An Introduction to Macroeconomics with

Household Heterogeneity

Dirk Krueger1
Department of Economics
University of Pennsylvania

March 6, 2018

1I wish to thank Orazio Attanasio, Richard Blundell, Juan Carlos Conesa,

Jesus Fernandez-Villaverde, Robert Hall, Tim Kehoe, Narayana Kocherlakota,
Felix Kubler, Fabrizio Perri, Luigi Pistaferri, Ed Prescott, Victor Rios-Rull and
Thomas Sargent for teaching me much of the material presented in this manuscript.
c by Dirk Krueger. All comments are welcomed, please contact the author at
dkrueger@econ.upenn.edu.
ii
Contents

I Introduction 1
1 Overview over the Monograph 3

2 Why Macro with Heterogeneous Households (or Firms)? 7

3 Some Stylized Facts and Some Puzzles 9

3.1 Household Level Data Sources . . . . . . . . . . . . . . . . . . 9
3.1.1 CEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.2 SCF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.3 PSID . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.4 CPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.5 Data Sources for Other Countries . . . . . . . . . . . . 11
3.2 Main Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.1 Organization of Facts . . . . . . . . . . . . . . . . . . . 12
3.2.2 Aggregate Time Series: Means over Time . . . . . . . . 12
3.2.3 Means over the Life Cycle . . . . . . . . . . . . . . . . 16
3.2.4 Dispersion over the Life Cycle . . . . . . . . . . . . . . 18
3.2.5 Dispersion (Inequality) at a Point in Time . . . . . . . 19
3.2.6 Changes in Dispersion over Time . . . . . . . . . . . . 19
3.2.7 Other Interesting Facts (or Puzzles) . . . . . . . . . . . 19

4 The Standard Complete Markets Model 23

4.1 Theoretical Results . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.1 Arrow-Debreu Equilibrium . . . . . . . . . . . . . . . . 25
4.1.2 Sequential Markets Equilibrium . . . . . . . . . . . . . 34
4.2 Empirical Implications for Asset Pricing . . . . . . . . . . . . 39
4.2.1 An Example . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3 Tests of Complete Consumption Insurance . . . . . . . . . . . 50

iii
iv CONTENTS

4.3.1 Derivation of Mace’s (1991) Empirical Specifications . . 51

4.3.2 Results of the Tests . . . . . . . . . . . . . . . . . . . . 54
4.3.3 The Problem of Measurement Error . . . . . . . . . . . 57
4.3.4 Other Empirical Issues and Solutions . . . . . . . . . . 59
4.3.5 Self-Selection into Occupations with Differential Ag-
gregate Risk . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4 Appendix A: Discussion of the No Ponzi Condition . . . . . . 67
4.5 Appendix B: Proofs . . . . . . . . . . . . . . . . . . . . . . . . 69
4.6 Appendix C: Properties of CRRA Utility . . . . . . . . . . . . 71
4.6.1 Constant Relative Risk Aversion . . . . . . . . . . . . 72
4.6.2 Constant Intertemporal Elasticity of Substitution . . . 73
4.6.3 Homotheticity and Balanced Growth . . . . . . . . . . 75

II The Standard Incomplete Markets Model (SIM) 77

5 The SIM in Partial Equilibrium 81
5.1 A Simple 2 Period Toy Model . . . . . . . . . . . . . . . . . . 82
5.1.1 Household Decision Problem . . . . . . . . . . . . . . . 82
5.1.2 Comparative Statics . . . . . . . . . . . . . . . . . . . 83
5.2 The General Model with Certainty Equivalence . . . . . . . . 85
5.2.1 Nonstochastic Income . . . . . . . . . . . . . . . . . . 86
5.2.2 Stochastic Income and Quadratic Preferences . . . . . 90
5.3 Prudence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3.1 A Simple Model and a General Result . . . . . . . . . 103
5.3.2 A Parametric Example for the General Model . . . . . 114
5.4 Liquidity Constraints . . . . . . . . . . . . . . . . . . . . . . . 118
5.4.1 The Euler Equation with Liquidity Constraints . . . . 119
5.4.2 Precautionary Saving Due to Liquidity Constraints . . 121
5.4.3 Empirical Tests of Liquidity Constraints . . . . . . . . 122
5.5 Prudence and Liquidity Constraints: Theory . . . . . . . . . . 123
5.5.1 Finite Lifetime . . . . . . . . . . . . . . . . . . . . . . 124
5.5.2 Infinite Horizon . . . . . . . . . . . . . . . . . . . . . . 125
5.6 Prudence and Liquidity Constraints: Computation . . . . . . . 139
5.6.1 IID Income shocks . . . . . . . . . . . . . . . . . . . . 141
5.6.2 Serially Correlated, Mean-Reverting Income Shocks . . 141
5.6.3 Permanent Shocks . . . . . . . . . . . . . . . . . . . . 143
5.6.4 From Policy Functions to Simulated Time Series . . . . 148
CONTENTS v

5.6.5 Implementation of Specific Algorithms . . . . . . . . . 149

5.7 Prudence and Liquidity Constraints: Empirical Implications . 150
5.7.1 Time Series Implications: How Does Consumption Re-
spond to Income Shocks . . . . . . . . . . . . . . . . . 150
5.7.2 Cross-Sectional Implications: How Does Consumption
Inequality Evolve over the Life Cycle . . . . . . . . . . 154
5.8 Appendix A: The Intertemporal Budget Constraint . . . . . . 154
5.9 Appendix B: Derivation of Consumption Response to Income
Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

6 A Digression: Stochastic Earnings or Wage Processes 159

6.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.3 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.3.1 RIP vs. HIP . . . . . . . . . . . . . . . . . . . . . . . . 165
6.3.2 Size of ρ . . . . . . . . . . . . . . . . . . . . . . . . . . 167

7 The SIM in General Equilibrium 169

7.1 A Model Without Aggregate Uncertainty . . . . . . . . . . . . 170
7.1.1 The Environment . . . . . . . . . . . . . . . . . . . . . 170
7.1.2 Theoretical Results: Existence and Uniqueness . . . . . 178
7.1.3 Computation of the General Equilibrium . . . . . . . . 183
7.1.4 Qualitative Results . . . . . . . . . . . . . . . . . . . . 184
7.1.5 Numerical Results . . . . . . . . . . . . . . . . . . . . 185
7.2 An Incomplete Markets Model with Unsecured Debt and Equi-
librium Default . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.2.1 Model Overview . . . . . . . . . . . . . . . . . . . . . . 189
7.2.2 Institutional Details of Bankruptcy . . . . . . . . . . . 189
7.2.3 Household Problem in Recursive Formulation . . . . . 190
7.2.4 Production Firms . . . . . . . . . . . . . . . . . . . . . 192
7.2.5 Financial Intermediaries . . . . . . . . . . . . . . . . . 192
7.2.6 Stationary Recursive Competitive Equilibrium . . . . . 192
7.2.7 Characterization of Household Default Decision . . . . 194
7.2.8 Characterization of Equilibrium Loan Interest Rate Func-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.2.9 Bringing the Model to the Data: Calibration and Es-
timation . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.2.10 Quantitative Predictions of the Model . . . . . . . . . 196
vi CONTENTS

7.3 Unexpected Aggregate Shocks and Transition Dynamics . . . 196

7.3.1 Definition of Equilibrium . . . . . . . . . . . . . . . . . 197
7.3.2 Computation of the Transition Path . . . . . . . . . . 199
7.3.3 Welfare Consequences of the Policy Reform . . . . . . 201
7.4 Aggregate Uncertainty and Distributions as State Variables . . 203
7.4.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.4.2 Computation of the Recursive Equilibrium . . . . . . . 209
7.4.3 Calibration . . . . . . . . . . . . . . . . . . . . . . . . 211
7.4.4 Numerical Results . . . . . . . . . . . . . . . . . . . . 214
7.4.5 Why Quasi-Aggregation? . . . . . . . . . . . . . . . . . 217
7.4.6 Rich People are Not Rich Enough . . . . . . . . . . . . 218

