Lecture 12: Testing GARP,

aggregating demand, etc.

1 Where are we?

• Thursday, we proved some more properties of Hicksian demand;
saw the Slutsky Equation, which lets us infer the derivatives of Hicksian demand from Mar-
shallian demand,
and therefore gives us a way to empirically test our model of utility maximization;
and we saw the condition that finite data needs to satisfy, GARP,
to be consistent with profit maximization

• Today, I’ll say a little more about Afriat’s theorem,

a little about empirical tests of GARP,
and hit a few other topics to wrap up consumer theory

• Thursday and next Tuesday, we’ll talk choice under uncertainty

• First... any questions?

196
2 Rationalizing Finite Data (GARP)
• We wrapped up last lecture with Afriat’s Theorem:1

Theorem (Afriat’s Theorem). Let (pi , xi ), i = 1, 2, . . . , n be a finite number of observations of

price vectors and consumption bundles. The following conditions are equivalent:

1. There exists a locally nonsatiated utility function that rationalizes the data

2. The data satisfy GARP

3. There exist numbers {U i , λi }ni=1 , λi > 0, such that for every {i, j},

Ui ≤ U j + λj pj (xi − xj )

4. there exists a LNS, continuous, concave, monotonic utility function that rationalizes the data

• So, for any finite set of observations,

if they satisfy GARP, they’re compatible with rational choice;
and if they fail GARP, they’re not

1
For the proof, see Varian (1982), “The Nonparametric Approach to Demand Analysis,” Econometrica 50(4), or
the papers by Afriat that he cites.

197
 • Why do we need GARP, not just WARP?

• That is, isn’t the pairwise relation enough, like in producer theory with WAPM?
Why do we need to extend it with transitivity?

• Example:2

prices demand
(1, 1, 2) (1, 0, 0)
(2, 1, 1) (0, 1, 0)
(1, 2, 1.5) (0, 0, 1)

• The first observation implies (1, 0, 0) % (0, 1, 0);

but we never see (0, 1, 0) chosen when (1, 0, 0) is affordable, since at the second observation,
(1, 0, 0) is outside of the budget set,
so we never get a direct contradiction

• The second observation implies (0, 1, 0) % (0, 0, 1);

but again, we never see (0, 0, 1) chosen when (0, 1, 0) is affordable

• Finally, the third observation implies (0, 0, 1)  (1, 0, 0),

but we never see (1, 0, 0) chosen when (0, 0, 1) is affordable

• So we need transitivity to make the argument –

we know the data can’t be rationalized,
and they fail GARP, but they don’t fail WARP

• (With the producer problem, the firm was always choosing from the same production set as
prices changed,
so a pairwise comparison of choices was enough.
With the consumer problem, a consumer has the same preferences across observations,
but both prices and wealth vary,
so in some sense the consumer is facing different problems at each price vector)

2
Borrowed from Ichiro Obara’s lecture slides from UCLA

198
• Afriat’s theorem gives us a funny second result: if data can be rationalized at all,
it can be rationalized by a utility function that’s monotone and concave

• Now, if u is concave, then it’s quasi-concave,

and if u is quasi-concave, then the preferences it represents are convex

• So this is saying, any data that’s consistent with rational choice,

is consistent with monotone, convex preferences

• So... does this mean rational preferences have to be monotone and convex?

• No, of course not

• You can have whatever preferences you like –

they may be monotone, they may be convex, they may not be

• What this is saying is that with observed choice data,

we’ll never be able to detect whether your preferences have these properties or not

• Whatever choices you make,

we could find monotone and convex preferences that would have led you to the same choices

• Let’s see why

• First, suppose you had non-monotone preferences

• That is, suppose there were two bundles x and x0 , with x0  x but x % x0

• How would we observe this? We would need to see you choose x when x0 was also available

• But that means finding a consumer problem (p, w) such that x ∈ x(p, w) while p · x0 ≤ w

• But since x  x0 , if p · x0 ≤ w, p · x < w

• And we know that with LNS preferences, you’ll never choose x when it doesn’t exhaust your
budget

