Let us examine a second paradox.

In the Chain Store Game, a Dutch monopolist called Alban Heym

has branches in 20 towns. The chain faces 20 potential competitors, one in each town, who can choose
“enter” or “stay out.” They do so sequentially and one at a time. If a potential competitor chooses “stay
out,” he receives a payoff of 1, while the monopolist receives a payoff of 5. If he chooses “enter,” he
receives a payoff of either 2 or 0, depending on the monopolist’s response to his action. The monopolist
must choose between two strategies, “cooperative” or “aggressive.” If the monopo- list chooses the
former, he and the competitor receive a payoff of 2, and if he chooses the latter, each player receives a
payoff of 0. The last round of the game is depicted in Figure 4.15.

The subgame-perfect equilibrium is easy to see. Entering into a price war, Alban Heym can only lose.
The competitor knows this, and since he has a higher payoff when the market is shared than when he
stays out of it, he will enter. (“enter,” “cooperative”) is the equilibrium reached by backward induc- tion
in this game.

But the managers of Alban Heym think they can outwit the compet- itor. They reason that if Alban
Heym demonstrates being tough by playing “aggressive” in early rounds, potential entrants will be
deterred and the chain store can reap higher profits (cf. Selten 1978). A game theorist might respond:
“Look, your threat is not credible. You will most certainly not enter a price war in the final round—there
is nothing to gain from it. Thus, the competitor will enter for sure. In the penultimate round, there is again
no reason to fight. It will be costly, and it has no effect on the final round. Continue to reason thusly, and
you will see that deterrence is not a valid strategy.”

Indeed, the managers of Alban Heym made a mistake. If “aggressive” is a preferable strategy in early
rounds of the game, then this should be reflected in the payoffs. The managers analyzed a different game,
not the Chain Store Game. However, one can show that if there is a small amount

Competitor

Stay out

Enter

(1, 5)

Alban Heym

Co-operative

Aggressive

(2, 2)

(0, 0)

Figure 4.15 The Chain Store Game

Game Theory 73

of uncertainty about the payoffs, it may be rational for a monopolist to build up a reputation by fighting
entry initially. This leads to another refinement of the subgame-perfect equilibrium called sequential
equilibrium (Kreps and Wilson 1982).
 The idea is that at least one player’s profile is determined exogenously and unobservable to the other
player. In the Chain Store Game, for instance, the monopolist may be “weak” or “strong” depending on
whether his preferences are as in Figure 4.15 or the converse, as on the right-hand side of Figure 4.16.
The game is solved by finding for a player i, at each information set after which i is to play, the best reply
to the strategies of the other players. In this game, Alban Heym can build a reputation of being tough and
thereby deter potential entrants from trying to enter the market.

In each case considered above, the “refinement” of the Nash equilibrium consisted in adding a
structural feature to the game which changed its nature. In no case was the refinement justified on the
basis of considerations of what would be reasonable in a situation to do. One consequence is that the
refinements often do not reduce indeterminacy but rather add to it. Both the Centipede and the Chain
Store Game show that in order to settle on specific equilibria one has to weaken common knowledge
assumptions, which means that one is moving away from attempting to defeat indeterminacy.

Game theory, understood as a theory of rational decision-making, is thus highly problematic. The Nash
equilibrium is ill-justified. Even if it were justified, it would solve few problems because most games
have multiple Nash equilibria. Thus far, the refinement program has produced few results that can be
defended from the point of view of rationality and that have helped to reduce indeterminacy.

Nature

p 1–p

Competitor

Competitor

Stay out

Enter Stay

Enter out

Alban Heym

Co-operative

Aggressive

Figure 4.16 The Chain Store Game with Uncertainty

Alban Heym (1, 5)

(1, 5)

Aggressive Co-operative

(2, 2)

(0, 0) (2, 0) (0, 2)

74 Rationality
Game Theory as Explanatory Theory It has sometimes been argued that game theory, as theory of rational
choice, would be a very good candidate for an empirical theory of social phenomena (see for instance
Grüne-Yanoff and Lehtinen 2012). A theory which predicts that people act rationally is self-fulfilling.
People who accept the theory and therefore predict that other people behave rationally have an incentive
to act rationally themselves. A theory which predicts that other people behave irrationally does not have
this benefit. Agents acting on such a theory have all the more reason to deviate from the theory because
doing so will improve their performance. Rational-choice theories therefore have a stabilizing influence
on social phenomena.

