Introduction and Methodology

Mark Dean

Lecture Notes for Fall 2015 PhD Class in Decision Theory - Brown University

1 Introduction1

This is a course primerily designed to study behavioral economics through the lens of decision
theory. Decision theory means di¤erent things to di¤erent people - however most people would
probably go for a de…nition along the lines of “decision theory is the axiomatic development of
single person choice theory”.2 What does this mean? Well the middle bit is simple - decision
theory tends to deal with environments in which there is only one actor, so we don’t have to worry
about the strategic considerations that arise when several agents are interacting (this is the realm
of game theory). The last bit means that decision theory is in general concerned with developing
models of how people make choices. This is perhaps a little bit more controversial these days than
when it was written - partly because the line between what is a ‘choice’ and what is not is a bit
more blurred than it once was, and partly because some of the techniques that are commonly used
in decision theory have proved useful in analyzing non-choice data - such as fMRI data. We will
discuss some examples of this towards the end of the course.

What about the ‘axiomatic’ bit? This means that decision theory is, in general, concerned
with proving representation theorems. A representation theorem starts with a set of axioms (or
propositions) about the behavior of a data set, and shows that these axioms are equivalent to
1
For more on the methodology behind decision theory, It is worth reading the introduction to Notes on the Theory
of Choice. It is also worth reading "The Foundations of Positive and Normative Economics: A Handbook" edited
by Andrew Caplin and Andrew Schotter. In particular the article on ’The Case for Mindless Economics’by Gul and
Pessendorfer (also available on the web if you search for it)
2
Most people would agree with this because David Kreps says it - in the introduction to Notes on the Theory of
Choice

1
some model of decision making (the representation). This representation will generally include
some concepts that are not directly observable (such as utility), making it not immediately obvious
how to test this model. The role of the representation theorem is to link the model containing
unobservable features to a set of observable, testable axioms. It is probably easiest to demonstrate
what is meant by this using a couple of examples (which you are probably familiar with).

2 Two Familiar Representation Theorems

To discuss what decision theory is, and why it is powerful, it is going to be useful to have some
concrete examples to work with.

2.1 Example 1: Choice and Preference Maximization

Our …rst example is going to ask the following question: under what circumstances can we think
of a decision maker (DM) as a preference maximizer? In other words, when can the choices of a
DM be represented as resulting from the maximization of a complete, transitive, re‡exive binary
relation on X? This is an example that you should have come across before, and is an important
one, as almost all economics builds on the assumption that people maximize stable, well behaved
preferences.

In order to answer this question, we are going to prove a representation theorem: we are going
to write down a set of conditions, or axioms on the choices people make, and show that these
axioms are equivalent to the statement that people are preference maximizers: if a DM satis…es
these axioms then we can think of them as a preference maximizer. If they don’t then we cannot.
In order to do this, we are going to have to be more formal about what we mean by the various
parts of the above sentence.

First of all, choices over what? We are going to think of choices from subsets of a grand set
X. Initially, we will assume that X is …nite, and that an object in the set is just an object
- it doesn’t have any other characteristics. They could be fruit, musical instruments, stocks,
in…nite consumption streams, lotteries or anything else. All we are going to know about each
object is the label that identi…es it in the set. We will relax both these assumptions later in

2
the course.

What do we mean by choices? For now, let us imagine that we observe the choices that our
DM makes from every subset of X. We will represent these choices by a complete choice
correspondence C : 2X =? ! 2X =? such that C(A) A for all A 2 2X =?. Note that here we
are allowing the decision maker to choose more than one option from any given choice set.
This is a technically useful assumption (because it allows us to deal with indi¤erence), but
an observationally very dubious one - a point that we will come back to later. Notice also
that we are assuming that we observe our DM choose once (and only once) from each subset
of X. This is also a strong assumption, and one that we will again relax later.

What do we mean by a preference maximizer? Formally, a binary relation B is a subset of

X X. We will use the symbol to represent our binary relation, so x y means that
fx; yg 2 B. We demand that our decision maker behaves according to a binary relation that
has certain properties:

De…nition 1 A transitive binary relation is one in which x y z implies x z. A

re‡exive binary relation is one such that x x for all x 2 X. A binary relation that is
transitive and re‡exive is called a preference relation (or a preorder). A complete binary
relation is one for which, for any x; y 2 X, either x y or y x or both.

We will think of the complete preference relation as representing ’weak preferences’ - i.e.
x y means that ’x is at least as good as y’. Our behavioral model is that people have
preferences that are well behaved (in the sense of being complete, transitive and re‡exive),
and these preferences govern their choices. (It will become obvious why we think of such
preferences as well behaved).

