Estimating Discount Functions

with
Consumption Choices over the
Lifecycle

David Laibson, Andrea Repetto, and

Jeremy Tobacman

Current Draft: August 2005

Hyperbolic discounting

A model of self-control.

People are patient in the long run, but impatient

in the short run.

We think we should: Exercise more, smoke less,

and eat more healthily.

But we postpone until tomorrow the decision to

go running, quit smoking, and forgo the Ben and
Jerry's.
 Another example:

I prefer a full day o in 101 days to half day o in 100

days.

But I prefer half day o right now to a full day o

tomorrow.

These preferences are time inconsistent:...

Once it gets to be 100 days from now, I once again

prefer instantaneous grati cation

... and induce a self-control problem

 An intergenerational discount function introduced by
Phelps and Pollak (1968) provides a particularly tractable
way to capture the intrapersonal e ects described above.

Quasi-hyperbolic discounting (Laibson, 1997), present-

biased preferences (O'Donoghue and Rabin, 1999),
quasi-geometric discounting (Krusell and Smith 2000):

1; ; 2; 3 ; :::

Ut = u(ct) + u(ct+1) + 2 u( c 3 u( c
t+2 ) + t+3 ) + :::

For exponentials: =1

Ut = u(ct) + u(ct+1) + 2u(ct+2) + 3u(ct+3) + :::

For hyperbolics: <1

h i
Ut = u(ct)+ u(ct+1) + 2 u(c 3 u(c
t+2 ) + t+3 ) + :::
Self-reports about saving.

Baby boomers report median target savings rate

of 15%.

Actual median savings rate is 5%.

76% of household's believe they should be saving

more for retirement (Public Agenda, 1997).

Of those who feel that they are at a point in their

lives when they \should be seriously saving al-
ready," only 6% report being \ahead" in their
saving, while 55% report being \behind."

Consumers report a preference for at or rising

consumption paths.
Further evidence: Normative value of commitment.

\Use whatever means possible to remove a set

amount of money from your bank account each
month before you have a chance to spend it."

Choose excess withholding.

Cut up credit cards, put them in a safe deposit

box, or freeze them in a block of ice.

\Sixty percent of Americans say it is better to

keep, rather than loosen legal restrictions on re-
tirement plans so that people don't use the money
for other things."

Social Security.

Christmas Clubs (10 mil. accounts).

Outline

1. Hyperbolic Discounting

2. Facts

3. Model

4. Estimation Procedure

5. Results

6. Conclusion
1 Consumption-Savings Behavior

Substantial retirement wealth accumulation (SCF)

Extensive credit card borrowing (SCF, Fed, Gross

and Souleles 2000, Laibson, Repetto, and Tobac-
man 2000)

Consumption-income comovement (Hall and Mishkin

1982, many others)

Anomalous retirement consumption drop (Banks

et al 1998, Bernheim, Skinner, and Weinberg 1997)
2 Data

Statistic me seme
% borrowing on `Visa' ? 0.68 0.015
(% Visa)

borrowing / mean income 0.12 0.01

(mean Visa)

C-Y comovement 0.23 0.11

(CY )

wealth
median 50-59 income 3.88 0.25
wealth
weighted mean 50-59 income 2.60 0.13
(wealth)
Three moments on previous slide (wealth, % Visa,
mean Visa) from SCF data. Correct for cohort,
household demographic, and business cycle ef-
fects, so simulated and empirical hh's are anal-
ogous. Compute covariances directly.

C-Y from PSID:

ln(Cit) = Et 1 ln(Yit) + Xit + "it (1)

C drop from PSID

RETIRE + X
ln(Cit) = Iit it + "it (2)

\A Debt Puzzle": only \% Visa" and \wealth"

\JEP paper": add \liquid share" and \% low

liquid wealth"
3 Model

Recent consumption papers use simulations

Rich environments, eg with income uncertainty

and liquidity constraints

Literature pioneered by Carroll (1992, 1997), Deaton

(1991), and Zeldes (1989)

Gourinchas and Parker (2001) use method of sim-

ulated moments (MSM) to estimate a structural
model of life-cycle consumption
3.1 Demographics

Mortality, Retirement (PSID), Dependents (PSID),

HS educational group

3.2 Income from transfers and wages

Yt = after-tax labor and bequest income plus govt

transfers (assumed exog., calibrated from PSID)

yt ln(Yt): During working life:

yt = f W (t) + ut + W
t (3)

During retirement:

yt = f R (t) + R
t (4)
3.3 Liquid assets and non-collateralized debt

Xt + Yt represents liquid asset holdings at the

beginning of period t:

Credit limit: Xt Yt

= :30; so average credit limit is approximately

$8,000 (SCF).
3.4 Illiquid assets

Zt represents illiquid asset holdings at age t:

Z bounded below by zero.

Z generates consumption ows each period of

Z.

Conceive of Z as having some of the properties

of home equity.

Disallow withdrawals from Z ; Z is perfectly

illiquid.

Z stylized to preserve computational tractability.

