Deal (see Table 5). Clearly, prospect theory with a dynamic but sticky
reference point is a plausible explanation for this path-dependent pattern. Still, we stress that
our analysis of prospect theory serves merely to explore and illustrate one possible
explanation, and that it leaves several questions unanswered. For example, we have assumed
homogeneous preferences and no subjective probability transformation. The empirical fit may
improve even further if we would allow for heterogeneous preferences and probability
weighting. Further improvements may come from allowing for a different curvature in the
domains of losses and gains, from allowing for different partial adjustment after gains and
losses, and from stakes-dependent curvature and loss aversion. We leave these issues for
further research.
VI. Experiments
The previous sections have demonstrated the strong effect of prior outcomes or path-
dependence of risk attitudes. Also, the amounts at stake seem to be important, with a stronger
propensity to deal for larger stakes levels. Prior outcomes and stakes are, however, highly
confounded within every edition of the game show: unfavorable outcomes (opening high-
value briefcases) lower the stakes and favorable outcomes (opening low-value briefcases)
raise the stakes. The stronger the effect of stakes, the easier it is to explain the weak
propensity to Deal of losers, but the more difficult it is to explain the low Deal percentage
of winners. To analyze the isolated effect of the amounts at stake, we conduct a series of
classroom experiments in which students at Erasmus University play Deal or No Deal. We
consider two variations to the same experiment that use monetary amounts that differ by a
factor of ten, but draw from the same student population.
Both experiments use real monetary payoffs to avoid incentive problems (see, for
example, Holt and Laury, 2002). In order to compare the choices in the experiments with
those in the original TV show and to provide a common basis for comparisons between the
37
two experiments, each experiment uses the original scenarios from the Dutch edition.7 At the
time of the experiments, only the first 40 episodes are available. The original monetary
amounts are scaled down by a factor of 1,000 or 10,000, with the smallest amounts rounded
up to one cent. Despite the strong scaling, the resulting stakes are still unusually high for
experimental research. Although the scenarios are predetermined, the subjects are not
deceived in the sense that the game is not manipulated to encourage or avoid particular
situations or behaviors. Rather, the subjects are randomly assigned to a scenario generated by
chance at an earlier point in time (in the original episode). The risk that the students would
recognize the original episodes seems small, because the scenarios are not easy to remember
and the original episodes are broadcast at least six months earlier. Indeed, the experimental
Deal or No Deal decisions are statistically unrelated to which of the remaining prizes is in
the contestants own briefcase.
We replicate the original game show as closely as possible in a classroom, using a game
show host (a popular lecturer at Erasmus University) and live audience (the student subjects
and our research team). Video cameras are pointed at the contestant, recording all her actions.
The game situation (unopened briefcases, remaining prizes and bank offers) is displayed on a
computer monitor in front of the stage (for the host and the contestant) and projected on a
large screen in front of the classroom (for the audience). This setup is intended to create the
type of distress that contestants must experience in the TV studio. Our approach seems
effective, because the audience is very excited and enthusiastic during the experiment,
applauding and shouting hints, and most contestants show clear symptoms of distress.
All our subjects are students, about 20 years of age. A total of 160 business or
economics students are randomly selected from a larger population of students at Erasmus
University who applied to participate in experiments during the academic year 2005 2006.
Although each experiment requires only 40 subjects, 80 students are invited to guarantee a
large audience and to ensure that a sufficient number of subjects are available in the event that
some subjects do not show up. Thus, approximately half of the students are selected to play
the game. To control for a possible gender effect, we ensure that the gender of the subjects
matches the gender of the contestants in the original episodes.
At the beginning of both experiments we hand out the instructions to each subject,
consisting of the original instructions to contestants in the TV show plus a cover sheet
7
Original prizes and offers are not available when a subject continues play after a Deal in the TV episode. The
missing outcomes for the prizes are selected randomly (but held constant across the experiments), and the bank
offers are set according to the pattern observed in the original episodes.
38
explaining our experiment. Next, the games start. Each individual game lasts about 5 to 10
minutes, and each experiment (40 games) lasts roughly 5 hours, equally divided in an
afternoon session with one half of the subjects and games, and an evening session with the
other half.
Small-Stake Experiment
In the first experiment, the original prizes and bank offers from the Dutch edition are
divided by 10,000, resulting in an average prize of roughly 40 and a maximum prize of
500.
The overall level of risk aversion in this experiment is lower than in the original TV
show. Contestants on average stop later (round 6.9 versus 5.2 for the TV show) and reject
higher percentage bank offers. Still, the changes seem modest given that the initial stakes are
10,000 times smaller than in the TV show. In the TV show, contestants generally become risk
neutral or risk seeking when only thousands or tens of thousands of Euros remain at stake.
In the experiment, the stakes are much smaller, but the average contestant is clearly risk
averse. This suggests that the effect of stakes on risk attitudes in this game is relatively weak.
By contrast, the effect of prior outcomes is very strong; witness for example the (untabulated)
Deal percentages (for round 2 9 combined) of 3, 21 and 19 for loser, neutral and
winner, respectively.
The first column of Table 9 shows the maximum likelihood estimation results. The
estimated utility function exhibits the same pattern of extreme IRRA as for the original
shows, but now at a much smaller scale. See, for example, the CC of 0.072 for a 50/50
gamble of 0 or 1,000. It follows from Rabins (2000) observation that plausible levels of
risk aversion require much lower initial wealth levels for small-stake gambles than for large-
stake gambles. Indeed, initial wealth is estimated to be 11 in this experiment, roughly a
factor of 10,000 lower than for the original TV sample. As for the original episodes, the
model errs by predicting that the losers and winners would stop earlier than they actually do.
Prospect theory with a sticky reference point fits the data substantially better than the
expected utility model, both in terms of the log-likelihood and in terms of the hit percentage.
Large-Stake Experiment
The modest change in the choices in the first experiment relative to the large-stake TV
show suggests that the effect of stakes is limited in this game. Of course, the classroom
39
Table 9
Experimental Results
The table shows the maximum likelihood estimation results for our choice experiments. The first
column (Small stakes) displays the results for the experiment with the original monetary amounts
in the Dutch TV format of Deal or No Deal divided by 10,000, the second column (Large stakes)
displays the results for the experiment with prizes scaled down by a factor of 1,000, and the third
column (Pooled) displays the results for the two samples combined. Panel A shows the results for
expected utility theory. Panel B shows the results for prospect theory. The results are presented in
the same format as the results in Table 6 and Table 8, respectively.
Small stakes Large stakes Pooled