Our question is, under what circumstances can we …nd some complete preference relation
such that choice is equal to the set of maximum objects in each set according to that preference
ordering: the DM chooses the best objects according to that preference ordering. In other
words, we want to …nd some such that, for all A 2 2X =?

C(A) = fx 2 Ajx y 8 y 2 Ag

Note that we are allowing for the possibility that two objects are indi¤erent - that x y and

3
 y x. In this case, and if both objects are preferred to all other objects in some choice set,
then we want both objects to be ‘chosen’(in some not very well de…ned sense).

We can now de…ne our problem more formally. The aim of our representation theorem is to
…nd some conditions on the choice function C such that we can …nd some preference relation that
rationalizes the DM’s choices (i.e. the DM chooses the best objects according to those preferences).
Note here that the concept of observability is crucially important. We assume that we can observe
choices but cannot observe preferences. If we could observe preferences then it would be easy to
test whether people are maximizing preferences - all we would have to do is look and see whether
the item they chose in each set was the most preferred item. Instead, we are completely agnostic
about what this preference relation is, we just want the DM to be behaving in a manner consistent
with some preference relation.

So what are the relevant conditions? At this point, assuming you have taken the graduate
micro class, you should be screaming ‘WARP’!3 . And you would be right. However, it is going to
be convenient for us to break WARP down into two pieces (as originally done by Armartya Sen):

Axiom 2 (Property ) If x 2 B A and x 2 C(A), then x 2 C(B)

Axiom 3 (Property ) If x; y 2 C(A), A B and y 2 C(B) then x 2 C(B)

It is worth stopping and thinking for a minute about these two axioms. The …rst is equivalent
to the independence of irrelevant alternatives, and is very intuitively appealing. It says that, if you
choose an alternative x from a larger set, then take some objects out of that set that are not x,
then you should still choose x from the smaller set. This is clearly a property you would expect a
‘rational’ decision maker to obey - if they choose x from the larger set, they are telling you that
they prefer x to all of the other objects in that set. They should therefore prefer x to any objects
in a subset of that larger set.

What about the second property? The …rst thing to note is that this property only has bite in
the case of choice correspondences. If C is single valued, then the condition x; y 2 C(A) can never
3
The Weak Axiom of Revealed Preference (WARP) states:

Axiom 1 (WARP) If fx; yg 2 A \ B, x 2 C(A), and y 2 C(B) then x 2 C(B)

4
hold, so this axiom will be satis…ed trivially. In the case of a choice correspondence, this condition
says that, if x and y are chosen from a set, and y is chosen from a superset of that set, then x must
also be chosen from that superset. Again, this makes sense if we think of a rational decision maker.
If x and y are chosen together, then the DM must be indi¤erent between them. If y is chosen from
some other set, then it must be at least as good as anything else in that set, and therefore so must
x. (Note, why does property restrict itself to B’s which are supersets of A? Surely this property
should hold for any B? What is going on here?)

You should convince yourself that, properties and between then are equivalent to WARP.

We are now in a position to state and prove our representation theorem. You are probably
familiar with the proof, but we will go through it again in order to highlight some points

De…nition 2 A choice correspondence C : S ! 2X =? for some S 2X =? is rationalized by a

preference relation if, for every A 2 2X =?, it is the case that

C(A) = fx 2 Ajx y 8 y 2 Ag

Theorem 1 For any …nite set X and complete choice correspondence C : 2X =? ! 2X =?, there
exists a complete preference relation that rationalizes that choice correspondence if and only if C
satis…es property and .

Proof. The …rst thing to do is note that this proof must come in two parts, as we are making two
claims: this comes from the fact that the statement is ”if and only if ’, so we have to show (i) that
and imply that we can …nd a rationalizing preference relation and (ii) any rationalizable choice
function satis…es and . We will start with the former, as this is the more tricky bit (in fact, we
have already argued informally for the latter.)

Proof (axioms imply representation). We will break the proof down into the following steps

1. Generate a candidate binary relation. Our claim is that, if the choice correspondence
satis…es and , then it is rationalizable by some complete preference relation. The …rst stage
of the proof is to describe such a relation, which we will then shows does the necessary job.
We will de…ne the relationship using choices from two objects we will say that x D y if and
only if x 2 C(fx; yg), so x is ‘weakly preferred’ to y (according to our candidate preference

5
 relation) if it is chosen from the set containing x and y only. We will stretch this de…nition
somewhat by saying that x D x, as x is de…nitionally chosen from the set fxg.

