Notes For Econ202A: Consumption: Pierre-Olivier Gourinchas UC Berkeley

Notes for Econ202A:

Consumption
Pierre-Olivier Gourinchas
UC Berkeley

Fall 2014

Pierre-Olivier
c Gourinchas, 2014, ALL RIGHTS RESERVED.
Disclaimer: These notes are riddled with inconsistencies, typos and omissions. Use at your
own peril.
1 Introduction
Where the second part of econ202A fits?
• Change in focus: the first part of the course focused on the big picture: long run growth,
what drives improvements in standards of living.
• This part of the course looks more closely at pieces of models. We will focus on four
pieces:
– consumption-saving. Large part of national output.

– investment. Most volatile part of national output.
– open economy. Difference between S and I is the current account.
– financial markets (and crises). Because we learned the hard way that it matters a
lot!
2 Consumption under Certainty

2.1 A Canonical Model
A Canonical Model of Consumption under Certainty
• A household (of size 1!) lives T periods (from t = 0 to t = T − 1). Lifetime

preferences defined over consumption sequences {ct }Tt=1 :
T
X −1
U= β t u(ct ) (1)
t=0
where 0 < β < 1 is the discount factor, ct is the household’s consumption in period
t and u(c) measures the utility the household derives from consuming ct in period t.
u(c) satisfies the ‘usual’ conditions:
– u0 (c) > 0,
– u00 (c) < 0,
– limc→0 u0 (c) = ∞
– limc→∞ u0 (c) = 0
• Seems like a reasonable problem to analyze.
2
2.2 Questioning the Assumptions
Yet, this representation of preferences embeds a number of assumptions. Some of these
assumptions have some micro-foundations, but to be honest, the main advantage of this
representation is its convenience and tractability. So let’s start by reviewing the assumptions:
• Uncertainty. In particular, there is uncertainty about what T is. Whose T are we

talking about anyway? What about children? This is probably not a fundamental
assumption. We will introduce uncertainty later. This is not essential for now.
• Aggregation. AggregatePconsumption expenditures represent expenditures on many

different goods: ct = i pi,t ci,t over commodities i (where I am assuming that
aggregate consumption is the numeraire). If preferences are homothetic over individual
commodities, then it is possible to ‘aggregate’ preferences of the form u(c, p) into an
expression of the form u(c) where c = p.c
• Separation. Other arguments enter utility: labor supply etc... The implicit assumption
here is that preferences are separable over these different arguments: u(c) + v(z).
• Time additivity. The marginal utility of consumption at time t only depends on

consumption expenditures at that time.
– What about durable goods, i.e. goods that provide utility over many periods?
Distinction between consumption expenditures (what we pay when we purchase
the goods) and consumption services (the usage flow of the good in a given
period). The preferences are defined over consumption services but the budget
constraint records consumption expenditures. Stock-flow distinction.
– What if utility depends on previous consumption decisions, e.g. u(ct , Ht ) where
Ht is a habit level acquired through past consumption decisions? Habit for-
mation would correspond to a situation where ∂Ht /∂cs > 0 for s < t and
∂ 2 u/(∂c∂H) > 0. In words: past consumption increases my habit, and a higher
habit increases my marginal utility of consumption today. Internal habit.
– What if utility depends on the consumption of others, e.g. u(ct , C̄t ) where
C̄t is the aggregate consumption of ‘others’ (catching up with the Joneses).
External habit. As the name suggests, external habit implies an externality of my
consumption on other people’s utility that may require corrective taxation).
• Intertemporal Marginal Rate of Substitution. Consider two consecutive periods t

and t + 1. The IMRS between t and t + 1 seen from period 1 is β t+1 u0 (ct+1 )/β t u0 (ct ).
The same IMRS seen from time t is βu0 (ct+1 )/u0 (ct ). The two are equal! Key property
that arises from exponential discounting (Strotz (1957)). Example: 1 apple now, vs 2
apples in two weeks. Answer should not change with the time at which we consider
the choice (period 1 or period t). Substantial body of experimental evidence suggests
that the present is more salient then exponential discounting.
3
P −1 t
Suppose instead that U = u(c0 ) + θ Tt=1 β u(ct ) with 0 < θ < 1 represent the
lifetime preferences of the household in period 1. Notice that θ only applies to future
utility (salience of the present). quasi-hyperbolic discounting (see Laibson (1996)).
The problem is that preferences become time-inconsistent: next period, the household
would like to re-optimize if given a chance. Not the case with exponential discounting
(check this):
t−1
X T
X −1
s ∗
maxct ,ct+1 ,...cT −1 β u(cs ) + β s u(cs )
s=1 s=t
2.3 The Intertemporal Budget Constraint

Since there is no uncertainty, all financial assets should pay the same return (can you explain
why?). Let’s denote R = 1 + r the gross real interest rate between any two periods, assumed
constant. The budget constraint of the agent is:
at+1 = R(at + yt − ct )
at denotes the financial assets held at the beginning of the period, and yt is the non-financial
income of the household during period t. [Note that this way of writing the budget constraint
assumes that interest is earned ‘overnight’ i.e. as we transition from period t to t + 1.]
We can derive the intertemporal budget constraint of the household by solving forward for at
and substituting repeatedly to get:
T
X −1
a0 = R−1 a1 − y0 + c0 = ... = R−t (ct − yt ) + R−T aT
t=0
Since the household cannot die in debt T , we know that aT ≥ 0 and the intertemporal budget
constraint becomes:
TX−1 T
X −1
−t
R ct ≤ a0 + R−t yt
t=0 t=0
T
X −1 T
X −1
R−t ct ≤ a0 + R−t yt (2)
t=0 t=0
Interpretation:
• the present valuePof consumption equals initial financial wealth (a1 ) + present value of
−1 −t
human wealth ( Tt=0 R yt ).
• the term on the right hand side is the economically relevant measure of total wealth:
financial + non-financial.
• the combination of time-additive preferences and an additive intertemporal budget
constraint is what makes the problem so tractable (Ghez & Becker (1975))
4
2.4 Optimal Consumption-Saving under Certainty
Optimal Consumption-Saving under Certainty.
The problem of the household is to maximize (1) subject to (2):
T
X −1
max β t u(ct )
−1
{ct }T
t=0 t=0
s.t.
T
X −1 T
X −1
R−t ct ≤ a0 + R−t yt
t=0 t=0
We can solve this problem by setting-up the Lagrangian (where λ > 0 is the Lagrange
multiplier on the intertemporal constraint):
T −1 T −1 T −1
!
X X X
t −t −t
L= β u(ct ) + λ a0 + R yt − R ct
t=0 t=0 t=0
The first order condition for ct is:
u0 (ct ) = (βR)−t λ
[Note: from this you should be able to infer that the IBC will hold with equality). Can you
see why?]
Interpretation:
u0 (ct ) = (βR)−t λ
• β captures impatience, i.e. the preference for the present. Makes us want to consume
now.
• R determines the return on saving. A higher R makes us want to consume later (is
that really the case? More later....)
• Marginal utility will be decreasing over time if βR > 1 and increasing otherwise.
• Since marginal utility decreases with consumption, this implies that consumption will
be increasing over time when βR > 1 and decrease otherwise.
• When βR = 1 the two forces balance each other out and consumption becomes flat.
• Note that this gives us some key information on the slope of the consumption profile
over time, but not on the consumption level.
5
2.5 A Special Case: βR = 1
u0 (ct ) = λ
This implies that consumption is constant over time: ct = c̄. Substitute this into the intertem-
poral budget constraint to obtain:
T −1 T −1
X 1 − R−T 1 − βT X
R−t c̄ = c̄ = c̄ = a0 + R−t yt
1 − R−1 1−β
t=0 t=0
from which we obtain:

T −1
!
1−β X
−t
c̄ = a0 + R yt
1 − βT t=0
Observe:
• consumption is a function of total wealth.
• the marginal propensity to consume is (1 − β)/(1 − β T ) and converges to 1 − β when

the horizon extends (T → ∞).
• if β = 0.96 (a reasonable estimate), this gives R = 1/0.96 = 1.0416. Then we should

consume about 4% of total wealth every period.
2.6 The Permanent Income Hypothesis

Friedman’s (1957) Permanent Income.
T −1
!
1−β X
c̄ = a0 + R−t yt
1 − βT t=0
• This is Friedman’s permanent income hypothesis. Individual consumption is not

determined by income in that period, but by lifetime resources, unlike Keynesian
consumption functions of the form ct = a + byt .
• Friedman actually defines permanent income as the right hand side of this equation.
This is the annuity value of total resources.
• This implies that consumption should not respond much to transitory changes in
income, since these will not affect much permanent income, but should respond if
there are changes in your permanent income.
– you earn an extra $200 today

– you just got tenured and learn that starting next year, your income will double.
– you learn that you won $10m at the state lottery
6
2.7 Understanding Estimated Consumption Functions
Keynes (1936) argues that ‘aggregate consumption mainly depends on the amount of aggre-
gate income,’ ‘is a stable function,’ and ‘increases less than proportionately with income.’
In other words, Keynes argues for a consumption function of the type ct = a + byt .
Empirically, it matters whether we look (a) in the cross section or (b) in the time series.
This looks quite different from Friedman’s permanent income which we can write as
ct = ytp where ytP is simply permanent income.
Yet, Friedman’s PIH can account for the empirical observations. Observe that we can
write yt = ytP + ytT where ytT is the transitory component of income. An OLS regression of
consumption on income yields:
cov(ct , yt ) cov(ytP , yt ) var(ytP )
b̂ = = = <1
var(yt ) var(yt ) var(yt )
â = E(ct ) − b̂E(yt ) = (1 − b̂)E(ytP )
ci,t ct
45◦ 45◦
whites
blacks
in the cross section in the time series
0 yi,t 0 yt
Figure 1: Keynesian consumption functions and the PIH
T : lower b̂.
• In the cross section: more variations from yi,t
• in the time series, more variation from ytP : higher b̂

• in the cross section: lower intercept â if lower E(ytP )
7
2.8 The LifeCycle Model under certainty
Modigliani and Brumberg (1954) consider the lifecycle implications of the previous model.
Suppose that people live T periods (from 1 to T ) and that β = R = 1. (Note: this is a
stronger assumption than βR = 1)
The PIH model tells us that consumption is constant and equal to the permanent income
of the agent: c = c̄ = y P .
This is irrespective of the income profile {yt } over time. Suppose now that the agent
works for N < T periods, earning income y, then retires.
The household saves y − c̄ when working, then dissaves −c̄ when retired.
human wealth
y
total wealth
c̄
wealth
N T time N T time
Figure 2: The Baseline Life-Cycle Model
Equations for the simple Modigliani-Brumberg (1954) lifecycle model:

N
c̄ = y
T
T −N
at = t y for t ≤ N
T
N
at = (T − t)y for N ≤ t ≤ T
T
ht = (N − t)y human wealth
N
wt = at + ht = (T − t)y for all t ≤ T
T
8
Assume there is no growth. Then, we have the following (aggregating across cohorts):
N
w̄ = y(T + 1)
2T
N N
h̄ = y(N + 1) ; ā = y(T − N )
2T 2T
Note:
• The household runs total wealth to 0
• Human wealth runs out at t = N . It is supplemented by financial wealth
• We can have positive financial wealth even if there is no bequest motive.
• The ratio of human to financial wealth h̄/ā does not depend on income y (it is equal
to (N + 1)/(T − N ).
• The details of the social security system matter. This describes a fully funded system
(or even more precisely, what should happen if there is NO social security system and
no bequest motive). What if we have a society where the young take care of the old
(China); or an unfunded system where the government taxes the young to support the
old? What happens to
– consumption profiles?
– income?
– private saving?
2.9 Saving and Growth in the LifeCycle Model

How does growth affect saving in the lifecycle model?
• Start with zero growth: the age-profile = cross-section. Aggregate wealth is constant
and aggregate saving equals 0. The young save, and the old dissave
• population growth: more young saving, so saving increases with population growth.
• productivity growth: more complex and depends on how productivity growth affects
each cohort’s income:
– If productivity growth is across cohorts (i.e. each cohort’s income is constant but
younger cohorts have a higher income profile) then productivity growth increases
saving. (Why? b/c the young save more than the old dissave)
– but instead if productivity growth increases income over a worker’s lifetime, then
young workers may decide to borrow against higher future income in middle age.
In that case, faster growth can reduce savings.
9
c̄
N T time
Figure 3: The Life-Cycle Model when Income grows over time
2.10 Interest Rate Elasticity of Saving