III Complete Market Models with Frictions 221

8 Limited Enforceability of Contracts 225
8.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
8.2 Constrained Efficient Allocations . . . . . . . . . . . . . . . . 228
8.3 Recursive Formulation of the Problem . . . . . . . . . . . . . 229
8.4 Decentralization . . . . . . . . . . . . . . . . . . . . . . . . . . 230
8.5 An Application: Income and Consumption Inequality . . . . . 232
8.6 A Continuum Economy . . . . . . . . . . . . . . . . . . . . . . 237

9 Private Information 245

9.1 Partial Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . 245
9.1.1 Properties of the Recursive Problem . . . . . . . . . . 249
9.2 Endogenous Interest Rates in General Equilibrium . . . . . . . 253
9.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
9.3.1 New Dynamic Public Finance . . . . . . . . . . . . . . 253

IV Conclusions 255
 Part I

Introduction

1
Chapter 1

Overview over the Monograph

This monograph is meant to be an introduction into the research field of

macroeconomics with household heterogeneity. It is concerned with con-
structing, computing and empirically evaluating dynamic stochastic macroe-
conomic models in which individual households differ according their abili-
ties, wages, incomes, preferences or market opportunities in a way such that
the household sector cannot be aggregated into a representative agent. In
these models, therefore, aggregate allocations and prices will depend on the
cross-sectional household distribution of household characteristics. We will
argue below that the macroeconomic implications of these models differ, often
(but not always) in a quantitatively significant way, from their representative
agent counterpart. In addition, these models are in principle suitable to ask
and answer positive and normative distributional questions about which the
representative household paradigm is silent by construction.
In the next chapter we will briefly summarize selected empirical obser-
vations on wages, earnings, income, consumption and wealth from cross-
sectional, household level data sets that have motivated this literature, and
briefly describe the data sources from which these observations are drawn.
We will also document some empirical puzzles that we aim at explaining with
the models to be developed below. Note that to call an empirical finding a
puzzle requires to take a stand on the standard economic theory relative to
which the finding is puzzling; on of the goals of the course is to develop
extensions or, if needed, radical departures, of standard theory to explain
the empirical puzzles. The remainder of the monograph is then devoted to
the construction, analysis and applications of theoretical models aimed at
explaining these facts that also can be used for applications (e.g. to the

3
4 CHAPTER 1. OVERVIEW OVER THE MONOGRAPH

positive or normative analysis of fiscal policy).

In order to provide a taxonomy of the models discussed in this mono-
graph, consider a world in which an infinitely-lived households is faced with
income risk and aims at maximizing expected lifetime utility. Let s ∈ S de-
note a state of nature, where S is a finite set. For ease of exposition1 let the
shock s be iid over time, with probabilities π(s). Household income is given
by y(s). Households desire smooth consumption in the presence of stochastic
income fluctuations, and the key distinguishing feature of the models dis-
cussed in this monograph is the set of financial assets that households have
access to, and the extent to which they can go short in these assets. On one
extreme, households have no access to any formal or informal ways to smooth
income fluctuations, and thus their consumption fully inherits the stochastic
properties of the income process. I will call these households hand-to-mouth
consumers, and the associated (non-) market structure financial autarchy.
On the other extreme households may be able to trade a full set of state-
contingent short-term bonds, so that their budget constraint reads as
X
c+ q(s0 )a0 (s0 ) ≤ y(s) + a(s) (1.1)
s0 ∈S

where c is current consumption, y(s) is current income, a0 (s0 ) is the number

of bonds bought today that pay of one unit of consumption in state s0 tomor-
row, and q(s0 ) is its price. If households face no other trading constraints2
I will call the resulting model the standard complete markets model. The
next chapter is devoted to the review of its theoretical predictions and em-
pirical tests. With this market structure (and under suitable assumptions
on household preferences) household consumption is fully insured against
household-specific (idiosyncratic) income shocks and the household sector
can be represented by a representative consumer. As a result, with this mar-
ket structure macroeconomics can proceed by ignoring underlying household
heterogeneity induced by income shocks. This model therefore serves as an
important point of departure and comparison for models in which household
heterogeneity does matter for macroeconomic outcomes.
The remainder of this monograph is then devoted to models whose market
structure lies in between those of financial autarchy and complete markets,
1
In this introduction, as in most of this monograph, my focus is not on utmost gener-
ality, but rather on clarity of exposition without (hopefully) compromising on rigor.
2
One obviously needs some constraints on short-sales to rule out Ponzi schemes the
presence of which shall always be assumed in this monograph even if not explicitky stated.
 5

and in which households will be able to partially, but not fully (self-)insure
against random income fluctuations. In part II I will discuss the perhaps most
important workhorse model in this literature. In the standard incomplete
markets model (SIM) households have access only to a single, one period
risk-free bond, so that their period budget constraint reads as
c + qa0 ≤ y(s) + a (1.2)
and the shortsale of bonds a0 might be limited by a constraint of the form
a0 ≥ −Ā. In chapter 5 I will discuss the theoretical properties of various
versions (with alternating assumptions on the household utility function, the
stochastic nature of the income process, and the tightness of the borrowing
constraint) of the model in which the price of the bond q and thus the interest
rate 1 + r = 1/q is exogenously given. (and as a consequence we can analyze
the behavior of one household in isolation). After a brief digression in chapter
6 that discusses the properties of the main driving of this class of models, the
stochastic process for earnings, I will then incorporate the decision problems
from chapter 5 into a dynamic general equilibrium model in which interest
rates and wages are determined endogenously in the labor and capital market.
This is done in chapter 7. There, individual households’ consumption and
saving decisions are aggregated to obtain aggregate labor and capital supply,
firms’ decisions are aggregated to obtain aggregate labor and capital demand,
and wages and interest rates move to clear both markets. Depending on
whether individual uncertainty averages out in the aggregate (no aggregate
uncertainty) wages and interest are constant over time or are themselves
stochastic processes (presence of aggregate uncertainty), leading to severe
computational problems when computing these models. The aggregation
of individual decisions also leads to (possibly time-varying) cross-sectional
consumption and wealth distributions; thus these models are possibly useful
for the study of the effects of redistributive and social insurance policies such
as progressive taxation, unemployment insurance, welfare or social security.
Common to all these models is the assumption of the absence of explicit
insurance arrangements in an environment in which mutual insurance is po-
tentially quite beneficial. In part III I will discuss a strand of the literature
that aims to explain the stylized facts from chapter 3with models that depart
directly from the complete markets model, without a priori ruling out explicit
insurance contracts (as the standard incomplete market model does). Chap-
ter 8 discusses models in which imperfect consumption insurance arises due
to the assumed imperfect enforceability of insurance contracts. If households
6 CHAPTER 1. OVERVIEW OVER THE MONOGRAPH

are given the option to default, with the punishment, say, of being expelled
into financial autarchy forever after (and thus becoming the hand-to-mouth
consumers discussed above), full consumption insurance might not be incen-
tive compatible. I will argue that this results in a model in which households
face a budget constraint of the form (1.1), exactly as in the standard complete
market model, but now in addition face state-contingent shortsale constraints
on the state contingent bonds a0 (s0 ) ≥ Ā(s0 ) that are potentially very tight
and do not permit the implementation of the full consumption insurance al-
location. Finally I will briefly discuss, in chapter 9 models in which perfect
consumption insurance might be impossible to be implemented since individ-
ual incomes are not publicly observable. Consequently a mechanism designer
of consumption insurance contracts will have to respect the incentives of
households to mis-report their incomes and to construct a partial-insurance
consumption insurance contract that induces all households to report their
incomes truthfully.
A short, necessarily subjective assessment of where this area of research
stands and where it might be headed will conclude this monograph.
Chapter 2

Why Macro with

Heterogeneous Households (or
Firms)?