A. Expected utility theory
0.019 (0.000) 0.002 (0.001) 0.002 (0.001)
0.000 (1.000) 0.000 (1.000) 0.000 (1.000)
W 11 (0.920) 50 (0.930) 0 (1.000)
0.306 (0.000) 0.294 (0.000) 0.354 (0.000)
MLL -0.342 -0.337 -0.351
LR 10.17 (0.017) 10.14 (0.017) 9.37 (0.025)
Hits 81% 83% 80%
No. 231 234 465
CC (0/101) 0.953 0.995 0.995
CC (0/102) 0.583 0.953 0.953
CC (0/103) 0.072 0.588 0.586
CC (0/104) 0.007 0.074 0.074
CC (0/105) 0.001 0.007 0.007
CC (0/106) 0.000 0.001 0.001
B. Prospect theory
2.307 (0.000) 2.678 (0.000) 2.518 (0.000)
0.732 (0.000) 0.695 (0.000) 0.693 (0.000)
1 1.045 (0.000) 1.024 (0.000) 1.023 (0.000)
2 -0.119 (0.000) 0.019 (0.000) 0.013 (0.250)
3 -0.086 (0.000) -0.046 (0.000) -0.049 (0.000)
0.267 (0.000) 0.196 (0.000) 0.218 (0.000)
MLL -0.275 -0.265 -0.272
LR 40.94 (0.000) 44.04 (0.000) 83.29 (0.000)
Hits 87% 88% 87%
No. 231 234 465
experiment is not directly comparable with the TV version, because, for example, the
experiment is not broadcast on TV and uses a different type of contestant (students). Our
second experiment therefore investigates the effect of stakes by replicating the first
experiment with larger stakes.
The experiment uses the same design as before, with the only difference being that the
original monetary amounts are divided by 1,000 rather than by 10,000, resulting in an average
prize of roughly 400 and a maximum prize of 5,000 extraordinarily large amounts for
experiments. For this experiment, 80 new subjects were drawn from the same population,
excluding students involved in the first experiment.
40
Based on the strong IRRA in the first experiment, the expected utility model would
predict a much higher risk aversion in this experiment. However, the average stop round is
exactly equal to the average for the small-stake experiment (round 6.9), and subjects reject
similar percentage bank offers (the highest rejected bank offer averages 82.5 percent versus
82.4 percent for the small-stake experiment). Therefore, the isolated effect of stakes seems
much weaker than suggested by the estimated IRRA in the individual experiments.
The second column of Table 9 displays the maximum likelihood estimation results.
With increased stakes but similar choices, the expected utility model needs a different utility
function to rationalize the choices. In fact, the estimated utility function seems scaled in
proportion to the stakes, so that the 50/50 gamble of 0 or 1,000 now involves
approximately the same CC as the 50/50 gamble of 0 or 100 in the small-stake experiment.
By contrast, for prospect theory, the estimated parameters are roughly the same as for the
small-stake version and a substantially better fit is achieved relative to the implementation of
expected utility theory.
In both experiments, risk aversion is strongly affected by prior outcomes, which are
strongly related to the level of stakes within the experiments, but the stakes do not materially
affect risk aversion across the experiments. Since the stakes are increased by a factor of ten
and all other conditions are held constant, the only plausible explanation seems that prior
outcomes rather than stakes are the main driver of risk aversion in this game.
Pooled Sample
The last column of Table 9 shows the results for the pooled sample of the two
experiments. As noted above, the choice behavior in the two samples is very similar, despite
the large differences in the stakes. The important role of the stakes in the individual samples
and the weak role across the two samples lead to two very different utility functions. Stakes
appear to matter more in relative terms than in absolute terms. Combining both samples will
cause problems for the expected utility model, since the model assigns an important role to
the absolute level of stakes. Using a single utility function for the pooled sample indeed
significantly worsens the fit relative to the individual samples. The prospect theory model
does not suffer from this problem because it attributes the low Deal propensity of losers and
winners in each sample to the slow adjustment of a reference point that is proportional to the
stakes in each sample. In this way, the model relies on changes in the relative level of the
stakes rather than the absolute level of the stakes. Whether outcomes are gains or losses
depends on the context. An amount of 100 is likely to be placed in the domain of gains in the
41
small-stake experiment (where the average prize is roughly 40), but the same amount is
probably placed in the domain of losses in the large-stake experiment (with an average prize
of roughly 400).
