1

Introduction
GAME THEORYIS ABO T WHAT HAPPENS when genes, or nations-
interact. Here are some examples: Tennis players decidingwhether toerve
to the left or right side of the court; the only bakery in town offering a dis-
counted price on pastriesjust before it closes; employees deciding howhard
to work when the bos is awa ; an Arab rug seHer deciding how quickly to
lower his price when haggling with a tourist; rival drug firms investing in a
race to reach patent; an e-<:ommerce auction company leanring which fea-
tures to add to its website by trial and error, real estate developers guessing
when a downtrodden urban n ighborhood will spring back to life; San Fran-
ds 0 coromuters deciding which route to work will be quickest when the Bay
Bridge is closed; Laroelara m n in Indonesia deciding whether to join the
day's whale hunt, and how to divide the whale if they catch one; airline
workers hustling to get a plane away from the gale on time; MBAs decid-
ing what their degree will gnal to prospective employers (and whether
quilting after the first year of their two-y ar program to join a dot-eom
startup signals guts or stupidity); a man framing a memento from when
he first met his wife, a a gift on their first official date a year Jat r (they're
happily married now!); and p ople biddingfor art or oil leases, or for knick-
knacks on eBay. These examples illustrate, respectively, ultimatum games
(bakery, Chapter 2), gift ex ange (employees, Chapter 2), mixed equilib-
rium (tennis, Chapter 3), Tuni ian bazaar bargaining (rug seller, Chapter
4), patent race games (patents, Chapter 5), learning (e-eommercc, Chap-
ter 6), stag hunt games (whalers, Chapter 7), games (airlines,
Chapter 7), order-statistic games (developers, Chapter 7), signaling (MBAs
and romance, Chapter 8), auctions (bidding, Chapler 9).
1
2 1 Introdueti.()n
1 ImroductUm
In all of these situations, a person (or firm) mu t anticipate what oth-
ers will do and what other will infer from the person's own actions. A
game i a mathematical x-ray of the crucial features of these situations.
A game consists of the "strategies" each of several "players" have, with pre-
cise rules for the order in which players choose strategies, the information
they have when they choo e, and how they rate the desirability (or "util-
ity") of resulting outcomes. An appendix to this cha ter describes the basic
mathematics of game theory and give some references for further reading.
Game theory has a very clear paternity. Many of its main features ere
introduced by von eumann and Morgenstern in 1944 (following earlier
work in the 1920s by von eumann, Borel, and Zermelo). A few years later,
]ohn ash proposed a "solution" to the problem of bow rational pIay rs
would play, now called Nasb equilibrium. Nash's idea, based on the idea of
equilibriumin a physical system, as that players would adjust their strategies
until no player could benefit from changing. All players are then choosing
strategies that are best (utility-maximizing) responses to aU the other players'
strategies. Important steps in the 1960s were the realization that behavior
in repeated sequences of one-shot games could differ substantially from
behavior in one- hot game, and theori s in which a player can have private
information about her values (or "type"), provided all players know the
probabilities of what those types might be. In 1994, Nash, John Harsanyi,
and Reinhard Selten (an active experimenter) shared the obel Prize in
Economic Science for their pathbreaking contributions.
In the past fifty year, game theory has gradually become a standard
language in economics and is increasingly used in other odal sciences (and
in biology). In economics, gam theory is used to analyze behavior of firms
that worry about what their competitors will do.
1
Game theory i also good
for understanding how workers bebave .in firms (sucb as the reacti n of
CEOs or salespeople to incentive contracts), the spread ofsocial conventions
such as ~ g u a g e 'and fashion, and which genes or cultural practices will
spread.
The power of game theory is its generality and mathemati al precision.
The same basic ideas are used to analyze all the games--tennis, bargaining
for rugs, romance, whale-hunting-described in the first paragraph of this
chapter. Game theory is also b Idly precise. Suppose an Arab rug seDer
can always buy more rugs cheaply, an interested tourist values the rugs at
somewhere between $10 and $1000, and the seller has a good idea of bow
IGame theory fills the conceptual gap between a single monopoly. which need not worry bOUl what
other firms and consumers wiU do because it has monopoly power; and "perfect competition," in wbieb
no firm i5 big enough for competitors to worry abouL Game theory is used to study the intermediate case.
"oligopoly," in which there are few enough fimls that each company should anticipate what the othen
will do.
impatient the tourist is but isn't
rug. Then game theory tells yol.:
at, and exa.aly how quickly he s:
haws. In experimental re-creati
half-right and half-wrong: it's w
but the rate at which experim
amazingly close to the rate that
It is important to distingtlli
onomy of strategic situations, ~
periodic table of elements in c
ematical derivation of what pia:
likely to do in games.
2
Game m(
limited its spread outside econ(
and guesses rather than careful
games. This book aims to corn:
scribing hundreds ofex:perimer
results are used to create bebavi
about what players aauaay do.
tion, mistakes, limited foresigb
learning to analytical game m(
sophical perspective). Behavior
economics, an approach to ecc
to suggest ways to weaken ratie:
Camerer and Loewenstein, 200
Because the language of g.
unify many parts of social sdeJ
psycbologists, socioJogists, phil
development, and others. But
precisely defined in a game: '
doem't have to payyou back, bi
would, you trus her. Ifshe pays
gives a way to measure trust, a
places (including Bulgaria,. Sou
The spread ofgame theory
from the misconception that yo
it, and from the fact that most p
weD ground d in observation. ~
iTo be ptecise, thisbook is only about "no
binding agreements about what to do, 50 the
complementary branch of game theory whie.
blodlng agreements.
1 Introduction
1 Introduction 3
firm) must antidpatewhat oth-
m the person' own action. A
ial features of these situations.
several "players" have, with pre-
oose strategies, the information
rate the desirability (or "util-
) this chapter describes the basic
e references for further reading.
I. Many of its main features were
tern in 1944 (following earlier
and Zermelo). Afew years later
Iroblem of how rational players
idea, based on the idea of
tyers would adjust their strategies
19. All players are then choosing
responses to all the other play rs'
:ce the realizatio.n that behavior
could differ substantially from
1 which a player can have private
J provided all players know the
. In 1994, Nash, John Harsanyi,
ater) shared the obel Prize in
:ontributions.
as gradually become a standard
Ised in other social sciences (and
lSed to analyze behavior of firms
ill do.} Game theory is also good
1 firms (such as the reaction of
, the spread f so ial conventions
genes or cultural practices will
Jity and mathematical precision.
!l the games-tennis, bargaining
)ed in the first paragraph of this
ise. Suppose an Arab rug seller
:rested tourist values the rugs at
he eller has a good idea of how
monopoly, which nOt worry about what
y power, and competition: in which
is ""ed to study intermediate case.
I company should anticipate what the others
impatient the tourist is but isn't sure how much the tourist likes a particular
rug. Then game theory tells you exactly what price the seller should start out
at, and exactly how quickly he should cut the price as the tourist hems and
haws. In experimental re-creations of this kind of rug-selling, the theory is
half-right and half-wrong: it's wrong about the opening prices sellers state,
but the rate at which experimental sellers drop their prices over time is
amazingly close to the rate that game theory predicrs (see Chapter 4).
It is important to distinguish games from game theory. Games are a tax-
onomy of strategic situations, a rough equivalent for social science of the
periodic table of elements in. chemistry. Analytical game thwry is a math-
ematical derivation of what play r with different cognitive capabilities are
likely to do in games.
2
Game theory is often highly mathematical (which has
limited its spread outside economics) and is usually based on introspection
and guesses rather than careful observation of how people actually play in
games. Thi book aims to correct the imbalance of theory and facts by de-
scribing hundreds of experim nts in which people interact strategically. The
results are used to create beha 'oralgame theory. Behavioral game theory is
about what players actually do. It expands analytical theory by adding erno-
.on, mistakes, limited fore ight, doubts about how smart others are, and
learning to analytical game theory (Colman. in press, gives a more philo-
sophical perspective). Behavioral game theory is one branch of behavioral
conomies, an appr ach to econo.mics 'Which uses psychological regularity
to sugge t ways to weaken rationality assumptions and extehd theory (see
Camerer an Loewenstein, 2003).
Because the language of game theory is both rich and crisp, it could
unify many parrs of social science. For example, trust is studied by social
psychologists, ociol gists, phiIos phers, economists interested in economic
development, and other. But what is trust? This slippery concept can be
precisely defined in a game: ould you lend money to omebody who
doesn't have to pay you back, but might feel morally obliged to do so? lfyou
would, you trust her. Ifshe pays you back. she is trustworthy. This definition
gives a way to measure trust, and has been used in experiments in many
places (including Bulgaria, South Mrica, and Kenya; see Chapter 3).
The spread of game theory outside ofeconomics has suffered I believe,
rom the misconception that you need to know a lot of fancy math to apply
it, and from the fact that most predictions of analytical game theory are nOl
well grounded in observation. The need for empirical regularity to inform
2To be prctise, this book ;,only about "noncooperative"game theory--that is, when playerscann.ol malle
binding agreemen LS about what to do, s<l th.ey OOtl$t gu."'" what others will do. Cooperativ game w<:ory is a
complementary branch of game theory which deals with how players dMde the spo.ib after they AAVl: made
binding
"
4 1 Introduction
1.1 What Is Game Theory GoodFt
game theory has been recognized many times. In the opening pages of their
seminal book, von eumann an Morgenstern (1944, p. 4) WTote:
the empirical background of economic science is definitely inadequate.
Our knowledge of the r levant fa ts of economics is incomparably
smaller than that commanded in ph sics at th time when mathemati-
zation of that subject was achjev d.... It would have been absurd in
physics to expect Kepler and ewton without Tycho Brahe-and there
is no reason to hope for an easier d velopment in econ mics.
This book is focused on experiments as empirical background. arne
theory bas also been tested u ing data that naturally occur in field ettings
(particularly in clearly structured situations such as auctions). But experi-
mental control is particularly u eful because game theory predictions often
depend sen itively on th choic play rs have, bow they value outcomes,
what they know, the order in which lhey move, and so forth. As Ora ord
(1997, p. 207) explain:
Behavior in gam.es i notoriously sensitive to details of the environmenl,
so that trategic models carry a heavy morma 'onal burden, which is
often compounded in the field by an inability to ob erve all rel vant
variables. Important advances in experimental technique over th past
three d cades allow a control that often gives experiments a decisive
advantage in identifying the relationship b tween behavior and envi-
ronment. ... For many questions, [experimental data are] the most
important source of empiri 1information we have, and [they are] un-
likely to be less reliable than casual mpiricism or introspection.