• Empirical Fact: lots of observed and unobserved heterogeneity among

households (Heckman, Wolpin and many other micro labor people)
• Models: Household heterogeneity affects:
– Aggregate quantities (Aiyagari ’94) and prices (Huggett ’93) and
their evolution (Krusell and Smith ’98). But how much?
– Answers to normative questions (e.g. Lucas ’87 vs Krusell-Smith
’99 and Krebs ’07 on the cost of business cycles)
– Answers to counterfactual policy (effect of temporary tax cuts, see
Heathcote ’05, Kaplan and Violante ’12, ’13)
• Many questions involve distributional aspects that cannot be studied
in rep. agent world
– What explains aggregate trends in inequality (e.g. Krueger and
Perri ’06)
– Aggregate and redistributive consequences of change in tax pro-
gressivity, social security (e.g. Auerbach and Kotlikoff ’87)
• Note: strong overlap in questions with empirical micro, especially labor
economics.

7
8CHAPTER 2. WHY MACRO WITH HETEROGENEOUS HOUSEHOLDS (OR
Chapter 3

Some Stylized Facts and Some

Puzzles

In this chapter we will discuss the main stylized facts that heterogeneous
agent macro models are designed to explain. We are mainly interested in four
main economic variables, labor earnings, income, wealth and consumption

3.1 Household Level Data Sources

3.1.1 CEX
For the US, the only household level data set with extensive information
about a wide range of consumption expenditures is the Consumer Expen-
diture Survey (CEX) or (CES).1 The CEX is conducted by the U.S. Bu-
reau of the Census and sponsored by the Bureau of Labor statistics. ¿From
1980 onwards the survey is carried out on a yearly basis. The CEX is a
so-called rotating panel: each household in the sample is interviewed for
four consecutive quarters and then rotated out of the survey. Hence in each
quarter 20% of all households is rotated out of the sample and replaced by
new households. In each quarter about 3000 to 5000 households are in the
sample, and the sample is representative of the U.S. population. The main
advantage of the CEX is that it contains very detailed information about
1
In class I will also discuss some European data sources, such as the British Fam-
ily Expenditure Survey (FES) and the German Socio-Economic Panel (SOEP) and the
Einkommens- und Verbrauchsstichprobe.

9
10 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES

consumption expenditures. Information about income and wealth is inferior

to the Survey of Consumer Finances (SCF), the Current Population survey
(CPS) and the Panel Study of Income Dynamics (PSID), also the panel di-
mension is significantly shorter than for the PSID (one household is only
followed for 4 quarters). For further information and the complete data set
see http://www.stats.bls.gov/csxhome.htm.

3.1.2 SCF
With respect to income, the PSID as well as the SCF contains data that are
supposedly of higher quality than the income data from the CEX. The SCF is
conducted in three year intervals; the four available surveys are for the years
1989, 1992, 1995 and 1998. It is conducted by the National Opinion Research
center at the University of Chicago and sponsored by the Federal Reserve sys-
tem. It contains rich information about U.S. households’ income and wealth.
In each survey about 4,000 households are asked detailed questions about
their labor earnings, income and wealth. One part of the sample is represen-
tative of the U.S. population, to give an accurate description of the entire
population. The second part over-samples rich households, to get a more
precise idea about the precise composition of this groups’ income and wealth
composition. As we will see, this group accounts for the majority of total
household wealth, and hence it is particularly important to have good infor-
mation about this group. The main advantage of the SCF is the level of detail
of information about income and wealth. The main disadvantage is that it
is not a panel data set, i.e. households are not followed over time. Hence
dynamics of income and wealth accumulation cannot be documented on the
household level with this data set. For further information and some of the
data see http://www.federalreserve.gov/pubs/oss/oss2/98/scf98home.html.

3.1.3 PSID
The Panel Study of Income Dynamics (PSID) is conducted by the Survey
Research Center of the University of Michigan and mainly sponsored by the
National Science Foundation. The PSID is a panel data set that started
with a national sample of 5,000 U.S. households in 1968. The same sam-
ple individuals are followed over the years, barring attrition due to death or
non-response. New households are added to the sample on a consistent basis,
making the total sample size of the PSID about 8700 households. The income
3.1. HOUSEHOLD LEVEL DATA SOURCES 11

and wealth data are not as detailed as for the SCF, but its panel dimension
allows to construct measures of income and wealth dynamics, since the same
households are interviewed year after year. Also the PSID contains data on
consumption expenditures, albeit only food consumption. In addition, in
1990, a representative national sample of 2,000 Latinos, differentially sam-
pled to provide adequate numbers of Puerto Rican, Mexican-American, and
Cuban-Americans, was added to the PSID database. This provides a host of
information for studies on discrimination. For further information and the
complete data set see http://www.isr.umich.edu/src/psid/index.html

3.1.4 CPS
The Current Population Survey (CPS) is conducted by the U.S. Bureau of
the Census and sponsored by the Bureau of Labor Statistics. In its annual
March supplement detailed information about household income is collected.
The survey started to gather information about household income in 1948,
but comprehensive information about household income and income of its
members is available only since 1970’s. The main advantage of the CPS is
its sample size: in each year it contains a representative sample of 40,000 to
60,000 households. However, no information about consumption or wealth
information is collected. Also, this survey, like the SCF is a purely cross-
sectional data set without panel dimension as it does not follow individual
families over time. Fore more details see http://www.bls.census.gov/cps.

3.1.5 Data Sources for Other Countries

• Family Expenditure Survey (FES) for the U.K. Very similar in scope
to the American CEX.

• Socio-Economic Panel (SOEP) for Germany. Very similar in scope to

the American PSID.

• Einkommens- und Verbrauchsstichprobe for Germany. Similar to CEX,

but less frequent.

• Italian Survey of Household Income and Wealth (SHIW). Biannual

panel data for income, consumption and wealth.

• Eurosystem Household Finance and Consumption Survey (by ECB)

12 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES

• Many other national data sets for other countries. See RED 2011, Vol.
1.
• Luxembourg Income Study as well as the Luxembourg Wealth Study

3.2 Main Findings

3.2.1 Organization of Facts
• Variables of interest: w, h, wh, y, c, a
• Dimensions of the data
µ σ ∆σ
age a means (1) See next box → var. over lc (2)
time t RA macro ineq. levels (3) Ineq. trends (4)
• Follow a cohort over time as she ages: Life cycle means (1) and cross-
cohort variances (2)
• Cross-section of population at point in time (3)
• Evolution of the cross-section over time (inequality trends), (4)
• Household debt and default

3.2.2 Aggregate Time Series: Means over Time

The focus of this monograph is to explain household behavior and its im-
plications for the macro economy. The main economic choices individual
households make are how much to work, and how much of the resulting
income to consume and save.
In figure ?? we plot real GDP and real total household consumption
expenditures, on a logarithmic scale, for the U.S. from 1964 to 2012. We
observe that 60-70% of GDP is used for private consumption expenditure.
This fraction is fairly stable, but displays an upward trend starting in the
80’s and continuing through 90’s. Associated with this increase is a decline
in the personal savings rate sp , which is defined as
Cp
sp = 1 −
Ypdi
3.2. MAIN FINDINGS 13

Real GDP and Consumption Per Capita in the U.S., 1964−2012

GDP
10.8 Consumption
Log Real GDP and Consumption p.c.

10.6

10.4

10.2

10

9.8

9.6

9.4

1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Figure 3.1: Aggregate Consumption over Time

where Cp is total private consumption expenditures and Ypdi is personal dis-

posable income.
Consumer durables make up about 13% of total nominal consumer ex-
penditures, fraction slightly increasing. Fraction devoted to nondurables de-
clined, now about 30%. Expenditures for consumer services about 60% of
total consumer expenditures now, increasing. BUT: use of nominal data
misleading because of changes in relative prices. Now: deflate components
by their relative price (data from 1987). Consumers seem to reallocate con-
sumption towards durables, away from nondurables. Reconciliation with first
graph: relative price of services has increased. Consumption is smoother
over the business cycle than GDP; Indicates households’ (limited) ability to
isolate their consumption path from income fluctuations induced by the busi-
ness cycle. Nondurable consumption least volatile, then services; purchases
of consumer durables fluctuate most over the cycle.