VII. Conclusions
The behavior of contestants in game shows cannot always be generalized to what an
ordinary person does in her everyday life when making risky decisions. While the contestants
have to make decisions in just a few minutes in front of millions of viewers, many real-life
decisions involving large sums of money are neither made in a hurry nor in the limelight.
Still, we believe that the choices in this particular game show are worthy of study, because the
decision problems are simple and well-defined, and the amounts at stake are very large.
Furthermore, prior to the show, contestants have had considerable time to think about what
they might do in various situations, and during the show they are encouraged to discuss those
contingencies with a friend or relative who sits in the audience. In this sense, the choices may
be more deliberate and considered than might appear at first glance. Indeed, it seems plausible
that our contestants have given more thought to their choices on the show than to some of the
other financial choices they have made in their lives such as selecting a mortgage or
retirement savings investment strategy.
What does our analysis tell us? First, we observe, on average, what might be called
moderate levels of risk aversion. Even when hundreds of thousands of Euros are at stake,
many contestants are rejecting offers in excess of 75 percent of the expected value. In an
expected utility of wealth framework, this level of risk aversion for large stakes is hard to
reconcile with the same moderate level of risk aversion found in small-stake experiments
both ours, and those conducted by other experimentalists. Second, although risk aversion is
moderate on average, the offers people accept vary greatly among the contestants; some
demonstrate strong risk aversion by stopping in the early game rounds and accepting
relatively conservative bank offers, while others exhibit clear risk-seeking behavior by
rejecting offers above the average remaining prize and thus deliberately entering unfair
gambles. While some of this variation is undoubtedly due to differences in individual risk
attitudes, a considerable part of the variation can be explained by the outcomes experienced
by the contestants in the previous rounds of the game. Most notably, risk aversion generally
decreases after prior expectations have been shattered by eliminating high-value briefcases or
after earlier expectations have been surpassed by opening low-value briefcases. This path-
dependent pattern occurs in all three editions of the game, despite sizeable differences in the
42
initial stakes across the editions. Losers and winners generally have a weaker propensity
to Deal than their neutral counterparts.
The relatively low risk aversion of losers and winners is hard to explain with expected
utility theory and points in the direction of reference-dependent choice theories such as
prospect theory. Indeed, our findings seem consistent with the break-even effect (losers
becoming more willing to take risk due to incomplete adaptation to prior losses), and the
house-money effect (a low risk aversion for winners due to incomplete adaptation to prior
gains). A simple version of prospect theory with a sticky reference point explains the Deal or
No Deal decisions substantially better than expected utility theory. These findings suggest
that reference-dependence and path-dependence are important, even when the decision
problems are simple and well-defined, and when large real monetary amounts are at stake.
Of course, we must be careful with rejecting expected utility theory and embracing
prospect theory. We use the flexible expo-power utility function, which embeds the most
popular implementations of expected utility theory, and find that this function is unable to
provide an explanation for the choices of losers and winners in this game show. However, a
(nonstandard) utility function that has risk seeking segments and depends on prior outcomes
could achieve a better fit. Such exotic specifications blur the boundary between the two
theories, and we therefore do not reject or accept one of the two.
Our main finding is the important role of reference-dependence and path-dependence,
phenomena that are often ignored in implementations of expected utility theory. Previous
choice problems are a key determinant of the framing of a given choice problem. An amount
is likely to be considered as large in the context of a game where it lies above prior
expectations, but the same amount is probably evaluated as small in a game where it lies