But of course this argument has any bite only to the extent that game theory is successful as a theory of
rationality. As we saw in the last section, it is not. Not all is lost yet, though. Perhaps game theory isn’t so
good a theory of rationality but it might nevertheless be a useful predictive and explanatory theory.
Perhaps the justification of playing a Nash equilib- rium strategy in this or that refinement is wanting, but
if people play the strategies the theory predicts anyway, who cares? As long as the theory has empirical
content, or, more carefully, as long as it is useful for mode- ling empirical phenomena, it may have its
virtues. In this section we will see that even as explanatory or predictive theory, game theory is very
problematic.

In order to make predictions (only a theory that predicts empirical phenomena can also explain them),
any theory must in one way or another be brought to bear on empirical phenomena. Most genuine theories
tran- scend the realm of the observable, i.e., they contain terms that refer to unob- servable states of
affairs. If the arguments that were given here concerning the nature of preferences are correct, one of the
core elements of game theory refers to something unobservable. Carl Hempel called the principles that
connect a theory’s theoretical (or unobservable) vocabulary with an observa- tional vocabulary “bridge
principles” (Hempel 1966). We need bridge prin- ciples for a theory to have empirical content.

The architecture of games is given by their payoff structure. As we saw in the previous section, the
payoffs are utility indices indicating preference ranking. This is unfortunate. If they indicated material
outcomes, the game theorist could straightforwardly determine which game is being played in a given
situation because material outcomes are observable (perhaps not liter- ally, but for all intents and
purposes of the game theorist). The usual story that goes along with what I have called the Eurobonds
Game is one about a proposal made to two prisoners, which is why the game is normally known as
Prisoner’s Dilemma:

In the Prisoner’s Dilemma, two prisoners ... are being interrogated separately. If both confess, each is
sentenced to eight years in prison; if both deny their involvement, each is sentenced to one year. If just
one confesses, he is released but the other prisoner is sentenced to ten years. (Rasmusen 2006: 20;
footnote suppressed)

The corresponding matrix could be written as shown in Figure 4.17.

Prisoner 2

Confess Deny

1renosirP

Confess Deny (8 years, 8 years) (0, 10 years) (10 years, 0) (1 year, 1 year)

Figure 4.17 The Prisoner’s Dilemma in Game Form

 The advantage of presenting games in this way—that application to empir- ical situations is easier—is
frustrated by the fact that game theory could not make any predictions without knowledge of players’
preferences. In some cases, what people prefer may be relatively straightforward, such as here. It is very
reasonable to assume that almost everybody has the following prefer- ence ranking: acquittal > 1 year > 8
years > 10 years. In other words, it is reasonable to assume that utility is strictly decreasing by number of
years in prison and therefore that the Eurobonds Game of Figure 4.2 is an adequate transformation of the
game form of Figure 4.17 (for the notion of a game form, see Weibull 2004).

What is reasonable to assume in one case should not be blindly accepted as a rule more generally. In
other words, one should allow the utility function:

U = U(M),

where M designates the material outcomes of a game, to vary between people and across situations.
Although in early experimental applications of game theory, subjects were assumed to care only about
their own material gains and losses, this is clearly not a substantive hypothesis of game theory as such. As
game theorist and experimenter Ken Binmore remarks:

Actually, it isn’t axiomatic in economics that people are relentlessly selfish. ... Everybody agrees that
money isn’t everything. Even Milton Friedman used to be kind to animals and give money to charity.

(Binmore 2007: 48)

76 Rationality

Thus, institutions, social and cultural norms and other immaterial facts may affect people’s valuations of
the material outcomes of a game. It is by no means obvious for instance how people rank the different
outcomes of the Ultimatum Game (Figure 4.4, above). If a strong fairness norm is at work, players might
rank (material) outcomes as follows (5, 5) > (0, 0) > (6, 4) > (4, 6) > (7, 3), etc., which is not at all strictly
increasing in material outcomes.

There is now a growing literature on what functional form U might take. For instance, in Fehr and
Schmidt’s theory of fairness, the utility function for two-person games has the following form (Fehr and
Schmidt 1999: 822):

U i (x) = x

max{x

−x

i
, 0} −

max{x

−x

j , 0}, i ≠ j,

where the x’s are player i’s and j’s monetary payoffs and and are param- eters measuring how
disadvantageous and advantageous inequality affects a player’s utility. In Cristina Bicchieri’s
formulation, social norms are explicit arguments in the function (Bicchieri 2006: 52):

U i (s)=π

i (s) − k

max

−j

max

m≠j

{π

(s

−j

, N j (s

−j

)) − π

(s), 0},
where s = (s

≠L

−j for player i relevant norm. , and s

2 , A ..., k i norm ≥ s n

) 0 is is a a strategy constant profile, representing π

i (s) is the a material payoff function player’s sensitivity to the for player i is represented S for i

, where the other L −i

ӬS

players.