Summary Statistics on Consumption Growth (1959-96)

Nodur. & Serv. Dur.
Annual Growth Rate p.a. 2.3% 4.8%
Std. Dev. of Gr. Rate 1.8% 6.9%
14 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES

Share of Consumption in GDP (Personal Income), 1964−2012

C/GDP
C/Pers. Inc.

0.8
Consumption Share

0.75

0.7

0.65

0.6
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Figure 3.2: Consumption Share over Time

Use of Personal Consumption in the U.S., 1964−2012

0.6

0.5
Consumption Shares

0.4
Nondurable Goods
Durable Goods
0.3 Services

0.2

0.1

0
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Figure 3.3: Components of Household Consumption over Time

3.2. MAIN FINDINGS 15

Real Consumption in the U.S., 1995−2012

0.6
Real Consumption Shares

0.5

0.4
Nondurable Goods
Durable Goods
0.3 Services

0.2

0.1

0
1996 1998 2000 2002 2004 2006 2008 2010 2012
Year

Figure 3.4: Components of Household Consumption over Time, Goods-

Specific Prices

Personal Saving Rate, 1964−2012

15
Personal Saving Rate

10

0
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Figure 3.5: Household Saving Rate over Time

16 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES

Figure: Labor Income and Net Worth by Age, SCF 2007 ($1,000)

1200.00 120.00

1000.00 100.00

800.00 80.00

600.00 60.00

400.00 40.00

Net Worth (left axis)

200.00 20.00
Labor Income (right axis)

0.00 0.00
20-29 30-39 40-49 50-59 60-69 70 or more
Age Group

Figure 3.6: Income and Net Worth over the Life Cycle

3.2.3 Means over the Life Cycle

In this section we document mean life cycle profiles of labor earnings, wealth
(net worth) and consumption. In figure 3.2.3 we plot the average (across
households of a specific age) labor income and net worth against household
age.2
As we see from this figure, in the data, mean labor income (right scale)
has a hump shape over the life cycle and peaks at age of around 45 to 50.
The hump is very significant in size: average earnings at age 45 is almost 2.5
times as at age 25. And at age 65 households on average earn only about 60-
70% of the earnings of the typical household age 45. W also observe that net
worth (left scale) is accumulated over the life cycle, and then is decumulated
in retirement. However, average net worth is still very significantly positive
at old age.
In figure 3.2.3 we plot consumption over the life cycle. The figure contains
two plots, both based on average consumption across households of a given
age. One plot measures household consumption as the total consumption
expenditure of a household, the other aims at constructing per capita con-
sumption in the household, by dividing household consumption expenditures
2
This plot is derived from data from the 2007 SCF, and reproduced from Glover et al.,
2014.
3.2. MAIN FINDINGS 17

Expenditures, Total and Adult Equivalent

4500
T otal
Adul t Equi val ent

4000

3500

3000

2500

2000

1500
20 30 40 50 60 70 80 90
Age

Figure 6: Consumption over the Life Cycle

Figure 3.7: Consumption over the Life Cycle

by a household equivalence scale that controls for household size.3

We observe that consumption follows qualitatively the same hump shape
over the life cycle as labor income does. It appears that consumption tracks
income over the life cycle.
Excess sensitivity puzzle: consumption excessively sensitive to pre-
dicted changes in income:
However, in the data family size is also hump-shaped over the life cycle,
and the figure shows that about half of the hump is gone after family size is
controlled for, leaving a smaller (but still significantly positive) hump to be
explained by deviations from the most basic life cycle model.
If one follows an average household over its life cycle4 , two main stylized
facts emerge (see Attanasio (1999) for the detailed figures). First, disposable
income follows a hump over the life cycle, with a peak around the age of
45 (the age of the household is defined by the age of the household head).
This finding is hardly surprising, given that at young ages households tend to
3
This figure is taken from Fernandez-Villaverde and Krueger (2007). Similar findings
are presented by Attanasio et al. (1999) and Gourinchas and Parker (2002), among others.
4
How to construct such an average household in the absence of panel data sets which
lack a sufficient time series dimension is a challenging problem. The solution, the so-called
pseudo panel method, will be discussed in the second half of the course.
18 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES

obtain formal education or training on the job and labor force participation
of women is low because of child bearing and rearing. As more and more
agents finish their education and learn on the job as well as promotions occur,
average wages within the cohort increase. Average personal income at age
45 is almost 2.5 times as high as average personal income at age 25. After
the age of 45 personal income first slowly, then more rapidly declines as more
and more people retire and labor productivity (and thus often wages) fall.
The average household at age 65 has only 60% of the personal income that
the average household at age 45 obtains.
The second main finding is the surprising finding. Not only personal
income, but also consumption follows a hump over the life cycle. In other
words, consumption seems to track income over the life cycle fairly closely.
This is one statement of the so-called excess sensitivity puzzle: consump-
tion appears to be excessively sensitive to predicted changes in income. In
fact, the two standard theories of intertemporal consumption allocation we
will consider in the next section both predict that (under specific assumptions
spelled out explicitly below) consumption follows a martingale and current
income does not help to forecast future consumption. The hump-shaped con-
sumption age profile apparently seems to contradict this hypothesis. Later
in the course we will investigate in detail whether, once we control for house-
hold size (which also happens to follow a hump shape), the hump-shape in
consumption disappears or whether the puzzle persists. Figure ?? (taken
from Fernandez-Villaverde and Krueger, 2007) documents the life cycle pro-
file of consumption, with and without adjustment for family size by household
equivalence scales. The figure is derived using a synthetic cohort analysis, a
technique from Panel data econometrics that allows us to construct average
life cycle profiles for households that we do not observe over their entire life
(we will talk about this technique in detail below). The key observation from
this figure is that consumption displays a hump over the life cycle, and that
this hump persists, even after controlling for family size. The later observa-
tion is not entirely noncontroversial, and we will discuss below the different
positions on this issue.

3.2.4 Dispersion over the Life Cycle

Deaton and Paxson (1994) and Storesletten et al. (2004a) document how
inequality of income and consumption evolves over the life cycle. Figure 1 of
Storesletten et al. (2004a) plots, against age of the households, the variance
3.2. MAIN FINDINGS 19

of log earnings and log consumption. The figure shows that as a cohort ages,
the distribution of earnings and consumption within a cohort fans out.5 Note
that the increase in consumption inequality is substantially less pronounced
than the increase in the dispersion of earnings.
Cross-sectional dispersion (or inequality) has also changed dramatically
over time. Heathcote et al. (2004) and many others have documented a
strong upward trend in wage and even more pronounced in household earn-
ings inequality over the last 30 years. The inequality in hours worked has
remained fairly stable, and consumption inequality has increased by sub-
stantially less than earnings inequality (see Krueger and Perri (2006) and
the discussion in Attanasio et al. (2007)). Finally, it appears that wealth
inequality has increased as well in the 1980’s and since the remained fairly
constant, to the extent the wealth data are available and can be trusted (see
Favilukis, 2007). A substantial body of literature has sprung up in the last
few years trying to explain these trends with structural macro models with
heterogenous agents, of the type studied in these notes.
Heathcote, Perri and Violante (2009)

3.2.5 Dispersion (Inequality) at a Point in Time

Minneapolis FED QR pieces by Victor and Vincenzo. Show Lorenz curves
and tables with inequality statist

3.2.6 Changes in Dispersion over Time

Wages, earnings and income: Heathcote, Perri and Violante (2011), con-
sumption: Krueger and Perri (2006, wealth more speculative: Favilukis
(2007)

3.2.7 Other Interesting Facts (or Puzzles)

Other intersting puzzles in this literature include:
5
When working with household data that does not contain a long panel dimension
nd thus one cannot literally follow the same cohort and uses repeated cross-sections of
households, oe runs into the well-known identfication problem that time, cohort and age
effects are not separately identified (since age, time and cohort birth year are prefectly
colinear). There are varios ways to deal with this problem which I will discuss in class.
20 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES

Excess Smoothness Puzzle

If the stochastic income process of households is only difference stationary
(say, it follows a random walk) then a shock to current income translates
(more than) one to one into a shock to permanent income and hence should
induce a large shock to consumption. With difference-stationary income
processes consumption should be as volatile as current income, but we saw
that, at least on the aggregate level, consumption is smoother than income.
Is in fact consumption “too smooth”. This is the excess smoothness puzzle.