2. Show that D is a complete preference relation. So we have de…ned a binary relation.

Great. However, our theorem demands that choices be rationalized by a complete preference
relation - i.e. a complete, transitive, re‡exive binary relation. We next need to show that D
has these properties. Re‡exivity is easy - in fact we de…ned D explicitly so that it is re‡exive.
Completeness is also relatively straightforward. By de…nition, C(fx; yg) is either fxg, fyg
or fx; yg. Thus, by the construction of D either x D y, y D x or both. Finally, we need to
show transitivity, which we will do by contradiction. Imagine there exists x; y; z 2 X such
that x D y D z but not x D z. This implies

x 2 C(fx; yg)

y 2 C(fy; zg)

x 2
= C(fx; zg)

This in turn implies that z 2 C(fx; zg). We can now show that we must have a violation of
either property or property . Consider the set fx; y; zg. If x 2 C(fx; y; zg), then the fact
that x 2
= C(fx; zg) is a direct violation of property . If y 2 C(fx; y; zg), then by property ,
y 2 C(fx; yg) = fx; yg. Property then implies that x 2 C(fx; y; zg), which we have already
shown leads to a violation of . If z 2 C(fx; y; zg), then by z 2 C(fy; zg) = fx; zg, and so
by y 2 C(fx; y; zg). Again, we have already shown that this leads to a violation. However,
as C(fx; y; zg) is nonempty, one of these cases must occur, and so a failure of transitivity
implies a failure of either or :

3. Show that D rationalizes C. We now need to show that, for all sets, our DM chooses as
if they are maximizing D. In other words, for some arbitrary A 2 2X =? we need to show
that C(A) = fx 2 Ajx D y 8 y 2 Ag. As we are proving the equality of two sets, this in itself
takes two stages

(a) C(A) fx 2 Ajx D y 8 y 2 Ag. Say x 2 C(A). Take any y 2 A. We need to show that
x D y - in other words that x 2 C(fx; yg). However, this follows directly from property
. Thus, anything that is chosen from A must be ’preferred’ to everything else in A

(b) C(A) fx 2 Ajx D y 8 y 2 Ag. Say x D y 8 y 2 A. Then, x 2 C(fx; yg) for all y 2 A.
Now C(A) must be non-empty, so either x 2 C(A) (in which case we are done), or

6
 y 2 C(A) for y 6= x. By property , this implies that fx; yg = C(fx; yg), and so by
property , x 2 C(fx; yg)

This shows that properties and are su¢ cient for rationalizability

Proof (representation implies axioms). The next thing that we have to do is show the ‘only
if ’part of the statement - that rationalizability implies properties and . In other words, we have
to show that if there is a complete preference relation such that C(A) = fx 2 Ajx y 8 y 2 Ag,
then C must obey properties and

Property : x 2 C(A) implies that x y 8 y 2 A. Thus x y 8 y 2 B for any B A.

Thus, as, rationalizes choice from B, then it must be the case that x 2 C(B)

Property : If x; y 2 C(A), then x y. If y 2 C(B), then y z 8 z 2 B. Thus, by

transitivity, x z 8 z 2 B, and, as rationalizes choice in B, x 2 C(B)

This completes the proof of our …rst representation theorem.

One …nal thing to note. The preference relation that rationalizes a complete set of choice data
is unique. For completeness we will prove this claim as well:

Theorem 2 Let C be a choice correspondence that satis…es properties and . There is one and
only one preference relation that rationalizes C

Proof. The fact that there is such a preference relation we have already proved. We will prove
uniqueness by contradiction. Say 1 and 2 both rationalized C ,and 1 6= 2. Without loss of
generality, this implies that there exists an x and y such that x 1 y but not x 2 y. But the
former statement implies that x 2 C(fx; yg) while the latter implies x 2
= C(fx; yg), a contradiction.

2.2 Example 2: Preferences and Utility Maximization

The second example we are going to work through is another that you should be thoroughly familiar
with: Under what circumstances is it possible to represent ‘preferences’ with a numerical utility

7
function (note that the language here has become a bit di¢ cult as we have de…ned a ‘preference
relation’ already. What we really mean is: under what circumstances can we represent a binary
relation numerically). To put matters more formally, we want to …nd a utility representation for a
binary relation

De…nition 3 A binary relation on a set X has a utility representation if there exists a utility
function u:X ! R such that

u(x) u(y)

if and only if x y

for all x; y 2 X

Again, this is a pretty fundamental question, as almost all of economics uses utility functions,
rather than preference relations as their basis. This is because we have a load of cool tools to
work with utility functions and not very many cool tools to work with binary relations (though
note that most of these cool tools require the utility function to be di¤erentiable - something that
we will not say anything about at the moment). It is also one that you almost certainly already
know the answer to - the properties that we require (if X is …nite) are completeness, re‡exivity
and transitivity. Note that it is no coincidence that these are the properties that we used to de…ne
‘well behaved’preferences.

Before we proceed - note that we have changed our assumptions about observability. Here, we
are assuming that preferences are observable, but utility is not. We shall come back to this point
later.