• So the exact observation we would need to learn that your preferences are not monotone, is
impossible to get if your preferences are LNS

• (To put it another way, the non-monotonicity in your preferences would be in a region where
you weren’t choosing anyway, so we’d have no way to observe it from choices)

199
 • Similarly, suppose you have non-convex preferences

• Convex preferences say that if x % y and x0 % y, then xt = tx + (1 − t)x0 % y as well

• So if preferences are not convex, there must be some x, x0 , y, and t such that
x % y, x0 % y, and y  xt

• Or by transitivity, x  xt and x0  xt

xt
x’

x x

• Now, let’s think about what we’d need to observe to figure out that both x  xt and x0  xt
xt xt
x’ that x  xt , we’ll need a data point where x is chosen, even though xt is in
x’ • First, to learn
the interior of the budget set

• And second, to learn that x0  xt , we’ll need a data point where x0 is chosen, even though xt
is in the interior of the budget set

x x x x

xt xt xt xt
x’ x’ x’ x’

• But in the first case, if x is on the budget line and xt is in the interior, then x0 is in the
interior too
So we can only observe x being chosen if x  x0

• And likewise, we can only see x0 chosen in the second case if x0  x

• So with LNS preferences, there’s simply no way to generate the data we would need to learn
that your preferences aren’t convex

• They might be, they might not be, we just won’t know

200
3 Testing GARP
• Christopher Chambers and Federico Echenique (2016), Revealed Preference Theory, Cam-
bridge University Press (Econometric Society Monographs)
ch. 5 is all about “practical issues” of testing GARP,
and an overview of the empirical literature

• One challenge is that GARP is a binary condition –

however many observations you have, your data collectively either satisfy it or don’t;
and in a rich dataset, you always expect some noise,
so you would expect it to be violated

• There’s a need, therefore, to find a way to quantify whether violations are significant –
either statistically, or economically

• Afriat proposed an index, now called Afriat’s Efficiency Index,

to measure “how close” a consumer was to satisfying GARP,
basically saying how much we need to “deflate” an individual’s choices by to satisfy GARP,
with 1 being data that satisfies GARP already, and lower numbers meaning larger violations

• Others have modified this index

• Chambers and Echenqiue propose a different idea, the “money pump index,”
which measures the severity of a GARP violation by how much money someone could make
by arbitraging your “irrational” choices

• (If your choices violate GARP, it’s because you sometimes “overpay” for a bundle you could
have bought cheaper a different time;
they measure how much money someone could make basically buying this bundle cheap and
selling it to you expensive,
swapping around your choices)

• Under some assumptions, they turn this measure of “how irrational” you are into a statistical
test

201
• Another issue is power

• People often use cross-sectional data,

looking at observations of different households at the same time

• In those cases, wealth (or expenditure) varies a lot more than prices

• This means budget sets are usually nested,

and it’s almost impossible for GARP to be violated,
so using GARP as a test has very little power

• I’ll mention one solution in a minute

• A few interesting empirical papers they mention:

– Dowrick and Quiggin (1994), “International Comparisons of Living Standards and Tastes:
A Revealed-Preference Analysis,” American Economic Review 84.1
cross-country comparisons (60) using 1980 national consumption data –
consistent with GARP, seven consumption categories
– Landsburg (1981), “Taste Change in the United Kingdom, 1900-1955,” Journal of Po-
litical Economy 89.1
time series data, weakness of test given rising income
– Blundell, Browning and Crawford (2003), “Nonparametric Engel Curves and Revealed
Preference,” Econometrica 71.1
cross-sectional data, introduced Engel curve adjustment to solving rising-income problem

• And a few experimental ones:

– Andreoni and Miller (2002), “Giving According to GARP: An Experimental Test of the
Consistency of Preferences for Altruism,” Econometrica 70.2
(testing whether overly generous behavior in a “dictator game” can be explained by
“rational” altruistic preferences – basically, yes)
– Harbaugh, Kraus and Berry (2001), “GARP for Kids: On the Development of Rational
Choice Behavior,” American Economic Review 91.5
(everyone’s pretty close to rational; 11 year olds are more rational than 7 year olds, and
about as rational as college-age adults)
– Choi, Kariv, Müller and Silverman (2014), “Who Is (More) Rational?” American Eco-
nomic Review 104.6 – linking “rationality” to socioeconomics
(men, well-educated, and rich choose “more rationally”;
higher Afriat efficiency index is a predictor of higher household wealth)

202
4 Aggregating Demand
• Other than the experimental papers, most of the papers above use aggregate data –
they observe total consumption for a group of individuals (maybe even an entire country),
but test for rationality as if it were the demand of a single individual

• So... is this a valid exercise?