The response of consumption and savings to changes in interest rates is an important question.
Think about:
• the transmission of monetary policy (changing the real interest rate)
• changes to the tax code that affects rates of returns on savings. you have seen in the
first part of this course how changes in savings can affect growth rates temporarily (if
growth is exogenous) and potentially permanently (if growth is endogenous)
Consider the first-order condition again:
u0 (ct ) = (βR)−t λ
Rewrite it in two consecutive periods and eliminate λ. This is the Euler equation under
certainty:
u0 (ct ) = (βR) u0 (ct+1 )
Consider CRRA preferences: u(c) = c1−θ /(1 − θ). [We have already seen these preferences
when solving the Ramsey-Cass-Koopmans problem: θ represents both the CRRA coefficient
and the inverse of the (IES).]
Substitute to get:
ct+1 /ct = (βR)1/θ
• if 1/θ = 0 (Leontieff) then ct is flat regardless of the interest rate. No substitutability
• for 1/θ < 1: weak substitution effects
• for θ = 1 income and substitution effects cancel out (log preferences)
10
• for 1/θ > 1: strong substitution effects
• if 1/θ → ∞ then ct becomes very responsive to the interest rate. In the limit, con-
sumption growth becomes so responsive that the interest rate R will have to stay ‘close’
to 1/β to ensure that consumption growth does not become too extreme.
ct+1 /ct = (βR)1/θ

In general, consumption growth should be responsive to changes in the interest rate. On can
rewrite:
d ln(ct+1 /ct ) = (1/θ)d ln R = (1/θ)d ln(1 + r) ≈ (1/θ)dr
An increase in the real interest rate by 100bp should increase consumption growth by 1/θ%.
But the analysis is incomplete: we need to figure out by how much consumption itself
changes.
What about the overall effect?

To simplify things, let’s consider first a two-period problem (with a0 = 0)
c1 /c0 = (βR)1/θ
c1 y1
c0 + = y0 +
R R
Let’s start with a simple case where y1 = 0:
We have:
• a substitution effect. Keeping the utility level constant, the change in interest rate leads
us to substitute consumption today for consumption tomorrow: c0 falls, c1 increases.
• an income effect: the budget line rotates around (y0 , 0). This means more consumption
can be afforded in each period. This increases c0 and c1 .
The effect on c1 is unambiguous. The effect on c0 is ambiguous. When 1/θ > 1 the
substitution effect dominates so that c0 falls.
Mathematically:
c0 (1 + (βR)1/θ /R) = y0
2.10.1 The 2-period case with y1 = 0

Now, let’s consider what happens when y1 6= 0
• the budget curve rotates around (y0 , y1 ).
• in addition to the income and substitution effects, there is a wealth effect: future income
is worth less to the household. This reduces c0 and c1 .
11
B
C
A
c1
c0 y0
Figure 4: An increase in interest rates when y1 = 0
• the net effect often depends on whether the household is a net borrower or lender.
– if c0 = y0 and c1 = y1 then there is no income and wealth effect (why? because

the initial consumption bundle remains on the new budget line). So there is only
a substitution effect and c0 falls.
– if the household is initially a saver (i.e. c0 < y0 ) then the income effect dominates
the wealth effect and the overall effect on c0 is indeterminate
– if the household is initially a borrower (c0 > y0 ) then the income effect is weaker
than the wealth effect. Therefore c0 falls.
Mathematically:
c0 (1 + β 1/θ R1/θ−1 ) = (y0 + y1 /R)
2.10.2 The 2-period case with y1 6= 0 and c0 = y0

The substitution effect is the only effect. c0 falls.
2.10.3 The 2-period case when c0 < y0 (lender)

Ambiguous. Income effect stronger than wealth effect.
12
B
c 1 = y1
A
c 0 = y0 h0 = y0 + y1 /R
Figure 5: An increase in interest rates when c0 = y0 and c1 = y1
2.10.4 2-period case when c0 > y0 (borrower)

wealth effect stronger than income effect. c0 falls.
Savings and Interest Rates: Recap:
The literature often considers the case θ = 1 as a benchmark, where income and
substitution effects cancel out, leaving the saving rate independent of the interest rate;
Furthermore, empirical estimates of the elasticity of intertemporal substitution suggest

relatively low numbers for 1/θ, especially since there are also income and wealth effects.
It is tempting to conclude from this that (a) the slope of consumption growth and (b) the
level of consumption will be largely unaffected by changes in the interest rate. However, this
answer can be misleading for a number of reasons:
• lifetime horizon. But even if the IES is small, it can have a large impact over a lifetime
(Summers 1981).
• This omits the wealth effect. Even if income and substitution effects cancel out, a
change in interest rates affects human wealth, and this leads to a change in consumption
in the PIH-LC model.
• Finally, the nature of the change in interest rates matters. For instance, a change in
interest rates due to tax changes may be offset somewhere else to leave government
13
c1 B
y1
A
c0 y0 h0 = y0 + y1 /R
Figure 6: An increase in interest rates for a saver (c0 < y0 )
revenues unchanged. In that case, there is no income effect and only the substitution
effect. This might not be very helpful if the IES is small anyway.
2.11 The LifeCycle Model under Certainty Again

Consider now the case where R and β differ from 1. In addition, suppose that a0 = 0 and
that y is constant as before.
The Euler equation with CRRA preferences implies:
ct = (βR)t/θ c0
Substituting into the budget constraint, we obtain:

T −1
1 − (βR1−θ )1/θ X −t
c0 = R yt
1 − (βR1−θ )T /θ t=0
1 − (βR1−θ )1/θ 1 − R−N
c0 = y0
1 − (βR1−θ )T /θ 1 − R−1
Suppose that βR > 1 so that consumption grows over time, even it θ is low. If the horizon
T is long enough relative to the working period, consumption must be much higher at the end
14
y1
A
c1 B
C
y0 c0 h0 = y0 + y1 /R
Figure 7: An increase in interest rates for a borrower (c0 > y0 )
of life than at the beginning: the agent must accumulate a large stock of wealth. Aggregate
wealth and saving may be highly responsive to changes in interest rates. See Summers 1981.
3 Consumption under Uncertainty

Last class we looked at the consumption model under certainty. The model provides important
insights:
• consumption is a function of total wealth (permanent income)
• the slope of the consumption profile is controlled by the discount rate, the interest rate
and the intertemporal elasticity of substitution
• in a lifecycle environment, there is a substantial amount of life-cycle wealth accu-
mulation. In the simple model, the amount of wealth is ā/y = N/(2T )(T − N ) =
40/(120)(20) = 800/120 = 6.66
• the elasticity of aggregate saving to the interest rate is complex.
The model needs to be extended to allow for uncertainty. Precautionary saving is another
reason why households decide to save. We start with the canonical model, augmented for
uncertainty.
15
THE A MERICAN ECONOMIC RE VIE W SEPTEMBER 1981
g equation (5) it is apparent that TABLE 1-THE INTEREST ELASTICITY

OF AGGREGATE SAVINGS
Value of r
S ( ) WL ) .04 .06 .08
WL r-n-g
y= .5
71r 3.71 2.26 2.44
uting for C/ WL in (6) yields the ag- S/WL .121 .274 .451
e savings function: y=O
71r 3.36 1.89 1.87
S/ WL .068 .142 .210
y= -.5
SrL (r -r ( (g-r)T' 71r
S/ WL
3.09
.049
1.71
.096
1.54
.135
Y= -I
71r 2.87 1.59 1.37
e(( 1)-g-n)T ) S/ WL .038 .073 .099
y= -2
71r 2.38 1.45 .122
S/ WL .028 .048 .063
y= -5
l -n-g] (g-r)(e( -r)T, 1) 71r .741 1.09 1.18
S/ WL .014 .019 .025
Note: The calculation assumes n =.015, g =.02, T'= 50,

T=40, and 8 .03. The savings rate is measured as a
I -e -NT )(r-n-g) )_ n g fraction of labor income.
The results
Figure 8: Tableuniversally
1 from Summers (1981)a high in-
support
oteworthy that (7) shows that the life terest elasticity. In the plausible logarithmic
hypothesis gives rise to a steady-state utility case, the interest elasticity of the sav-
3.1 The Canonical Model
ate savings function which may be ings rate varies from 3.36 at 4 percent to 1.87
ented by a variable propensity to set-up
3.1.1 the save at 8 percent. This case also generates the
labor income, and a zero-savings pro- most reasonable values for the aggregate sav-
out of capital income. The
Thehousehold
life cycle has the following preferences over consumption sequences:
ings rate. The table demonstrates the un-
esis thus gives rise to a savings func- importance of"X the
∞ elasticity of substitution
#
hich is quite different than that usually betweenUpresent= E and β t u(c t )|Ω0consumption. For
future (3)
ed in growth models which allow dif- example, at an interestt=0 rate of .06, the elastic-
savings propensities out of different ity of saving varies only from 2.26 when
Notice two differences with the model under certainty:
of income. y- 1/2 to 1.09 when y =-5. The insensitiv-
clear from (7) that the relationship
1. First I assume thatitytheofhorizon
the elasticity to the
is infinite. This level toofshow
is mostly y reflects
you how to use solve
n savings and the interest ratetheismodel
com-in thatthecase.fact the "reduction
that substantial
Nothing rests oninthathuman wealth"
hypothesis and I will point out
nd depends on all of the otherasparame- effect is much more
we go where things might be different if important
we have a finite the sub-
that horizon. Formally, you
the model. In Table 1 the savings rate,that households
may think stitutioncare aboutoftheir
effect interest changes.
offsprings The
and apply thebasic
same discount rate.
d as S/ WL, and interest elasticity of conclusion, a significant long-run interest
ate savings Tjr' evaluated at various elasticity of aggregate savings, is quite robust
of the interest rate, are reported for to changes in all16of the parameter values.
ble parameter values. It is assumed that While very low values of y could generate
tion grows at a 1.5 percent per an- low or even negative savings elasticities, they
productivity increases by 2 percent per would also give rise to unrealistic savings
, and that individuals live fifty-year propensities, unless the other parameter val-
2. The term E[.|Ω0 ] captures expectations conditional on information available at time
t = 0. This information set is denoted Ω0 . This is also an important assumption. It
implies that preferences are separable over states and over time.. To see this, suppose
that there are St possible states of the world in period t and that each of them has
probability (as of time 0) given by π s,t . Then we can write the utility as:
∞ X
X
U= β t π s,t u(cs,t )
t=0 s∈St
This double separation imposes strong structure on preferences, but it simplifies tremen-
dously the analysis.
3. To lighten the notation, I will indifferently write Et [.] or E[.|Ωt ] to indicate conditional
expectation as of time t.
The household budget constraint takes the same form as before, except that now, I will
suppose that households face some uncertain interest rate R̃t+1 and an uncertain future
income ỹt . In this notation, the˜indicates that a variable is stochastic (as seen from previous
periods). The budget constraint then takes the form:
at+1 = R̃t+1 (at + ỹt − ct ) (4)
3.1.2 Recursive Representation