Lack of Decumulation Puzzle

Household level wealth data show that a significant fraction of very old house-
holds still hold a large portfolio of financial and real estate assets. Why don’t
these households decumulate their assets and enhance their consumption, as
standard life-cycle theory predicts?

Drop of Consumption Puzzle

As people retire, their consumption drops by about 15% on average

Portfolio Allocation Puzzle

The median wealth US household does not own stocks, but holds its major
fraction of the wealth portfolio concentrated in its own home and the rest
in low-return checking or savings accounts. This is despite the fact that re-
turns to equity and returns to human capital (i.e. the households’ wage) are
roughly uncorrelated for this fraction of the population. In addition, Gross
and Souleles (2000) document that a significant fraction of the population
has simultaneously high-interest credit card debt and liquid, low return as-
sets such as a significantly positive checking account balance. Tobacman
(2007) documents that a substantial fraction of the U.S. population repeat-
edly borrows using payday loans, which carry an annualized interest rate of
about 7000%.

Debt and Default

The US legal system allows private households to file for personal bankruptcy
under Chapter 7 or Chapter 11 of the US Bankruptcy Code. In particular,
3.2. MAIN FINDINGS 21

under Chapter 7 households are discharged of all their debts, are not re-
quired to use any of their future labor income to repay the debt and can
even keep their assets (financial or real estate) below a state-dependent ex-
emption level. Whereas about 1% of all households per year file for personal
bankruptcy, White (1998) computes that currently at least 15% of all US
households would financially benefit from filing for bankruptcy. The fraction
of U.S. households filing for bankruptcy has increased sharply over the last
decade. So has the extent of uncollateralized debt, as a fraction of disposable
household income (the same is very much true for collateralized debt). Also,
charge-off rates of lenders (the fraction of loans the lender does not recover),
have increased substantially (see Livshits et al., 2007). Show pictures on
trends of debt and default (both uncollateralized and collateralized) from
Livshits et al.
Any given model will likely not be able to resolve all these puzzles at
once, and some models will abstract from some of the issues altogether, but
the “stylized facts” of this section should be kept in mind in order to guide
extensions of the models presented next.
22 CHAPTER 3. SOME STYLIZED FACTS AND SOME PUZZLES
Chapter 4

The Standard Complete

Markets Model

Let the economy be populated by N individuals, indexed by i ∈ I = {1, 2, . . . N }.

Each individual lives for T periods, where T = ∞ is allowed. In each period
there is one nonstorable consumption good. Each individual household has
a stochastic endowment process {yti } of this consumption good. Risk in this
economy is represented by a current aggregate shock (or “event”) st . We will
assume that st ∈ S, where S is a finite set of cardinality K, and denote by
st = (s0 , . . . st ) the event history of this economy, respectively. Each event
history st has a probability πt (st ) > 0 of occurring, and all households share
common and correct beliefs about these probabilities. Using slight abuse of
notation let by S t = S ×S ×. . .×S denote the t+1-fold Cartesian product of
S such that st ∈ S t . Also, in order to make some of the results a little easier
to derive, take the event s0 in period 0 as given (i.e. not random); obviously
which event the economy starts with is irrelevant for the discussion to follow.
Individual endowments are then functions of the underlying event histo-
ries1 , that is yti = yti (st ). Note that so far we did not make any assumptions
about the stochastic process governing {st }; in particular the process need
not be a Markov processes, and the individual income processes implied by
our formulation need not be iid across agents or time.
A consumption allocation (ci )i∈I = {(cit (st ))i∈I }Tt=0,st ∈S t maps aggregate
1
The assumption that S is finite is to avoid having to worry about measurability of the
consumption allocations below.
Note that we could as well have taken as fundamental sources of risk the individual
income processes {yti } and defined st as st = (yt1 , . . . , ytN ).

23
24 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

event histories st into consumption of agents i ∈ I at time t. Agents have

preferences over consumption allocations that are assumed to permit an ex-
pected utility representation
X
ui (ci ) = πt (st )Uti (ci , st ) (4.1)
st ∈S t

Furthermore, unless otherwise noted, we will assume that the utility function
is additively time-separable and that agents discount the future at common
subjective time discount factor β ∈ (0, 1), so that the period utility function
takes the form Uti (ci , st ) = β t U i (cit (st ), st ) and thus expected lifetime utility
is given by:
X T X
ui (ci ) = β t πt (st )U i (cit (st ), st ) (4.2)
t=0 st ∈S t

By ρ = β1 − 1 let denote the subjective time discount rate (i.e. β = 1+ρ1

).
t
The presence of s allows for the potential presence of preference shocks
to individual agents’ utility from consumption. For example, the arrival of
a new child (modeled as an exogenous stochastic process driven by {st })
may reduce utility from a given household consumption level, provided that
members care about per capita consumption (and cit (st ) denotes consumption
expenditures of household i). For future reference we now define a feasible
allocation and a Pareto-efficient allocation.

Definition 1 A consumption allocation {(cit (st ))i∈I }Tt=0,st ∈S t is feasible if

cit (st ) ≥ 0 for all i, t, st (4.3)

N
X XN
cit (st ) = yti (st ) for all t, st (4.4)
i=1 i=1

Definition 2 A consumption allocation is Pareto efficient if it is feasible and

there is no other feasible consumption allocation {(ĉit (st ))i∈I }Tt=0,st ∈S t such
that

ui (ĉi ) ≥ ui (ci ) for all i ∈ I (4.5)

ui (ĉi ) > ui (ci ) for some i ∈ I (4.6)
4.1. THEORETICAL RESULTS 25

4.1 Theoretical Results

The main distinction between the complete markets model and the standard
incomplete markets model is the set of assets that agents can trade to hedge
against individual income risk, and hence the individual budget constraints.
In the standard complete markets model we suppose that there is a complete
set of contingent consumption claims that are traded at time 0, prior to
the realization of income (or any other type of) risk. The individual Arrow
Debreu budget constraints take the form
T X
X T X
X
t
pt (s )cit (st ) ≤ pt (st )yti (st ) (4.7)
t=0 st ∈S t t=0 st ∈S t

where pt (st ) is the period 0 price of one unit of period t consumption, deliv-
ered if event history st has realized.

4.1.1 Arrow-Debreu Equilibrium

Definition 3 An Arrow Debreu competitive equilibrium consists of alloca-
tions {(cit (st ))i∈I }Tt=0,st ∈S t and prices {pt (st )}Tt=0,st ∈S t such that

1. Given {pt (st )}Tt=0,st ∈S t , for each i ∈ I, {cit (st )}Tt=0,st ∈S t maximizes (4.2)
subject to (4.3) and (4.7)

2. {(cit (st ))i∈I }Tt=0,st ∈S t satisfies (4.4) for all t, st .

Now let us make the following assumption

Assumption 4 The period utility functions U i are twice continuously differ-

entiable, strictly increasing, strictly concave in its first argument and satisfy
the Inada conditions

lim Uci (c, st ) = ∞ (4.8)

c→0
lim Uci (c, st ) = 0 (4.9)
c→∞

where Uci is the derivative of U with respect to its first argument.

