Demand Functions, Income E¤ects and Substitution E¤ects:

Theory and Evidence

David Autor
14.03 Fall 2004

1 The e¤ect of price changes on Marshallian demand

A simple change in the consumer’s budget (i.e., an increase or decrease or I) involves a parallel shift of
the feasible consumption set inward or outward from the origin. This economics of this are simple. Since
px
this shift preserves the price ratio py ; it typically has no e¤ect on the consumer’s marginal rate of
Ux
substitution (MRS), Uy ; unless the chosen bundle is either initially or ultimately at a corner solution.

A rise in the price of one good holding constant both income and the price of other goods has economically
more complex e¤ects:

1. It shifts the budget set inward toward the origin for the good whose price has risen. In other words,
the consumer is now e¤ectively poorer. This component is the ‘income e¤ect.’

2. It changes the slope of the budget set so that the consumer faces a di¤erent set of market trade-o¤s.
This component is the ‘price e¤ect.’

Although both shifts occur simultaneously, they are conceptually distinct and have potentially di¤erent
implications for consumer behavior.

1.1 Income e¤ect

First, consider the “income e¤ect.”What is the impact of an inward shift in the budget set in a 2-good economy
(X1 ; X2 ):

1. Total consumption? [Falls]

2. Utility? [Falls]

3. Consumption of X1 ? [Answer depends on normal, inferior]

4. Consumption of X2 ? [Answer depends on normal, inferior]

1
1.2 Substitution e¤ect

In the same two good economy, what happens to consumption of X1 if

p1
"
p2

but utility is held constant?

In other words, we want the sign of

@X1
Sign jU =U0 .
@p1
X1
Provided that the axiom of diminishing MRS applies, we’ll have p1 jU =U0 < 0.

In words, holding utility constant, the substitution e¤ect is always negative.

By contrast, as we established above, the sign of the income e¤ect is ambiguous,

@X1
? 0;
@I

depending on whether X1 is a normal or inferior good.

6.1#1
y

I/py
This section of budget
set becomes unfeasible

-(px1/py)
-(px2/py)

I/px2 I/ px1 x

1.3 Types of goods

The fact that the substitution e¤ect is always negative but the income e¤ect has an ambiguous sign gives rise
to three types of goods:

2
 @X @X
1. Normal good: @I > 0; @px
jU =U0 < 0: For this type of good, a rise in its price and a decline in income
have the same e¤ect–less consumption.
6.1#2
y

Income effect
Normal good
I/py

S
C2 Substitution effect
C1

U1
U2

I/px2 I/px1 x

Although we only observe the movement from C1 to C2 , we can conceive of this movement as having two
parts: the movement from C1 to S (substitution e¤ect) and the movement from S to C2 (income e¤ect).

@X @X
2. Inferior good: @I < 0; @px
jU =U0 < 0: For this type of good, the income and substitution e¤ects are
countervailing. Why countervailing? Even though both derivatives have the same sign, they have opposite
e¤ects because a rise in price reduces real income–thereby increasing consumption through the income
e¤ect even while reducing it through the substitution e¤ect.

3
 y 6.1#3

I/py
Inferior good
S

C1

C2

I/px2 I/px1

Here, the substitution e¤ect is the S C1 and the income e¤ect is C2 S.

@X @X
3. Strongly inferior good (‘Gi¤en’ good). @I < 0; @px
jU =U0 < 0: Similar to a conventional inferior good,
the income and substitution e¤ects are countervailing. But what’s special about a Gi¤en good is that the
income e¤ect dominates the substitution e¤ect (in some range). Hence, a rise in the price of a Gi¤en good
causes the consumer to buy more of it–demand is upward sloping. Even though a price increase reduces
demand due to the substitution e¤ect holding utility constant, the consumer is e¤ectively so much poorer
due to the income loss that her demand for the inferior good rises.
y 6.1#4

I/py
Giffen good
S
C1

C2

I/px2 I/px1

4
 The notion of a Gi¤en good is a theoretical curiousity. It’s hard to imagine a case where when the price
of a good rises, demand increases. But theory says such goods can exist. We’ll look at evidence on this in
the Jensen and Miller paper.

Question: The price of gasoline rises just about every summer, as does the gallons of gas consumed per
household. Is gas a Gi¤en good?