The problem is to maximize (3) subject to (4), and any other restriction on consumption
and asset choices, for a given initial level of wealth a0 . For instance, we know that we only
consider positive consumption: ct ≥ 0.
We have also already discussed the fact that the household will not be allowed to run
Ponzi-like schemes:
lim β T aT ≥ 0
T →∞
This constraint holds in the uncertain case, along all possible consumption sequences (tech-
nically, it holds almost surely).1 But there might be other constraints on assets holdings. For
instance, the household may be prevented from borrowing beyond a certain limit:
at ≥ a
At time t, at is a state variable of the household consumption problem, in the sense that
it is pre-determined by the previous actions of the households and is beyond its control.
1
It is not sufficient that the No-Ponzi condition holds in expectation, that is E0 [limT →∞ β T aT ] ≥ 0. If this
were the case, then there would be possible paths with non-zero measure where the No-Ponzi condition would
be violated. Along these paths, lenders would have to agree to provide an infinite level of consumption to the
household. Note also that if the NPC holds a.s., then it holds in expectation, while the reverse is obviously not
true.
17
We are going to assume that income and return realizations are iid, so that ỹt and R̃t are
not state variables of the household problem. This is mostly to keep notations simple. It
would be quite straightforward to extend the set-up to a case where ỹ and R̃ have a Markov
structure.
Remark 1 In some situations, it is easier to use cash-on-hand xt as the state variable, where
xt is defined as : xt = at + ỹt . xt represents the resources available for consumption
and saving to the household, after the realization of current income. The budget constraint
becomes:
xt+1 = R̃t+1 (xt − ct ) + ỹt+1
Since financial assets at are the sole state variable, we can write the value function that
maximizes the utility of the agent as a function of the state variable a:
"∞ #
X
t
v(a0 ) = max{ct }∞ t=0
E0 β u(ct )
t=0
Given the nicely recursive structure of the problem, we write the Bellman equation as
follows. Suppose that the level of assets is a in a given period. Consumption that period
must satisfy:
v(at ) = maxct ∈Ct u(ct ) + βEt [v(at+1 )]

s.t.
at+1 = R̃t+1 (at+1 + ỹt − ct )
where Ct denotes the set of permissible consumption choices at time t. Notice that it is
the same value function that enters on both sides of this equation. So, one way to think about
the household problem is that the Bellman equation defines the value function as fixed point
of a functional equation. There are various theorems that establish existence and uniqueness
of this fixed point, when the Bellman equation is well-behaved -as is the case here.
Remark 2 If the functional equation is contraction mapping, then the Bellman equation has
a unique solution AND this solution can be found by iterating on the value function. This
provides a convenient (if not especially rapid) way to characterize numerically the value
function (value function iteration).
3.1.3 Optimal Consumption and Euler Equation

We start by assuming that the solution is interior to the set Ct . The first-order condition of
the above problem yields:
u0 (ct ) = βEt [v 0 (at+1 )R̃t+1 ]
18
Let’s now consider what happens when there is a small change in at on the household
value function v(at ). To calculate v 0 (at ), let’s take a full derivative of the Bellman equation.2
The total variation is:
v 0 (at )da = u0 (ct )dct + βEt [v 0 (at+1 )R̃t+1 (da − dc)]
where dc denotes the change in optimal consumption for a given small change in a. Regroup-
ing terms, we obtain:
v 0 (at )da = (u0 (ct ) − βE[v 0 (at+1 )R̃t+1 ])dc + βE[v 0 (at+1 )R̃t+1 ]da
The first term on the right hand side is zero from the first-order condition of the problem. So
we are left with:
v 0 (at ) = βEt [v 0 (at+1 )R̃t+1 ]
This is a straightforward application of the Envelope Theorem.
Combining the first order condition and the Envelope theorem, we conclude that:
u0 (ct ) = v 0 (at )
Substituting back into the first order condition, we obtain the well-known Euler Equation
under uncertainty:
u0 (ct ) = βEt [R̃t+1 u0 (ct+1 )] (5)
What is the intuition for the Euler equation? A variational argument might help. Suppose
that we reduce consumption from the optimal path in period t by , and increase consumption
by R̃t+1 next period (so that we are back on the optimal consumption path after period t + 1).
The marginal disutility (as of time t) of reducing consumption in t is u0 (ct ). The marginal
increase utility from higher consumption in t + 1 (as of time t) is βEt [R̃t+1 u0 (ct+1 ). For a
small the two should be equal (otherwise the proposed consumption is not optimal to start
with).
Note that the discount rate β and the interest rate R̃t+1 still play opposing force on
consumption growth, so the insights from the certainty case do carry over to the uncertain
case. But we now also have to take into account uncertainty over future returns and future
marginal utility.
Remark 3 The derivation above assume that consumption is interior. What would happen if
consumption is at the boundary. For instance, suppose that we impose the conditions that
0 ≤ c ≤ a + y (how should we interpret this condition?). What form does the Euler equation
take?
2
We are assuming that the value function is differentiable, which is not always the case. See Stokey, Lucas
and Prescott (1983) for more details on this.
19
3.2 The Certainty Equivalent (CEQ)
The Euler equation provides some important insights into consumption behavior, but in its
general form, it is not very tractable. We now make a number of simplifying assumptions,
following Hall (1978).
First, we assume that there is no uncertainty in interest rates, so R̃t = R. Moreover, we
assume that there is no tilt in consumption profiles, that is βR = 1.
Second, we will consider a very particular form of preferences:
u(c) = αc − γc2 /2; γ > 0, α > 0
In other words, preferences are quadratic over consumption. These preferences are very
weird from a number of points of view:
• even if α > 0, utility turns negative for sufficiently large consumption
• these preferences admit negative consumption (they definitely violate Inada’s condi-
tions)
So why would we want to make these crazy assumptions? Two possible justifications are:
• we could think of these preferences as a second order approximation of utility for more
general utility functions. If we think about it this way, then it would suggest that this
may not be such a bad approximation for relatively small changes in consumption over
time.
• these preferences have the important property that marginal utility is linear, or equiva-
lently that the second derivative is constant: u0 (c) = α − γc and u00 (c) = −γ.
Let’s make these two assumptions and substitute into the Euler equation (5) to obtain:
ct = Et [ct+1 ] (6)
The stunning result here is that consumption follows a Random Walk. This means that
changes in consumption are unpredictable. To see how stunning it is, recall that if we had no
uncertainty (and βR = 1), then we would get
u0 (ct ) = u0 (ct+1)
and so consumption would be constant over time, and therefore entirely predictable. Instead,
once we introduce uncertainty, consumption becomes entirely unpredictable!
To see what is going on, it helps to solve for the level of consumption in the CEQ case.
To do this, let’s first derive the Intertemporal Budget Constraint. First recall that the dynamic
budget constraint is:
at+1 = R(at + ỹt − ct )
20
Let’s solve this sequence forward for a given sequence of consumption and income realization:
∞
X
−
a0 = R 1(a1 + c0 − ỹ0 = ... = R−t (ct − ỹt ) + lim R−T aT
T →∞
0
With the No-Ponzi condition, the last term has to be positive, so the intertemporal budget
constraint takes the form:
X∞ X∞
R−t ct ≤ a0 + R−t ỹt
0 0
Notice that this intertemporal budget constraint does not have an expectation term: it has
to hold along any possible realization of income and consumption: it holds almost surely.
But if it holds almost surely, then we are allowed to take expectations and the following also
holds:
∞
X ∞
X
E0 [ R−t ct ] ≤ a0 + E0 [ R−t ỹt ]
0 0
The next step is to observe that we can move the expectation inside the summation, and
use the fact that under the random walk hypothesis, the following holds:
E0 ct = E0 Et−1 ct = E0 ct−1 = ... = c0
where the second term follows from the Law of Iterated Expectations, the third one from the
fact that consumption follows a random walk at t and the last one from iterating the argument.
It follows that consumption at time t = 0 must satisfy:
∞
1 X
c0 = a0 + E0 [ R−t ỹt ]
1 − R−1
0
"∞ #!
X
−t
c0 = (1 − β) a0 + E0 R ỹt
0
where we used the assumption that R−1 = β and the definition of xt in the second line.
What this tells us is that consumption follows the PIH in expectation. The term in parenthesis
on the right hand side is expected total wealth, where the expectation is over future labor
income.
This is why the model is called the ‘certainty equivalent’ model: as far as consumption
decisions are concerned, the household behaves as if future income was certain and equal
to its expected value. The source of this behavior can be traced back to the assumption of
quadratic utility. Note that the Euler equation in the CEQ model is:
ct = Et ct+1
21
what this tells us is that the household is smoothing consumption, but taking future consump-
tion as if it were certain and equal to its expected value. But if you retrace your steps, you
will see that this result arises from the Euler equation in general form:
u0 (ct ) = Et u0 (ct+1 )
and the fact that marginal utility is linear when utility is quadratic: u0 (c) = α − γc.
Anticipating on the next lecture, this tells you that this result will not hold in the more
general case where marginal utility is not linear.
Why are changes in consumption unpredictable, while the consumption level itself seems
to follow a minor modification of the PIH? To see what is going on, consider consumption in
two consecutive periods, t and t + 1:
"∞ #!
X
ct = (1 − β) at + Et R−(s−t) ỹs
s=t
∞
" #!
X
ct+1 = (1 − β) at+1 + Et+1 R−(s−(t+1)) ỹs
s=t+1
Take the difference and substitute at+1 = R(at + ỹt − ct ) to obtain:

∞ ∞
!
X X
ct+1 − ct = (1 − β) at+1 − at + R−(s−(t+1)) Et+1 ỹs − R−(s−(t)) Et ỹs
s=t+1 s=t
∞ ∞
!
X X
−(s−(t+1)) −(s−(t))
= (1 − β) R(at + ỹt − ct ) − at + R Et+1 ỹs − R Et ỹs
s=t+1 s=t
∞ ∞
!
X X
= (1 − β) (R − 1)at + Rỹt − Rct + R R−(s−t) Et+1 ỹs − R−(s−(t)) Et ỹs
s=t+1 s=t
∞ ∞
!
X X
= (1 − β) −R R−(s−t) Et ỹs + R R−(s−t) Et+1 ỹs
s=t+1 s=t+1
using the expression for ct and the fact that βR = 1, we obtain finally:
∞
X
ct+1 − ct = (R − 1) R−(s−t) (Et+1 ỹs − Et ỹs ) (7)
s=t+1
Notice that the term in the summation on the right hand side of (7) is Et+1 ỹs − Et ỹs , that
is, the revision in expectations about future income. Of course, this revision is unpredictable
as of period t, otherwise it would already have been incorporated in the current expectation
Et ỹs !
22
This gives us a very nice result: the change in consumption is related to the news the
household receives about future income. We will see also that it provides us with a way to
test the certainty equivalent model.
Remark 4 You can check that if you take expectations as of time t on both side of this
equation, you recover Et ct+1 − ct = 0.
Example 1 Consider the case where income follows an AR(1) process:
ỹt+1 = ρỹt + η t+1 ; 0<ρ≤1
Then we can easily check that Et ỹs = ρs−t ỹt . Substituting back into (7), we obtain after
some easy manipulations:
1−β
ct+1 − ct = η
1 − ρβ t+1
Since ρ ≤ 1, ct+1 − ct ≤ η t+1 , that is, consumption in general responds less than 1 for 1 to
a change in income. The case where consumption moves 1 for 1 is when ρ = 1, i.e. income
itself is a random walk.
3.3 Tests of the Certainty Equivalent Model

3.3.1 Testing the Euler Equation
The literature up until Hall (1978) used to attempt to derive closed form solution for
consumption (i.e. a consumption function) and estimate it. But a closed form solution
for the consumption function is often not available. So instead, the literature would try to
identify the determinants of consumption and estimate empirically the relationship between
consumption and its determinants. This would not allow for a rigorous test of the theory. In
addition, the regression typically faced a serious problem of identification since income (the
most common right hand side variable) is not exogenous.
Instead, Hall argued that we can test the theory by directly testing the first-order condition
of the model, i.e. the Euler equation. Under rational expectation, any variable can be
expressed as the sum of its conditional expectation and an innovation term, orthogonal to
any information available at time t:
ct+1 = Et ct+1 + t+1 = ct + t+1
where Et t+1 = 0 and the second equality uses (6). So the theory implies that ct contains all
the relevant information necessary to predict ct+1 .
Under the null hypothesis that the theory is correct, a regression of the form:
ct+1 = a + b ct + c xt + εt+1 (8)
23
where xt is any variable available at time t to the household should yield:
â = 0; b̂ = 1; ĉ = 0
What is important in that regression is that it does not matter whether yt is exogenous or
not (the key problem with the consumption function estimation approach). The key test is
whether ĉ = 0 or not. If we find some variables, known as of time t that can help predict
next period’s consumption after controlling for current consumption, then the theory has to
be incorrect.
Remark 5 Notice that the theory does not say that consumption should not react to current
income. In other words, if we run the regression
ct+1 = a + b ct + c xt + d yt+1 + εt+1
there is no presumption that dˆ should be equal to 0.
Remark 6 Notice that given (7), we know that