26 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

It is then straightforward to argue that in this economy any competi-

tive equilibrium allocation is the solution to the social planners problem of
maximizing
XN
max αi ui (ci ) (4.10)
{(cit (st ))i∈I }T
t=0,st ∈S t i=1

subject
PNto (4.3) and (4.4), for some Pareto weights (αi )N i
i=1 satisfying α ≥ 0
and i=1 αi = 1. First, observe that the assumption is (more than) sufficient
to establish the first welfare theorem, and thus we know that any compet-
itive equilibrium allocation is Pareto efficient. Second, any Pareto efficient
allocation is the solution to the social planner problem in (4.10), for some
Pareto weights (see, e.g. MasColell et. al., chapter 16; this result requires
other parts of assumption 1, especially concavity).

Characterizing Efficient Allocations: Perfect Consumption Insur-

ance
Characterizing the solutions to the maximization problem in (4.10), and thus
characterizing the properties of all efficient (and thus competitive equilib-
rium) allocations is straightforward. Attaching Lagrange multipliers λ(st )
to the resource constraint and ignoring the non-negativity constraints on
consumption we obtain as first order necessary conditions for an optimum

αi β t πt (st )Uci (cit (st ), st ) = λt (st ) (4.11)

for all i ∈ I. Hence for i, j ∈ I

Uci (cit (st ), st ) αj

= (4.12)
Ucj (cjt (st ), st ) αi

for all dates t and all states st . Hence in any efficient allocation (and thus
in a market economy with a complete set of contingent consumption claims
being traded) the ratio of marginal utilities of consumption of any two agents
is constant across time and states. Also those agents, ceteris paribus (i.e. if
they had the same utility function and the same preference shocks), with
higher relative Pareto weights will consume more in every state of the world
because the utility function is assumed to be strictly concave.
Thus the benevolent social planner whose aim to insure consumption of
all households over time and across states of nature finds it optimal to keep
4.1. THEORETICAL RESULTS 27

the ratio of marginal utilities of consumption of any two agents constant

across time and states. Since the social planner faces no constraints other
than the resource constraint it is natural to define as perfect consumption
insurance of an allocation the property (4.12) of a constant ratio of marginal
utilities:
Definition 5 A consumption allocation {(cit (st ))i∈I }Tt=0,st ∈S t is said to satisfy
perfect consumption insurance if the ratio of marginal utilities of consumption
between any two agents is constant across time and states of the world.
Thus, the complete markets model exhibits perfect consumption insur-
ance, in the sense defined above. With additional assumptions we obtain
an even stronger (and hence easier to test empirically) implication of the
complete markets model.
Assumption 6 All agents have identical CRRA utility, and either prefer-
ence shocks are absent or utility is separable consumption and preference
shocks, i.e.  c1−σ −1
i t 1−σ
+ v i (st ) if σ 6= 1
U (c, s ) = (4.13)
log(c) + v i (st ) if σ = 1
Here σ ≥ 0 is the coefficient of relative risk aversion.
We discuss the properties of the CRRA utility function further in Ap-
pendix 4.6
With this assumption equation (4.12) becomes
 i  σ1
cit (st ) α
j t = (4.14)
ct (s ) αj
i.e. the ratio of consumption between any two agents is constant across time
and states. P This, in particular, implies that there exist weights (θi )i∈I with
θi ≥ 0 and i∈I θi = 1 such that
X
cit (st ) = θi yti (st ) ≡ θi yt (st ) = θi ct (st ) (4.15)
i∈I
t i t
is aggregate income in the economy and ct (st ) =
P
where yt (s ) = i∈I yt (s )
i t
P
i∈I ct (s ) is aggregate consumption. The weights are given in closed form
by
1
αiσ
θi = P 1 ≥ 0 (4.16)
N
j=1 jα σ
28 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

Note that with logarithmic preferences (σ = 1) we have that θi = αi , i.e. the

share of aggregate consumption an agent i receives from the social planner
corresponds to the Pareto weight the planner attaches to this agent.
That is, with separable CRRA utility any efficient consumption allocation
(and thus any complete markets competitive equilibrium allocation) has the
feature that individual consumption at each date, in each state of the world is
a constant fraction of aggregate income (or consumption). Note that it does
not imply that individual consumption is constant across time and states of
the world, because it still varies with aggregate income (a variation against
which the agents cannot be mutually insured by the social planner). It also
does not imply that consumption among agents is equalized. ¿From (4.14)
we see that the level of consumption of agent i will depend positively on the
Pareto weight of that agent.

Determining the “Right” Pareto Weights

So far we have characterized efficient allocations, given arbitrary Pareto
weights (αi )i∈I of the social planner, and we have argued that all competitive
equilibrium allocations share this characterization. Although not necessary
for the empirical tests of these implications we now briefly describe how to
construct competitive equilibrium consumption allocations, that is, how to
find the the “right” Pareto weights. This section describes Negishi’s (1960)
method for doing so.2 Start from an arbitrary efficient consumption alloca-
tion cit (st , α), indexed by Pareto weights α = (αi )i∈I . For given α there is no
reason to believe that each household can afford the corresponding allocation
1
αiσ
cit (st ) = θi ct (st ) = P 1 ct (st ) = cit (st , α).
N
j=1 αj
σ

This was irrelevant in the social planner problem, but is required for a com-
petitive equilibrium allocation. But in order to compute how much such
2
Negishi’s goal was to prove existence of equilibrium. He did so by first arguing that all
solutions to the social planner problem satisfy the elements of the definition of a competi-
tive equilibrium apart from household optimality (with prices equal to Lagrange multipliers
on the resource constraints of the social planner problem). He then showed that there exist
Pareto weights such that the associated solution to the social planner problem solves the
household maximization problem (again with prices given by the Lagrange multipliers on
the resource constraint). These Pareto weights are given by the solution to the system of
equations (4.18) below.
4.1. THEORETICAL RESULTS 29

an allocation costs we need the appropriate prices. It turns out that the
Lagrange multipliers (that is, the shadow prices) λt (st , α) on the resource
constraints from the social planner problem are appropriate. So for each
agent i define the transfer functions as
XX
ti (α) = λt (st ) cit (st , α) − yti (st ) .
 
(4.17)
t st ∈S t

Thus ti (α) is the value of lifetime consumption net of the value of lifetime
income, where income and consumption at each history st is valued at its
social value λt (st , α). For our economy, from equation (4.11) the Lagrange
multipliers are given by

λt (st , α) = αi β t πt (st )Uci (cit (st ), st )

= αi β t πt (st )cit (st , α)−σ

−σ
= αi β t πt (st ) θi ct (st )
" N #σ
X 1
= αjσ β t πt (st )ct (st )−σ
j=1

In a competitive equilibrium the value of each agent’s consumption allocation

has to not exceed (and with strictly increasing utility, has to equal) the value
of that agent’s income stream. The “right” α’s are then those that insure
that
ti (α) = 0 for all i. (4.18)
P i
Note that since we normalized P i α = 1 there are only N − 1 unknowns.
But it is easy to show that t (α) = 0, and thus the system (4.18) has only
have N − 1 independent equations as well.
Once one has solved for α∗ from these equations, the competitive equilib-
rium allocations are given by cit (st , α∗ ) and equilibrium prices are given by
λt (st , α∗ ), as we will confirm below. Note that solving for competitive equi-
libria using the Negishi approach then amounts to solving the social planner
problem for arbitrary weights α, and solving a system of N − 1 equations in
N − 1 unknowns, which may be substantially easier than solving for the com-
petitive equilibrium directly. But also note that in general equations (4.18)
can be quite messy, and in particular, can be highly nonlinear in the α’s, so
that typically no analytical solution to the system (4.18) is available.
30 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

Equation (4.17) also makes clear that the Pareto weights that make the
transfer functions ti (α) equal to zero depend on the properties of the indi-
vidual income processes {yti }. As a consequence, the equilibrium level and
thus share θi of consumption of each agent i depends crucially on the value
of her endowment process.