• Should we expect demand from a group of rational individuals to satisfy GARP,

as it would if it were the choices of a single rational individual?

• In producer theory, we saw that production aggregates –

if each firm is maximizing profits based on its own production set,
you showed on the first homework of the semester,
it will look just like profit maximizing behavior of a single firm.

• Does the same thing hold for consumers?

• The answer is, generally no, but yes in a few special cases.

• For general utility functions, individual consumers can be perfectly rational,

but generate data that is not rationalizable at the aggregate level

• However, demand does aggregate in two special cases,

which you’ll be exploring on the next homework

• (The problem is summed up well in the intro to the Blundell/Browning/Crawford paper:

“...If we do reject revealed preference conditions on aggregate data,
we have no way of assessing whether this is due to a failure at the micro level
or to the inappropriate aggregation across households that do satisfy the integrability condi-
tions but who have different nonhomothetic preferences.”)

203
 • The reason demand doesn’t aggregate has to do with varying wealth

• In producer theory, each firm had a fixed production set,

and prices were varying across observations

• In consumer theory, each individual has fixed preferences,

but prices and individual budget levels can vary across observations

• But because Marshallian demand has wealth effects,

aggregate demand depends on the distribution of wealth, not just total wealth,
which wouldn’t be picked up in aggregate data

• Easiest to understand if we think about two consumers,

one who loves steak, one who loves sushi

• Suppose we have two observations

• At the first observation, steak guy had a lot of money, sushi guy didn’t,
so aggregate demand was for a lot of steak and a little sushi

• At the second observation, steak guy had a little money, sushi guy had a lot,
so aggregate demand was for a lot of sushi and a little steak

• But that’s basically independent of prices

• If the price of steak was a little higher on the first day,

and the price of sushi was a little higher on the second day,
aggregate demand would appear irrational –
the two consumers together demanded lots of steak when steak was expensive,
and lots of sushi when sushi was expensive,
violating GARP – even though both consumers were perfectly rational3

3
We can generate a similar counterexample even with consumers who have identical preferences. Consider two
individuals with utility functions u(x1 , x2 ) = x101 + ( x102 )2 . You can verify that optimal behavior is to spend one’s
10p2
whole budget on good 2 if w > p12 and one’s whole budget on good 1 otherwise. Consider two observations, one
at prices p = (10, 9) when the consumers have budgets w1 = 80 and w2 = 70; and a second at prices (9, 10) when
the consumers have wealth w1 = 120 and w2 = 30. Aggregate demand would be (15, 0) at the first observation and
(3 31 , 12) at the second, which we can confirm would fail GARP.

204
• Again, this contrasts with producer theory –
on the very first homework, you all proved that if each individual firm is maximizing profits,
industry data will be rationalizable as well

• The difference is that the consumer problem has wealth effects –

consumers behave differently at different wealth levels –
which means that the distribution of wealth among consumers, not just the aggregate wealth
level, matters

• At our two observations, total wealth was the same;

but in this case, consumers rationally demand good two only when they’re sufficiently wealthy,
so shifting the distribution of wealth shifted demand,
in a way that was inconsistent with GARP

205
4.1 so when does demand aggregate?
• So, when does demand aggregate?
when can we aggregate individual demand choices,
and get something consistent with a single person’s choices according to rational preferences?

• Since the “problem” we just saw had to do with wealth effects,

it’s not surprising that if we eliminate wealth effects, we can get demand to aggregate

• There are two ways we can do this:

1. give each consumer quasilinear utility (not necessarily the same quasilinear utility)
2. give each consumer identical, homothetic preferences
(preferences are homothetic if they “scale up and down” –
if x % y if and only if λx % λy for any λ > 0)

• There’s a more general condition, but these are the two main cases that satisfy it.