∞
X
t+1 = (R − 1) R−(s−t) (Et+1 ỹs − Et ỹs )
s=t+1
One could think that this would provide another way to test the theory. For instance, when
income follows an AR(1) process as above, we know that the consumption innovation is given
by:
1−β
t+1 = η
1 − ρβ t+1
so the innovation to consumption t+1 and the innovation to income η t+1 are linked in a very
precise way. However, this could be exploited only if the household learns about the change
in its income as it happens. If instead, the household learns about a change in its income
before it is realized, this is when consumption will change, and not when the actual change in
income occurs. Unless the econometrician has information on when the information becomes
available to the household (more on this below), then the relationship above will not be
terribly useful. Testing the first order condition remains valid, however, since any information
known at time t to the household should not help predict future consumption.
Hall (1978) tests the CEQ model using aggregate quarterly data on non-durable real
consumption per capita and real disposable income per capita. The results (see attached
table) suggest that indeed lagged income is not helpful in predicting future consumption (on
top of lagged consumption).
24
This content downloaded from 169.229.32.36 on Thu, 16 Oct 2014 04:12:50 AM
~~~~~~
H0
0
0 :
All use subject to JSTOR Terms and Conditions
P4~~~~~~~~~0
~~~40
0 CZ
Q e
(O
0 -e
~ ~ ~ ~ ~~c
C CL.
~
.9 I*
X
dz
00
W.)
om
.
.'
co "I-T
0li
-98
CO
r-CO
C
I
..I
0)
I~
m.~~~~CD
m.
0-4
C4
v
o
C'
s
co I
014
0'4
r
Figure 9: Table 3 in Hall (1978).
3.3.2 Allowing for time-variation in interest rate: the log-linearized Euler equation
The regression (8) imposes that the gross real interest is constant and equal to the inverse
of the discount factor. It also imposes that preferences are quadratic. We can relax both
assumptions yet obtain a result very similar to the CEQ, as long as we are looking at small
deviations around the equilibrium.3
To see how this is done, consider the Euler equation of the general model:
u0 (ct ) = βEt [R̃t+1 u0 (ct+1 )]
Assume that R̃t+1 is known as of time t. This would be the case if R̃t+1 is the return on
a one-period risk free bond between t and t + 1. Assume further that preferences are CRRA
so that u0 (c) = c−θ , with θ > 0.
The Euler equation takes the form:
c−θ −θ
t = βRt+1 Et [ct+1 ]
We can rewrite this as follows:
1 = βRt+1 Et [c−θ θ
t+1 ct ]
1 = exp(−ρ + rt+1 )Et [exp(−θ∆ ln ct+1 )]
0 = −ρ + rt+1 + ln Et [exp(−θ ln(ct+1 /ct )]
where we define ρ = − ln β and rt = ln Rt+1 and where the third line takes logs. Assume
now that ∆ ln ct+1 is conditionally normally distributed. Then, the Euler equation takes the
form:
1
0 = −ρ + rt+1 − θEt ∆ ln ct+1 + θ2 Vt ∆ ln ct+1
2
3
Recall that we motivated the CEQ as a second order approximation of preferences around the equilibrium.
Instead of taking a second order approximation of preferences then solving for optimal smoothing, we can take a
first order condition of the first order condition of the general consumption-saving problem.
25
where Vt ∆ ln ct+1 is the conditional variance of consumption growth.4 If consumption
growth is not conditionally normally distributed, this expression is a second-order approxi-
mation.
Re-arranging, we obtain:
1 1
Et ∆ ln ct+1 ≈ (rt+1 − ρ) + θVt ∆ ln ct+1 (9)
θ 2
If we ignore the conditional variance term, and assume that the interest rate is equal to
the discount rate (rt = ρ), then we obtain an expression similar to the CEQ:
Et ln ct+1 ≈ ln ct .
log-consumption follows a random-walk.5
If the interest rate is not constant, but we still ignore the variance term, we recover that
expected consumption growth depends on the difference between the interest rate and the
growth rate, scaled by the IES 1/θ: Et ∆ ln ct+1 = 1/θ(rt+1 − ρ).
As we will see a bit later, the variance term captures the precautionary savings component
of consumption growth. It is always positive, increasing the growth rate of consumption.
For now, let’s assume that the variance term is either zero, or constant. The log-linearized
Euler equation leads to the following empirical specification:
ln ct+1 = a + b ln ct + c xt + d rt+1 + εt+1
and, if the equation is correctly specified, the point estimate dˆ should be the Intertemporal
Elasticity of Substitution 1/θ.6
Equations of that form have been estimated in literally hundred of papers. The goals of
these regressions are usually two-fold:
ˆ the IES from the coefficient on rt+1
1. estimate 1/θ from d,
2. test the orthogonality restriction that information available at time t does not predict
consumption growth: ĉ = 0. For example, expected income growth Et ∆ ln ỹt+1
should not help predict consumption growth.
The literature typically finds:

4
This results from the fact that if x is distributed N (µ, σ) then E[exp(x)] = exp(µ + 1/2σ 2 ).
5
The original CEQ model states that consumption in levels follows a random walk. The log-linearized result
states that it is log-consumption that follows a random walk. The two are not very different for small deviations.
Moreover, a random walk in logs is probably a better empirical specification given that consumption (and its
innovations) grow over time. Campbell and Mankiw (1989) test the CEQ in logs.
6
The constant a captures the sum of the impatience terms ρ/θ and the –constant– precautionary term
θ/2Vt ∆ ln ct+1 and so does not provide useful information.
26
1. estimates of 1/θ between 0 and 0.2, i.e. very low estimates of the sensitivity of
consumption growth to the interest rate. Recall however, that aggregate saving can still
be quite sensitive to the interest rate because people have long lifetimes (so the effects
build up over long periods)
2. ĉ is positive and significant when using Et ∆ ln ỹt+1 as a regressor. This means

that we reject the strict CEQ restriction. Expected income growth predicts expected
consumption growth. This is sometimes referred to as the ‘Excess Sensitivity of
Consumption’. The conclusion is we reject the joint assumptions that (1) the Euler
equation is true; (2) the utility function is CRRA; or (3) the linearization is accurate;
Remark 7 The excess sensitivity is in response to variables that consumption should not
respond too. In addition to excess sensitivity, there is an excess smoothness puzzle, whereby if
income changes are very persistent, then innovations to consumption should be more volatile
than innovations to income. One can see this is we assume the following process for income:
∆ỹt+1 = µt+1 + γµt , with γ > 0, i.e. a MA(1) for income change. Then one can show that
t+1 = (1 + γ)µt+1
Deaton (1987) observed that despite γ > 0, consumption innovations appear less volatile
than income innovations.
3.3.3 Campbell and Mankiw (1989)

This is an example of a paper estimating an equation similar to (9). They start from the
baseline CEQ model (consumption is a random walk). Rather then simply test the null that all
consumers are CEQ consumers, so that aggregate consumption also follows a random walk,
they specify an alternative where a fraction λ of consumers are ‘hand-to-mouth’ with cht = yth
while the remaining consumers are CEQ with ∆crt+1 = rt+1 . Aggregate consumption change
is then the sum of the consumption change of the hand-to-mouth consumers and of the CEQ
consumers:
∆ct+1 = λ∆yt+1 + (1 − λ)t+1
Since t+1 is not observed, it is treated as the residual of the regression.
However, t+1 is likely to be correlated with income changes, so the equation above
cannot be estimated directly. Instead, income changes need to be instrumented, using as
instruments any lagged variables (hence orthogonal to t+1 ) and good predictors of income
growth. Equivalently, this consists in constructing a measure of predicted income change
Et ∆yt+1 and regressing realized consumption changes on predicted income changes.
This approach provides an alternative to the null that all consumers are CEQ. It
even allows to quantify the share of consumers that are hand-to-mouth, as measured
27
by λ. The estimates indicate that λ is quite high, on the order of 0.48 in the US (see Table 10).7
In addition, they run the regression including the real interest rate (see Table 11). They
use as real interest rate the 3-month T-bill rate over the quarter minus the rate of change in
the PCE deflator. They instrument the real interest rate using lagged real interest rates (since
the regression should use ex-ante real interest rates).
The specification becomes:
∆ct+1 = λ∆yt+1 + (1 − λ)/θrt+1 + (1 − λ)t+1
Notice that the coefficient on the real interest rate is not the IES, but the IES multiplied by
the share of CEQ consumers. Hence, λ = 0, the coefficient is going to be biased downwards.
Figure 10: Campbell and Mankiw (1989), Table 2.
7
Note that Campbell andThis
Mankiw in fact log
content downloaded fromconsumption
169.229.32.36 on when
Thu, 16 they estimate
Oct 2014 their regression, so in fact they
04:11:12 AM
are working with a specification close to the log-linearized CEQ.
28
Figure 11: Campbell and Mankiw (1989), Table 5.
3.3.4 Household level data: Shea (1995), Parker (1999), Souleles (1999) and Hsieh
(2003)
While early papers (like Hall (1978) or Campbell and Mankiw (1989)) tested the CEQ using
aggregate data, the literature quickly moved to testing the CEQ using household level data.
There are a number of reasons why this is more satisfying:
• aggregate consumption does not fluctuate much. By contrast, household level con-
sumption can fluctuate a lot more
• we can hope to identify more precisely expected future income for some categories of
workers (Shea (1995)) and so we get cross sectional variation that adds to the power of
the tests
• we can use ‘natural experiments’ where households learn about future income in
a measurable way, and then measure the response of consumption (Parker (1999),
Souleles (1999), Hsieh (2003)).
But household level data can also be problematic: accurate time series data on consumption
and income for a given household is hard to obtain. One needs to worry about sampling
weights, sample attrition...
• Shea (1995, AER): uses the PSID. PSID is a panel with reasonable income data, but
only food consumption. It also oversample poor households. Shea matches head of
This content downloaded from 169.229.32.36 on Thu, 16 Oct 2014 04:11:12 AM
households to union contracts andto uses
All use subject publicly
JSTOR Terms available information on union wage
and Conditions
growth to construct a measure of expected wage growth. He assigns respondents to

unions with national or regional bargaining: trucking, postal service, railroads... or
29
lumber in the Pacific Northwest, shipping on the East Coast.... He matches other
respondents to dominant employers in some areas (e.g. worker in the automobile
industry living near flint will be assigned to G.M.). Ends up with 647 observations
from 285 households. Finds that expected wage growth predicts expected consumption
growth.
• Parker (1999) uses the Consumption Expenditure Survey, used to construct the con-
sumption weights for the Consumer Price index. Not a panel, but repeated cross
sections. Each household stays in the CEX for 4 quarters, which yields 5 interviews,
with income and demographic data collected in interviews 2 and 5. The CEX has
excellent coverage of consumption, but relatively poor measures of income and also
has a very limited time dimension for each household. Parker uses the fact that different
households will hit the cap at different times during the calendar year and therefore
will see a drop in payroll taxes in different quarters. Crucially, the household knows
in advance whether he/she will hit the cap in a given quarter, so consumption should
not respond. Yet, he finds a predictable response of consumption. The elasticity of
consumption to predictable income is around 0.5.
• Souleles (1999) looks at income tax refunds, using the CEX. Since income tax refunds
are known in advance (taxpayers know their income of the previous year and calculate
the size of the refund when they file), consumption should not change when the refund
is received. The CEX survey asks the household about tax refunds in the second and
final interviews. Souleles finds that consumption responds to the tax refund. In a more
recent paper, Parker, Souleles, Johnson and McClellan (2013, AER) add a module to
the CEX questionnaire during the period of the 2008 stimulus payment (a tax cut).
Cross sectional variation arises from the random timing of receipts, based on the last
two digits of the recipient SSN. The amount of the tax cut was also well publicized
and known in advance ($300-$600 for individuals, $600-$1200 for couples....). They
find that households spent 12-30% of the tax cut.
• Hsieh (2003) instead exploits the annual payments of the Alaska Permanent fund to all
Alaskan residents. The amount disbursed is large and known in advance. Using the
CEX again, Hsieh finds that there is little evidence of a change in consumption during
the quarter of the receipt of the fund. Variation comes from changes over time in the
size of the transfers as well as cross sectional variation in family size. Importantly,
Hsieh finds that –while consumption does not respond to the payment of the Alaskan
Permanent Fund– it does respond to tax refunds, suggesting that there is an issue of
‘saliency’ of the change in income.
Why does expected income growth predict consumption growth?
• Consumption and leisure are substitutes (Aguiar and Hurst 2005 JPE). Home production
and work-related expenditures. Relevant when thinking about drop in consumption at
retirement. Food expenditures drops because home production (not recorded in food
expenditures) rises dramatically.
30
400 THE AMERICAN ECONOMIC REVIEW MARCH 2003
TABLE 2—RESPONSE OF CONSUMPTION TO ALASKA PFD
dlog(Nondurable dlog(Durable
consumption) consumption)
(1) (2) (3) (4) (5) (6)
PFD t # Family Size h 0.0002 !0.0167 !0.0034 !0.1659 !0.1741 !0.1488
Family Income h (0.0324) (0.0336) (0.0328) (0.0878) (0.0916) (0.0890)
Controls for:
Family size No No Yes No No Yes
Year dummies No Yes No No Yes No
Number of observations 806 806 806 806 806 806
Notes: Dependent variable is log(C IV /C III ). Standard errors are in parentheses. All regres-
sions are ordinary least squares (OLS) and include a quadratic in age and changes in the
number of children and adults in the household.
quarterly income, and z contains variables