below prior expectations. For contestants who expected to win hundreds of thousands, an
amount of 10,000 probably seems small; the same amount is likely to appear much
larger when thousands or tens of thousands were expected.
To isolate the effect of the amounts at stake, we conducted two series of choice
experiments that use a homogeneous student population and mimic the TV show as closely as
possible in a classroom. We find that a tenfold increase of the initial stakes does not
materially affect the choices. Moreover, the choices in the experiments are remarkably similar
to those in the original TV show, despite the fact that the experimental stakes are only a small
fraction of the original stakes. Consistent with the TV version, the break-even effect and the
house-money effect also emerge in the experiments. These experimental findings reinforce
our conclusion that choices are strongly affected by previous outcomes. The combination of
43
(i) a strong effect of variation in stakes caused by a subjects fortune within a game and (ii) a
weak effect of variation in the initial stakes across games calls for a choice model that
properly accounts for the context of the choice problem and its dynamics.
This study has focused on episodes from the Netherlands, Germany, and the United
States, because these episodes have a very similar game format. For further research, it would
be interesting to collect more international data in order to obtain more degrees of freedom to
analyze the effect of prior outcomes in greater detail and to examine the role of the cultural,
social or economic background of the contestant. It would also be interesting to further extend
our choice experiments. While the stakes are much smaller, experiments do allow the
researcher to control contestant characteristics, rules and situations, and to more closely
monitor contestants and their behavior. Our experiments were designed to mimic the TV
studio and used real monetary payoffs, but further experiments may also take place in the
behavioral laboratory and employ some sort of random-lottery incentive system to reduce the
costs.
Epilogue
Following the success of Deal or No Deal in the Netherlands, the game show was
sold to dozens of countries worldwide. Other research groups have investigated episodes of
editions other than those used in this study. Their analyses employ not only different datasets,
but also different research methodologies and different (implementations of) decision
theories, and the results sometimes seem contradictory. Reconciling the seemingly disparate
results will be a valuable exercise, but is beyond the scope of this study. We will limit
ourselves at this point to a synopsis of the available studies, which are presented below in
alphabetical order, and some concluding remarks.
Using the UK edition, Steffen Andersen et al. (2006a) estimate various structural choice
models, assuming a homoskedastic error structure and accounting for forward-looking
behavior. Their expected utility estimates suggest CRRA and initial wealth roughly equal to
average annual UK income; their rank-dependent expected utility estimates indicate modest
probability weighting along with a concave utility function; their prospect theory estimates
indicate no loss aversion and modest probability weighting for gains, using several plausible
specifications of the reference point. Andersen et al. (2006b) study the UK television shows
and related lab experiments using a mixture model in which decision makers use two criteria:
one is essentially rank-dependent expected utility, and the other is essentially a probabilistic
44
income threshold. They find evidence that both criteria are used in the game show and that lab
subjects place a much greater weight on the income threshold.
Guido Baltussen, Thierry Post and Martijn J. van den Assem (2007) compare various
editions of DOND. Their sample includes editions from the same country that employ very
different initial sets of prizes. Comparing editions from the same country can separate the
effect of current stakes and prior outcomes without introducing cross-country effects, in the
same way as changing the initial stakes in our experiments. Consistent with reference-
dependence and path-dependence, they find that contestants in large- and small-stake editions
respond in a similar way to the stakes relative to their initial level, even though the initial
stakes are widely different across the various editions.
Pavlo Blavatskyy and Ganna Pogrebna (2007a) show that Italian and UK contestants do