−i

,S

is player i’s strategy set by the function and S

−I

the set of strategy N i profiles : L

–I

Generally speaking, though, economists are loath to make any substantial assumptions about people’s
utility functions. They instead believe that one can learn people’s preferences in one situation and use that
knowledge to make predictions about what the same people will prefer in another situa- tion. Learning
about what people prefer in choice situations is called “pref- erence elicitation.” Often this is done by
having experimental subjects play a subgame of a game of interest. Suppose we are interested in eliciting
the players’ preferences in the Eurobonds Game. First we have to write down the game in game form with
material outcomes. For simplicity, let us assume that the main material outcome is growth rates. In
extensive form, the game could then look as shown in Figure 4.18.

To elicit the South’s preferences, we have it choose over strategies in the subgames (Figure 4.19).

To elicit the North’s preferences we simply switch roles (we can do so because the material payoffs are
symmetrical). If both North and South prefer “profligate” to “frugal” in both subgames we know that
Figure 4.2 is a correct rendering of the game, and we can use game-theoretic tools to solve it and make a
prediction.
 The problem with this elicitation procedure is that it assumes that pref- erences are quite strongly
context-independent. What that means, and why this is not always a reasonable and safe assumption to
make, can be

Game Theory 77

illustrated with the following game, called the “Trust Game” (Figure 4.20). Here an investor can choose
between keeping his investment or transferring it to a trustee. If he does the latter the money is
quintupled. The trustee then decides whether to keep the money or return half of it to the investor.

Profligate

North

South

Figure 4.18 The Eurobonds Game in Game Form

Profligate

Profligate

Frugal

Frugal

Profligate Frugal

(1%, 1%) (5%, 0%) (0%, 5%)

(3%, 3%) South South Frugal

Profligate Frugal

(1%, 1%) (5%, 0%) (0%, 5%) (3%, 3%) Figure 4.19 Two Subgames of the Eurobonds Game in Game Form

Investor

Keep money

Invest

(€10, €10)

Trustee

Keep money

Return investment

(0, €50)
(€25, €25)

Figure 4.20 The Trust Game in Game Form

78 Rationality

Consider the subgame beginning at the second decision node. On its own this subgame is in fact a
version of the so-called Dictator Game in which a player decides whether or not to split a pie between
herself and a second player. Unlike in the Ultimatum Game (recall Figure 4.4), in the Dictator Game the
second player does not have the option to “reject,” and the first player therefore does not have to fear
punishment. It is thus in the subgame of the Trust Game. We can then expect people to choose similarly
in both games, and indeed, the assumption that people choose similarly is implicitly made in using this
elicitation procedure.

Suppose an individual chooses “Keep money” in the subgame when played on its own. Is it reasonable
to expect her to do the same in the full game? When the full game is played, other norms may affect
players’ decisions than when only the subgame is played. If I am trusted I may wish to reward the trusting
partner by returning his investment, even at a material cost to me, and even when the investor has no way
to punish me for a selfish decision. Or I may act on equity considerations. Unlike in the Dictator Game, in
the Trust Game the investor helps to produce the pie, and I might therefore think that she deserves a
return or that I am obliged to pay her back because it is her money.

Be the normative considerations as they may, subjects do in fact choose differently in the Dictator
Game and in the Investment Game (see for instance Cox 2004). This means that their preferences over
outcomes are not context-independent. An aspect of the relevant context in this case is whether another
player has moved first and thereby helped to create the opportunity to “reward” or “pay back.” Without
context independence, this particular elicitation procedure is invalid.

There are other problems with this procedure. Francesco Guala points out that it cannot be used to test
games in which reciprocity matters (Guala 2006). That this is correct is not hard to see: if my being
“nice” (in the sense of moving in such a way as to give a co-player a higher payoff than I could have done
had I chosen differently) depends on my co-player’s being “nice”; i.e., if my preferences depend on my
co-player’s preferences, one cannot learn about these preferences in situations where other players do not
exist.

It is important to note, however, that game theory is not wedded to this particular elicitation procedure,
or any other procedure for that matter. Weibull 2004, for instance, proposes for a slightly more
complicated version of a Dictator Game in which four outcomes have to be ranked that “the
experimentalist asks subject A to state her ordinal ranking of the four plays for each of the 24 possible
(strict) ordinal rankings that B may have, and likewise for subject B.” The experimenter could then find a
matching pair, inform each player about the other’s preference and make a prediction on that basis.