1. House of value H , mortgage of size M .

2. Consumption ow of H; minus interest cost of

M; where = i (1 ) :

3. =) net consumption ow of H M
(H M ) = Z: We've explored di erent possi-
bilities for withdrawals from Z before.
3.5 Dynamics

Let ItX and ItZ represent net investment into as-

sets X and Z during period t

Dynamic budget constraints:

Xt+1 = RX (Xt + ItX )

Zt+1 = RZ (Zt + ItZ )
Ct = Yt ItX ItZ

Interest rates:
(
RCC if Xt + ItX < 0
RX = ; RZ = 1
R if Xt + ItX > 0

h i
Three assumptions for RX ; ; RCC :

Benchmark: [1:0375; 0:05; 1:1175]

Aggressive: [1:03; 0:06; 1:10]
Very Aggressive: [1:02; 0:07; 1:09]
3.6 Time Preferences

Discount function:

f1; ; 2; 3 ; :::g

= 1: standard exponential discounting case

< 1: preferences are qualitatively hyperbolic

Null hypothesis: =1

T
X
Ut(fC gT=t) = u( C t ) + u( C ) (5)
=t+1
In full detail, self t has instantaneous payo function
Ct + Zt 1
nt 1
u(Ct; Zt; nt) = nt
1
and continuation payo s given by:
T +N
X t
i i 1
j=1 st+j (st+i) u(Ct+i; Zt+i; nt+i):::
i=1
T +N
X t
i i 1
+ j=1 st+j (1 st+i) B (Xt+i; Zt+i)
i=1

nt is e ective household size: adults+(.4)(kids)

Zt represents real after-tax net consumption ow

st+1 is survival probability

B ( ) represents the payo in the death state

3.7 Computation

Dynamic problem:

max u(Ct; Zt; nt) + EtVt;t+1( t+1)

ItX ,ItZ
s:t: Budget constraints

t = (Xt + Yt; Zt; ut) (state variables)

Functional Equation:

Vt 1;t( t) =
fst[u(Ct; Zt; nt)+ EtVt;t+1( t+1)]+(1 st)EtB ( t)g

Solve for eq strategies using backwards induction

Simulate behavior

Calculate descriptive moments of consumer be-

havior
4 Estimation

Estimate parameter vector and evaluate models wrt data.

me = N empirical moments, VCV matrix =

ms ( ) = analogous simulated moments

q ( ) (ms ( ) me) 1 (m (
s ) me)0, a scalar-
valued loss function

Minimize loss function: ^ = arg min q ( )

^ is the MSM estimator.

Pakes and Pollard (1989) prove asymptotic con-

sistency and normality.

Speci cation tests: q (^) 2 (N #parameters)

5 Results

Exponential ( = 1) case:
^ = :8459 (0:025); q ^; 1 = 436

Hyperbolic case:
(
^ = :703 (0:1093)
^ = :958 (0:0068) q ^; ^ = 67

h i
(Benchmark case: RX ; ; RCC = [1:0375; 0:05; 1:1175])
Punchlines:

estimated signi cantly below 1.

Reject = 1 null hypothesis with a t-stat of 6.

Speci cation tests rejects the exponential model.

Benchmark Exponential Hyperbolic Data Std err
ms( ^ ; ^)
ms(1; ^) ^ = :703
Statistic: ^ = :846 me seme
^ = :958
% V isa 0.67 0.63 0.68 0.015

mean V isa 0.15 0.17 0.12 0.01

CY 0.29 0.31 0.23 0.11

wealth -0.05 2.69 2.60 0.13

q (^) 436 67
Robustness
h i
Benchmark: RX ; ; RCC = [1:0375; 0:05; 1:1175]
h i
Aggressive: RX ; ; RCC = [1:03; 0:06; 1:10]
h i
Very Aggressive: RX ; ; RCC = [1:02; 0:07; 1:09]

Benchmark Aggressive Very Aggressive

exp
^ .846 .930 .923
(:025) (:001) (:002)

q ^; 1 436 278 64

hhypi
^; ^ [:958; :703] [:944; :815] [:932; :909]
(:007) ; (:109) (:001) ; (:014) (:002) ; (:016)

q ^; ^ 67 45 33
 Aggressive
Exponential Hyperbolic Data Std err
(3%; 6%; 10%)
ms( ^ ; ^)
ms(1; ^) ^ = :815
Statistic: ^ = :930 me seme
^ = :944
% V isa 0.44 0.65 0.68 0.015

mean V isa 0.08 0.16 0.12 0.01

CY 0.10 0.22 0.23 0.11

wealth 2.50 2.61 2.60 0.13

q (^) 278 45
Very Aggressive
Exponential Hyperbolic Data Std er
(2%; 7%; 9%)
ms( ^ ; ^)
ms(1; ^) ^ = :909
Statistic: ^ = :923 me seme
^ = :932
% V isa 0.58 0.65 0.68 0.015

mean V isa 0.12 0.15 0.12 0.01

CY 0.14 0.19 0.23 0.11

wealth 2.53 2.66 2.60 0.13

q (^) 64 33
6 Conclusion

Structural test using the method of simulated mo-

ments rejects the exponential discounting null.

Speci cation tests rejects the exponential model.

Quantitative results are sensitive to interest rate

assumptions.

Hyperbolic discounting does a better job of match-

ing the available empirical evidence on consump-
tion and savings.