Theorem 3 Let X be a …nite set. A binary relation on X has a utility representation if and
only if is a complete preference relation.

Proof. Again, we have two things to prove here, as this is an if and only if statement. Again, we
will begin by showing that the axioms imply the representation, which is the more di¢ cult direction.

Proof (axioms imply representation). We will proceed using induction on the size of the set
X. That is, we will show that (i) it is true for jXj = 1 and (ii) if it is true for jXj = n 1 then it

8
is true for jXj = n. The case of jXj = 1 is trivial (though note that it uses re‡exivity), so we will
move onto the second part of the proof. Let X be a set of size n, and let be a complete preference
relation on x. Remove object from the set X, which we will denote x . Now note that X=x is a
set of size n 1 and · Thus, there is a
induces a complete preference relation on X=x (yes?).
function v : X=x ! R such that v(x) v(y) if and only if x y. We will use this to construct a
utility function u on X. We will set u(x) = v(x) for all x 2 X=x . This utility function will clearly
represent on X=x in the sense that u(x) u(y) if and only if x y for all x; y 2 X=x . Thus,
all that remains to do is to is to set u(x ) and show that the utility function works here to. There
are 4 cases.

1. x x and x x for some x 2 X=x . In this case, we set u(x ) = u(x). Now, note that,
for any y 6= x

u(x ) u(y)

if and only if u(x) u(y)

if and only if x y

if and only if x y

The third line follows from the fact that x and y 2 x 2 X=x , and so by the inductive hypothesis
u represents the relationship between these two. The last line follows from transitivity. Using
the same technique it is possible to show that u(y) u(x ) if and only if y x

2. x y for all y 2 X. (for the next three cases we will assume that there is no x 2 X=x such
that x x and x x ). In this case we set

u(x ) = max v(y) + 1

y2X=x

Now, for any y 6= x, u(x ) > u(y) and x y. By assumption y x for no y 6= x

3. y x for all y 2 X: In this case, we set

u(x ) = min v(y) 1

y2X=x

Now, for any y 6= x, u(y) > u(x ) and y x. By assumption x y for no y 6= x

9
 4. There exists at least one y 2 X=x such that y x and z 2 X=x such that x z. In
this case, de…ne two sets: X = fx 2 X=x jx x g and X = fx 2 X=x jx xg. Note that
these two sets are disjoint (as we have ruled out the possibility that x x and x x for any
x 6= x ), and that, for any x 2 X and y 2 X , x y but not y x (x y follows directly
from transitivity. If y x, then x y x , which we have ruled out by assumption). This
in turn implies that
min v(x) > max v(y)
x2x y2X

We will therefore set

1 1
u(x ) = min v(x) + max v(y)
2 x2x 2 y2X

Thus, for any x 6= x

u(x ) u(x)

if and only if x 2 X

if and only if x x

Similarly

u(x) u(x )

if and only if x 2 X

if and only if x x

This completes the …rst part of the proof.

Proof (Representation Implies Axioms). This direction is relatively simple. Say that
is a binary relation on X and that u : X ! R is a utility representation of that function. Then
u(x) u(x) implies x x (re‡exivity), for any x; y either u(x) u(y) or u(y) u(x) implying
either x y or y x (completeness), and that x y z implies u(x) u(y) u(z), and so
x z (transitivity).

Finally, note that the utility function that can represent a complete preference relation is not
unique. It is unique only up to strictly increasing transformation. This means that, if the function

10
u represents a set of preferences, then the function v will represent the same preferences if and only
if v is a strictly increasing transform of u.

Theorem 4 Let u : X ! R be a utility representation for a complete preference relation . Then

v : X ! R will also represent if and only if there is a strictly increasing function T such that

v(x) = T (u(x)) 8 x 2 X

Proof. To show the if part, note that, if v is a strictly increasing transform of u then

v(x) v(y)

if and only if u(x) u(y)

if and only if x y

To show the only if part, note that if v is not a strictly increasing transform of u, then there
exists an x and y such that u(x) > u(y) but v(x) v(y). u(x) > u(y) implies that it is not the case
that y x. Thus, v does represent .

This uniqueness result is important, as it tells us how much information is in the utility function.
In this case, it is telling us that it is only the ordinal (ordering) information that is important -
that the utility number is bigger than another. The magnitude of those di¤erences are meaningless.
It is therefore meaningless to say things like ’the utility of x is twice that of y’, because we could
just as well use another utility function in which the utility of x is a million times that of y, or one
where the utility of x is 1% higher than that of y. Any utility function that preserves the same
ordering properties will do the job.