not exhibit lower risk aversion when the probability of a large prize is small, and they
interpret this as evidence against the overweighting of small probabilities. Blavatskyy and
Pogrebna (2007b) find that the fit and relative performance of alternative decision theories
depends heavily on the assumed error structure in the Italian and UK datasets. Pogrebna
(2008) finds that Italian contestants generally do not follow nave advice from the audience.
Blavatskyy and Pogrebna (2006a) analyze the UK, French and Italian editions, which
sometimes include a swap option that allows contestants to exchange their briefcase for
another unopened briefcase. Blavatskyy and Pogrebna (2006b) conduct a nonparametric test
of ten popular decision theories using the UK and Italian edition.
Matilde Bombardini and Francesco Trebbi (2007) use the Italian edition to estimate a
structural dynamic CRRA expected utility model and find that the risk aversion is moderate
on average and shows substantial variation across individual contestants. They also find that
contestants are practically risk neutral when faced with small stakes and risk averse when
faced with large stakes. Accounting for strategic interaction between the banker and the
contestant (the Italian banker knows the contents of the unopened briefcases) does not change
their conclusions.
Fabrizio Botti et al. (2007) estimate various structural expected utility models for the
Italian edition, assuming that contestants ignore subsequent bank offers and compare the
current bank offer with the set of remaining prizes. They find that the CARA specification fits
the data significantly better than the CRRA and expo-power specifications, and they also
report a gender effect (males are more risk averse) and substantial unobserved heterogeneity
in risk aversion.
45
Cary A. Deck, Jungmin Lee, and Javier A. Reyes (2008) estimate structural CRRA and
CARA expected utility models for Mexican episodes of Deal or No Deal. They consider
both forward-looking contestants and myopic contestants who look forward only one game
round, and they vary the level of forecasting sophistication by the contestants. They find a
moderate level of average risk aversion and considerable individual variation in risk attitudes,
with some contestants being extremely risk averse while others are risk seeking.
Using the Australian edition, Nicolas de Roos and Yianis Sarafidis (2006) estimate
structural dynamic CARA and CRRA expected utility models using random effects and
random coefficients models. Their models produce plausible estimates of risk aversion, and
suggest substantial heterogeneity in decision making, both between contestants and between
decisions made by the same contestant. They also find that rank-dependent expected utility
substantially improves the explanatory power. In addition to these main-game results, they
also investigate contestants choices in special Chance and Supercase game rounds,
which are specific for the Australian edition. Risk attitudes elicited in these additional game
rounds seem to be similar to risk attitudes elicited in the main game. Also using Australian
data, Daniel Mulino et al. (2006) estimate a structural dynamic CRRA expected utility model.
Their estimates reveal moderate risk aversion on average and considerable variation across
contestants. They also find that risk aversion depends on contestant characteristics such as age
and gender, but not on wealth. Like De Roos and Sarafidis, they investigate the choices in the
Chance and Supercase rounds, but they do find a difference in risk attitudes between
these special rounds and the main game.
Clearly, Deal or No Deal can be studied for several research purposes and with a
variety of methodologies and theories, and different studies can lead to different, sometimes
opposing conclusions. Some final remarks may be useful to evaluate the existing studies and
to guide further research. First, to analyze risk attitudes without the confounding effect of
ambiguity and strategic insight, it is useful to analyze the basic version of the game. Of
course, the more exotic versions with special game options and informative bank offers are
interesting for other purposes, as demonstrated in some of the above studies. Second, to
disentangle the effect of the amounts at stake and the effect of previous outcomes, it is useful
to analyze multiple game show editions or choice experiments with different initial amounts
at stake. Within one edition or experiment, current stakes and prior outcomes are perfectly