Proposition. Suppose there are n consumers, and consumer i has indirect utility function

vi (p, wi ) = ai (p) + b(p)wi

Then aggregate demand

n
X
X = xi (p, wi )
i=1

is the same as the demand of a single consumer with indirect utility function

n
!
X
V (p, W ) = ai (p) + b(p)W
i=1
P
when W = i wi .

• vi = ai (p) + b(p)wi is called the “Gorman Form”

• note that ai can differ across i, but b(·) cannot

• So basically, if your consumers have indirect utility functions which are linear in wealth and
all have the same slope,
then you can add up all their demand, treat them like a single person, and everything works

• But otherwise, you generally can’t

206
• The two special cases I mentioned give indirect utility functions that take this Gorman form,
and they’re both preferences you’ll work with on Homework 6

• Homothetic preferences are preferences where x % y if and only if λx % λy for every λ > 0

• You’re showing that homothetic preferences can be represented by a utility function which is
homogeneous of degree 1, i.e., u(λx) = λu(x)

• If utility is homogeneous of degree 1, indirect utility turns out to be linear in wealth,

v(p, w) = ṽ(p)w

• So if all your consumers had identical preferences, and those preferences were homothetic,
then indirect utility takes the Gorman form and demand aggregates

• And you’ll also show that this holds for quasilinear utility,
since quasilinear preferences lead to an indirect utility function

v(p, w) = a(p) + w

meaning consumers don’t even have to have the same quasilinear preferences;
so as long as each consumer’s preferences are quasilinear, demand aggregates

• But those are the main cases where this works;

otherwise, individual wealth effects create enough of a problem that “market demand” doesn’t
have to behave like the choices of a single, rational representative consumer

207
5 Recovering Preferences from x(p, w)
• We said earlier that if Marshallian demand is homogeneous of degree 0,
satisfies Walras’ Law, and gives a Slutsky matrix that’s symmetric and negative semidefinite,
then it’s consistent with rational choice

• So if we observe Marshallian demand x(p, w) at every (p, w),

and it satisfies these conditions,
can we recover preferences that rationalize it? Yes!

• Recall first that Marshallian demand is determined by preferences, not utility –

the solution to the consumer problem is the same for any u representing the same preferences

• So if we start with demand, the best we can hope to do is recover ordinal preferences,
not a unique cardinal utility function

• Which leaves us with a degree of freedom – we can normalize cardinal utility however we
want, since we’re going to recover the same preferences however we do it

• So let’s fix a reference price vector p0 ,

and normalize utility such that v(p0 , w) = w, or e(p0 , u) = u

• (That is, since there’s no natural scale for utility,

we can pick a reference price vector, and just define the cardinal level of utility as the amount
of money it takes to afford it at prices p0 )

• Now fix a value of u; for each good i, we know that

∂e
(p, u) = hi (p, u) = xi (p, e(p, u))
∂pi

• Since xi (·, ·) is a known function (it’s observed),

this is a differential equation where the unknown variable is e

• And since we can do this for each good i,

we have a system of k differential equations in e,
plus the initial condition e(p0 , u) = u

208
 • It turns out that symmetry of the Slutsky matrix is exactly the condition that this system
has a solution (also known as “integrability” conditions);
and negative semi-definiteness guarantees that the solution e is concave (which it has to be)

• Since we can do this separately for each value of u,

we’ve now got a concave expenditure function e, defined for all p and u,
that’s consistent with our observed Marshallian demand

• So, if we start with Marshallian demand, we can recover an expenditure function e(p, u) that
generates it

• Once we know the expenditure function e(p, u), we can recover preferences from that

• MWG do it one way, very much analogous to how we constructed the outer bound Y O from
the profit function in producer theory4

• An alternative I prefer5 is, once we have the expenditure function, to recover the indirect
utility function as the solution to

e(p, v(p, w)) = w

• And then once we have v, define the utility function by

u(x) = min v(p, 1) subject to p·x=1

p0

• The logic is that if x is optimal at prices p and wealth w = 1,

then p · x = 1 (by Walras’ Law) and u(x) = v(p, 1);
and for any other price vector p0 such that p0 · x = 1,
x is feasible, so v(p0 , 1) ≥ u(x)

• So for any x that’s optimal for some price level, u(x) must be minp:p·x=1 v(p, 1)

• (For x that are never chosen, this may be optimistic relative to the truth –
there could be x that are never chosen that have u(x) < v(p, 1) for every p with p · x = 1 –
this just gives a utility function that’s consistent with the observed choices.)