Of course, it is important to ask how well the results of experiments
with (mostly) college students playing for a couple of hours for modest fi-
nancial stakes generalize to workers in firms, companies creating corporate
strategy, diplomats negotiating, and. so forth. But these doubts about gen-
eralizability are a demand for more elaborate experiments, not a dismissal
of the experimental method per se. Experimenters have studied a few di-
mensions of generalizability-particularly the effects of playing for mare
money, which are usuallysmall. But more ambitious experiments with teams
of players, complex environments, communication, and overlapping gener-
ations
3
would enhance generaliz biIity further. and people sh uld do more
of them.
'See SchOller and Sopher (2000).
1 ~ 1 What Is Gf
Is game theory meant to predic
what.? The theorist's answer is lh
simply "analytical," a body ofan
players with various degrees of r
way theory says, their behavior h
more than finding that cashiers s
arithmetic.
In practice, however, the to
predict, and also to xplain (or
good example ofall three uses of,
about the rules of the auction an
such as an oil lease or a paintin
rational bidders win pay.
Theory can help explain wb'
than others. For example, in "S(
bidder buys the object being aUCI
bid. Under some conditions the
revenue for sellers than traditio
bidder pays what she bid. But se,
Reilly, 2000). Why? Game thea
bidder pays a price other than w:
auctions are vulnerable to r n a n i ~
artificial bid to force the high bil
How well does auction theo
lematic: Because bidders' valua
tell whether they are bidding op
tested. Fortunately, there are roo
Kagel and Levin, in press). The
private-value auctions in which e
the object (and doesn' t care ho,",
ably close to the amounts they;
mapping values into bids is nonl
In cornmon-value auction
same faT everyone, but is uncer
example-different oil compani
bUl aren't sure how much oil is th
optimistic about the value of th
4 [n some domailU of social science, th se
~ n l unfolded are called "analytical n ~
5See Chen. and Plott (1998) and the seale'
1 Introduction
1.1 What Is Game Theary GoodFqr1 5
es. In the opening pages of their
tern (1944, p. 4) wrote:
science is definitely inadequate.
of economics is incomparably
lC at the time when mathemati-
. It would have been absurd in
rithout Tycho Brahe-and there
lopm nt in economics.
as empirical background. Game
: naturally occur in field settings
lS such as auctions). But experi-
,e game theory predictions often
have, how they value au tcomes,
n ve, and so forth. As rawford
ive to details of the environment,
informational burden, which is
inability to observe all relevant
imental te hnique over the pa t
:en gives experiments a decisive
hip between behavior and envi-
xperirnentaI data ar ] the most
ltion we have, and [they are] un-
lpiricism or introspection.
well the results of exp riments
a couple of hours for modest fi-
lS, companies creating corporate
-tho Bu th se doubts about gen-
'ate experiments, not a di mi at
riroenters have studied a few di-
the effects of playing for more
mbitious experiments with team
.nication, and overlapping gener-
ther, and people should do more
1.1 What Is ame Theory Good For?
Is game theory meant to predict what people do, to give them advice, or
what? The theorist's answer is !:hat game theory is none of !:he above-it is
simply "analytical," a body ofanswers to ma!:hematical questions about what
players with various degrees of rationality will do. If people don't play the
way theory says, their behavior bas not proved the mathematics wrong, any
more than finding that ca hiers sometime give the wrong change disproves
arithmetic.
In practice, however, the tools of analytical game theory are used to
predict, and also to explain (or "postdict,,4) and prescribe. Auctions are a
good example ofall three uses ofgame theory. Basedon precise assumptions
about the rule of the auction and the way in whic.h bidders value an object,
su h as an oil lease or a painting, auction theory then derives how much
rational bidders will pay.
Theory can help explain why some types of auction are more common
than others. For example, in "second-price" or Vickrey auctions the high
bidder buys the object being auctioned at a price equal to the second-highest
bid. Under some conditions these auctions should, in theory, raise more
revenue for sellers than traditional first-price auctions in which the high
bidder pays what she bid. But second-price auctions are rare (see Lucking-
Reilly, 2000). Why? Game theory offers an explanation: Since the h.igh
bidder pays a price other than what she bid in a second-price auction, such
auctions are vulnerable to manipulation by the seller (who can sneak in an
artificial bid to force the high bidder to pay more).
How well does auction theory predict? Tests with field data are prob-
lematic: Because bidder' valuations are usually hidden, it is difficult to
tell whether they are bidding optimally, although some predictions can be
tested. Fortunately, there are many careful experiments (see Kagel, 1995;
Rag I and Levin, in press). The results of these experiments are mixed. In
private-value auctions in which each player has her own personal value for
the object (and doesn't care bow much others value it), people bid remark-
ably dose to the amounts they are predicted to, even when the function
mapping values in to bids is nonlinear and counterintuitive.
5
In common-value auctions the value of the objec is essentially the
same for everyone, but i uncertain. Bidding for leases on oil tracts is an
example-different oil companies would all value the oil in the same way
but aren't sure how much oil is there. In these auctions players who are most
optimistic about the value of the object tend to bid the highest and win.
4 In SOme d o m a i n ~ of social science. these kinds ofgame-theoretic "stories" about how ..... institution or
event unfolded are called "analytical narratives .....d are proving incre3SlUgly popular (Bates e[ a1.. 1998).
SSee Chen ;md Plott (1998) and the sealed-bid mew . m res Ils In Chapu:r 4.
6 1 Introduaion
1.2 ThreeExamples
The problem is that, if you win, it means you were much mo.re O?timistic
man any other bidder and probably paid more than the object 1S worth
a possibility called the "winner's curse." Analytical game theory assu.mes
rational bidderswill anticipate the winner's curse and bid very conservatIvely
to avoid it. .Experiments show th t players do not anticipate the winner's
curse, so winning bidders generally pay more than they should.
Perhaps the most important modern u e of auction theory is to pre-
scribe how to bid in an auction, or how to design an auction. The shin-
ing triumphs of modern auction theory e recent auctions of to
telecommunications companies. In several auctions i.n different countries,
regulatory agencies decided to put airwave spectrum up for auction. An auc-
.on raises government revenu and ideally, ensures that a public resourc
ends up in the hands of the firms that are best able to create value from
it In most countries, the aucti were de ign. d in collaborations among
theorists and experimental "testbedding" that helped detectunanticipat d
weaknesses in proposed designs (like using a wind tunnel to test the d 'go
of an airplane wing, or a "tow-tank" pool to ee which ship designs sink and
which float). The designs that emerged were not exactly copied from books
on auction theory. Instead, theorists spent a lot of time pointing out how
motivated bidders could exploit loopholes in de ign - proposed by la rs
and regulators, and using the results oftestbedding to improve designs. Auc-
tion designers opted for a design that gay bidders a chance to learn from
potential mistakes and from wat bing other, rather than a simpler
bid" design in which bidders simply mail in bids and the Federal Com-
munications Commission opens the env lope and announces the highest
on s. One of the most powerful and surprising ideas in auction theory-
equivalence"-is that some types of auctions will, in theory, raise
the same amount of revenue as other auctions that are quite different in
structure. (For example, an "English" aucti n, in which prices are raised
slowly until only one bidder remains, is revenue-equivalent to a sealed-bid
"\lickr y" auction, in which the hi hest bidder pays what the second-highest
bidder bid.) But when i came to designing an auction that actual compa-
nies would participate in with billions of dollars on the line, the auction
design rs were not willing to bet that behavi()T would actu.ally be equivalent
in different types of auctions, despite what theory predicted. Their d sign
choices reflect an implicit theory ofactual behavior in games that is probably
closer to the ideas in this book than to standard theory based on unlimited
mutual rationality. otice that, in this process of de ign and prescription,
guessing accurately how players will actually behave-good predicti n-is
crucial.
6
6Howard Rai!fa poinled this out many times. calling game theory "asymmetrically normative.'
Evenifgame theory is not aIw;
tive opportunity. Just as evangeJis
moral codes, the fact that players
give helpful advice, Simply mappi
extremely efu1 because it teUs F
ular book for busines . managers,
(1996) draw attention to the bare:
actions, and outcomes. Both are
a more theoretical book. They c.
working with managers convincec
game theory is more helpful
Game theory is often used t
game theory is used to figure ou'
interaction, SO a person or camp
their advantage. (This is a kind 0
bow to improve an existing situati
1.2 Th
This chapter illustrates the basic.
perimental approach with three
detail in later chapters): ultimatu
dination games, and "beauty con
these games show how behavioral
more accurately by extending ana
feel about the payoffs other playe
learning.
The three game use a reci
reported in this book: pick a gam(
bold prediction or a vague predic-
are particularlyuse.ful because onI
make a prediction. If the predicD.
at fault, and the results usually
better.
In the experiments, games ar
game theory rarely specifies how.
Subjects make a simple choice, an
of other subjects combine to de
1 These design bel hCllVily on tht I:O@
oCatrategic thinking wbkh tl'anSGCnd idiosynaa:
domain spedBc then the basic enterprise lhls t
1 Introduction
1.2 Three Examples 7
you were much more optimistic
more than the object is worth,
\naLytical game theory assumes
curse and bidvery conservatively
l do not anticipate the winner's
,re than they should.
use of auction theory is to pre-
to design an auction. The shin-
~ e recent auctions of airwaves to
I auctions in different countries,
spectrum up for auction. An auc-
Iy, ensures that a public resource
e best able to create value from
~ s i g n e d in collaborations among
that helped detect unanticipated
~ a wind tunnel to test the design
) see which ship designs sink and
re not exactly copied from books
.t a lot of time pointing out how
s in designs proposed by lawyers
bedding to improve designs. Auc-
e bidders a chance to learn from
:rs, rather than a simpler "sealed-
1 in bids and the Federal Com-
.opes and announces the highest
>rising ideas in auction theory-
s of auctions will, in theory, raise
tions that are quite different in
etion, in whoch prices are raised
_venu -equivalent to a sealed-bid
Ider pays what the second-highest
ng an auction that actual compa-
. dollars 00 the line, the auction
lviOT would actually be equivalent
it theory predicted. Their design
oehavior in games that is probably
ndard theory based on unlimited
ocess of design and prescription,
ally b have-good prediction-is
une theory "asymmetrically nonn.tive:
Even ifgame theory is not always accurate, descriptive failure isprescrip-
tive opportunity. Just as evang lists preach because people routinely violate
moral codes, the fact that players violate game theory provides a chance to
give helpful advice. Simply mapping ocial situations into types of games is
extremely useful because it tells peopl what to look out for. In their pop-
ular book for business managers, Co-opetiticm, Brandenburger and Nalebuff
(1996) draw attention to the bare t bones ofa game-players, information,
actions, and outcomes. Both are brilliant theorists who CO'Uld have written
a more th oretical bo k. They chose not to because teaching MBAs and
working with managers convinced them that teaching the basic elements of
gam theory is more helpful.