Figure 12: Hsieh typical Table
for (2003): household2 in my sample by slightly
changes in the number of adults, number of more than 20 percent (see Table 1), the point
children, and a second-order polynomial in age estimate of the elasticity of nondurable con-
of the• head of the household
Households supporttolots capture the fact insumption
of dependent suggests that
mid-life (children, the Permanent
college...). Hence Fundthey
that household consumption is generally not flat payments increased household consumption by
over thehavelife no choice
cycle. The but to have
amount high
of the consumption
payment 0.004when percentincome is high
(roughly 4 cents) in the fourth
received by each household is calculated as the quarter of the year.
• households
product of PFD andare liquidity
family size. constrained
The main in-and impatients The estimate (more on this
in the basiclater)
specification in the
dependent variable is the percentage increase in first column is identified both by differences in
• some consumers
a household’s income in the arefourth
‘rule of thumb’
quarter dueconsumers
the size(Campbell
of the payment and Mankiw
across time 1989)
and across
to payments from the Permanent Fund, and the families of different sizes. The second column
• welfareofcosts
key parameter of not
interest is "optimizing
1 which measures
constantlyinare second
Table order for year effects and thus
2 controls
the elasticity of consumption to household in- identifies the effect of the Permanent Fund only
come. The dependent variable is the change in from the cross-sectional variation in family size.
household
3.3.5 A consumption
Detour: GMM (in logs) from the third
Estimation Although one should interpret these estimates
quarter to the fourth quarter of the year. As with caution since there are clearly reasons to
previously
With CRRA mentioned,
preferences, Alaskan residents
the general re- toexpect
equation the seasonal
be estimated is pattern of consumption to
ceived their dividend payments in the fourth quar- differ between families of different sizes, the
ter of the year. Under the certainty-equivalent c−θ = βE [R̃pointc−θ estimate
] of the elasticity of consumption
version of the LC/PIH (or a versiont of the t ist+1 stillt+1
essentially zero. The specification in the
LC/PIH in which the expected variance of con- third column controls for family size and thus
This is
sumption is aconstant),
non-linear "1 equation
should bewith equal at toleast two
onlyparameters: θ and β.across
uses the variation The general
time inideathe
of the Hansen and Singleton (1982) Generalized
zero. Method
amount of theofpayment
Moments (GMM)the
to identify estimation
consump-
−θthe re-
The first
method column
is to defineinzt+1
Table= 2β R̃presents
t+1 ct+1 − c−θ tion effects of the dividend payments. Once
t . Then, under the null that the model is correct,
sults of the first set of
Et zt+1 = 0. In other words, excess sensitivity tests for
zt+1 is unpredictable again,
basedone on should be cautious
any variable in tinterpreting
dated or earlier.
nondurable consumption.10 The point estimate these numbers, since the seasonal pattern of
Suppose we consider such a variable w i,t . Then we must have E[z w ] = 0. The sample
of "1 is positive, but economically and statisti- PT −1consumption may t+1 have i,tchanged over time.
analog
cally of this moment
insignificant; condition
it indicates that a is: vi = t=0Nonetheless,
10-percent wi,t zt+1 = the 0. Ifpoint
we have J such
estimate of thevariables,
income
increase
then we in canhousehold
estimate theincome increasesof con-
parameters the model elasticity v 0 Ωv where
of consumption
by minimizing is still veconomically
‘stacks’ all
sumption by 0.002 percent. Since the dividend 0 and statistically insignificant.
the moment conditions: v = (v1 , v2 , ...vJ ) and where Ω is some weighting matrix. This
payments increased the quarterly income of the The last three columns in Table 2 present
method is very general (and is one of the reasonsestimates Hansen of received the Nobel
the response prize in 2013).
of expenditures on
durables to the Permanent Fund payments using
10
All the regressions also include a constant. I do not use
the three excess sensitivity tests. The coefficient
the CEX’s sampling weights, although the results are vir- estimates are small but marginally significant.
tually identical if the weights are used. Surprisingly, the point estimates indicate that
31
Fig 1.—Percentage change in food expenditure, predicted food consumption index, and time spent on food production for male household
by three-year age ranges. Data are taken from the pooled
Figure 13:1989–91
Aguiar and 1994–96
& Hurst crossfigure
(2005), sections
1. of the CSFII, excluding the oversample of low-
households. The sample is restricted to male household heads (1,510 households). All series were normalized by the average levels for househol
aged 57–59. All subsequent years are the percentage deviations from the age 57–59 levels. See Sec. IV for details of data and derivation
consumption index
4 Moving beyond the CEQ
The CEQ model provides some interesting insights, mainly consumption smoothing (i.e.
the fact that consumption is going to respond to permanent changes more than to transitory
ones). Yet the model is soundly rejected.
In particular, consumption growth responds to income growth. The model also predicts
that consumption profile should be dictated by the interest rate and the rate of time preference,
not by the actual timing of income. This is also rejected: Carroll and Summers show that
consumption profiles track income profiles for different occupations or in different countries.
See figures 14 and 15.
Most households have relatively little wealth, so their consumption will mostly track their
income. Yet, the use their modest saving to ‘buffer’ income shocks. We are interested in
understanding these ‘buffer stock’ households.
32
322 Christopher D. Carroll and Lawrence H. Summers
Some Grade School Some High School Finished High School
71,25i.””a
1 50
1.W
t
:“,h
2’W
1.M
T
1.M
1D.75- 4 10.75- 1
25- 30- 35. 40 45- 50- 55- 60 25- M- 3s- 40- 45- 50- s5- 6a 7.5 30- 35- 4 0 45- Y)- 55- 60-
29 34 39 44 49 54 59 64 29 34 39 44 49 Y 59 64 29 34 39 44 49 54 59 64
Some College Finished College
0.75 h1-1 0.75

25- 30- 35- 40- 45- x)- 55- M 25 30 35- 4 0 45- 50- 55- 60-
29 34 39 44 49 54 r) 64 29 34 39 44 49 54 59 64
--b- Disposable Income d Consumplion
Fig. 1 0 . 7 ~ Income and consumption profiles by educational group, 1960-61

CES
Source: Calculations by authors using CES tapes.
Figure 14: Carroll & Summers (1991), figure 10.7a

cational and occupational groups. While the issue deserves further research,
our tentative conclusion is that parallel movements in income and consump-
tion cannot be explained by family size considerations.
Another explanation of the consumptionlincome parallel was provided by
Ghez (1975). Using the 1960 CES, Ghez 33 prepared a figure for all consumers
similar to our figures 10.7 and 10.8 for subcategories of consumers and sought
to explain the observed close correlation between income and consumption
using a “family production function” model of the type advocated by Becker
323 Consumption Growth Parallels Income Growth
Craftsmen operators hofessionals

1.75 1.15 1.75
1.50 1.50 1.50
1.25 1.25 1.25
1.00 1.M l.W
0.15 0.75 0.75
25- 30- 3% 40- 45- 50- 55- 60 25- 3 0 35- a 45- 50- 55- 60 25- 30- 35- 40- 45- 50- 55- a-
29 34 39 44 49 54 59 64 29 34 39 44 49 54 59 64 29 34 39 44 49 54 59 64
Unskilled Labor Clerical Managers

175 1.15 1.75
150 1.50 1.50
1.25 1.25 1.25
1.w 1.W I .XI
0.15 0.15 0.15
25- 30 35- 40- 45- 50- 55- M 25- 30- 35- 45- 50- 55- 6a 25- 30- 35. 40- 4s- 50- 55- 60-
29 34 39 44 49 54 59 64 29 34 39 44 49 54 59 64 29 34 39 44 49 54 59 64
Service SaleS Self-Employed

1.75 1.75 1.75
1.50 1.50 I .50
1.25 1.25 1.25
1.W 1.M 1.M
D.75 0.15 0.15
25- 30- 35- a 45- 50- 55- 60 25- 30. 35- 40- 45- m 55- M 25- 30- 35. 40- 45- 50- 5s- a-
29 34 39 44 49 54 59 64 29 34 39 44 49 54 59 64 29 34 39 44 49 54 59 64
-b- Disposable income COnSUmP~On
Fig. 10.7b Income and consumption profiles by occupational group, 1960-61

CES
Source: Calculations by authors using CES tapes.
The consumer will Figure 15: Carroll

therefore 34 consuming
& Summers
be observed (1991), figure 10.7b
more during those peri-
ods of life when he works most and earns the most income. To be more spe-
cific, this model would suggest that busy executives late in life would be more
likely to have a maid to do housekeeping chores and more likely to send out
4.1 Precautionary Saving
If we move away from linear marginal utility (i.e. quadratic utility), we open the door to
precautionary saving. In that case, the household will care about higher moments of future
consumption. Suppose that u0 (.) is convex, i.e. that u000 (.) is positive. In that case, we know
from the convexity of u0 that:
Et [u0 (ct+1 )] ≥ u0 (Et [ct+1 ])
and the inequality is strict when there is uncertainty about future consumption. This implies
that we cannot have ct = Et [ct+1 ]. If this were the case, then :
u0 (ct ) = u0 (Et [ct+1 ]) < Et [u0 (ct+1 )]
which would violate the Euler equation. Consumption ct has to be lower than in the CEQ
case , that is
ct < Et [ct+1 ]
Uncertainty depresses current consumption and thus raises saving. This is known as
precautionary saving (or saving for a rainy day).
u0 (c)
u0 (cl )
E[u0 (c)] A
u0 (E[c]) B
u0 (ch )
cl E[c] ch c
Figure 16: Convex marginal utility
35
For general preferences, Kimball (1990) showed that what matters for the precautionary
motive is the concavity of −u0 (.) (or equivalently the convexity of u0 (.)). He defined a
coefficient of relative prudence as:8
u000 (c)c
CRP = −
u00 (c)
Remark 8 For CRRA preferences u0 (c) = c−θ and the coefficient of relative prudence is
constant and equal to:
θ(1 + θ)c−θ−1
CRP = − =1+θ
−θc−θ−1
To see the role that prudence plays, consider the case where βR = 1 (so that a CEQ
household would want to keep consumption flat over time) and let’s perform a second-order
Taylor expansion of the Euler equation around ct :
E[u0 (ct+1 )] ≈ E[u0 (ct ) + u00 (ct )(ct+1 − ct ) + 1/2u000 (ct )(ct+1 − ct )2 ]
≈ u0 (ct ) + u00 (ct )E[(ct+1 − ct )] + 1/2u000 (ct )Et [(ct+1 − ct )2 ]
Now, substitute into the Euler equation (with βR = 1) to obtain:
u0 (ct ) = u0 (ct ) + u00 (ct )E[(ct+1 − ct )] + 1/2u000 (ct )Et [(ct+1 − ct )2 ]
Rearrange to solve for expected consumption growth:
ct+1 − ct u000 (ct )ct ct+1 − ct 2 ct+1 − ct 2