Competitive Equilibrium Prices and the Representative Agent

After this detour, let’s turn back to the characterization of efficient (and
hence equilibrium) allocations. The growth rate of consumption between
any two dates and states is given from (4.15) as

cit+1 (st+1 ) ct (st+1 )

   
log = log (4.19)
cit (st ) ct (st )

that is, if agents have CRRA utility that is separable in consumption, then
in an efficient (competitive equilibrium) allocation individual consumption
growth is perfectly correlated with and predicted by aggregate consumption
growth. In particular, individual income growth should not help to predict
individual consumption growth once aggregate consumption (income) growth
is accounted for. This result is the starting point of the most basic empiri-
cal tests of perfect consumption insurance, see e.g. Mace (1991), Cochrane
(1991), among others.
To obtain Arrow Debreu prices associated with equilibrium allocations
we obtain from the consumer problem of maximizing (4.2) subject to (4.7)
that
pt+1 (st+1 ) πt+1 (st+1 ) Uci (cit+1 (st+1 ), st+1 )
= β (4.20)
pt (st ) πt (st ) Uci (cit (st ), st )
Under assumption 2 this becomes
−σ
pt+1 (st+1 ) πt+1 (st+1 ) cit+1 (st+1 )

= β (4.21)
pt (st ) πt (st ) cit (st )
−σ
πt+1 (st+1 ) ct+1 (st+1 )

= β (4.22)
πt (st ) ct (st )

and hence equilibrium Arrow Debreu prices can be written as functions of

aggregate consumption only. We then have the following
4.1. THEORETICAL RESULTS 31

Proposition 7 Suppose allocations {(cit (st ))i∈I }Tt=0,st ∈S t and prices {pt (st )}Tt=0,st ∈S t
form an Arrow-Debreu equilibrium. Then under assumption 2 the allocation
{ct (st )}Tt=0,st ∈S t defined by
X
ct (st ) = cit (st ) (4.23)
i∈I

and prices {pt (st )}Tt=0,st ∈S t form an Arrow Debreu equilibrium for the repre-
sentative agent economy in which the representative agent has an endowment
process given by X
yt (st ) = yti (st ) (4.24)
i∈I
and CRRA preferences with parameter σ.
The fact that the equilibrium of the representative agent economy has
consumption allocations given by (4.23) is of course trivial and follows di-
rectly from the market clearing condition. The content of this proposition
lies in the statement that in the representative agent economy (with the rep-
resentative household having the same CRRA utility function as agents in
the heterogeneous agent economy) has the same equilibrium Arrow-Debreu
prices as the economy with I consumers that might differ in their income
processes in an arbitrary way. Thus, in order to derive Arrow-Debreu prices
(and hence all other asset prices) in an economy with complete markets, it
is sufficient to study the corresponding representative agent economy. Most
of the consumption based asset pricing literature since Lucas (1978) (and
indeed most of model-based macroeconomics) has indeed employed models
with a representative consumer. As the previous result suggests, if financial
markets are complete (and sufficiently strong assumptions on agents’ utility
functions are made), then the abstraction from household heterogeneity is
innocuous from the perspective of macroeconomic research. It should also
be noted that for this proposition assumption 2 can be weakened3 , although
3
We have assumed idential CRRA utility functions and time discount factors across all
agents. Koulovatianos (2007), building on the large literature on aggregation in dynamic
models, such as Chatterjee (1994) and Caselli and Ventura (2000), gives necessary and
sufficient conditions on individual utility functions to obtain a representative consumer.
See his Theorem 1 and 2.
Maintaining identical time discount factors the period utility function has to be either
of power or exponential form (with heterogeneity in the CARA, but not in the CRRA
possible). If the time discount factor is allowed to vary across agents, only exponential
utility gives rise to the result stated here.
32 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

the construction of the utility function of the representative agent may more
involved.

Endogenous Labor Supply

The previous discussion took as exogenous a stochastic income process and
thus abstracted from endogenous labor supply. Now we will argue that un-
der the appropriate assumption on the utility function the results from the
previous section go through essentially unchanged if agents make endogenous
labor supply decisions. However, we will also present an example that shows
that nonseparabilities between consumption and leisure in the utility func-
tion may alter the empirical predictions of the complete markets model in a
qualitatively important way.
Instead of stochastic income shocks let households now face stochastic
productivity (wage) shocks wti (st ). We assume that a household produces
wti (st ) units of the consumption good per unit of time worked. In a compet-
itive equilibrium individual productivity and real wages will coincide, and
thus wti (st ) will also denote the stochastic wage process that household i
faces in a competitive equilibrium.
As long as markets are complete, equilibrium allocations can still be de-
termined by solving a social planner problem with appropriate social welfare
weights. Now the resource constraints read as

N
X N
X
cit (st ) = wti (st )lti (st )
i=1 i=1

and the period utility function is given by U (cit (st ), lti (st )), where lti (st ) ∈ [0, 1]
is the fraction of the time a household works.
The key efficiency conditions now read as

Uc (cit (st )lti (st )) αj

= (4.25)
Uc (cjt (st ), ltj (st )) αi
Ul (cit (st )lti (st ))
− i t i t
= wti (st ) (4.26)
Uc (ct (s ), lt (s ))

The first condition is the familiar efficient risk sharing condition across house-
holds. The second condition governs the efficient allocation of consumption
4.1. THEORETICAL RESULTS 33

and leisure for an arbitrary household i. We can divide equations (4.26) for
any two agents to arrive at

Ul (cit (st )lti (st )) αj wti (st )

= (4.27)
Ul (cjt (st ), ltj (st )) αi wtj (st )

For given welfare weights α equations (4.25) and (4.27) determine relative
consumption and leisure allocations between households i and j.
Let us consider two important classes of period utility functions U (.) in
the following two examples.

Example 8 Suppose U (cit (st ), lti (st )) is additively separable between consump-
tion and leisure (e.g. U (cit (st ), lti (st )) = v(cit (st )) − g(lti (st )) where v is strictly
concave and g is strictly convex. In this case the implications for efficient
consumption risk sharing are exactly the same as in the case with exogenous
labor supply:
v 0 (cit (st )) αj
= .
v 0 (cjt (st )) αi
In particular, if v is of CRRA form, then as before consumption of each
agent i equals a constant fraction θi of aggregate production. The efficient
allocation of labor supply, in this case, is characterized by

g 0 (lti (st )) αj wti (st )

= .
g 0 (ltj (st )) αi wtj (st )

Thus, since g is strictly convex more productive households work harder. But
although labor supply does respond to idiosyncratic productivity shocks wti (st ),
as before the share of consumption of agent i does not.

Example 9 U (cit (st ), lti (st )) is nonseparable. Then in general also consump-
tion shares respond to idiosyncratic wage shocks (which in turn suggests that
one has to be cautious when testing perfect consumption insurance). We
demonstrate this through a second example. Suppose the period utility func-
tion is of familiar Cobb-Douglas form
1−σ
[cit (st )ν (1 − lti (st ))1−ν ] −1
U (cit (st ), lti (st )) = .
1−σ
34 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

Exploiting the efficiency conditions (4.25) and (4.27) we find that

 σ1  (1−ν)(1− σ1 )
cit (st ) αi wti (st )

=
cjt (st ) αj wtj (st )
 σ1  1−(1−ν)(1− σ1 )
1 − lti (st ) αi wj (st )

= .
1 − ltj (st ) αj wti (st )

Since σ > 0 and ν > 0 the second equation implies that more productive
households (holding Pareto weights constant) consume less leisure and work
more, that is, if wi goes up relative to wj in a particular node, then household
i’s labor supply increases, relative to that of household j. The first equation
implies that more productive households also consume more if and only if σ >
1. If this condition is satisfied, it is easy to verify from the utility function that
Uc,1−l < 0, that is, then the marginal utility of consumption falls with higher
leisure. Recall from equation (4.25) that it is efficient to keep the ratio of
marginal utilities of consumption constant across time and states. If leisure of
household i goes down relative to j and Uc,1−l < 0, then for fixed consumption
Uci /Ucj would rise. Thus to keep Uci /Ucj constant, consumption of household i
has to rise relative to household j in that node: it is efficient to compensate
household i for her higher labor supply with larger consumption. If σ < 1
marginal utility of consumption increases with leisure, and the reverse logic
is true. Finally, if σ → 1 then the utility function is additively separable in
consumption and leisure

U (cit (st ), lti (st )) = ν log cit (st ) + (1 − ν) log 1 − lti (st )

and case 1 applies. Regardless of the size of σ, with nonseparabilities between

consumption and leisure (that is, as long as σ 6= 1), in the efficient perfect
risk sharing allocation individual consumption of household i responds to id-
iosyncratic productivity shocks {wti (st )} for that household, a fact that has to
be accounted for in the empirical tests of the perfect consumption insurance
hypothesis.