Game theory is often used to prescribe in a subtler way. Sometimes
game theory is used to figure out what it is likely to happen in a strategic
interaction, so a person or company can then try to change the game to
their advantage. (This is a kind of engineering approach too, since it asks
how to improve an existing situation.)
1.2 Three Examples
This chapter illustrates the basics of behavioral game theory and the ex-
perimental approach with thr e examples (which are discussed in more
detail in later chapters): ultimatum bargaining, "continental divide" coor-
dination games, and ''beauty contest" guessing games. Experiments using
these games show how behavioral game theory can explain what people do
more accurately by extending analytical game theory to include howplayers
feel about the payoffs other players receive, limited strategic thinking, and
learning.
Th three games use a recipe underlying most of the experiments
reported in this book: picka game for bicb standard game theory makes a
bold prediction or a vague prediction that can be sharpened. Simple games
are particularly u eful because onlyone or two basic principles are needed to
make a prediction. If the prediction is wrong, we know which principles are
at fault, and the results usually suggest an alternative principle that predicts
better.
In the experiments, gam s are usually posed in abstract terms because
game theory rarely specifies how adding realistic details will affect behavior.
Subjects make a irople choice, andknow how their choices and the choices
of other subjects combine to determine monetary payoffs? Subjects are
? These design choice. bet heavily on the cognitive prclIUlIJption that people are using generic principlo
ofStrategic thinking which tralJseend idi05)'T1cnllic dlffi:rences. in verbal descriplioJ1Jl ofgame$, Ifchoice- are
domain specific then the b:uic en terpn.e this book: desai'bes b incomplete; varying game labels lO evoke
8 1 lntroduaion
1.2 Three Examples
actually rewarded based on their rrnance because we are inter sted in
extrapolating the resulrs to n rurally occurring games in which players have
substantial financial incentives. The game are usually repealed beca we
are interested in equilibration and learning over time. An appendix to this
chapter describes some key de ign choices experimenters make and why
they matter.
1.2.1 Example 1: Ultimatum Bargaining
I once took a cruise with some friends and ap"hotographer took our picture,
unsolicited, a we boarded the boat. en we disembarked hours later, the
photographer tried to sell us the picture for $5 and refused to negotiate. (His
refusal was credible b cause several oth r groups stood around deciding
whether to buy their pictures, also < r $5. Ifhe caved in and cut the pri e, it
would be evidentto all others and he would lose alotmore than the discount
to us since he would have to offer the discount to everyone.) Being good
game theorisrs, we balked at the price and pointed out that the picture was
worthless to him (one cheapskate offered $1). He rejected our insulting
offer and refused to bac.k down.
The game we played with the photographer was an "ultimatum game,"
which is the simplest kind of bargaining. In an ultimatum game there is
some gain from exchange and one player makes a Lake-it-or-Ieave-itoffer of
how to divide that gain. Our piC-cur pre umably bad no value to him and
was valuable to us (worth more than $5 in sentimental value). A price is
imply proposing a way to divide the gains from exchange between our true
reservation price and his cost. His offer to sell for 5 was an ultimatumoffer
because he refused to negotiate.
In laboratory ultimatum game Uke this, two players, a Proposer and
a Responder, bargain over some amount, say $10 (the sum used in many
experiments). The $10 represents the value of the gain to exc.hange (or
"surplus") that would be lost if the trade wasn't made. he Proposer 0 ers
x to the Responder, leaving herself $1 0 - x. The Responder can either take
the offer-then the Responder gets x and the Propo er gets $10 - x--or
reject it and both get nothing.
Because the ultimatum game is so simple, it is not a good model of
the protracted process of most naturally occurring bargaining (and isn'l
intended to be). It is the right model ofwhat happened to us after the cruise,
domain-spedfic reasoning;s the next step. The study by Cooper et aL (J 999) of rateheteffeclS in productivity
games using Chinese factory managers--who face such effeclS in plannt:d economics--i$ a good enmple
(see Chapter 8).
and what happens in the wanin
on the courthouse steps before;
step in much bargaining, and h
complicated situations (see Ch2
imple games test game-the
Ultimatum games, and related
people feel about the allocation
The analytical game theory
First assume players are "self-ir
the most moneyfor themselves.
will accept the smallest money
anticipates this, and wants to ge
$0.25 and keep 9.75. In forme
positive amount) is the "subgan
Proposer has all the bargaining
a [erested Responder will
To many people the lopsid
lyrical game theory (with self-in
is considered unfair, the way pe
are willing fO take cosdyactiom
the cruise-picture example, off<
rapher offered added $4 to au
thought this was unfair to him,
though everyone suffers-he e
we would like t.o own). The lab (
Responders put their money wi
seem unfair? If so, will Pr0POSI
stubbornly make unfair offers?
In dozens of experiments
Proposers offer $4 or $5 out of$
Offers of $2 or less are rejected
much less than half is un.f.Ur aI
punish the Proposer who beha:
a study by Hoffman, McCabe,
amount being offered to the R
frequency of offers of different
bar is the number of offers that
8 Note also that every offer is a "Nash e<
optimal offer if the Proposer think.i the Respc
if the Proposer belie..es 11.. she will nCV1'"r I3U
be part of a Nash equllibrium.)
and what happens in the waning minutes before a labor strike is called or
on the courthouse steps befor a lawsuit goes to trial It is a model of the last
step in much bargaining, and hence is a building block for modeling more
ompli ated ituations (see Chapter 4).
Simple games te t game-theoretic principles in the clearest possible way.
Ultimatum games, and relat d games, also are useful for measuring how
peopl feel about th allocations of money between themselves and others.
The analytical ga e theory approach to ultimatum bargaining is this:
First a ume layers are "self-interested"; that is, they care about earning
th most money for th mselv s.lfplayers are self-interested, the Responder
will accept the smallest money amount offered, say $0.25. If the Proposer
anti ipates this, and wants to get the most sbe can fOT herself. she will offer
$0.25 and keep $9.75. In formal terms, offering 0.25 (and acc pling any
p sitive amount) is the "subgame perfect equilibrium".8 By going first, th
Proposer has all the bargaining power and, in theory, can exploit it b cause
a self-interested Responder will take whatever she can get.
To many people, the lopsided distribution of the 10 predicted by ana-
lytical game theory (with self-interest) seems unfair. Because the allocation
is considered unfair, the way people actually bargain shows whether people
are willing to take costly actions that express their concerns for fairness. In
the cruise-picture example, offering $1 instead of the $5 price the photog-
rapher offered added $4 to our urplu and ubtraeted $4 from his. If he
thought this was unfair to him, he could reject it and earn nothing (even
though everyon suffers-he earns no money and we don't get a picture
we would like to own). The lab experiments simulate this simple game. Will
Re ponders put their money where their mouths are and reject offers that
eem unfair? If so, will Proposers anticipate this and make fair offers, or
stubbornly make unfair offer ?
In dozens of experiments conducted in several different countriesj
Proposers offer 4 or $5 out of lOon average, and offers do not vary much.
Offers of 2 or less are reject! d about half the time. The Re ponders think
much less than half is unfair and are willing to reject such small offers, to
punish the Propo r who behaved so unfairly. Figure 1.1 shows data from
a study by Hoffman, McCabe, and Smith (1996a). The x-axis shows the
amount being offered to the Responder, and the y-axis shows the relative
frequency of offers of different amounts. The dark part of each frequency
b is the number of offers that were rejected. Most offers are close to half
1 Introduction
IDce because we are interested in
'ring games in which players have
; are usually rep ated because we
Ig over time. An appendix to mis
:s experimenters make, and why
tum Bargaining
a pbot grapher took our picture,
I we disembark d hours later, me
.$5 and refused to negotiate. (His
r groups stood around deciding
f he caved in and cut the price, it
I lose a lot more than the di count
:count to everyone.) Bing good
pointed out that the picture was
:l $1). He rejected our insulting
apherwas an Qultimatum game,"
In an ultimatum game there is
makes a take-it-or-Ieave-i t offer of
JJllably had no value to him and
in sentimental value). A price is
from xchange between our true
ell for $5 was an ultimatum offer
his, two player, a Proposer and
say $10 (the sum used in many
lue of th.e gain to exchange (or
rasn't made. The Proposer offers
c. Th R spond r can either take
:l the Proposer gets $10 - x-or
mple, it is not a good model of
occurring bargaining (and isn't
It happened to us after the cruise,
1.2 Three Examples 9
",r et aI. (1999) ofratchet effects in productivity
cr.. in planned econ.omier--is 11 good example
8 ote aJso that e v e ~ y offer is a' ash equilibrium" or JllUfUaI besl-resporue paltern because % is me
optimal offer if the Proposer thinks the Responder will rejecuny other offer. (This belid"may be wrong but.
if the Proposer believes il, she wiU never rake an action thal disconfimu her bel.icf, so the wrong belief can
be part ofa Nash equilibrium.)
10 1 IntroduetUm 1. 2 Three Examples
0.3

c:
.,
::>
0.2 (T
.,
It:
$100 pie
0.1
ccept
Reject
0.0 -----,--'--L..f"""----'P----""I'------''''I'''L....--L.t-'---+-'---,-----'
o 110 11-20 21-30 31-40 41-50 5HiO 61-70
Proportion of pie offered (%)
Figure 1.1. Offers and rejections in highr and low-stakes ultimatum garnes. ()urce: Based
on data from Hoffman, McCabe, and mith (1996a).
and low offers are often reject d. Figure 1.1 also shows that the arne pattern
of results occurs when stake were multiplied by ten and Arizona students
bargained over 100. (A couple of ubjects rejected $30 offers!) The same
basic r wt has been replicated with 400 take (List and Cherry, 2000) in
Florida and in countrie with low dispo able income. in luding Ind nesia
and Slovenia, wher modest s es byAmeri .an standards represent e era!
weeks' wages.