4
Assign utility by u(x) ≥ u if and only if p · x ≥ e(p, u) for every p  0. The proof this works is on MWG p 77
5
from Varian (1992), Microeconomic Analysis (3rd ed.), pp 129-130

209
• Since we’re inferring preferences from choices,
there’s no guarantee we match the original preferences on bundles that never get chosen;
but we are guaranteed that we’ve found preferences that generate the Marshallian demand
function we started with.

• (Since we’re doing all this under the assumption that Marshallian demand is single-valued
and differentiable,
we’re more or less thinking about the case where preferences are strictly convex.
Like with production, if preferences are strictly convex,
then every strictly-positive bundle will sometimes be optimal,
and so we’ll fully recover the “right” preferences this way.
If preferences have some non-convexities, we can never learn about preferences over the bun-
dles that are never optimal,
all we can do is find a set of preferences that would rationalize the observed Marshallian
demand.)

210
6 Welfare effects of price changes

Skipped in lecture – please read on your own!

(Just some definitions you’re supposed to know)

• We’ll wrap up consumer theory thinking about measuring the welfare effect of price changes

• So, some big policy thing happens, and it results in some price changes from p0 to p1

• How do we measure the results?

• In some sense, what people care about is

v(p1 , w) − v(p0 , w)

the change in utility they can achieve

• But that can’t give us a sensible answer, because it depends on cardinal utility –
different utility functions representing the same preferences give different answers,
and we know we have no way to infer peoples’ cardinal utility from choices

• Instead, we can focus on the change in expenditure functions

• That is, if we can calculate

e(p1 , ·) − e(p0 , ·)

we can say, how much more or less money does it take now,
to achieve a particular level of utility?

• This is based purely on preferences – not the utility representation –

so it’s a better-posed question;
and since the answer is in dollars, it allows us to add up across people

• Of course, there’s still the question of what target utility level to use,
and there are two obvious candidates – the original one, and the new one

211
6.1 Compensating Variation
• Let’s consider a “good” change – a shift to lower prices, meaning it’s now cheaper to afford
the same utility level as before, or we can now achieve a higher level than we could prior

• Compensating variation measures how much more cheaply we can afford the old level of
utility now: if we let u0 = v(p0 , w) and u1 = v(p1 , w), then

CV = e(p0 , u0 ) − e(p1 , u0 )

• CV also answers the question, now that prices have changed,

how much money can we take away from you, and leave you as well off as you were before

• Since e(p0 , u0 ) = w, we can rewrite the previous as

w − CV = e(p1 , u0 )

so it follows that
v(p1 , w − CV ) = v(p1 , e(p1 , u0 )) = u0

so after a good price change, this is how much money you could lose and still be as well-off
as you were before the change6

6.2 Equivalent Variation

• Equivalent Variation measures how much more cheaply you can now afford the new level
of utility, relative to what it would have cost you at the old price level

• That is,
EV = e(p0 , u1 ) − e(p1 , u1 )

• Or, how much money could we have given you instead of the price change,
to make you equally happy?

• That is, since e(p1 , u1 ) = w,

w + EV = e(p0 , u1 )

and therefore
v(p0 , w + EV ) = v(p0 , e(p0 , u1 )) = u1

6
Since e(p0 , u0 ) = w = e(p1 , u1 ), CV can also be written as CV = e(p1 , u1 ) − e(p1 , u0 ), because we’re again
measuring how much money could you give up right now (at the new prices) and still afford your old lifestyle.