3 What Have We Just Done?

These are two important representation theorems, and the proofs contain some of the tricks that
you will see again in more complicated settings. However, they are also theorems that you have
probably come across before. One of the reasons that I wanted to put them on the table is so that
we can think about the structure that these theorems have in common. This is also going to allow
us to discuss two di¤erent philosophical approaches to decision theory. Both of these approaches
have the same …rst stage:

11
Stage 1: De…ne the primitive ’data set’. The …rst job in constructing a representation theorem
is to think about the properties of the observations to which you are going to apply your
axioms.

In the …rst case above, we took as our observations the choices made by the DM from di¤erent
sets of objects. We assumed that the objects were just objects - they didn’t have any other
characteristics. This is an assumption we will change later on. Instead, we may assume
that the objects of choice are probability distributions (lotteries), or consumption streams or
bundles of goods. We assumed that there was only a …nite number of them. We also assumed
that we observed DMs choose once from every subset of a grand choice set, but only once from
each choice set. Finally, we assumed that we could observe choice correspondences, rather
than just choice functions. All these assumptions played an important role in the nature of
the representation theorem that we eventually ended up with.

In the second example, the primitive of the representation function was the preferences of the
DM, rather than choices. This might seem a little puzzling - in what way can we think of
preferences as data? This is an issue that we will come back to. However, note again that we
have choices to make - for example again about the properties of the objects over which the
preferences were de…ned.

In most cases, the primitive data sets that decision theorists deal with either come in the form of
choices or preferences. However, there are many di¤erent variations. For example, one might
take as one’s primitive observation of choices from a choice set and a reference point (i.e.
rather than C(A), we consider C(A; z), the choice from A when the reference point is z.) Or
we might consider choices from a choice set after the decision maker has thought about the
problem for a certain length of time (i.e C(A; t), the choice from A after the DM has thought
about the problem for length of time t). We will consider both of these examples during the
course. One could also consider data sets of a completely di¤erent nature - for example one
model we will consider is on in which the data set is a function (z; p), which we interpret
as the amount of the neurotransmitter dopamine released when a prize z is obtained from a
lottery p. Any and all of these data sets are amenable to ‘decision theoretic’analysis.

What is the next stage of the theory? Well, at the end of the day, we are going to end up with
a representation theorem linking a set of axioms concerning the data set to a model of that data

12
set. The question is, which of these comes …rst? Do you start o¤ by thinking of a set of intuitively
plausible axioms, and then show that these axioms imply some model of behavior in that data set?
Or do you start o¤ with a model of what is going on in the data set, then …nd a set of axioms
that capture the behavioral implications of that model? I think that the traditional approach has
been the former: axioms come …rst, with the model being derived from those axioms. However, in
my view, the most useful way to use decision theory is the latter - it allows you to say something
concrete about the observable implications of your model. We will come back to this point below.
For now we shall simply note the two di¤erent approaches.

Stage 2 Version 1 - The Traditional Approach: De…ne a set of axioms These are a set of
statements concerning the data set. In the …rst case above our axioms were properties and
- simple, testable and intuitively plausible statements about how people make choices. In
the second, our axioms were completeness, transitivity and re‡exivity. Again, easily testable
and intuitively plausible statements about the primitive data set (in this case preferences).

Stage 2 Version 2 - The Alternative Approach: De…ne a behavioral model. The alter-
native approach is to next think about a plausible model to explain what is going on in our
data set. In our …rst example, the model is that people are making choices in order to maxi-
mize some well behaved preference relation. Notice that here we assume that preferences are
unobservable, so it is not immediately obvious how to test this model using our data set. In
the second case, our model is that people have preferences that are derived from the process
of utility maximization. Again, in this case we are assuming that utility is not directly ob-
servable (though preferences are), meaning that the observable implication of this model are
once again unclear.

Stage 3: The representation theorem. The next stage, whether one is coming from the tra-
ditional or the alternative approach is to prove a representation theorem. This is a theorem
that links together a set of axioms to a behavioral model. In our …rst example, we showed
that preference maximization is equivalent to properties and . In the second, we showed
that completeness, transitivity and re‡exivity were the same as the existence of a utility rep-
resentation. Note that, in both cases, these theorems are ‘if and only if’ - the axioms are
necessary and su¢ cient for the representation. This is the gold standard of these theories - it
means that the axioms are exactly equivalent to the behavioral model. In our …rst example, if

13
 properties and hold then there is some set of preferences that rationalize the choice data.
If the do not, there is no such preference relation. Thus, and are the exact observable
implications of the model of preference maximization.

Stage 4: Uniqueness result. In both our examples, we …nished by proving a uniqueness result.
In the …rst case we showed that there was only one preference relation that could rationalize
a set of choices. In the second case we showed that there were many utility functions that
could represent a set of preferences, but they would all be linked by strictly increasing trans-
formations. This is an important step of the process, as it tells us how ’seriously’to take our
representation. In our second example, we should take the ordinal information in the utility
function very seriously, but not the cardinal information.