E[ ] = − 00 Et [( ) ] = CRP Et [( ) ]
ct u (ct ) ct ct
The slope of consumption growth is controlled by the coefficient of relative prudence

and something that looks like the conditional variance of consumption growth (it’s not quite
the variance since ct 6= Et [ct+1 ]).
Everything else equal, precautionary saving tends to increase expected consumption

growth. How does this modify the Euler equation estimation?
To answer this question, let’s consider again the log-linearized Euler equation (9) rewritten
below for convenience:9
1 1
Et ∆ ln ct+1 ≈ (rt+1 − ρ) + θVt ∆ ln ct+1
θ 2
8
There is also a coefficient of absolute prudence, defined as −u000 (c)/u00 (c).
9
Technical observation: in the log-linear Euler equation the conditional variance is scaled by θ while in the
derivation above, it is scaled by the CRP 1 + θ. The reason for the difference is that in the log-linear Euler
equation there is an adjustment for the concavity of the log (Jensen’s inequality).
36
This gives the following empirical specification:
1 1
∆ ln ct+1 = (rt+1 − ρ) + θVt ∆ ln ct+1 + εt+1
θ 2
Suppose for the time being that we ignore the precautionary term in this expression,
i.e. that we lump it with the error term of the regression εt+1 and estimate the first-order
log-linear Euler equation. This would be valid if we satisfy the orthogonality condition
that the precautionary term is orthogonal to the interest rate. Carroll (1997) [death to the
log-linearized Euler equation!] argues that this is unlikely to be the case. The reason is
that both the interest rate rt+1 and the conditional variance of consumption growth are
endogenous objects and are likely to interact in the general equilibrium of the economy.
This implies that the precautionary saving term is an omitted variable that is likely to be
correlated with the equilibrium interest rate.
To see how this works most simply, suppose that the economy is on a balanced growth
path where households face some level of idiosyncratic risk, but no aggregate risk.10 Along
such a balanced growth path, the growth rate of aggregate consumption must equal the growth
rate of aggregate income which is certain, since there is no aggregate risk, and which we
denote gy = ∆ ln yt+1 . If households are ex-ante identical, facing the same amount of risk
etc... they will all choose the same expected consumption growth, therefore also equal to gy .
It follows from the second order log-linearized Euler equation that:
1 1
gy ≈ (rt+1 − ρ) + θVt ∆ ln ct+1
θ 2
This equation tells us that in the aggregate equilibrium, the variance of idiosyncratic
consumption growth and the interest rate will be related by:
θ2
rt+1 = θgy + ρ − Vt ∆ ln ct+1
2
A higher amount of uninsurable idiosyncratic risk (as measured by Vt ∆ ln ct+1 ) will be
associated with a lower real riskfree interest rate rt+1 . Carroll concludes that estimating the
first-order log-linearized Euler equation (i.e. without controlling for precautionary saving) is
likely to be seriously misspecified. This can explain in particular why the estimated IES is
very low.
A way to address this critique would be to incorporate directly in the regression a term
that controls for the importance of precautionary saving, i.e. for the term 2θ Vt ∆ ln ct+1 in
the regression. This is what Dynan (1993) does by adding proxies for income uncertainty.
But it is difficult to obtain such estimates in the first place, and if we try to instrument for the
precautionary saving motive, we have to be careful to find instruments that for precautionary
saving that are independent from the interest rate, not an easy task.
10
Implicitly this requires that some risk sharing opportunities are not exploited. Otherwise, households would
like to diversify their idiosyncratic risk away. Technically, such models are called Bewley models, after Bewley
(1977). For a seminal Bewley model, see Aiyagari (1994).
37
4.2 The Buffer Stock Model
Intuitively, precautionary saving tilts-up consumption profiles and therefore leads to more
saving and wealth accumulation. Consider a household that faces income uncertainty. If that
household has a high wealth level, then heuristically income uncertainty should not matter
much and therefore consumption should not be too different from the certainty equivalent
framework (CEQ). We know that in that case, what controls the slope of the expected
consumption profile (and therefore of subsequent wealth) is whether βR is smaller of greater
than 1.
• If βR > 1, the household is patient and would like to save. In that case, the precaution-
ary and smoothing motive push in the same direction: eventually, the household will
manage to accumulate enough assets to insure against income fluctuations. In fact, if
βR > 1 an infinitely lived household would accumulate an unbounded level of assets.
• If βR = 1, the argument is a bit more subtle, but the result is the same. Here,
the household would like to smooth marginal utility. It will be able to do this by
accumulating an unbounded amount of wealth.11 The upshot is that if βR ≥ 1 the
model is not terribly interesting: the household would just accumulate vast amounts
of wealth, enough to be indifferent to the impact of income fluctuations on marginal
utility. This is neither interesting nor realistic!
• The last case is when βR < 1. In that case, the household is impatient. A CEQ
household would run choose to consume more today and run down assets. But by
running down assets, it increases the strength of the precautionary saving motive since
income fluctuations are more likely to impact marginal utility. So this case present
an interesting tension: on the one hand, the household would like to save to smooth
fluctuations in marginal utility. On the other hand, it wants to consume now and prefers
not to accumulate wealth. The result from this tension is that the household will aim
to achieve a certain target level of liquid wealth, but not more. Once households have
accumulated this target level of wealth, consumption will tend to track income at high
frequency (even in response to predictable income change), thus potentially explaining
the excess sensitivity puzzle. It can also explain why consumption tracks income at
low frequency (explaining the Carroll-Summers (1991) empirical patterns in figures
14 and 15). This is the buffer-stock model.
Let’s flesh the details of that model out. Consider a household with standard preferences:
∞
X
U = max β t u(ct )
{ct }
t=0
11
This result is formally established by Schechtman (1975) and Bewley (1977). See Deaton (1991) for a
discussion.
38
and with a budget constraint:
at+1 = R(at + ỹt − ct )
The household faces a constant interest rate R but a stochastic income stream {ỹt },
where we assume for simplicity that ỹt is independently identically distributed every period.
We assume that βR < 1 so that, if there was no uncertainty, the household would prefer
to consume now and would run down assets over time, and even borrow against future income.
How much would the household borrow? If ymin ≥ 0 is the lowest possible realization
of income every period, then it is immediate to show that the household would not be able
to run its asset levels below amin = −ymin /(R − 1).12 If the household borrowed a larger
amount at any point in time, there would be a strictly positive probability that it would not be
able to repay. In other words, amin is the natural borrowing limit faced by the household. It
is the present value of the lowest possible income the household would receive from now
on, and at ≥ amin . Of course, it is possible that the household faces a stronger liquidity
constraint than the natural borrowing limit, if access to credit markets is limited. This is a
relevant feature of the world since many people face limited access to credit markets.
In order to fix ideas, we are going to consider an extreme case where the household
cannot borrow at all. That is, we impose the restriction that:13
at ≥ 0
If there was no uncertainty, the solution to the household consumption-saving problem

would be quite straightforward: it would run down initial assets a0 , then set consumption
equal to income. With uncertainty, this is not going to be optimal for the reasons discussed
above: it would leave the household exposed to too much fluctuations in marginal utility.
Therefore, there should be some target level of liquid wealth that the household would
like to revert to.
We can write the income fluctuations problem as (see Deaton (1991)):

∞
X
U = max β t u(ct )
{ct }
t=0
subject to:
at+1 = R(at + ỹt − ct )

at ≥ 0
ct ≥ 0
12
This condition derives from the intertemporal budget constraint and the requirement that consumption
remain positive.
13
This corresponds to the natural borrowing limit if ymin = 0.
39
It is useful to express the problem in terms of ‘cash on hand’ xt , defined as the amount
of liquid resources the household has access to at the beginning of the period:
x t = a t + yt
The constraints of the problem become:
xt+1 = R(xt − ct ) + ỹt+1

0 ≤ ct ≤ xt
Let’s define v(xt ) the value function of this problem. We can write the associated Bellman
equation:
v(xt ) = maxct u(ct ) + βEt [v(xt+1 )]

s.t.
xt+1 = R(xt − ct ) + ỹt+1
0 ≤ ct ≤ xt
The first order condition associated with this Bellman equation is:
u0 (ct ) = βREt [v 0 (ct+1 )] + λt
where λt is the Lagrange multiplier associated with the constraint ct ≤ xt .14 The comple-
mentary slackness condition is:
λt (xt − ct ) = 0
For the usual envelope reasons, the marginal value of cash on hand satisfies:
v 0 (xt ) = βREt [v 0 (xt+1 )] + λt = u0 (ct )
It follows that:
• when the credit constraint does not bind, the usual Euler equation holds:
u0 (ct ) = βREt [u0 (ct+1 )]
• when the credit constraint binds, λt > 0 and ct = xt and
u0 (xt ) > βREt [u0 (ct+1 )]
We can summarize both cases as follows:
u0 (ct ) = max βREt [u0 (ct+1 )], u0 (xt )

(10)
The credit constraint ct ≤ xt operates in two ways:

14
Technically there is another Lagrange multiplier associated with the constraint ct ≥ 0, but this one will
never bind as long as the Inada conditions are satisfied, so we ignore it here.
40
1. If the household is constrained at time t, it is forced to consume less than desired.
2. The credit constraint also matters, even in periods where it does not bind directly,
because of the likelihood that it will bind in the future. Technically, this is encoded in
Et [u0 (ct+1 )]. The curvature of marginal utility leads the household to save more to
reduce the likelihood of being constrained in the future.
This model cannot be solved in closed form. Instead, we have to resort to numerical
techniques to characterize optimal consumption behavior. Denote ct = f (xt ) the optimal
consumption rule followed by the household. It is not a function of time because the problem
is recursive and stationary. We can then rewrite the Euler equation as:
u0 (f (xt )) = max βREt [u0 (f (xt+1 ))], u0 (xt )

(11)
xt+1 = R(xt − f (xt )) + ỹt+1 (12)
The problem becomes one of solving for the function f (.). The right hand side of equation
(14) defines a functional equation:
T (f )(x) = u0−1 max βRE[u0 (f (x+1 ))], u0 (x)

x+1 = R(x − f (x)) + ỹ+1

where u0−1 is the inverse of the marginal utility (assumed well defined). The optimal
consumption rule is then a fixed point of the operator T (f ):
f (x) = T (f )(x)
Not surprisingly, the regularity condition that ensures that this operator has a unique
fixed point is βR < 1, i.e. precisely the requirement that the household is impatient.15
Moreover, this fixed point can be obtained by iteration. Suppose that we have a candidate
consumption function c(x) = f n (x). Then we can construct f n+1 (x) as
f n+1 (x) = T (f n )(x)
i.e. as the solution of:
u0 (f n+1 (x)) = max βREt [u0 (f n (x+1 ))], u0 (x)

(13)
n
x+1 = R(x − f (x)) + ỹ+1 (14)
The sequence f n (x) converges uniformly to f (x), i.e. limn→∞ ||f n (x) − f (x)|| = 0
where ||.|| is some Euclidean distance.
This is called Euler equation iteration.16
Figure 17 shows the optimal consumption rule for this problem for the case where
ymin > 0. It has the following properties:
15
Technically, this condition ensures that the operator T (f ) is a contraction mapping.
16
Another approach, called value function iteration works with the value function v(xt ) that solves the Bellman
equation.
41
Certainty Equivalent Buffer Stock Model
forward looking much less forward looking
retirement saving households will not save for retirement at age 20
consumption and income paths independent Once you have your buffer, gc ≈ gy
interest rate elasticity small effect of interest rate
uncertainty does not matter uncertainty matter
Table 1: Comparing CEQ and Buffer Stock Models
• consumption is a function of x, not y.
• below a certain threshold level x∗ , the household prefers to consume all its assets:
c = x. This is because the current marginal utility of consumption is very high.
• above x∗ , the consumption rule is concave, and always below the certainty equivalent
consumption
• we can represent expected consumption growth Et ∆ ln c as a function of cash on hand

x. It is a decreasing function:
– for low levels of wealth, precautionary saving dominate, cash on hand will
increase and consumption is expected to grow.
– For high levels of cash on hand, consumption grows at rate βR < 1 so cash on
hand decreases.
– The target level of cash on hand can be defined as that level that remains constant
(in expectations), i.e. the level x∗∗ such that E[xt+1 |xt = x∗∗ ] = x∗∗ . Carroll
(2012) shows that that expected consumption growth is below 1 and above βR at
the target level of cash-on-handx∗∗ .17
• Even if c is a function of x, once x is close to its target, c will move together with y: if
y is expected to decline, then consumption will decline once x declines (not before):
predictable movements in y will translate into movements in c.
Figure 18 reports the dynamics of the buffer stock model, as computed by Carroll (2012).
We can summarize the two models as in table 1:
4.3 Consumption over the Life Cycle