There are man interpretacio fwhatcaus Responder to rej ctsub-
rantia! um ( e Chapter 3). There is little doubt that orne player define a
fair plitof 10asclo tohalfandhav apreferenceforbei g eat dfairly.
Such rejection are vidence of "ne ative reciprocity": Responders r cipro-
cate unfair behavior by har . g the person who treat d them unfairly, at a
sub cantia! cOst to th m elves (provided the unfair Proposer is harmed more
than they are). 'egati e r cipr city is evident in other so' dom' ,even
wh n monetary take are high-jilted boyfriends who accost their exe ,
ugl divorces that COSt p ople large sums, impul iv street crimes cau ed by
a stranger allegedly "di respecting" an assailant, the failure f partie in le-
gal "nuisance case " to renegoti
could benefit (Farnsworth,
This explanation for ultim::
fairness preferences came fron
man exp rience in our ancestr.
brain mechanism. orin the intf
which cause people to get angry
ting angry had survival value wI'
in a small group ( ee Frank, 1
cultures cr ate different tanda
ne of kin relations or the del
strangers (compared with sharir
dards are tran mined socially C
children.
Remarkable evidence for th,
byele n anthropologists who c.
ture in Africa, the Amazon, Pal
( ee Cbapter 2). In some of the
ing fairly was necessary. Propos.
equivalent of 1.50 out of 10):
fer. Ironically, these simple soc.
behav exactly as game theory I
Note that in ultiJ
strategic principles underlying I
The Responder simply decides'
ing, or wants to get a mall bar
fact that a Responder rejects me:
but it does nOl mean she is not
ries attempt to explain rejectiOl
balance a p rson' desire to ha\
cate tho e wh have treated tht
Such functions have a long pedi
1890 ). Economi ts ha e resi te
introduce a n fa tor in the u
theories striv to explain r suit
Having a lot of d ta from differ
po ible and imposes discipline.
9My sisterJeannine told me Ih l in At.I:lJ
"high-rolkr" tomen &t.ealing luxuriow 1mI
rooms after losing :u the casioo to their lOUI
9 My sister Jeannine told m that in Atlantic City the 'n sometimes have problems with lucrative
"high-roller" customers Stealing luxurious to cls, robes. and other items from their (complimentary) h tel
rooms afler losing at the casinos In their mmds these losers e simply tUing thing3 they have paid for.
gal "n - ance c e" to r negotiate after a courtjudgment ven when both
could benefit (Farnsworth 1999), and so on.
9
This expl ation fi r ultimatum reje tions begs the question of where
wne prefer nces came from. A popular line of argument is that hu-
man experience in Ot ancestral past created evolutionary adaptations in
brain mechani ms, or in the interaction ofcognitive and emotional tem,
which ca e pe pie to get an y when they are pushed around because get-
tin angry had urvival valu when ople interacted with the same p ople
in a mall group ( Frank, 1988), A different line of argument i that
u1ture creat different tandards of Urn ,perhap owing to the clo
ness of kin relation or the degree of anonymous market exchange with
rranger ( ompar d with h' among relatives), and these cultural tan-
dard e tran mi tted dally through oral traditions and socialization of
childe n.
Remark ble evidenc for the cultural standards viewcomes frOOl a study
byel yen anthropologists who condu ted ultimatum games in primitive cul-
tures in Africa, the Amazon, Papua New Guinea, Indonesia, and Mongoli
( e Chapter 2). In orne of these cultures, p ople did nOl think th l har-
ing ai ly as nece . Proposers in th cultures off, r d very little (the
equival Dt 1.50 out of 10) and Responders accepted virtualJy ev ryof-
fer, Iro i ly th imple cietie are the only known populations who
behave exactly as game theory pr dicts!
Jote th t rejecti n in ultimatum games do not necessarily rej Cl the
Ira gic principle derlyin game theory (for example, WeibulJ, 2000).
The e ponder impl d cides whether she wants both playe to g t noth-
'n , or wants [0 get a mall har when the Peopo r gets much more. The
fact th taRe pond r rej cts aus she is not maximizing her own earnings,
but it doe not mean he is not capable of strategic thinking. Recent th er
n att mpt to explain reje dons using social preti renee function which
balanc a per n' de ire to have more mon with their desire to reciprer
cat tho e who hav tr ated m fairly or unfairly, or to achieve quality,
Such function hav a long p digree (traceabl at 1 t to Edgeworth in the
18 s). Economi ts hav resi ted them becau it ems to b to to
introduce a n w fa r in the utility function for each game. But the new
theories triv to explain re ults in different gam s with single function,
H ving a lot of data from differenrgames to \ ork with makes this enterprise
ible and impos disciplin,
1 Introduction
31-40 41-50 51-60 61-70
Ie offered (%)
')-Stakes ultimatum games. Source: Based
I).
Lal 0 show that th same palt rn
Lied by ten and aona students
:s r je ted 30 offers.) Th am
) stake (List and eh rry, 2000) in
ble' come, including Indone ia
'ncan tandard repre eot veral
.tcause R P nder tor tiect ub-
~ doubt that orne player d fine a
reference for being treated fairly.
reciprocity": R ponder recipro-
,n who treated them unfairly, at a
e unfai.r Propo er is harmed mor
eDt in other social domain, ev n
oyfriends wh acco t their exe ,
impul ive eet crime cau d by
ailant, e failur of partie in le-
1.2 Three Examples 11
12 1 Introduction
1.2 Three Examples
Table 1.1. Payoffs in continenl,al div
1 45 49 52 55 56
2 48 53 58 62 65
3 48 54 60 66 70
4 43 51 58 65 71
5 35 44 52 60 69
6 23 33 42 52 62
7 7 18 28 40 51
8 -13 -1 11 211 37
9 -37 -24 -11 3 18
10 -65 -51 -37 -21 -4
11 -97 -82 -66 -49 -31
12 -IllS -117 -100 -82- -61 -
13 -173-1 6-137-118 -96 -
14 -217 -198 -179 -158 -134-1
Source: Van HlJYck, Baualio, and Cook (1
low number if you think most
high number if you think most
sure what others will do, pick a n
from 23 to 82 (hedging your be
In the experiments, player
The groups play together fiftee
median was, compute your eat
own choice and the median), an
think for a minute about what yc
over the course of playing fiftee
The payoffs have the prope
number is slightly below 7, her
number smaller man the guess
will be 7, your best response is 5
initially low, responding to low I
reach 3. Three i an equilibriur
everyone chooses 3, the median
of 3 is to choose 3. If players w(
by moving away. (The payoff f
Table 1.1.)
The new theories make smprising newpr dictions. For xample, when
there are two or more Propa ers, there i no way for anyone of them ingle-
handedly to earn more money and limit inequality. As a result some theories
predict that both Proposers offer almost everything to the Responder even
though they do care about equality. (If there had been two photographers
on that damn boat, we would have gotten our picture for $1.)
ew ocial preference theories houJd prove useful in analyzing bargain-
ing, tax policy, the strong tendency a t nantfarmers to share crop earnings
equally with landowners (Young and Burke, 2001), and wage-setting (par-
ticularly the reluctance of finns to cut wages in hard times, which is puzzling
to economists who assume changes in the price of labor will equaliz supply
and demand, and other phenomena) .
1.2.2 Example 2: Path-Dependent Coordination in
"Continental Divide" Games
In coordination games, pI ayerswant to conformto what others do (although
they may have different ideas about which conformist convention is best).
For example, in California there is an ongoing struggle over the physical
location of the "new media" firm, such as internet provi ion of film and
entertainment. ew media people could gravitate toward Silicon Valley,
where web geeks congregate, or toward Hollywood and outhern California,
where many movies and TVshows ar produced. Which geographical region
is the better location depends on whether you think. the location of internet
firms is central, and "content" producers should follow them, or whether
the internet is merely a di tribution channel and content pro 'deTS are
king.l
This economic tug-of-war can be modeled by game in which players
choose a location, and their earnings clep nd on the location they choose
and the location most other people choose. A game with this fI vor has
been studied by Van Huyck, Battalio, and Cook (1997). Tabl 1.1 5h ws the
payoffs (in cents). In this game, players pick numbers from 1 to 14 (think.
of the numbers as correspo ding to physical locations-low numbers are
Hollywood and high numbers are Silicol) Valley). The matrix in Table 1.1
shows the row player's payoff from choosing a number when the median
number everyone in a group picks--the middl number-is the number
in the different columns. If you choose 4, for example, and the median
is 5, you earn a healthy pay ff of 71; but if the median is 12 you eam
-14 (bankruptcy!). The basic payoff structure implies you should pick a
10 Of course, this e ~ = p l e is undermined by the fact that cyberspace is everywhere and nowhere, so
content prOviders might be able to stay pUt in the swank Hollywood Hills and still do business 'in" Silicon
Valley without moving.
Choi e 1 2 3 4 5
1 Introduction
1.2 Three Examples 13
I{ predictions. For example. when
10 way for anyone of them single-
equality. As a result some theorie
verything to the Responder even
ere had been two photographers
our picture for $1.)
prove useful in analyzingbargain-
ntfarmers to share crop earnings
ke. 2001), and wage-setting (par-
_s in hard times, which is puzzling
price oflaborwill equalize supply
dent Coordination in
Ie" Games
tformtowbatothers do (although
h. conformist convention is best) .
19Oing struggle over the physical
as internet provision of film an.d
1 gravitate toward Silicon Valley,
lllywood and Southern California,
luced. Which geographi al region
you think the location ofintemet
: should follow them, or whether
umel and content providers are
jeled by a game in which players
Jend on the I cation they choose
'0 e. A game with this flavor has
C ok (1997). Table 1.1 howsthe
lick number from 1 to 14 (think
sical locations-low numbers are
l Valley). The matrix in Table 1.1
Ising a number when the median
middle number-is the number
4, for example, and the median
,ut if the median is 12 you earn
uctur implies you should pick a
thaI cyberspace. is evcrY"'here and nowhere, so
lollywood Hills and still do business "in" Silicon
low Dumber if you think most others will pick low numbers. and pick a
high number if you think most others will pick high numbers. If you aren't
sure what oth rs will do, pick anumber such as 6, which gives payoffs ranging
from 23 to 82 (hedging your bet) .