4 Why Was This A Good Idea?

So now we have established what a representation theorem is: an equivalence result between a
set of observable axioms and a representation - a behavioral model that may rely on unobservable
elements. We also know what the 4 steps that are involved in developing a representation theorem.
What we have yet to cover is why this is an interesting thing to do. Again, we are going to discuss
two approaches The …rst, which I described as the ‘traditional approach’above, sees the axioms
themselves as interest. This could be for a number of reasons. - they could be seen as ‘self evidently
true’(i.e. axiomatic in the standard meaning of the word). They could be seen as justi…able as the
de…nition of rational behavior (e.g. someone who has intransitive preferences is by de…nition being
irrational). They could be justi…ed as capturing an essential element of a certain type of behavior
(Gul and Pessendorfer took this approach when they thought about temptation and self control.
They posited that the essential behavioral characteristic that de…ned temptation and self control
is that a person with self control issues may sometimes choose to restrict their own choice sets - i.e.
choose to have a smaller, rather than a larger choice set. This property they formalized as the set
betweenness axiom - allowing for the preference of smaller over larger choice sets in a structured
way.)

For all these reasons, people may be interested in the axioms that govern behavior. But this
does not explain why they might be interested in representation theorems. Why do they care what
sort of models are equivalent to these axioms? I think that there are a few reasons. One is that,

14
if (for example) you do believe that these axioms are capturing rational choice, then you might be
interested in what behavioral models are ’rational’. We have shown that utility maximization is
’rational’, but it turns out that other models will also lead to ’rational’outcomes. A second reason
is more practical - choice functions and binary relations are not very easy to work with, while utility
functions are! We have all sorts of mathematical tricks that we can use to …nd the maximum of
a utility function (i.e. the toolkit of static optimization) that just don’t work on binary relations.
Thus, if we believe that people are preference maximizers, it is very useful to know that we can
treat them as utility maximizers (though to use the power of our static optimization toolkit we also
need the utility function to be di¤erentiable, which is not guaranteed by anything we have done so
far, not least because we have only covered the case of a …nite choice set.) Similarly, it is useful
to know that my behavioral assumptions allow me to model people as expected utility maximizers,
exponential discounters etc. (though, again, to use these tricks, we need to know something about
the di¤erentiability of the resulting utility functions, something we have yet to say anything about).

While I can see some of the power of these arguments, this does not, in general, represent
the way that I use decision theory. Instead, I tend to go in the other direction. I start with
some behavioral model of how people make decisions, and use decision theory to understand the
observable implications of this model. In other words, I have some intuition about what sensible
decision making procedures, but I want some way to test whether this intuition is right. For
example, I may think that preference maximization sounds like a reasonable model of behavior,
but as a (social) scientist I would like to be able to test this. Due to the above result, I know that
means testing properties and :

Why do I need to go through the rigmarole of a representation in order to …nd the testable
implications? Because the models that we use to describe decision making tend to have unobservable
elements. Take for example the model of utility maximization. If I looked at objects and saw their
utility, then I wouldn’t need a representation theorem in order to derive testable implications - I
would just look and see whether people did in fact choose the highest utility object in each case
(consider a model of choice over amounts of money, where we assume that people choose more
money to less - we would not need a representation theorem to test this model! Or rather, any such
theorem would be trivial). However, because utility is not observable, it is not immediately obvious
how to test whether there is some utility function that rationalizes the data - i.e. that people are
acting as if they are utility maximizers.

15
 Is there an alternative? Yes, in fact there is. Take the model of utility maximization. Rather
than ask the question of whether there is some utility function that can rationalize our DMs choices,
we could make some assumptions about what that utility function should look like. For example,
if the objects of choice were lotteries over money, we could assume that people had a constant
relative risk aversion utility function, estimate the parameters of this utility function, and then see
how well the resulting estimated utility function explains the DM’s choices. In fact, this approach
is taken a lot by economists. What are the disadvantages of this approach? In my opinion there
are a few4 but here I would like to highlight two.

1. The resulting test is now a joint test of two hypotheses: that people maximize utility and
that utility is of the functional form that you have assumed. Moreover, if you think of more
general objects (such as teapots), it becomes very di¢ cult to see how one could sensibly
come up with a model that assigned utility based on the properties of that object (length of
spout?). In fact, there were lots of articles in early economics with titles such as ‘The Seven
Underlying Pillars of Utility’which tried to come up with general mappings from real world
properties to utility. This literature didn’t get very far.