correlated, and the two effects cannot be separated. Third, when using parametric structural
models, it seems important to analyze the robustness for the assumed mental frame and error
structure. For example, we found a relatively poor fit for models that assume that contestants
46
focus on the set of remaining prizes rather than the next rounds bank offer, and also for
models that assume that the error variance is equal for all choice problems, irrespective of the
level of the stakes or the variation in the prizes.
References
Abdellaoui, Mohammed, Carolina Barrios, and Peter P. Wakker. 2007. Reconciling Introspective Utility
with Revealed Preference: Experimental Arguments Based on Prospect Theory. Journal of Econometrics,
138(1): 356378.
Allais, Maurice. 1953. Le Comportement de lHomme Rationnel devant le Risque: Critique des Postulats et
Axiomes de lcole Amricaine. Econometrica, 21(4): 503546.
Andersen, Steffen, Glenn W. Harrison, Morten I. Lau, and E. Elisabet Rutstrm. 2006a. Dynamic Choice
Behavior in a Natural Experiment. Working Paper 06-10, Department of Economics, College of Business
Administration, University of Central Florida.
Andersen, Steffen, Glenn W. Harrison, Morten I. Lau, and E. Elisabet Rutstrm. 2006b. Dual Criteria
Decisions. Working Paper 06-11, Department of Economics, College of Business Administration,
University of Central Florida.
Baltussen, Guido, Thierry Post, and Martijn J. van den Assem. 2007. Risky Choice and the Relative Size of
Stakes. http://ssrn.com/abstract=989242.
Becker, Gordon M., Morris H. DeGroot, and Jacob Marschak. 1963. Stochastic Models of Choice
Behavior. Behavioral Science, 8(1): 4155.
Beetsma, Roel M. W. J., and Peter C. Schotman. 2001. Measuring Risk Attitudes in a Natural Experiment:
Data from the Television Game Show Lingo. Economic Journal, 111(474): 821848.
Binmore, Ken, John McCarthy, Giovanni Ponti, Larry Samuelson, and Avner Shaked. 2002. A Backward
Induction Experiment. Journal of Economic Theory 104(1): 4888.
Binswanger, Hans P. 1980. Attitudes toward Risk: Experimental Measurement in Rural India. American
Journal of Agricultural Economics, 62(3): 395407.
Binswanger, Hans P. 1981, Attitudes toward Risk: Theoretical Implications of an Experiment in Rural India.
Economic Journal, 91(364): 867890.
Blavatskyy, Pavlo, and Ganna Pogrebna. 2006a. Loss Aversion? Not with Half-a-Million on the Table!
Institute for Empirical Research in Economics (University of Zurich) Working Paper 274.
Blavatskyy, Pavlo, and Ganna Pogrebna. 2006b. Testing the Predictions of Decision Theories in a Natural
Experiment When Half a Million Is at Stake. Institute for Empirical Research in Economics (University of
Zurich) Working Paper 291.
Blavatskyy, Pavlo, and Ganna Pogrebna. 2007a. Forthcoming. Risk Aversion When Gains Are Likely and
Unlikely: Evidence from a Natural Experiment with Large Stakes. Theory and Decision.
Blavatskyy, Pavlo, and Ganna Pogrebna. 2007b. Models of Stochastic Choice and Decision Theories: Why
Both are Important for Analyzing Decisions. Institute for Empirical Research in Economics (University of
Zurich) Working Paper 319.
47
Bombardini, Matilde, and Francesco Trebbi. 2007. Risk Aversion and Expected Utility Theory: An
Experiment with Large and Small Stakes. http://faculty.chicagogsb.edu/francesco.trebbi.
Botti, Fabrizio, Anna Conte, Daniela Di Cagno and Carlo DIppoliti. 2007. Risk Aversion, Demographics
and Unobserved Heterogeneity. Evidence from the Italian TV Show Affari Tuoi. In Games, Rationality
and Behaviour; Essays on Behavioural Game Theory and Experiments, ed. Alessandro Innocenti and Patrizia
Sbriglia. Basingstoke: Palgrave Macmillan.
De Roos, Nicolas, and Yianis Sarafidis. 2006. Decision Making under Risk in Deal or No Deal.
http://ssrn.com/abstract=881129.
Deck, Cary A., Jungmin Lee, and Javier A. Reyes. 2008. Risk Attitudes in Large Stake Gambles: Evidence
from a Game Show. Applied Economics, 40(1): 4152.
Ellsberg, Daniel. 1961. Risk, Ambiguity, and the Savage Axioms. Quarterly Journal of Economics, 75(4):
643669.
Friedman, Milton, and Leonard J. Savage. 1948. The Utility Analysis of Choices Involving Risk. Journal
of Political Economy, 56(4): 279304.
Fullenkamp, Connel R., Rafael A. Tenorio, and Robert H. Battalio. 2003. Assessing Individual Risk
Attitudes Using Field Data from Lottery Games. Review of Economics and Statistics, 85(1): 218226.
Gertner, Robert H. 1993. Game Shows and Economic Behavior: Risk-Taking on Card Sharks. Quarterly
Journal of Economics, 108(2): 507521.
Hartley, Roger, Gauthier Lanot, and Ian Walker. 2006. Who Really Wants to be a Millionaire? Estimates of