See Gourinchas & Parker (2002) [GP]. Revisits the question of optimal consumption behavior:
• model with both lifecycle saving motive and precautionary saving motive
• structural estimation of the consumption function, i.e. not relying on Euler equation,
or reduced form consumption functions
17
For more on this, see Carroll (2012), “Theoretical foundations of buffer stock saving.”
42
1228 ANGUS DEATON
0
CN4
0
0
~~~~~~~~~=(/h+rx)/(l1+r)
0
-,0
E
cn 0 y is N(100,a), r=O.05, 6=0.10
0
0 coses are from top to bottom
p=2, a=1O and o-=15
?0 . . , H p=3, a=1 0 aond o- 1 5
0 40 80 120 160 200 240 280 320

income+assets
FIGURE 1.-Consumption functions for alternative utility functions and income dispersions.
p(x) > A(x) for x > x*, so that17:

Figure weDeaton
have(1991), figure 1
(13) c=f(x)=X, XsAX*,
• estimation based on household level data using income and consumption expenditures
c=f(x) Ax, x> x*.
The estimation procedure consists in constructing age-profiles of consumption based on
The consumption
micro function
data and estimating thereforeofhas
the parameters the general
the consumption shapethat
problems best replicate
shown in Figure 1,
these age profiles in the model.
shown there for y, distributed as N(A, a-), u = 100, r = 0.05, 8 = 0.10, and
A(c) = c-. Such figures appear in Mendelson and Amihud (1982) and are
4.3.1 The Model
"smoothed" versions of the piecewise linear consumptionfunctions derived in
theEach
certainty case
household byforHeller
lives andworks
T periods, Starrfor (1979) and Helpman
N periods. GP truncate(1981). Theatfigure
the problem
shows four different
retirement by writing: consumptionfunctionscorrespondingto the four combina-
tions of two values of p and o-;NX −1
they t
all beginN as the 45-degreeline, and diverge
U = E [
from it and one another at their respective
0 β u(ct ) + β values
VN (aN )]
of x*, all of which, in this
t=0
case, are a little below ,u, the mean value of income, shown as a vertical line.
Thesubject
otherto:
line in the figurewill be discussedbelow.
t+1 = R(at + yt − ct )
The general properties of athe solution are clear. Startingfrom some initial
level ofTheassets, VN (.)
functionthe summarizesreceives
household preferences from retirement
a draw of0 income.onwards, including
If the any
total value of
bequest motive. GP assume that preferences are CRRA: u (c) = c−θ . Further, they assume
assets and income is below the critical level x*, everythingis spent, and the
household goes into the next period with no assets. If the total is greater than
x*, somethingwill be held over, and43the new, positive level of assets will be
carriedforwardto be added to the next period's income. Note that there is no
presumptionthat savingwill be exactlyzero; consumptionis a functionof x, not
of y, and f(x) can be greater than, less than, or equal to y. Assets are not
E[ct+1 /ct ]
A
1
D
βR
x∗∗
Figure 18: The Buffer Stock Model
that labor income yt has a transitory and a permanent component:
yt = p t µt
pt = Gt pt−1 η t
where µt and η t are iid. GP assume that µt = 0 with some probability p. This is meant
to capture unemployment risk. One implication is that ymin = 0 so the natural borrowing
limit is amin = 0. However, with preferences that satisfy the Inada conditions, the household
will never choose to hit the borrowing limit.18 Otherwise, ln µt is N (0, σ 2µ ). ln η t is also
normally distributed N (0, σ 2η ). The variance of these shocks and the unemployment risk are
calibrated to household level data.
The problem features two state variables: cash on hand xt = at + yt and the permanent
level of income pt (since the latter conditions how large future income will be). In general,
the complexity of numerical problems grows exponentially with the number of state variables
(curse of dimensionality). Even with modern computers and parallelization techniques, we
cannot realistically solve problems with more than 1 or 2 state variables. Fortunately, the
assumptions of the problem allow to implement a normalization that reduces the number of
18
To see this, observe that if a = 0 in one period, then there is a strictly positive probability that the agent will
have zero consumption next period if the unemployment shock is realized.
44
56 p.-o. gourinchas and j. a. parker
Figure 19: Gourinchas & Parker (2002), figure 1
state variables. Define ẑ as the ratio of variable z to the permanent component of income:
ẑ = z/p. Then, we can rewrite the problem’s Euler equation as:
u0 (ĉt (x̂t )) = βREt u0 (Gt+1 η t+1 ĉt+1 (x̂t+1 ))

where
x̂t+1 = R(x̂t − ĉt )/(Gt+1 η t+1 ) + µt+1
Notice that the consumption rules are indexed by age, since households of different ages
face different remaining horizons. So unlike the infinite horizon model, we don’t need to look
for a fixed point of the Euler equation, but iterate backwards, starting from the consumption
rule at retirement, assumed to be linear in cash on hand:19
ĉN = γ 1 x̂N
For any value of the parameters, the resulting consumption functions can be evaluated
Figure
numerically (see figure 19). 1.—A tale of two households.
4.3.2 Estimating the Structural Model

Once the optimal consumption rules ĉt (x̂) are evaluated, GP construct the age-consumption
age 66 characterized by &0 = 0001 &1 = 0071. These parameter values
profiles C̄t by aggregating over the distribution of possible realizations of the state variables
are those
This implies that in 15
of our baseline estimates. If G and Z are constant, the finite horizon problem
19
principle we don’t need to ensure that the functional equation defines a contraction
−θ
mapping. If the horizon were infinite, the condition for buffer stock behavior would be βRE[(Gη) ] < 1
would converge to the infinite horizon one, as we move further away from retire-
ment. Consumption is always positive, increasing, 45 and concave in cash on hand.
Early in life, households exhibit ‘buffer stock’ behavior: for low levels of cash
on hand, typically less than the permanent component of their income x ≤ 1,
x and p: Z
ln c̄t = E[ln(ĉt (xt /pt )pt )] = ln(ĉt (xt /pt )pt )dFt (xt , pt ; ψ)
where dFt (., .) denotes the joint distribution, according to the model, of cash on hand
and permanent income at age t. In practice, this joint distribution is a complex object to
calculate. An important step is to construct the moments above by simulating a large number
of households (a Monte Carlo simulation). The simulated moment converges to the true
model moment as the size of the simulation increases.
The last step is in matching these simulated moments to the same moments constructed
on the household level data. The algorithm chooses the vector of structural parameters to
minimize the distance between P these simulated moments and the data moments.
If we denote gt (ψ) = 1/It i ln ci,t − ln c̄t (ψ) as the distance between simulated and
data consumption at age t, then the estimator minimizes
g(ψ)0 W g(ψ)
where W is a weighting matrix and g = (g1 , g2 , ..., gN −1 )0 .
The results of the estimation indicate the following:
• Consumption tracks income over the lifecycle
• The estimated parameters (θ and β) are quite reasonable with β = 0.96 and θ = 0.51,
with R − 1 = 3.44%. The model thus features “buffer stock” behavior in the sense that
agents want to keep a constant target level of cash on hand around 1.2 times permanent
income until around age 40.
• Around age 40, life cycle considerations kick in and saving increase markedly. At
that point, precautionary saving become less relevant since the higher liquid wealth
allows the household to smooth consumption. Hence in the latter phase of active life,
households behave like CEQ consumers with consumption growth controlled by βR.
• in the early part of their lifecycle, income is growing but households are unable to
borrow much. Hence precautionary saving dominates.
• The structural parameters are well identified precisely because the turning point at
which savings increase is determined by the relative strength of the two saving motives.
The lifecycle model with uninsurable labor income risk has become a workhorse to
evaluate quantitatively various policies. For instance, Scholz et al (2006) use it to evaluate
whether Americans are saving optimally. Their model features uncertain lifetimes, unin-
surable earnings, medical expenses, progressive taxation, government transfers and social
security benefits. They use the model to compare, household by household, wealth predictions
and find that the model accounts for more than 80% of the 1992 cross sectional variation in
46
see that the occupation and education groups with the most pronounced humps
in income present the most pronounced humps in consumption. Further, we can
formally reject the null hypothesis that the consumption profiles are flat. This is
essentially a now standard test of the linearized consumption Euler equation, as
studied by Attanasio and Weber (1995), Lusardi (1996).
Our profiles differ slightly from the results of Attanasio and Browning (1995)
and Attanasio and Weber (1995). These papers employ a larger set of preference
shifters: once controlling for these, consumption is smoother and the CEQ-LCH
is not rejected. In Attanasio and Weber (1995) and in the linear Euler-equation
approach generally used in micro data, precautionary effects are omitted so that
preference shifters absorb, correctly or incorrectly, variation in consumption that
we attribute to uncertainty. Clearly, allowing for enough preference variation can
Figure 3.—Household consumption and income over the life cycle, by education group.
Figure 20: Gourinchas & Parker (2002), figure 2
47
74 p.-o. gourinchas and j. a. parker
Figure 6.—Normalized target cash-on-hand by age.

Figure 21: Gourinchas & Parker (2002), figure 6. Target cash on hand over age.
households to smooth high frequency movements in income so that their behavior
morewealth.
closelyThemimics that
paper uses dataoffrom
certainty-equivalent consumers.
the Health and Retirement Study (HRS) supplemented with
We next decompose
restricted social securitytotal saving
data on and
earning wealth at
realizations each age
throughout life.into
The that
resultsdriven
indicateby
life-cycle considerations and that additional amount driven by the presence of
that most households save enough for retirement, especially give the fact that they contribute
uninsurable risk. Our
to social security previousretirement
and employer discussion
plans.might lead the reader to think that
agents have no concern for retirement when they are young and no concern
for labor
5 Asset income uncertainty later in life. This is incorrect since consumers are
Pricing
rational and perfectly foresee their retirement needs. First, we define saving in
liquid
5.1wealth as the discounted
The Canonical Model Againvariation in financial
with Multiple wealth from one period to
Assets
35
the next:
We now switch focus and use the canonical model to tell us about asset prices. Let’s consider
again the
S canonical
= W model.
− WNow assume
/R = Rthat the household
− 1/RW + Ycan−
invest
C in two assets:
i t i t+1 i t i t i t i t
• a riskless asset that pays a riskfree return Rt+1
Saving is equal to investment income—liquid and illiquid—plus labor income
minus consumption. From
• a risky asset that paysour empirical
a risky return Z̃profiles
t+1 in Figure 5, it follows that house-
holds save relatively
The problem little
of the and consume
household is: roughly their income on average early
in life. Second, at the estimated parameters, we ∞ compute the consumption path,
'CtLC (, that would occur if allUincome risks were pooled,
β t u(ct ) so that for all households
X
= max{ct ,ωt } E0
LC 36
Yt = E26 )Yt *, but the household’s environment t=0 otherwise remains unchanged.
Finally, weto:define life-cycle saving as the difference between total income and
subject
35 at+1 = (Rt+1 (1 − ω t ) + Z̃t+1 ω t )(at + ỹt − ct )
The discount comes from the assumption that income is received and consumption occurs at the
beginning of the period.
36
In order to do this, we input the consumption 48
rule at retirement as estimated in our benchmark
case. Our estimates imply that if households faced no risks after retirement, the age of death, N , is a
reasonable 87 years. That is, if we set preferences at our estimated values and N = 87, the standard