1.4 Marshallian and Hicksian demand

Alfred Marshall was the …rst economist to draw supply and demand curves. The ‘Marshallian cross’ is the
staple tool of blackboard economics. Marshallian demand curves are simply conventional market or individual
demand curves. They answer the question:

Holding income and all other prices constant, how does the quantity of good X demanded change with px ?
We notate this demand function as dx (px ; py ; I): Marshallian demand curves implicitly combine income
and substitution e¤ects. They are ‘net’demands that sum over these two conceptually distinct behavioral
responses to price changes.
6.1#5
y

I/py
Marshallian demand

U1

U2
U3

I/px3 I/px2 I/px1 x

One can also conceive of a demand curve that is composed solely of substitution e¤ects. This is called
Hicksian demand (after the economist J. R. Hicks) and it answers the question:

Holding consumer utility constant, how does the quantity of good X demanded change with px : We notate
this demand function as hx (px ; py ; U ): The presence of U as a parameter in the Hicksian demand function
indicates that this function holds consumer utility constant–on the same indi¤erence curve–as prices
change. Hicksian demand is also called ‘compensated’demand. This name follows from the fact that to

5
 keep the consumer on the same indi¤erence curve as prices vary, one would have to adjust the consumer’s
income, i.e., compensate them. For the analogous reason, Marshallian demand is called ‘uncompensated’
demand.
6.1#6
y

Hicksian demand

U0

I/px3 I/px2 I/px1 x

1.5 Relationship between Compensated and Uncompensated demand

These two demand functions are quite closely related (as show below). But they are not identical.

Recall from the prior lecture the Expenditure Function,

E(px ; py ; U );

which is the function that gives the minimum expenditure necessary to obtain utility U given prices px ; py :

For any chosen level of utility U , the following identity will hold:

hx (px ; py ; U ) = dx (px ; py ; E(px ; py ; U )):

In other words, for any chosen level of utility, compensated and uncompensated demand must equal
to one another. Another way to say this: Fix prices at px ; py . Fix utility at U . Use the expenditure
function to determine the income I necessary to attain utility U given px ; py . It must be the case that
hx (px ; py ; U ) = dx px ; py ; I .

Although these demand curves cross (by construction) at any chosen point, they do not respond identically
to a price change. In particular di¤erentiating the prior equality with respecting to px yields the following
equation:
@hx @dx @dx @E
= + : (1)
@px @px @I @px

6
Rearranging yields,
@dx @hx @dx @E
= : (2)
@px @px @I @px

In words, the uncompensated demand response to a price change is equal to the compensated demand
response (@hx =@px ) minus another term,
@dx @E
:
@I @px
This term deserves closer inspection.

The @dx =@I term should look familiar. It is the income e¤ect on demand for good X. But what is
@E=@px ?

Recall the expenditure minimization problem that yields E(px ; py ; U ): This problem looks as follows:

min px X + py Y s:t:U (X; Y ) U:

X;Y

The Langragian for this problem is:

$ = px X + py Y + (U U (X; Y )):

The …rst order conditions for this problem are:

@$
= px Ux = 0;
@X
@$
= px Uy = 0;
@y
@$
= U U (X; Y ):
@

The solutions to this problem will have the following Langragian multipliers:

px py
= = :
Ux Uy

As we know from the Envelope Theorem, at the solution to this problem,

@E
= :
@U
In other words, relaxing the minimum utility constraint by one unit, raises expenditures by the ratio of
prices to marginal utilities.

But what is @E=@px ? That is, holding utility constant, how do optimal expenditures respond to a minute
change in the price of one good? The answer is:

@E
= X:
@px

7
This follows directly from the envelope theorem for constrained problems. Since X and Y are optimally
chosen, a minute change in px or py will not a¤ect the optimal quantity consumed of either good holding
utility constant (as is always the case with the expenditure function). [See proof below.]

But a price increase will change total expenditures (otherwise utility is not held constant).

Since the consumer is already consuming X units of the good, a rise in price of 1 raises total expenditures
needed to maintain the same level of utility by X: This result is called “Shephard’s Lemma.”

An intuitive example. If you buy 10 bags of potato chips per day and the price of a bag of chips rises by 1
cent per bag, how much do we need to compensate you to hold utility constant? To a …rst approximation,
10 cents (it could never be more, it could actually be less). To hold utility constant given the price change,
your expenditures must rise by the price change times the initial level of consumption.

Question: Is the X obtained from @E=@px equal to hx or dx ; i.e., compensated or uncompensated demand?
Answer: hx :

Because the expenditure function holds utility constant, any demand function that arises from the expen-
diture function must also hold utility constant–and so is a compensated demand function.

So, to reiterate: The derivative of the Expenditure function with respect to the price of a good is the
Hicksian (compensated) demand function for that good.