Risk Aversion from Gameshow Data. Keele Economics Research Papers 2006/07.
Hersch, Philip L., and Gerald S. McDougall. 1997. Decision Making under Uncertainty When the Stakes Are
High: Evidence from a Lottery Game Show. Southern Economic Journal, 64(1): 7584.
Hey, John D., and Chris Orme. 1994. Investigating Generalizations of Expected Utility Theory Using
Experimental Data. Econometrica, 62(6): 12911326.
Holt, Charles A., and Susan K. Laury. 2002. Risk Aversion and Incentive Effects. American Economic
Review, 92(5): 16441655.
Johnson, Eric J., Colin Camerer, Sankar Sen, and Talia Rymon. 2002. Detecting Failures of Backward
Induction: Monitoring Information Search in Sequential Bargaining. Journal of Economic Theory, 104(1):
1647.
Kachelmeier, Steven J., and Mohamed Shehata. 1992. Examining Risk Preferences under High Monetary
Incentives: Experimental Evidence from the Peoples Republic of China. American Economic Review,
82(5): 11201141.
Kahneman, Daniel, and Amos Tversky. 1979. Prospect Theory: An Analysis of Decision under Risk.
Econometrica, 47(2): 263291.
Kbberling, Veronika, and Peter P. Wakker. 2005. An Index of Loss Aversion. Journal of Economic
Theory, 122(1): 119131.
Kszegi, Botond, and Matthew Rabin. 2006. A Model of Reference-Dependent Preferences. Quarterly
Journal of Economics, 121(4): 11331165.
Kszegi, Botond, and Matthew Rabin. 2007. Reference-Dependent Risk Attitudes. American Economic
Review, 97(4): 10471073.
48
Markowitz, Harry. 1952. The Utility of Wealth. Journal of Political Economy, 60(2): 151158.
Metrick, Andrew. 1995. A Natural Experiment in Jeopardy! American Economic Review, 85(1): 240253.
Mulino, Daniel, Richard Scheelings, Robert D. Brooks, and Robert W. Faff. 2006. An Empirical
Investigation of Risk Aversion and Framing Effects in the Australian Version of Deal or No Deal.
http://ssrn.com/abstract=926040.
Pogrebna, Ganna. 2008. Nave Advice When Half a Million Is at Stake. Economics Letters, 98(2): 148154.
Rabin, Matthew. 2000. Risk Aversion and Expected-Utility Theory: A Calibration Theorem. Econometrica,
68(5): 12811292.
Saha, Atanu. 1993. Expo-Power Utility: A Flexible Form for Absolute and Relative Risk Aversion.
American Journal of Agricultural Economics, 75(4): 905913.
Thaler, Richard H., and Eric J. Johnson. 1990. Gambling with the House Money and Trying to Break Even:
The Effects of Prior Outcomes on Risky Choice. Management Science, 36(6): 643660.
Tversky, Amos, and Daniel Kahneman. 1992. Advances in Prospect Theory: Cumulative Representation of
Uncertainty. Journal of Risk and Uncertainty, 5(4): 297323.
Von Neumann, John, and Oskar Morgenstern. 1944. Theory of Games and Economic Behavior. Princeton,
NJ: Princeton University Press.
Wooldridge, Jeffrey M. 2003. Cluster-Sample Methods in Applied Econometrics. American Economic
Review, 93(2): 133138.
49
* Post and Van den Assem hold positions at the Department of Business Economics, Erasmus
School of Economics, Erasmus University of Rotterdam. Post is Professor of Finance. Van
den Assem is Assistant Professor of Finance. E-mail: gtpost@few.eur.nl,
vandenassem@few.eur.nl. Baltussen is PhD candidate at the Tinbergen Institute and the
Erasmus School of Economics. E-mail: baltussen@few.eur.nl. Postal address: Erasmus
University of Rotterdam, P.O. Box 1738, 3000 DR, Rotterdam, the Netherlands. Thaler is
Ralph and Dorothy Keller Distinguished Service Professor of Behavioral Science and
Economics at the University of Chicago, Graduate School of Business. E-mail:
frthaler@chicagogsb.edu. Postal address: Graduate School of Business, University of
Chicago, 5807 S. Woodlawn Ave., Chicago IL 60637 USA. We thank Nick Barberis, Ingolf
Dittmann, Glenn Harrison, Phil Maymin and Peter Wakker, conference participants at BDRM
X 2006 Santa Monica, FUR XII 2006 Rome, EFA XXXIII 2006 Zurich, EEA XXI 2006
Vienna and EWGFM XL 2007 Rotterdam, seminar participants at the Erasmus University of
Rotterdam, the University of Zurich, the University of Groningen, the University of
Amsterdam and the University of Lugano, and anonymous referees for useful comments and
suggestions. We thank Monique de Koning, Endemol, TROS and Sat.1 for providing us with
information and/or recordings of Deal or No Deal, Marc Schauten for acting as game show
host in the experiments, and Nick de Heer and Jan-Hein Paes for their skillful research
assistance. Financial support by Tinbergen Institute, Erasmus Research Institute of
Management and Erasmus Center for Financial Research is gratefully acknowledged. Any
remaining errors are our own.
50