2. The process of deriving the axiomatic representation of a model gives you a complete list of
the implications of that model. It therefore makes it very clear what behavior is and is not
in line with the model. This makes it very clear how to design tests of your model, and how
the implication of your model relates to those of other models.

Note, however, that this doesn’t mean that I think that one should only use axiomatic methods
to test models, just that they do provide a useful additional tool. Without performing this step, it
is very easy to get confused about what you are really saying when you write down a new model,
as the following example5 illustrates.

Example 1 There has long been evidence that people’s behavior is reference dependent, in the
sense that what people choose is a¤ ected by what their reference point is (the classic example of
this is the endowment e¤ ect - if people are given a mug and asked if they want to exchange it for a
4
Which I highlight in “Axiomatic Methods, Dopamine and Reward Prediction Error” (with Andrew Caplin),
Current Opinion in Neurobiology, August 2008, 18(2): 197-202
5
This example is stolen from “The Case for Mindless Economics” by Gul and Pesendorfer

16
chocolate bar then most will keep the mug. If people are given the chocolate bar and asked if they
want to exchange it for the mug, most will stick with the chocolate bar). Generally, when we try to
model reference dependent preferences, we assume that we know what the DM’s reference point is
in any given situation. However, this is a strong assumption, and a recent paper6 took a di¤ erent
approach. It assumes that people had utility functions of the form U : X X ! R, where U (x; z)
is the utility of choosing alternative x when z is the status quo. They further assume that people
will choose from a set those objects which form a personal equilibrium. That is, the objects such
that, if that object is the status quo, then it is the preferred object in the choice set. In other words

C(A) = fx 2 AjU (x; x) U (y; x) 8 y 2 Ag

We will call this the general personal equilibrium (PE) model. The paper also provides a speci…c
version of the model, which adds two assumptions:

1. U has the following functional form:

X X
U (x; y) = uk (x) + (uk (x) uk (y))
k2K j2K

where K indexes the hedonic dimensions of the various objects and is an increasing function
with (0) = 0

2. ‘Status quo bias’

U (x; y) U (y; y)

) U (x; x) > U (y; x)

This condition states that if x is at least as good as y when y is the status quo, then x must
be strictly better than y when x is the status quo.

We will call a general PE model that satis…es these two assumptions a special PE model

The concept of a personal equilibrium seems like an interesting way of modelling reference de-
pendence, and one that may be worth studying. However, from just looking at the above description
of the model, it is hard to tell what the behavioral implications of the model are, and what the
di¤ erence is between the special and general PE models.
6
"A Model of Reference Dependent Preferences" By Koszegi and Rabin - 2005

17
 Unfortunately, Gul and Pessendorfer show that (i) the general and speci…c PE model have the
same implications, and (ii) both are equivalent to dropping the assumption of transitivity from the
standard rational model - in other words, a choice data set allows for a speci…c (or general) PE
model if and only if it can be rationalized with a binary relation that is complete (but not necessarily
transitive) . This means that the PE models are not necessarily particularly interesting unless you
have a richer data set.

We will now state this result in the form of a proposition

Proposition 1 Let C : 2X =? ! 2X =? be a choice function on a …nite X The following statements

are equivalent

1. (General PE model): There exists a general PE utility function U : X X ! R such that

C(A) = fx 2 AjU (x; x) U (y; x) 8 y 2 Ag

2. There exists a complete, re‡exive binary relation such that

C(A) = fx 2 Ajx y 8 y 2 Ag

3. (Special PE model) There exists a special PE utility function U : X X ! R such that

C(A) = fx 2 AjU (x; x) U (y; x) 8 y 2 Ag

Proof. The proof comes in three parts

Proof (1 implies 2). Say that C admits a general PE model. De…ne as x y if and only
if U (x; x) U (y; x). This must be complete, as, for any x; y, if C is a choice function then either
x 2 C(fx; yg) or y 2 C(fx; yg), and this implies that either U (x; x) U (y; x) or U (y; y) U (x; y).
Furthermore, note that, by assumption

C(A) = fx 2 AjU (x; x) U (y; x) 8 y 2 Ag

= fx 2 Ajx y 8 y 2 Ag

18
Proof (2 implies 3). Let n = jXj. Let x y indicate the asymmetric part of , and let
K = X X (i.e., the set of hedonic states are indexed by the cross product of X) For k 2 (w; z) 2 K,
de…ne the utility function
u(w;z) : X ! f 2; 0; 2; 3g

as 8 9
>
> 3 if x = y = z >
>
>
> >
>
>
> >
>
< 2 if x = w and w z =
u(w;z) =
>
> 2 if x = z and w z >
>
>
> >
>
>
> >
>
: otherwise ;

De…ne as follows 8 9
< 16nt if t 2 f 4; 3; 4g =
(t) =
: t if t 2 f 2; 0; 2; 3g ;

You should check that (t) satis…es the necessary properties.