life cycle model with a certain date of death implies the same value function at retirement as we
estimate (up to a constant).
612 journal of political economy
TABLE 1
Descriptive Statistics for the Health and Retirement Study (Dollar Amounts in
1992 Dollars)
Standard
Variable Mean Median Deviation
1991 earnings $35,958 $28,976 $39,368
Present discounted value
of lifetime earnings $1,718,932 $1,541,555 $1,207,561
Defined-benefit pension
wealth $106,041 $17,327 $191,407
Social security wealth $107,577 $97,726 $65,397
Net worth $225,928 $102,600 $464,314
Mean age (years) 55.7 4.7
Mean education (years) 12.7 3.4
Fraction male .70 .46
Fraction black .11 .31
Fraction Hispanic .06 .25
Fraction couple .66 .48
No high school diploma .22 .41
High school diploma .55 .50
College graduate .12 .33
Postcollege education .10 .30
Fraction self-employed .15 .35
Fraction partially or fully
retired .29 .45
Source.—Authors’ calculations from the 1992 HRS. The table is weighted by the 1992 HRS household analysis
weights.
random-effect assumptions
Figure 22: with homoskedastic
Scholz , Sheshadri errors to
& Khitatrakun estimate
(2006), equa-
Table 1
tion (1).
We estimate the model separately for four groups: men without a
college degree, men with a college degree, women without a college
degree, and women with a college degree. In the online Appendix B
we present details of the empirical earnings model and coefficient es-
timates from that model, and we describe our Gibbs sampling procedure
that we use to impute earnings for individuals who refuse to release or
who have top-coded social security earnings histories.6 Our approach is
appealing in that it uses information from the entire sequence of in-
dividual earnings, including uncensored W-2 data from 1980–91, to im-
pute missing and top-coded earnings.
Table 1 provides descriptive statistics for the HRS sample. Mean (me-
dian) earnings in 1991 of HRS households are $35,958 ($28,976),
though note that 29 percent of the sample was partially or fully retired
when interviewed in the 1992 HRS. The mean (median) present dis-
counted value of lifetime household earnings is $1,718,932
6
We repeated our central empirical analyses dropping individuals who refused to release
their social security records and generated results nearly identical to those reported in
the paper. Brief details are given in the sensitivity analysis, Sec. IVE.
Fig. 1.—Median defined-benefit pension wealth, social security wealth, and net worth (excluding defined-benefit pensions) by lifetime earnings decile
(1992 dollars).
Figure 23: Scholz , Sheshadri & Khitatrakun (2006), Figure 1
49
This content downloaded from 128.32.105.89 on Mon, 25 Aug 2014 18:54:33 PM
TABLE 2
Optimal Net Worth (Excluding Social Security and Defined-Benefit Pensions) and the Percentage of Population Failing to Meet
Optimal Targets (Dollar Amounts in 1992 Dollars)
Median Mean Percentage Median Median

Optimal Optimal below Median Median Social Defined-Benefit
Wealth Wealth Optimal Deficit Net Security Pension
Target Target Target (Conditional) Worth Wealth Wealth
Group (1) (2) (3) (4) (5) (6) (7)
All households $63,116 $157,246 15.6% $5,260 $102,600 $97,726 $17,327
No high school diploma 20,578 70,711 18.6 2,632 36,800 72,561 0
High school diploma 63,426 139,732 15.6 5,702 102,885 97,972 21,290
College degree 128,887 243,706 12.7 14,092 209,616 127,704 60,752
Postcollege education 158,926 333,713 13.2 23,234 253,000 129,320 152,781
Earnings decile:
Lowest 2,050 48,445 30.4 2,481 5,000 26,202 0
2nd 13,781 55,898 28.7 3,328 25,500 42,159 0
3rd 26,698 84,582 21.8 5,948 43,485 57,844 0
4th 43,566 123,441 19.4 4,730 75,000 77,452 14,830
Middle 53,709 128,285 16.9 6,979 90,000 94,929 29,497
6th 76,462 131,565 10.8 10,000 124,348 119,011 45,613
7th 80,402 154,891 9.9 11,379 128,580 133,451 56,033
8th 101,034 180,643 5.5 21,036 167,000 151,397 71,373
9th 136,075 238,186 4.4 5,206 226,000 163,639 104,657
Highest 238,073 463,807 5.4 25,855 393,000 202,659 126,998
Source.—Authors’ calculations as described in the text.
Figure 24: Scholz , Sheshadri & Khitatrakun (2006), Table 2

626 journal of political economy

Fig. 2.—Scatter plot of optimal and actual wealth. Observed net wealth is constructed
from the 1992 HRS. Optimal net worth comes from solving the baseline model described
in the text.
A. Are Households Preparing Optimally for Retirement?

Figure 2 gives a scatter plot of optimal net worth against actual net
Figure 25: Scholz , Sheshadri & Khitatrakun (2006), Figure 2
worth, for HRS households with optimal and actual net worth between
$0 and $1,000,000. The curved line gives a cubic spline of the median
values of observed and optimal net worth.25 If household net worth was
exactly the same as optimal net worth, the ordered pairs of actual and
50
optimal net worth for the HRS sample would map out the 45-degree
line. In fact, the ordered pairs cluster just below the 45-degree line. The
scatter plot gives striking visual evidence that most HRS households
have saved at or above their optimal retirement targets.
A second striking aspect of figure 2 is that it illustrates how a well-
This problem features two control variables: how much to consume ct , and how much to
invest in the risky asset (ω t ). Following the same steps as before (Bellman equation, first
order condition, envelope theorem), one can show that the following conditions hold:
h i
u0 (ct ) = βEt u0 (ct+1 )(Rt+1 (1 − ω t ) + Z̃t+1 ω t )
h i
Rt+1 Et u0 (ct+1 ) = Et u0 (ct+1 )Z̃t+1

The first condition is simply the usual Euler equation. The second one is an asset pricing
condition. In fact, combining the two equations, for any asset Z̃i,t , optimal portfolio allocation
requires that 0
u (ct+1 )
1 = Et β 0 Z̃i,t+1
u (ct )
How should we interpret this condition? Recall that we can write, for any two random
variables: E[XY ] = E[X]E[Y ] + cov(X, Y ). Substituting, we obtain:
h i
1 = Et Mt,t+1 Z̃i,t+1
h i
= Et [Mt,t+1 ] Et Z̃i,t+1 + covt Mt,t+1 , Z̃i,t+1
h i
Et Z̃i,t+1 = Rt+1 1 − covt Mt,t+1 , Z̃i,t+1
where we define Mt,t+1 = βu0 (ct+1 )/u0 (ct ), i.e. the intertemporal marginal rate of substi-
tution and the last line uses the fact that, from the Euler equation,
Et [Mt,t+1 ] Rt+1 = 1
The interpretation is now quite straightforward: an asset requires a high expected return
if it covaries negatively with the pricing kernel Mt,t+1 , i.e. if the return on the asset is low
when M is high. Now if we go back to the definition of the IMRS, we see that it is high
when consumption is low. In other words, an asset requires a premium if it offers a poor
return precisely at times when consumption is low.
The required excess return satisfies
h i
Et Z̃i,t+1 − Rt+1 = −Rt+1 covt Mt,t+1 , Z̃i,t+1
5.2 Stock Prices: a Present Value Formula

We can use the previous pricing equation to evaluate the value of a stock. Suppose we have
an asset with price Pt at time t, resale value Pt+1 at time t + 1 and a dividend dt+1 in period
t + 1. Then the return to the asset is:20
Pt+1 + dt+1
Zt+1 =
Pt
20
Here Pt is the price after the dividend in period t has been paid, i.e. an ex-dividend price
51
Substituting into the asset pricing equation, and solving forward, we obtain:
"∞ #
X
Pt = Et Mt,t+s dt+s
s=1
in other words, the stock price is the expected PDV of future dividends, evaluated using
the pricing kernel defined recursively as Mt,t+s = Mt,t+s−1 Mt+s−1,t+s .
5.3 The Equity Premium

Consider the pricing equation derived earlier:
h i
1 = Et Mt,t+1 Z̃i,t+1
Now, let’s log-linearize in the case where utility is CRRA, assuming that consumption
growth and asset returns are jointly lognormally distributed:
1 = exp(−ρ)Et [exp(−θ∆ ln ct+1 ) exp(ln Zi,t+1 )]

1 = exp(−ρ) exp (−θEt [∆ ln ct+1 ] + Et [ln Zi,t+1 ] + 1/2vart (−θ∆ ln ct+1 + ln Zi,t+1 ))
0 = −ρ − θEt [∆ ln ct+1 ] + Et [ln Zi,t+1 ] + 1/2vart (−θ∆ ln ct+1 + ln Zi,t+1 )
where the last line takes logs. Now taking the difference between a risky and the riskess asset,
we obtain:
Et [ln Zi,t+1 ] − rt+1 + 1/2vart (Zi,t+1 − Rt+1 ) = θcovt (∆ ln ct+1 , ln Zi,t+1 − Rt+1 )
Et [ln Zi,t+1 ] + 1/2vart (Zi,t+1 − Rt+1 ) − rt+1 = θcovt (∆ ln ct+1 , ln Zi,t+1 − Rt+1 )
The last term on the left hand side in this equation is a Jensen’s inequality term. The
left hand side measures the equity premium. Mankiw and Zeldes estimate that it is about 6
percentage points. The standard deviation of consumption growth is 3.6 percentage points
and the standard deviation of the excess return is 16.7 percentage points. The correlation of
consumption growth and the excess return is 0.40. It follows that the right hand side is equal
to θ × 0.40 × 3.6 × 16.7. To match the equity premium, the CRRA coefficient needs to be
about 25.
52

Notes For Econ202A: Consumption: Pierre-Olivier Gourinchas UC Berkeley

Uploaded by

Copyright:

Available Formats

Notes For Econ202A: Consumption: Pierre-Olivier Gourinchas UC Berkeley

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Notes For Econ202A: Consumption: Pierre-Olivier Gourinchas UC Berkeley

Uploaded by

Copyright:

Available Formats

Notes for Econ202A:

– consumption-saving. Large part of national output.

2 Consumption under Certainty

• A household (of size 1!) lives T periods (from t = 0 to t = T − 1). Lifetime

• Seems like a reasonable problem to analyze.

• Uncertainty. In particular, there is uncertainty about what T is. Whose T are we

• Aggregation. AggregatePconsumption expenditures represent expenditures on many

• Time additivity. The marginal utility of consumption at time t only depends on

• Intertemporal Marginal Rate of Substitution. Consider two consecutive periods t

2.3 The Intertemporal Budget Constraint

The first order condition for ct is:

from which we obtain:

• consumption is a function of total wealth.

• the marginal propensity to consume is (1 − β)/(1 − β T ) and converges to 1 − β when

• if β = 0.96 (a reasonable estimate), this gives R = 1/0.96 = 1.0416. Then we should

2.6 The Permanent Income Hypothesis

• This is Friedman’s permanent income hypothesis. Individual consumption is not

– you earn an extra $200 today

in the cross section in the time series

Figure 1: Keynesian consumption functions and the PIH

• in the time series, more variation from ytP : higher b̂

Figure 2: The Baseline Life-Cycle Model

Equations for the simple Modigliani-Brumberg (1954) lifecycle model:

• The household runs total wealth to 0

• Human wealth runs out at t = N . It is supplemented by financial wealth

• We can have positive financial wealth even if there is no bequest motive.

2.9 Saving and Growth in the LifeCycle Model

Figure 3: The Life-Cycle Model when Income grows over time

2.10 Interest Rate Elasticity of Saving

• if 1/θ = 0 (Leontieff) then ct is flat regardless of the interest rate. No substitutability

• for 1/θ < 1: weak substitution effects

• for θ = 1 income and substitution effects cancel out (log preferences)

ct+1 /ct = (βR)1/θ

What about the overall effect?

2.10.1 The 2-period case with y1 = 0

• the budget curve rotates around (y0 , y1 ).

Figure 4: An increase in interest rates when y1 = 0

– if c0 = y0 and c1 = y1 then there is no income and wealth effect (why? because

2.10.2 The 2-period case with y1 6= 0 and c0 = y0

2.10.3 The 2-period case when c0 < y0 (lender)

Figure 5: An increase in interest rates when c0 = y0 and c1 = y1

2.10.4 2-period case when c0 > y0 (borrower)

Furthermore, empirical estimates of the elasticity of intertemporal substitution suggest

Figure 6: An increase in interest rates for a saver (c0 < y0 )

2.11 The LifeCycle Model under Certainty Again

Substituting into the budget constraint, we obtain:

Figure 7: An increase in interest rates for a borrower (c0 > y0 )

3 Consumption under Uncertainty

g equation (5) it is apparent that TABLE 1-THE INTEREST ELASTICITY

Note: The calculation assumes n =.015, g =.02, T'= 50,

at+1 = R̃t+1 (at + ỹt − ct ) (4)

3.1.2 Recursive Representation

v(at ) = maxct ∈Ct u(ct ) + βEt [v(at+1 )]

3.1.3 Optimal Consumption and Euler Equation

v 0 (at )da = u0 (ct )dct + βEt [v 0 (at+1 )R̃t+1 (da − dc)]

u(c) = αc − γc2 /2; γ > 0, α > 0

• even if α > 0, utility turns negative for sufficiently large consumption

E0 ct = E0 Et−1 ct = E0 ct−1 = ... = c0