In the experiments, players are organized intO seven-'person groups.
The groups play together fifteen times. After each trial you learn what the
median was, compute your earnings from that trial (depending on your
own choice and the median) and play again. Since the game is complicated.
thinkfor a minute about what you would actually do and what might happen
over the course of playing fifte n times.
The payoffs have the proper(y that, if a player guesses that the median
number is slightly blow 7. her best response to that guess is to choose a
number smaller than the gue its If. For example, uyou think the m dian
. J be 7, y ur be t response is 5, which earns 83 cents. Thus, ifmedians are
initially low, responding to low medians wiU drive numbers lower until they
reach 3. Three is an equilib 'urn or mutual best-response point because, if
everyone choo e 3, th median will be 3 and yourbest response to a median
of 3 is to choose 3. IT players were to reach this point., nobody could profit
by moving away. (The payoff from this equilibrium is shown in italic in
Table 1.1.)
Table 1.1. Payoff. in "continental divide" experiment (cents)
Median choice
Choice 1 2 3 4 5 6 7 8 9 10 11 12 13 14
J 45 49 52 55 56 55 46 -59 -88 -105 -)17 -127 -135 -142
2 48 53 58 62 65 66 61 -27 -52 -67 -77 -86 -92 -98
3 48 54 60 66 70 74 72 1 -20 -32 -41 -48 -53 -58
4 43 51 58 65 71 77 80 26 8 -2 -9 -14 -19 -22
5 35 44 52 60 69 77 83 46 32 25 19 15 12 10
6 23 33 42 52 62 72 82 62 53 47 4! 41 39 38
7 7 18 28 40 5] 64 78 75 69 66 64 63 62 62
8 -13 -1 11 23 37 51 69 83 81 80 80 80 8] 82
9 -37 -24 -11 3 18 35 57 88 89 9]
92 94 96 98
10 -65 -51 -37 -21 -4 15 40 89 94 98 101 104 107 UO
11 -97 -82 -66 -49 -31 -9 20 85 94 100 105 no 114 119
]2 -133 -117 -100 -82 -61 -37 -5 78 91 99 106 112 U8 123
13 -173 -156 -137 -118 -96 -69 -33 67 83 94 103 110 117 123
1'1 -217 -198 -179 -158 -]34 -105 -65 52 72 8 95 104 112 120
Source: Van Huyck, Banalio, and Cook (1997).
14 1 Introduction
1.2 ThreeExamples
But there is another Nash equilibrium. If players guess that the median
will be 8 or above, they should choose numbers that are higher than their
guesses, until they reach 12; 12 is also a ash equilibrium because cboosing
12 gives the highest payoff if the median is 12.
Thi is a coordination game because there are two ash equilibria in
which everybody choo es the same strategy. Game theorists have struggled
for many decades to figure out which of many equilibria will result if there
are more than one.
This particular game ilJu trates proce e in nature and social systems
in which small historical accidents have a big long-run impact. A famous
xarnple i what chaos theorists call the "Lorenz effect"; Because weather
is a complex. dynamic system, the movement of a butterfly in China can
set in motion a complicated meteorological process that creates a Storm
in Bolivia. If that butterfly had just sat still, the Bolivians would be dryl
Another example is what social u1e rists call the "broken .window effect."
Anecdotal evidence suggests that, when there is a single broken window
in a community, neighbors feel less obligation to keep their yards clean,
replace their own broken windows, and put fresh paint on th ir houses.
Since criminals want to commit crimes in communities where neighbors
aren 't w ~ t c h f u l and other criminals are lurking (so the cops are busy), a
single broken window can lead to a spiralling process of social breakdo\vrl.
Policymakers love the broken window theory because it suggests an easy fix
to problem ofurbao decay-repair every window before the effect of a few
broken ones spreads throughout the community like a virus.
I call the game in Table 1.1 the "continental divide" game. The conti-
nental divide is a geographic line wbi h divides those parts of orth America
in which water will flow in one direction from the parts in which water flows
in the opposite direction. If you stand on the continental divide in Alaska,
and pour water from a canteen as I once did, some drops will flow north
to the Arctic Ocean and others will flow to the Pacific Ocean. Two drops of
water that start out infinitesimally clo e together in the canteen end up a
thousand miles apart
The game is called the continental divide game because medians below
7 are a "basin ofattraction" (in evolutionary game theory terms) for conver-
gence toward the equilibrium at 3. Medians above 8 are a basin of attraction
for convergence toward 12. The u sep trix" between 7 and 8 divides the
game into regions where players will "flow" toward 3 and players will flow
toward 12.
Whicb equilibrium is reached has important economic consequences.
The 12 equilibrium pays $1.12 for each player but the 3 equilibrium pays
only $0.60. On this basis alone, you might guess at players would choose
higher numbers in the hopes of reaching the more profitable equilibrium.
Before glancing ahead, ask yourself again what you think will happen. Ifyou
14
12
.-*" K..
10
.f
8
~
6
::E
4
2
0
2 3 4 5
Figure 1.2. Median choices in the U
Van Huyck, Battalio, and Cook (J 99i
have studied a lot of game thee
curiosity about what people acl
Figure 1.2 shows what be
groups started at a median at
Jow-payoff equilibrium at 3. Th
flowed to the high-payoff equil
The experin1ent has two in
gravitate toward the high-payol
up at lownumbers earn halfas r
again, or discussed Ihe game iJ
econd, the currents of histor'
initial conditions." Players who'
others who think 7 is their Iud
end up sucked into a whirlpool
a group whose median is 8 or h
One or two Chinese subjects (
could bring good fortune to ev
Ihe Bolivians.
No concept in analytical g
that some groups flow to 3 anc
more. Indeed, the problemofp
in games such as these may be j
cial conventions, communicatic
analogies playersdrawwilh exp
about lucky numbers could all
1 Introduction
1.2 Three Exampks 15
n. Ifplayer gue s that the median
umb rs that are higher than their
ash quilibrium because choosing
is 12.
ther are two ash equilibria in
: . Game theorists have truggl d
many equilibria . 1 re ult if ther
e e in na e and 0 . al sy tern
a big long-run impact. A famou
"Lorenz effect": B cau weather
ment of a butterfly in China can
process that creates a storm
still, the B livians would be dry!
i call the ubr ken .window effect."
there j a ingle broken window
igation to keep their yard clean,
put fre h paint on their house.
in ommunitie wh r neighbor
lurking ( 0 the cops are bu y), a
lling proce of ocial breakdown.
because it ugge ts an asy fix
-I window b ore the ffect of a few
ununity like a virus.
ltinental di 'de" game. Th onti-
vi e tho e parts of orch Ameri a
r m th parts in w 'ch water flow
1 the continental divide in Alaska,
e did, some drop will flow north
o the Pa . c Ocean. Two drop of
together in the canteen end up a
vi e game b cause medians blow
lry game theory t rm ) for onver-
n above 8 are basin ofattraction
trix" between 7 and 8 divides the
IW" to d 3 and players will flow
lportant economic ons quence .
player but the 3 equilibrium pay
It gue that players would choo
; the more equilibrium.
what you think will happ n. Ifyou
14,------------------------.
12 ...?.,:::=----... __o::::......... :.:, ... 6.-1
10

.....,._____::,....j
__
2+---------------------1
O+--.-...---.,--,--.,--,--.,--r--,---r--.--r---r----..,.----_1
2 3 4 5 6 7 8 9 10 11 12 13 14 15
Period
Frgure 1.2. Median clwices in the "continental divide" game.. &mree: Bas4d on dataftom
Van Huyck, Battalio, and Cook (1997).
have tu ied I t of e theory and till aren't ure what to expect, your
curio ity b ut what pe pi actually do. hould b piqued.
FIgUre].2 how what h ppened in ten experimental group. Five
group tarted at a median at 7 or below; all of them flow d toward the
low-payoff equilibrium at 3. The other five group tarted at or above and
flowed to the high-payoff equilibrium.
exp riment h two important findings. Frrst, p ople do tWt alway
gravitate toward the hi h-pa 0 equilibrium even though players who end
up at 10 bers earn halfas much. (Whether theywould if they could play
again, or i cd th e 'n advance, i an interesting open que tion.)
cod, ch currents of hi toryar tr ng, creating W xtreme en itivity to
initial condition." Playe who find them Ive in a group with two or thr e
other who think 7 is their lucky number, and choo it in the first peri d,
end up ucked into whirlpoolleadin to m asly 0.60 arning. Player in
a group whose median i 8 or high r end up earning almo t twice as much.
ne or two hine e subjects choosing 8-a lucky numb r for Chinese--
could bring good fortu e to ev ryone just as the butt rfly brought rain on
the Bolivians.
o co cept in analytical game theory gracefully accounts for the act
that some group flow to 3 and earn Ie ,whiJ other flow to 12 and earn
mor . lnd ed, the probl m ofpredicting which of many equilibria will re ul t
in game ch th e may be inher ntly un olvable by pure re ooing. 0-
cial conv nrion ,co munication, subtle features of th display of th game,
an 0 . es player dra\ .th xp riences they have had, and home pun id as
ab ullu ky numb r could all influ nce which equilibrium is r ached. As
16 1 IntrrxJ:ucUon
1.2 Three Examples
Schelling (1960) wrote, predicting what players will do in these games by
pure theory is like trying to prove that ajoke is funny without telling it.
1. 2.3 Example 3: "Beauty Contests" and Iterated Dominance
In Keynes's farnou book General Theory ofEmployment, Interest, and Money,
he draws an analogy between the stock market and a newspaper contest in
which people guess what faces others will guess ar most beautiful: "It is not
a case of choosing tho e which, to the best a one's judgment, are really
the prettiest, nor even those which average opinion genuinely thinks the
prettie t. We have reached the third degree, where we devote our intelli-
gences to anticipating what average opinion expects the average opinion
to be. And there are some, I believe, who practise the fourth, fifth, and
higher degrees" (1936, p. 156). This quote is perhaps no more apt than in
the year 2001 (when I first \\/Tote this) ,just after prices ofAmerican internet
stocks soared to unbelievabl heights inthe largest speculative bubble in his-
tory. (At one point, the market valuation of the e-tailer bookseller Amazon,
which had never reponed prfit, was worth more than all other American
book eHer combined.)