Let
X X
U (x; y) = uk (x) + (uk (x) uk (y))
k2K j2K

Next, we need to show that U (x; x) U (y; x) i¤ x y. To see this, …rst note that
X
2n uk (x) 2n
k6=(x;x)

And second note that, if we de…ne Kx;y = K= f(y; y); (x; x); (x; y); (y; x)g and note that for
k 2 Kx ;y
2 uk (x) uk (y) 2

Thus, it must be the case that

4n
X
uk (x) uk (y)
Kx;y
X
= (uk (x) uk (y))
Kx;y
4n

19
 Now assume that x y. This implies that (ux;y (y) ux;y (x)) 0 and (ux;y (y) ux;y (x)) 0.
Thus,

U (x; x) U (y; x)
X X
= (uk (x) uk (y)) (uk (x) uk (y))
k2K k2K
X
4n (uk (x) uk (y)
Kx;y

ufx;xg (x) ufx;xg (y)

ufy;yg (x) ufy;yg (y)

8n + 48n 3>0

Now, if y x, then (ux;y (y) ux;y (x)) = 0 and (ux;y (y) ux;y (x)) = 64n, so the above
becomes

U (x; x) U (y; x)
X X
= (uk (x) uk (y)) (uk (x) uk (y))
k2K k2K
X
4n (uk (x) uk (y)
Kx;y

ufx;xg (x) ufx;xg (y)

ufy;yg (x) ufy;yg (y)

(ux;y (y) ux;y (x))

8n + 48n 3 64n < 0

So we have the representation. The …nal thing that we need to check is that U (x; y) U (y; y) 0
implies that U (x; x) U (y; x) > 0. This follows from

U (x; x) U (y; x)

U (x; x) U (y; x) U (x; y) + u(y; y)

X
= (uk (y) uk (x)) + (uk (x) uk (y))
k2K
= 2 ( ( 3) + (3)) = 2(48n 3) > 0

Proof (3 implies 1). This is immediate

20
 In summary, representation theorems provide an important link between the models that we
have in our head (which often include latent, or unobservable variables) and the data on which we
will test these models. They tell us precisely what the testable implications of our models are for
any given data set, allowing us to understand whether a model is testable on a particular data set,
and whether two models do in fact have di¤erent implications. Thus, in my opinion, they play a
crucial role in the interaction between economic theory and data.

It is only fair to mention that there are issues in using axioms to derive testable implications of
our models. Axioms provide a very stark test: either the axioms hold, in which case the model can
explain the data, or they do not, and so it cannot. Thus, one mistaken choice, or one incorrectly
recorded outcome is enough to discard the entire model. Given that any actually data that we use
will invariably include many instances of both, we will presumably be in a situation in which we
will have to reject all feasible models that we come up with. While this is a serious issue, it is not
an insurmountable one: people are currently developing techniques to determine whether axioms
are ’approximately’correct, which we will discuss in a later class.

I …nish o¤ this section by adding two comments.

1. What makes a good set of axioms? Kreps suggests two technical requirements. One is
consistency - it should not be the case that there is no possible data set that can satisfy all
of the axioms. Obviously a set of axioms that fail this property are not very interesting! The
second is non-redundancy - it should be the case that, for each axiom in the set, it should be
possible to have a data set that satis…es all the other axioms in the system, but fails that
one. If this is not the case, then at least one of the axioms in the system is redundant, and
can be dropped.

In top of this, there are two less precise conditions that I would highlight. The …rst is testa-
bility. Given my emphasis on using decision theory to unpack the observable implications
of models of behavior, I think it is crucial that the axioms one comes up with should be
testable using the available data. The second property is what I will loosely term intuitive
content. A good axiom is one that intuitively captures the ‡avor of a particular concept.
For example, the set betweeness axiom of Gul and Pesendorfer really does seem to capture
something about temptation.

A further point is that, in order to get decision theorists excited, a set of axioms should have

21
 ‘emergent properties’. That is, it should be surprising that a (strong looking) representation
comes from a relatively weak set of axioms.

2. Is decision theory technically complicated? The short answer to this is ‘yes’. Decision
theoretic papers tend to be some of the more technically complicated in economics, requiring
an understanding of real analysis, functional analysis, measure theory etc. However, much of
the technical di¢ culty comes when one tries to extend representations from the case of …nite
X to in…nite (and, speci…cally uncountable X). Though we will do some of this during this
course, it is not in fact the area of decision theory I …nd most interesting. This is because I am
interested in testability, and the type of technical axioms required to support these extensions
are usually inherently untestable (e.g. some form of continuity). Thus, we will only cover this
stu¤ in a somewhat limited way.

22