212
6.3 CV versus EV
• We can think of CV as your maximum willingness to pay to “buy” a price change –
you could receive the price change, give up CV, and still be as well off as before

• EV might be more appropriate to measure how much better off you are when a change has
already happened –
“this change had the same utility benefit as if you had given me $50” –
while CV might be easier for considering lots of alternative policies, since it uses the common
baseline u0

• To see the difference more clearly, think about a dramatic example – how much has quality
of life improved since the Middle Ages

• Life expectancy was short – partly because a lot of things we take for granted today weren’t
available, like antibiotics and clean water

• If you were rich enough, you ate well, but there wasn’t a lot on TV

• We can think of prices of many as having come down a lot – perhaps even from infinity –
if we fix wealth at, say, an upper-middle-class level

• Compensating Variation asks, if you’re upper-middle-class today, how much money could you
give up, and still afford the quality of life of someone upper-middle-class in the Middle Ages

• Probably an awful lot – you could be pretty poor today, and still have a two-generations-old
iPhone, penicillin when you need it, and an apartment that stays warm in the winter,
but there should be some amount of money we could take away from you, to the point where
you’re destitute and starving,
and make you no better off than you would have been in the Middle Ages

• Equivalent Variation asks, if we went back in time to the Middle Ages, how much money
would we have had to give you, to let you afford your current 2019 level of utility

• Quite possibly infinite – there’s no amount of wealth in 1400 that would have bought you an
iPhone, or a Netflix account, or a life expectancy over 60
(The average life expectancy of English and Scottish kings from 1000 to 1600 was ∼ 50 years)

• As we’ll see, for normal goods, Equivalent Variation tends to be larger than Compensating
Variation

213
6.4 Changes in just one price
• Suppose that only the price of good i changes, with p1i < p0i

• We can calculate compensating variation as

Z p0i Z p0i
∂e
CV = e(p0 , u0 ) − e(p1 , u0 ) = (pi , p−i , u0 )dpi = hi (p, u0 )dpi
p1i ∂pi p1i

• And we can do the same with equivalent variation:

Z p0i Z p0i
0 1 1 1 ∂e
EV = e(p , u ) − e(p , u ) = (pi , p−i , u1 )dpi = hi (p, u1 )dpi
p1i ∂pi p1i

• Now, since price went down, we know u1 > u0

• We can show that if good i is a normal good (xi increasing in w), then hi is increasing in u,
and so EV > CV

• And if good i is an inferior good, then hi is decreasing in u, so EV < CV

214
6.5 One more measure – consumer surplus
• So, for changes in a single price, CV and EV can be calculated by integrating under the
Hicksian demand curve for that good, one at the old utility level and one at the new utility
level

• Of course, just for the hell of it, we could integrate under Marshallian demand instead;
we call that Consumer Surplus

• Consider the following, which shows the demand just for good i as a function of pi

(This diagram is for a normal good; for an inferior good, things would be reversed)

• This, I think, initially appealed because we can observe Marshallian demand but not Hicksian
demand, so we have a better shot at a direct calculation

• If i is either a normal or an inferior good, xi will be between the two Hicksian curves, so CS
will be between the other two, and maybe a reasonable estimate for either

• And if there are no wealth effects, then h(p, u) doesn’t depend on u, and EV, CV, and CS
would all be identical

• But consumer surplus is an interesting measure in its own right

215
• Consumer Surplus is a standard measure of how much extra value consumers get from a single
product – basically, the difference between the price they pay and the price they would have
been willing to pay

• If we think of the demand curve for one good as the aggregate demand, each point on the
demand curve represents a single consumer with that exact willingness to pay

• The area between a horizontal price line and the demand curve, then – the area above the
demand – is the consumer surplus

• And when a price goes down, the change in CS is clear visually

• Two points we already mentioned:

1. If good i is a normal good, then

EV ≥ ∆CS ≥ CV

so we can use EV and CV to put bounds on CS, or we can use CS to approximate either
of them
2. If preferences are quasilinear, then there are no wealth effects (Hicksian demand for good
i is the same at u1 as it was at u0 ), and so

EV = CV = ∆CS

216