A impLe game that captures the re oning Keynes had in mind is called
the "beauty contest" game (see Nagel, 1995, and Ho, amerer, and Weigelt,
1998). In a typical beauty contest game, each of N pI yers simuJtaneously
chooses a number Xi in the interval [0,100}. Take an average of the numbers
and multiply by a multiple p< 1 (say p= 0.7). The player whose number is
closest to this target (70 percent of the average) wins a fixed prize. Before
proceeding, think about what number you would pick.
The beauty contest game can be used to distinguish whether people
"practise the fourth, fifth, and higher degre s" of reasoning as Keynes
wondered. Here's how. Mo t players start by thinking, "Suppose the average
is 50". Then you shOUld choose 35, to be clo est to the target of 70 percent of
the average and win. But ifyou think all players will think this way the average
will be 35, so a shrewd player such as yourself (thinking one tep ahead)
should choose 70 percentof35, around 25. BUlifyou think all pLayers think
that way you hould choose 70 percent of 25, or 18.
In analytical gam theory, players do not stop this iterated re Ding
until they reach a be t-response point. But, since all players want to choose
70 percent of the average, if they all ch e the same number it must b
zero. (That i , jfyou solve the equation x* =0.7x*, you've found the unique
ash equilibrium.)
The beauty contest game provides a rough measure of the number of
steps of trategic think.i ng thal8ubjects are doing. It is called a "dominance-
solvable game" becau e it can be "solved"-i.e., an equilibrium can be
computed-by iterated applicat
one that yields a lower payoff th
of what other players do. Cho,
strategy because the highest po
you can always do better by choc
violates dominance by choosing
is 70 percent of70, or 49, so ChOI
obey one step of dominance. De
you to zero.
Many interesting games are
economics is Cournot duopoly. '
products to make. Since their I
determined by the total q u a n t i ~
is easy to show that there are ql
because flooding rhe market wil
cover fixed costs. Ifyou assume
somewhat lower quantities are I
!.his logic iteratively leads to a pl
In practice, it is unlikely th
steps of iterated thinking beca\.
(i.e., the amoUDt of informatio
one time). Consider embedded
mailman whose sister's boyfriCI
referred to at the end of the s
many people have forgotten wh
has only 0 much space. U Emhe
Dominance-solvable games are :
Iterated reasoning also reql
hard, and are thinking that you :
at a Caltech board of trustees m
known Ph.D. in finance) chose 1
the Nash equilibriumwas 0, but
was clever enough to do two ste
pick 17.5 (which is 70 percent 0
people who picked 17.5 or 18,
would pick high numbers, whil
good behavioral game theoryl (
What happens in beauty Cl
beauty contests with p=0.7 ~
II Seeing the SC1Jll:I'IGe on the wrltle.D ~
remember the rds and cannOt refer back ~
1 Introd1lction
1.2 Three Examples 17
players will do in these games by
oke is funny without telling it
;" and IteratedDominance
'if Employment, Interest, and Money,
arket and a newspaper contest in
guess are most beautiful: "It is not
est of one's judgment, are really
tge opinion genuinely thinks the
ree, where we devote our intelli-
cion expects the average opinion
10 practise the fourth, fifth, and
:e is perhaps no more apt than in
.after prices ofAmerican in[ernet
e largest speculative bubble in hi -
)f the e-taUer b okseller Amazon,
rth more than all other Am rican
ning Keynes had in mind is called
5, andHo Camerer, and Weigelt,
each of N players simultaneously
]. Take an average of the numbers
0.7). The player whose number is
lerage) wins a fixed prize. Before
u would pick.
:d to distinguish whether people
iegrees" of reasoning as Keynes
)y thinking, " uppose the average
osest to the target of70 percent of
yerswill think this way the average
urself (thinking one step ahead)
i. But ifyou think all players think
25, or lB.
, not stop thi iterated reasoning
t, ince all players ..vant to choose
ose the same number it mu t be
= O.7x, you've found the unique
~ o u g h measure of the number of
: doing. It is called a "dominance-
d"-i.e., an equilibrium can be
computed-by iterated application of dominance. A dominated slrategyis
ne that yields a lower payoff than another (dominant) strategy, regardle
of what other players do. Choosing a number above 70 is a dominated
strategy because the highest possible value of the target number is 70, so
you can always do better by choosing a number lower than 70. But ifnobody
violates dominance by cho ing above 70, then the highest the target can be
i 70 percent of 70, or 49, so choosing 49-70 is dominated ifyou think others
obey one step of dominance. Deleting dominated strategies iterativelyleads
you to zero. . , ,
Many interesting games are dominance solvable. A familiar example m
economics is ournot duopoly. 1\vo firms each choose quantities ofsimilar
products to make. Since their products are the same, the market price is
d termined by the total quantity they make (and by consumer demand). It
is easy to show that there are quantities so high that firms will lose money
becau flooding the market with a muchsupplywill drive prices too low to
cover fixed costs. Ifyou assu e your rivals won't produce that much, then
somewhat lower quantities are bad (dominated) choices for you. Applying
this logic iteratively leads to a precise solution.
In practice, it is unlikely that people perform more than a couple of
steps of iterated thinking because it strains the limits of working memory
(i.e., the amount 0 information people can keep active in their mind at
one time), Consider embedded sentences such as "Kevin's dog bit David's
mailman whose sister's boyfriend gave the dog to him." Who's the "him"
referred to at the end of the sentence? By the time you get to the end,
many p ople have forgotten who owned the dog because working memory
has only so much spac ,n Embedded sentences are difficult to under rand.
Dominance-solvable games are similar in mental complexity.
Iterated reasoning also requires you to believe that others are thinking
hard, and are thinking that you are thinking hard. When I played this game
at a Caltech board of tru tees meeting, a very clever board member (a well-
known Ph. . in finance) chose 18.1. Later he explained his choice: He knew
the ash equilibrium 0, but figured the average Caltech board member
was clever enough to do two steps of reasoning and pick 25. Then why not
pick 17.5 (which is 70 percent of25)? He added 0.6 so he wouldn't tie wiill
people who picked 17.5 or 18, and because he guessed that a few people
would pick high numbers, which would push the average up. Now that's
good behavioral game !.heory! (He didn't win, but was close.)
What happens in beauty contest ganles? Figure 1.3 shows choices in
beauty conte ts with P= 0.7 with feedback about the average given to
11 &eing the .etll.ence on the wnuen page makes it easier; Ir}' reading it aloud to somebody who must
remember the words and cannot refer back to them.
18 1 Introduction
1.2 Three Examples
Period
<1
1-10
11- 20
21- 30
31-40
41-60
51-60
61-70
71- 80 Choice
81- 90
91-100
(a)
100%
80%
60%
40%
20%
(t
Figure 1.3_ Convergence in lourstakes and high-stakes "beauty contest" games. aura: Un-
published data from Ho, Camerer, and W e i ~ l t .
subjeCl:5 after each often rounds (unpublished data from H Camerer, and
Weigel t). Bars show the relative frequency 0 choices in different number in-
tervals (on the side) across t n rounds (in front). The first histogra shows
results from games with low-s es payoffs (a 7 prize per period for seven-
person groups) and the second histogram shows re ull:5 from high-stakes
($28) payoffs.
First-round choices are around 21-40. A careful tatistical analysis indi-
cated that the median subject uses one or two steps of iterated dominance.
That is, most subjecl:5 roughly guess that the average will be 50 and choose
35, or guess that otherswill choose 35 and choose 25. Very few subjecl:5 cho e
the equilibrium of zero in the first round. In fact, they should not choose
zero. The goal is to be one step ahead of the average but no furthe and
choosing zero is being too smart for your own good!
Although the game-theore
1
initial choices, play rs are ill('
Behavioral game theory uses :
understand initial choice and
across rounds.
The beauty contest has bet
Chapter 5 for detail ), includ
12 Caltecl1.mdenls are a useful subjecl P'
YClU'$, the incoming 6rst-year dass bas a me<
the applicallU have been higher than the fl\
how the.e studeoIs play simple games C$tabl.
Generally they do n01 pla1 much differently
1 Introduction
100%
80%
60%
40'.
20%
<1
1- 10
11- 20
21- 30
3140
4150
51-50
70
Choke
,takes "beauty contest" games. Sr)U:rce: Un-
].2 Three Examples
PeriOd
1-10<1
11-20
21 30
31 40
41-50
51-60
61- 70
71 60 Choice
8190
91100
(b)
Figure 1.3 (continued)
60%
40%
20%
19
i hed data from Ho. Camerer, and
'of choices in different number in-
afront). The first histogram shows
S (a $7 prize per period for seven-
ill haws results from high-stakes
). A careful statistical analysis indi-
r two steps of iterated dominance.
the average will be 50 and choose
choose 25. Very few subjects chose
::I. In fact, they should not chao
f the average but no further and
. own good!
Although the game-theoretic equilibrium ofzero is a poor guess about
initial choices, players are inexorably drawn toward zero as they Jearn.
Behavioral gam theory uses a concept of limited iterated reasoning to
understand initial choices and a theory of learning to explain movement
across rounds.
The beauty contest has been replicated in dozens of subject pools (see
hapt r 5 for details), including Caltech undergraduates,12 trustees on
I! Caltech studenls are a useful subject pool because lhey are e>tll1lOrdinarily analytically skilled. In many
years. the incoming first-year clJL<S has a median math SAT score Df 800, Recently. lhe average ~ t Dres of
the appliwnJs have been higher than the average of those studenu who are t u a f I ~ d at HaTVard. Studying
how these tudents play simp.le games e.tabUshes whelher very analytica.l.tudenu can figure the games out.
Generally they do nDt play much differently lhw students at olher colleges.
20 1 Introductilm 1.3 Experimental ReguiLlriry and
the Caltech board (including a subsampJe of corporate CEOs) , econom-ics
Ph.D.s and game theorists, and readers of business newspapers (the F1'la1t-
cial Times in me United Kingdom, SfUktrum in Germany, and Expansrou in
Spain). The results in alll..hese groups aJ-e very similar: Players use 0-3 levels
of reasoning, and few subjects choose the Nash equilibl;um of zero. Com-
paring Figures 1.3(a) and 1.3(b) shows that increasing the prize by a fuctor
of four, leading to average earnings of$40 for a 45-minute experiment, has
only a small effect. (In me high-stakes condition there are more low-number
choices in periods 5--10).