Graphically the relationship between the two demand functions can be described as follows, according to
the type of good.
6.1#7
px
Normal good

Income effect
dx/dI >0

dx

hx

Qx

8
 6.1#8
px
Inferior good

Income effect
dx/dI < 0

hx

dx

Qx
6.1#9
Giffen good
px

dx

Income effect >>

Substitution effect

hx

Substitution effect Qx
Income effect

9
1.6 Proof of Shephard’s Lemma [optional]

Recall the dual of the consumer’s problem: minimizing expenditures subject to a utility constraint.

min px x + py y

s:t:U (x; y) v

$ = px x + py y + U (x; y) U ;

which gives E = px x + py y for U (x ; y ) = v .

Bear in mind the following …rst order conditions:

px py
= = :
Ux Uy

Now, calculate
@$ @x @x @y @y
= x + px Ux + py Ux :
@px @px @px @px @px
Substitute in for

@$ @x px @x @y py @y
= x + px Ux + py Uy = x:
@px @px Ux @px @px Uy @px

Hence,
@$ @E
= = x:
@px @px
Note that this x is actually hx px ; py ; U since utility is held constant.

1.7 Applying Shephard’s lemma

Returning to equation 2, we can substitute back in using Shephard’s Lemma to obtain:

@dx @hx @dx

= X:
@px @px @I

This identity is called the Slutsky equation.

It says that the di¤erence between the uncompensated demand response to a price change (the left-hand
side, @dx =@px ) is equal to the compensated demand response (@hx =@px ) minus the income e¤ect scaled
by the e¤ective change in income due to the price change (recalling that X = @E=@px ):

@dx
Notice also the economic content of the …nal term, @I X. The size of the income e¤ect on total demand
for good X in response to a change in px depends on the amount of X that the consumer is already
purchasing.

If the consumer is buying large quantities of X, an increase in px has a large income e¤ect. If the consumer
is consuming zero of good X initially, the income e¤ect of a change in px is zero.

10
 Applying the Slutsky equation to the three types of goods, it’s easy to see that:

– For a normal good ( @d

@I > 0), the income and substitution e¤ects are complementary.
x

– For an inferior good ( @d

@I < 0), the income and substitution e¤ects are countervailing.
x

@dx @hx
– For a Gi¤en good, the substitution e¤ect dominates: @I X> @px :

E¤ect of rise of px in two good economy (X; Y ).

Uncompensated Demand Compensated Demand

‘Marshallian’ ‘Hicksian’
Substitution: Substitution:
Consumption of X
Income: += Income: 0
Substitution: + Substitution: +
Consumption of Y
Income: += Income: 0
Consumer Utility 0

1.8 Uncompensated demand and the indirect utility function. [Optional]

We concluded above that the compensated demand function can be derived just by di¤erentiating the
expenditure function. Is there a similar trick for deriving the uncompensated demand function? Glad you
asked!

Recall the Lagrangian for the indirect utility function:

V = maxU (X; Y ) s:t: Xpx + Y py I;

x;y

$ = U (X; Y ) + (I Xpx Y py );
@$ @$ @$
= Ux px = = Uy py = =I Xpx Y py = 0:
@X @Y @

Now, by the envelope theorem for constrained problems:

@$ @V Uy Ux
= = = = :
@I @I py px

The shadow value of additional income is equal to the marginal utility of consumption of either good
divided by the cost of the good.

And by a similar envelope theorem argument:

@V @$
= = X: (3)
@px @px

Notice the logic of this expression. The utility cost of a one unit price increase in is equal to the additional
monetary cost (which is simply equal to X, the amount you are already consuming, times one) multiplied
by the shadow value of additional income.

11
 Returning to the potato chips example, a 1 cent price rise costs you 10 cents if you were planning to buy
10 bags. And the value of 10 cents in foregone utility is simply times 10 cents.

Putting together 3 and 4, we get the following expression:

@V (P; I)=@P
= X(P; I); (4)
@V (P; I)=@I

which is called Roy’s identity.

Roy’s identity is analogous to Shephard’s lemma above; both recover demand functions by di¤erentiating
solutions to the consumer’s problems with respect to prices. The di¤erence is that by di¤erentiating the ex-
penditure function, Shephard’s lemma gives the compensated demand function, whereas by di¤erentiating
the indirect utility function, Roy’s identity gives the uncompensated demand function.

We are now ready to put these tools to work...

2 Gi¤en goods in China (Jensen and Miller)

2.1 Context

In China, over 30% of the population survives on less than one dollar per day. (This info is from 2002,
when their paper was written; given China’s rapid growth, these facts are out of date).

The diet is very simple, consisting mostly of rice and noodles, plus some pork and other meat.

Most consumers get 70% of total calories from rice and meat alone.

Importantly for the study, regional preferences for rice versus noodles vary considerably (Table 1a).
In the South, rice is the staple.
In the North, noodles are the staple.

Meat is generally preferred to rice or noodles, but it is considerably more expensive. Meat typically
provides only one-third the calories or protein per Yuan as rice or noodles (Table 2).