Deal or No Deal (MATH) PDF

Uploaded by

Copyright:

Available Formats

Deal or No Deal (MATH) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deal or No Deal (MATH) PDF

Uploaded by

Copyright:

Available Formats

Deal or No Deal?

Decision Making under Risk in a Large-Payoff Game Show

Published in the American Economic Review, March 2008 (98:1), 3871

I. Description of the Game Show

bank offer 1 20 briefcases open 5 cases

bank offer 2 15 briefcases open 4 cases

bank offer 3 11 briefcases open 3 cases

bank offer 4 8 briefcases open 2 cases

bank offer 5 6 briefcases open 1 case

bank offer 6 5 briefcases open 1 case

bank offer 7 4 briefcases open 1 case

bank offer 8 3 briefcases open 1 case

bank offer 9 2 briefcases open 1 case

open contestants briefcase

maximum of eight subsequent rounds are 5, 4, 3, 2, 1, 1, 1, and 1. Accordingly, the number of

Mean Stdev Min Median Max

III. Preliminary Analysis

Unconditional Deal No Deal

Game Round (r)

Game Round (r)

More specifically, we classify a contestant in a given game round as a loser if BC r

Loser Neutral Winner

Analyzing the continuation value is more complicated. We elaborate on the continuation

Maximum Likelihood Estimation

B. Monetary bank offers

(9) ( xi ,r ) = (u( B( y)) cv( x

where () is the cumulative standard normal distribution function.5

where Ri is the last game round played by contestant i.

Netherlands Germany United States

Loser Neutral Winner

Reference Point Specification

Netherlands Germany United States

Small stakes Large stakes Pooled

You might also like