The limited iterated reasoning measured in these games provides one
explanation for persistence of phenomena such as the stock price bubbles
Keynes had in mind. Even if all investors foresee a crash, they do not
"backward induct" all tJ,e way to the present. They guess !.hat others will
sell a couple of steps before the crash, and plan to sell just before iliat
exodus. This reasoning process does not unravel all the way (because doubt
"reverberates"), which explains why bubbles can persist even if eveq'one
knows tJley will eventually burst. Allen, Morris, and Shin (2002) make their
argument precise and Camerer and Weigelt (1993) and Porter and Smith,
(1994) show that bubbles can happen in the !ab.
1.3 Experimental Regularity and Behavioral Game Theory
This book is a long answer to a question game theory students often ask:
'This theory is interesting ... but do people actually play this .....ay?" The
answer, not surprisingly, is mix.ed. There are no interesting games in which
subjects reach a predicted equilibrium immediately. And there are no games
socomplicated that subjects do not converge in the direction ofequilihlium
(perhaps quite close to it) wilh enough experience in the lab.
Consider the three examples above. In ultimatum bargaining, players
are far from the perfect equilibrium-asswning selfinterest, but they are
roughly in equilibrium when !.he Responder's preference for being treated
fairly is tLl.ken into account (because offers maximize expected profit given
observed rejection rates). Behavioral game tJleory explains these resuhs by
combining new theories of social utility with analytical game !.heory (see
Chapter 2). In the continental divide and beauty contest games, players Start
far from equilibrium and converge close to it in ten periods or so. Behavioral
game theory explains these resuhs using concepts of limited reasoning as
players first think about a game (see Chapter 5) and precise theories of
learning (sec Chapter 6).
Sherlock Holmes said, '1)atLl., dabl "I cannot make bricks .....-iiliout clay."
Experimental results are day for behavioral game theory. The goal is not to
game theory (a common reaction of psychologists and sociolo-
gists) but to improve it by establ
Without some sort of obsenoat:
in casual pseudo-empirical WOT
office discussions and using on
Biologists don't just ask 1 I .....
do itr They watch robins forag
part-time experimenter) Eric V.
effects of having too few datLl. 01
Without having a broad st
certain danger of
ematically elegant, yet hav
presentour empirical know
von Neumann and Morgen
teresting question why gam
to psychologisLS for infonr
processing processes used 1
Data are particularly importar
more than one equilibrium (se(
not perfecliy understood (see (
solve these problems.
Why has empirical observ
until recently? One possibility
to have "failed". In a 1952 RA1
eventual Nobel laureate Nash)
also did some experiments, the
reportedly discouraged Nash a
in data also suffered from the
puzzles were open for solutio'
From about 1970 onward, devt'
games of incomplete informatit
as principal-agent relations, c
U I Ihiuk \hue urI)' mad
W":ltegy"'lwHbria. In th."., glmlU. h-a
tiotu about which are difficult 10 I
It M:any "modem-ideas in beh:>.1'ior1tI pn
and left &lidc: or forg<)tletl.ln hill thdit N:uh
ollt>'lar til m<J><kmllYOI"Uonary game: thc:or1
_,Ilt to tu.,.., been reYivc:d br emplricilUaTO
andlUiffit (l9S?I.5c:hen (1978) c:mphasiwl
RllbOmtdn (1991). and5an
a -""",,1>.1 rnpCllUC tqUillbrium" n:rDon. loa
and Coc:rec: arul Holl (1999).
1 Introduction
1.3 Experimental &gularity and Behaviqral Game Theory 21
e of corporate CEOs), economics
of business newspapers (the Finan-
U1n in Germany, and Expansion in
very similar: Players us 0-3 levels
e ash equilibrium of zero. Com-
lat inc easing the pl'ize by a factor
ofor a 45-minute experiment, has
dition there are more low-number
ured in the e games provides one
la such as the stock price bubbles
)rs foresee a crash, they do not
~ s e n t . They gue that others will
and plan to sell just before that
mravel all the way (because doubt
bies can p rsist even if everyone
orris, and Shin (2002) make their
;elt (1993) and Porter and Smith,
the lab.
I Behavioral Game Theory
game theory students often ask:
'ople a tually play this way?" The
iT no interesting games in which
nediately. And there are no games
-ge in the direction ofequilibrium
lCperience in the lab.
In ultimatum bargaining, player
uming self-interest, but they are
;ler's preference for being treated
'S maximize expected profit given
Ie theory explains these results by
with analytical game theory (see
:>eauty contest games, players start
) it in ten periods or so. Behavi oral
concepts of limited reasoning as
tapter 5) and precise theorie of
not make britks without clay."
al game theory. The goal is not to
tion of psychologists and sociolo-
gists) b t to improve it bye tablishiog regularity, which inspires new theory.
Without some sort of observation, theoretical assumption are grounded
in casual pseudo-empiri al work-informal opinion poll in seminar and
office dis u ions and using one's own intuitions (a one-respondent poll).
Biologists don't just ask "IT I a robin foraging for food, bow might I
do it?" They watch robins forage, or ask somebody who has. Th orist (and
part-time e.xperimenter) Eric Van Damme among others, worries about the
effects of having too few data of this sort in game theory (1999, p. 204):
Without having a broad set of facts on which to theorize, there is a
certain danger of spending too much time on models that are math-
ematically elegant, yet have little connection to actual behaviour. At
present our empirical knowledge is inadequate (precisely the same word
von Neumann and Morgenstern used fifty year before!] and it is an in-
tere ring question why ganle theorists have nOl turned more frequently
to psychologists for information about the learning and information
processing processes used by humans.
Data e parti larly important for game theory because there is often
rna e than one equilibrium (see Chapter 7) and howequilibration occurs is
not perfe tly understood (s e Chapter 6). Pw-e mathematics alone will not
solve thes problems.
Why h empiri al ob ervation played a small role in game theory
until recently? One po sibility is that early experimentation was thought
to have "failed". In a 1952 RAND conference, several theorists (including
eventual obel laureate ash) gathered to think about game theory. They
also did some experiments, the results ofwhich did not confirm theory and
reportedly dis ouraged Nash and perhaps others (Nasar, 1998).13 Interest
in data also suffered from the fact that so many interesting mathematical
puzzles were open for solution in game theory for such a long time.!4.-
From bout 1970 onward, developments in the theory ofrepeated games,
games of incomplete information, and applications to important fields such
as principal-agent relations, contracting, and political science led to an
13 J think lIle.., early experinlentet'$ made a mume by COlltefi0"3.ling lOO much on g;unes wiLh moce<:l-
.lntegyequilibria. In lIlo.., g;unes, players have lowmonetary incentives and predietiolUdepend Oil asromJ>'
rion. aboul';'\( tastes, which are.difficult to measure orCVtm contrOl.
I. Many "modem" ideas in behavioralgame theorywere firn.propo;red earlyin the historyofgame llleory.
and left aside or forgotten. In his thesis Nash (1950) cieKribed a "mass action" interpreCitioll ofequllibrium
similar to modern evolurionary game theory (Weibull, 1995). Weigh<ed fictitious play (see Chapler 6), which
seem. to have been 'vcd by empirici.1S around 1995, is described in the amazingly insigbtful book by Luce
andRaiffa (1957). Selten (1978) empbasized howplayers perceive theg;une they play. " topic beingnmved by
Rubins<ein (l99I), Cameter (1999), and Samodsoo (2001), amongotbcrs. Rosenthal (1989) Mtptoposcd
a "quanlal response equiJihriuoo"-version, larer relined and applied by McKt.\vey and Palfrey (1995, 1998)
and Goeree and Holl (1999).
22 1 Introduc:aDn
1. 3 Experimental Regularity and
explosion of theory. There is no doubt that this pursuit has been extremely
insightful and necessary, but it was conducted with little empirical guidance
of any sorl There i also .little doubt that it is high time to raise the ratio of
observation to theory. It is also encouraging th t some theorists have turned
serious attention to modeling bounded or pro edural rationality formally
(e.g., Rubinstein, 1998).1
5
Of course, experimental data are only on component of behavioral
game theory. Detailed facts about cognitive mechanisms and field tests are
important too.1
6
The result of controlled experiments, field observation
and theorizing working together is summarized b Vmce Crawford (1997,
p.208):
The experimental evidence suggests that none of the leading theoreti-
cal framework for analyzing games-traditional non-cooperative game
theory, cooperative game theory, evolutionary game theory, and adap-
tive learning mode1s--gives a fully reliable account of behavior by itself,
but that most behavior can be understood in terms of a synthesis of
ideas from those frameworks, combined with empirical knowledge in
proportions that depend in predictabl ways on the environment
Rapid development ofbehavioral game th orywill depend on howscientists
react to data. Reactions vary.
If you are smitten by the leganc of analyti al game theory you might
take the data as simply showing wheth r subjects u derstood the game and
were motivated. If the data confirmgame theory, you might say. the subjects
must have understood; if the data disconrm. the subje ts must have not
understood. Resist this conclusion. The games are usually simple, and most
experimenters carefully control for understanding by using a quiz to be sure
subjects knowhowchoices lead to payoffs. Furthermore, by inferringsubject
understanding from data, th re is no way to falsify the theory. Physicists and
biologists would not have th same reaction if a theory about were
fal.sified by carefuJ experimentation ("The particles were confus dt") or if
birds didn't forage for food as predicted ("If they had more at stake [than
survival?] theywould get itrightl"). arne theorists "hould be similarlyopen-
minded to what behaving humans can teach th m about human behavior.
In fact, evidence cited as confirmation of game theory often supportS
a key elemen t of behavioral game theory-namely, that equilibration roay
take a long time, perhaps years or decades (and equilibration is therefore a
crucial component ofany theory). In the foreword to Roth and Sotomayor's
15This includes finite 3utOm3ta, E-equllibrium. evolutionary and dynamic theories, non-partitional
infonn.3tion structures, and so nn. Most of this work il not directly inspired or disciplined by d3ta, however.
16Roth's work on matching for coliege bowl If.une., sorority rush. and medJca.l residency are rare,
impressive examples (e.g... Roth and Xing, J994).
(1990) book about the theory of
cian Robert Aumann notes that
the Gale-Shapley [matehir;
use already since 1951 for t
United tatesi it had evolve.
more than halfa century...
down, the payoff is not five
have time to understand th,
fare quite well.