2.2 The ‘experiment’

Jensen and Miller (J&M) have extremely detailed data from the China Health and Nutrition Survey (CHNS)
for 1989, 1991, 1993. These data contain:

Food diaries on complete food intake over 3 day periods.

The market prices of all major food items in the local community.

They then make the following assumptions:

12
 Food prices at the community level vary exogenously. Sometimes up, sometimes down.

This variation could be due to any combination of supply and demand factors. Does this matter?

Households are price-takers so they simply face the market price.

The idea then is to look at household responses to price variation. Because J&M have panel data (i.e., same
households, di¤erent points in time), they can presumably hold individual tastes constant. So the idea is:

Same person

Same tastes

Di¤erent prices

Research question: What happens to consumption.

What’s the basic Gi¤en prediction for the change in quantity demanded for a change in price?

2.2.1 Experimental setup

Though Jensen and Miller do not do it this way, let’s set up the experiment with our potential outcome notation.
Consider two communities, j and k in the South.
Let Yjt equal rice consumption in households in community j in time t and similarly for Ykt .
Let Xjt = 1 if the price of rice is high and Xjt = 0 if the price of rice is low (in community j period t).
We can think of the experiment as being one in which ‘nature’ randomizes rice prices (X = f1; 0g) to
communities.
If this randomization is valid, the following should be true:

E Yjt0 jXjt = 0 0
= E Ykt jXjt = 0 ;

E Yjt1 jXjt = 1 1
= E Ykt jXjt = 1 :

That is, faced with the same prices, communities j and k have the same rice demands. If so, we could simply
compare rice demand (quantity consumed) in village j to rice demand in village k. The Gi¤en prediction is
that Yjt1 > Ykt
0
, that is village j consumes more rice than village k if the price of rice is high.
This isn’t entirely satisfactory, however, since village j and k may have slightly di¤erent underlying rice
demands. That’s where the di¤-in-di¤ design comes in.
Let’s say instead that

Yjt = j + Xjt T + t

Ykt = k + Xkt T + t;

13
where T is the ‘treatment’e¤ect of high prices on consumption, and t is a time e¤ect (let’s say that demand
varies seasonally).
Now, imagine you had the following two-by-two table:

t=1 t=2

Community j (treatment) Yj1 ; Xj1 = 0 Yj2 ; Xj2 = 1

Community k (control) Yk1 ; Xk1 = 0 Yk2 ; Xk2 = 0

Here, we take baseline rice data in period 1 when prices are low in both communities. In period 1, the price
rises in community j but not community k: We again make the comparison. What do we get?

Yj = Yj2 Yj1 = j j +T + 2 1 =T + 2 1

Yk = Yk2 Yk1 = k k + 2 1 = 2 1

Yj Yk = T.

Hence, we can identify the treatment e¤ect via this di¤-in-di¤ setup.
Notice an interesting thing about this ‘experiment:’ each household provides its own pre-post comparison
over multiple time periods. This means that Jensen and Miller are also assuming causal transience. (They
don’t have to assume temporal stability, however, since the control communities provide an estimate of the time
e¤ects: 1 ; 2 ::: T .)

3 Predictions and results

The variation in prices across communities provides the control group. In the simplest case, prices rise in
community 1, stay the same in community 2.

Would you have di¤erent Gi¤en predictions in South versus North?

Yes. You would only expect the staple food to be Gi¤en since only foods that compose a large part of the
budget share could have large income e¤ects on consumption (recall the Slutsky equation).
South–Rice could be Gi¤en
North–Noodles could be Gi¤en

Would you expect di¤erent behavior for low and high income households?
Yes. Staples foods are probably not large enough as a budget share for high income households to induce
Gi¤en behavior. You might expect Gi¤en behavior for low income but not high income households.

So, we have many contrasts here:

– Pre-post within households across communities with di¤erent price changes.

14
 – Di¤erent regional tastes for goods, giving a North-South prediction on which goods should be Gi¤en.

– Within-community, cross-income level variation in predictions. Only the poor should have Gi¤en
demand.

Having three types of contrasts makes for a potentially compelling experiment.

3.0.2 What they …nd

All of the key results are found in Table 3:

1. In both the South and North, rice and noodles are inferior (see row 4 of each panel), whereas pork is a
normal good.

2. Looking along the diagonals for poor households in the South, both noodles and pork have downward
sloping demand. But rice has upward sloping demand.

3. Looking along the diagonals for poor households in the North, both rice and pork have downward sloping
demand. But noodles have upward sloping demand.

4. For not poor households in both North and South, all goods have downward sloping demand.

These results seem to provide compelling evidence of Gi¤en demand.

Alternative interpretations?

15