Note that the "time to understa
yearsl
l7
Over such a span, a leal
Another reaction you may
design. Aumann, again, writes I
It is sometimes asserted tha
world," that people don'tre
scriptions. To back up sud
experiments using poorly J
derstand what they are
experiments represented tJ
Aumann is alluding to an earlie:
19705 which were not sensitiv(
This book largely ignores thos
in Chapter 3). The modern e:
from the past ten years--fully I
designed with them in mind. Su
students who are quizzed and I
Another reaction you are ill
to analytical game theory is t
than the experimenter create,
be tested and fal ified. Howev
bristle when they are made in
to ensure subject comprehen:
create games, and V3
for robusmess.
17 A similar point is made by Dixit and
batting averages converged in the 20th ce
hitting. Dilljt and Sk.ealh describe this as at
to play equilibrium strategies. But Ihe lea
\""-4l'l1ing theory i.just as imporW>t (or mO'
1 Introduction 1.3 Experimental Regularity and BeJuwioral Game Thetny 23
It this pursuit has been extremely
:ted with little empirical guidance
l is high time to raise the ratio of
g that some theorists have turned
,r procedural rationality formally
Iy one compone t of behavioral
Ie mechamsms and field tests are
I experiments, field observation,
Larized by Vmce Crawford (1997,
hat non of the leading theoreti-
traditional non-cooperative game
.utionary game theory, and adap-
able a count of behavior by itself,
rstood in terms of a synthesis of
led with empirical knowledge in
,Ie ways on the environment.
~ o r y w i l l depend on howscientists
malytical game theory you might
Jbjects understood the game and
:beory, you might ay, the subjects
firm, the subjects must have not
unes are usually simple, and most
;tandingby using a quiz to be sure
Furthermore, by inferring subject
.0 falsify the theory. Physicists and
III if a theory about particles were
e partides were confused!") or if
("If they had more at stake [tllan
:heorists hould b similarly open-
. h them about human behavior.
n of game theory often upports
-namely, that equilibration may
; (and equilibration is therefore a
oreword to Roth and Sotomayor's
ionary and dynamic theories, non'panitionaJ
tirectly inspired or disciplined by data, however.
IOrolity rush, and medical residency are HIre.
(1990) book about the theory ofmatehing markets, the brilliant mathemau...
ian Robert Aumann notes that
the Gale-Shapley [matching] algorithm had in fact been ill. practical
use already sine 1951 for the assignment of interns to hospitals in the
United S tes; it had evolv d by a triaJ..and-error process that panned
more than half a century.... in the real real world-when the chips are
down, the payoff i not fi: e dollars but a successful career, and people
have time to understand the situation-the predictions ofgam theory
fare quite well.
Note that the "time to understand the situation" Aumann refers to was fifty
yearsl
17
Over such a span, a learning or equilibration theory is essential.
Another r action you may have is to criticize det.ails of experimental
design. Aumann, again, writes (1990, p. xi):
It is ometime asserted thatgame theory is not "descriptive" of the "real
world," that peopl don't reallybehav according to game-theoretic pre-
scription. To back up su h assertions, some workers have conducted
experiments using poorly motivated subjects, subjects who do not un-
derstand what they e about and are paid off by pittances; as if such
experiments represented the real world.
Aumann is alluding to an earlier generation ofexperiments in the 1960s and
1970s which were n t sensitive to subject comprehension and incentives.
is book largely ignores those experiments (though some are described
in Chapter 3). The modern experiments described in this book-mostly
from the past t n y ars-fully respect concerns such as Aumann's and are
designed with them in mind. Subjects are typicallyanalytically skilled college
stud nts who are quiz.zed and highly motivat d.
Another reaction you are likely to have when behavior does not conform
to an.alyti al game theory is that subjects were playing a different game
than the experiment r created. ucll explanations are useful if they can
be test d and falsified. However, these explan.ations make experimenters
bristle when they are made in ignorance of me extraordinary care taken
to ensure SUbject comprehen 'ion, control for anonymity when trying to
create one-shot games, and variation in stakes and subject pool to check
for robustness.
17 A similar point is made by Dixit and Skcath (1999), StephenJay Could (1985) argued that baseball
batting averages converged in the 20th cenltlry because of dynamic adjustments in field, pitching, and
hitting. Dixit and keatb describe uus ,.. an "encouraging tale, drawn from rc:".u life, of bow players learn
to play equilibrium strdtegi(:$. But the learning wu on the order of decades, whidl mea1l5 a behavioral
learning theory isjust as important (or more $0) than an equilibrium concept.
24 1 Introduction
Ai.] Basi Game Theory
For example, a common interpretation of the fact that Responders
reject offers in ultimatum games is that the Re ponders think they might
be playing a repeated game be u e th ywill meet the Proposers again. But
experimenters go to great lengths t ensure that subjects won't meet again
and know that. For example, some exp rim ot rs paysubjectsoneata time,
with a short lag between each payment, and stand in the hall to be sure
ubjects don't wait for others to leave. Under these conditions, the faux-
repeated-game explanation of ultimatum re ults is simply wrong. Others
(such as the famously careful RayBattalio) are known to end an experiment
immediately ifa subject says something aloud that others hear. breaking the
experimenter's conu'ot The reaction that subjects are playing a different
game than the experimenter intended should disappear as mor theori ts
learn about what actually happens in lab ratorie and come to believe in
the quality of the data that are produc d.
Still another reaction you may have is that behavior which is not ra-
tion can't be modeled. For example. several year ago Abreu and Mat-
sushima (1992b) said experimental results are frequently in 'plicable by
"even approximately rational explanation." I d:isagree: Virtually all the re-
sults reported in this book can be accommodated by including b havioral
components-social utility, limited iterat d r asoning, and learning-into
analytical theory. They go on to ask, "Should we then give up the rationality
paradigm?" f course not. It is too useful as a ource of sharp predictions,
and it is often a good prediction of limitin behavior. Behavioral game t:he-
ory extends rationality rather than abandoning it. The last chapter of this
book hows how.
1.4 Conclusion
Thi chapter described three examples which illustrate experimental reg-
ularity, and hinted how that regularity is formalized in ehavioral game
theory.
In the ultimatum game, roposers typically offer close to half of a sum
to be divided, and Responders reject offers that are too low because they
dislike unfairness. The game is so simple that it is impossible to believe Re-
sponders rejecting money ar confused, and th result has been replicated
for very high stakes (up to 400 in Ameri a, and omparable sums in for-
eign countJ.ie ). According to beha 'oral game theory, Respond s reject
low offers because they like to earn money but di like unfair treatment (or
like being treated equally). In the continental divide game, players gravitate
toward equilibria over time andoften end up in P eto-inefficient equilibria
they could have avoided. Behavioral gam theory explains tho by assuming
that players aren't sure what to do (at the beginning of th game), so they
pick numbers in the middle; the
pIe statistical learning rules. In
do one or cwo steps of reasoninl
theory asswnes they keep goin
equilibrium.) And they learn c
results and describe other class
signaling. and auctions).
A
Al.I Be
This appendix introduces basic
the novice reader to understan
not have some other backgrou
understanding the experimenl
other books. A good introduce
(1999). More mathematical bOI
and Rubinstein (1995). Gintis (
ary theory and experimental ill
that are used ingraduate class s
and Tirole (1991).
Notation: Player i's strateg
for each player is denoted s =
removes player i's strategy {i.e.,
The utility of player j's payoffIi
Al
Deftniti6Tl Al.I.l The strategy s; ,
to any feasible strategy that the otht
The strategy I; is dominated ifther.
(
/I
'Uj 51' S.
18Thank& [0 Angela Hung for writing m
1 Introduction
AJ.l Basic Game Theory 25
ion of the fact that Responder
he Re ponders think they might
rill meet the Proposer again. But
re that subjects won't meet again
lenters pay subjects on at a time,
and tand in the hall to be sure
oder these condition the faux-
results is simply wrong. Others
are known to end an experiment
ud that other hear, breaking the
t subjects are playing a different
ould disappear as more theorists
)ratories and orne to believe in
is that behavior which i not ra-
:v ral years ago Abreu and Mat-
ts are frequently inexplicable by
." J disagre : VIrtually all the re-
nodat d by including behavioral
d reasoning, and learning-into
ld we then give up the rationality
as a ource of sharp predictions,
g behavior. Behavioral game the-
lDing it, The last chapter of thi
non
lich illustrate experimental reg-
formalized in behavioral game
ically 0 er dose to half of a sum
r that are too low becau e they
hat it is impo ible to believe Re-
ld the re ult bas been replicated
:a, and comparable urns in for-
game theory, Responder reject
but dislike unfair treatmen t (or
tal divide game, players gravitate
lp in Pareto-inefficient equilibria
theory explains this by assuming
beginning of the game). so uley
pick numbers in the middle; then they respond to history according to sim-
ple statistical learning rules. In the beauty contest game, seem to
do one or two steps of reasoning about others, then stop. (Analytical game
theory assume they keep going until they reach a mutual besc-response
equilibrium.) And they learn over time. Later. on
results and describe other class of games (rmxed equilibna, bargammg,
signaling, and auctions).
APPENDIX
AI.I Basic Game Theory
This appendix introduce b ic ideas in game theory.18 The goal is to equip
the novice reader to understand the gist of the rest of the book. If you do
not have orne other background in game theory, and are serious about
understanding the e perimental results described later, you should read
other books. A good introductory book (low on math) is Dixit and Skeath
(1999). More math matical books include Rasmusen (1994) and Osborne
and Rubinstein (1995). Gintis (1999) includes fresh material on evolution-
ary theory and experimental data. and (Ons of problems. The heavy tomes
that are used in graduate class at places such as Caltech includeFudenberg
and Tirole (1991).
Notation: Player i's strategy is denoted Sj. A vector of trategies, one
for ach player. is denoted S = S1, 'Sn' The part of this vector which
removes player i's strategy (Le., every other player'S strategy) is denoted S_j.
The utility of player its payoff from playing Sj is Uj(Sj. s_J.
A1.l.} Dominance
DefinitionAl.i.l The stratelfJ s1 is a dominant strategy ifit is a strict best response
to any feasible strategy that the others might play
The strategy s; is dominated if there exists s;' E Sj such that
18Thanks to Angela Hung fOr ,..,-iung much of wis appendix.