4.1.2 Sequential Markets Equilibrium

In the Arrow Debreu market structure trade in state-contingent consumption
claims takes place only once, at the beginning of time, prior to the resolu-
tion of any risk. With this market structure it is most straightforward to
4.1. THEORETICAL RESULTS 35

establish the connection between equilibrium and efficient allocations, and

to characterize the former. However, this market structure might appear
somewhat unrealistic and does not capture the trade of real-world financial
assets (such as bonds and stocks) over time.4 Therefore I now present a sec-
ond market structure and associated equilibrium concept, which I will call
sequential markets equilibrium, in which consumption and a carefully chosen
set of financial assets are traded at each node st of the event tree. Fortu-
nately, as discussed below, the set of equilibrium allocations in both market
structures will coincide, and thus the perfect risk sharing characterization of
consumption derived above will also apply to sequential markets equilibrium
allocations.
We assume that the set of financial assets consists of one period contingent
bonds, that is, financial contracts bought in period t, that pay out one unit
of the consumption good in t + 1, but only if a particular state st+1 ∈ S is
realized tomorrow.5 Let qt (st , st+1 ) denote the price, at node st , of a contract
that pays out one unit of consumption in period t+1 if and only if tomorrow’s
event is st+1 . These contracts are often called Arrow securities, contingent
claims or one-period state contingent bonds. Denote by ait+1 (st , st+1 ) the
quantity of the Arrow security bought (or sold) at node st by agent i that
pays off in node st+1 ≡ (st , st+1 ).
The period t, event history st budget constraint of agent i is given by
X
cit (st ) + qt (st , st+1 )ait+1 (st , st+1 ) ≤ yti (st ) + ait (st ) (4.28)
st+1

Note that agents purchase Arrow securities {ait+1 (st , st+1 )}st+1∈S for all con-
tingencies st+1 ∈ S that can happen tomorrow, but that, once st+1 is realized,
only the ait+1 (st+1 ) corresponding to the particular realization of st+1 pays off
and thus determines her asset position at the beginning of the next period.
We assume that all agents start their life with an asset position of zero, that
is, ai0 (s0 ) = 0. We can now state the following:
4
Admittedly, such assets could be introduced into the Arrow Debreu market structure
and could be priced in a straightforward manner, once equilibrium prices {pt (st )} for
state-contingent consumption claims have been determined.
5
A full set of one-period Arrow securities is sufficient to make markets “sequen-
tially complete”, in the sense that any (nonnegative) state-contingent consumption al-
location {cit (st )} is attainable with an appropriate sequence of Arrow security holdings
{ait+1 (st , st+1 )} satisfying all sequential markets budget constraints.
36 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL
  
Definition 10 A SM equilibrium is allocations { ĉit (st ), âit+1 (st , st+1 st+1 ∈S }Tt=0,st ∈S t ,
i∈I
and prices for Arrow securities {q̂t (st , st+1 )}Tt=0,st ∈S t ,st+1 ∈S such that

1. For i ∈ I given {q̂t (st , st+1 )}Tt=0,st ∈S t ,st+1 ∈S , for all i, {ĉit (st ), âit+1 (st , st+1 st+1 ∈S }Tt=0,st ∈S t
maximizes (4.2) subject to (4.28) and the constraints cit (st ) ≥ 0 and
ait+1 (st , st+1 ) ≥ −Āi (st+1 , q̂, ci )

2. For all t, st ∈ S t
I
X I
X
ĉit (st ) = yti (st ) (4.29)
i=1 i=1
I
X
âit+1 (st , st+1 ) = 0 for all st+1 ∈ S (4.30)
i=1

Several comments about this equilibrium definition are in order. First,

note that since at each node st as many different state-contingent bonds
are traded as there are events in S, we require a market clearing condition
for each of these Arrow securities. Second, we need to restrict the negative
position of state contingent bonds (that is, the amount of state-contingent
debt) by some upper bound Āi (st+1 , q, ci ) which for now I allow to be a general
function of the node st+1 , the equilibrium prices and the chosen consumption
allocation. In the absence of these constraints an infinitely lived household
would run Ponzi schemes: take out initial debt and roll it over indefinitely. As
a consequence the household budget set would be unbounded, the household
maximization problem would have no solution and thus no sequential markets
equilibrium would exist.6
Let us now discuss the No Ponzi scheme condition ait+1 (st , st+1 ) ≥ −Āi (st+1 , q̂, ci )
in greater detail. The objective is to come up with a specification for
Āi (st+1 , q̂, ci ) such that Ponzi schemes are ruled out, but that otherwise does
not impose borrowing constraints that are binding in equilibrium.7 With
tighter, potentially binding borrowing constraints the equivalence between
the associated sequential markets equilibria and Arrow Debreu equilibria
6
If agents have a finite life span T I will require them to die without debt, that is,
impose aT +1 (sT , sT +1 ) ≥ 0 for all sT +1 . Thus in this case Āi (sT +1 ) ≡ 0.
7
This discussion follows Wright (1987). For those interested in a generalization of the
result in the presence of long-lived assets, see Huang and Werner (2004).
4.1. THEORETICAL RESULTS 37

would break down. Appendix 4.4 discusses the choice of the No Ponzi con-
dition in greater detail; here we simply present a specification that is easy to
motivate, commonly used and gives rise to the desired equivalence result.
For a given Arrow securities price process q = {qt (st , st+1 )} define date
zero prices of state contingent consumption claims {v0 (st )} implied by q as
follows:

v0 (s0 ) = 1
v0 (st+1 ) = v0 (st )qt (st , st+1 ) = q0 (s0 , s1 ) ∗ . . . ∗ qt (st , st+1 ) ∗ 1. (4.31)

Using these prices derived from q we now state the No Ponzi scheme condition
as
−v0 (sT )aT (sT −1 , sT ) ≤ W (sT ) (4.32)
where ∞ X
X
T
W (s ) = v0 (st )yti (st ) (4.33)
t=T st |sT

is the Arrow-Debreu future wealth of the agent. Thus

W (st+1 )
Āi (st+1 , q) = (4.34)
v0 (st+1 )

We will assume that Āi (st+1 , q) is finite for all t, st+1 , which is a joint re-
striction on the income process {yti } and the Arrow security prices q under
consideration. The no Ponzi condition, stated this way, is sometimes called
the natural debt limit. It is easier to check as it only involves the endow-
ment process, but also prices (either Arrow securities prices or Arrow Debreu
prices).
We now have the following

Proposition 11 If (c, a, q) is a sequential markets equilibrium, then (c, p) is

an Arrow-Debreu equilibrium where prices are given as

pt (st ) = v0 (st ) (4.35)

Reversely, if (c, p) is an Arrow-Debreu equilibrium, then there exists asset

holdings a such that (c, a, q) is a sequential markets equilibrium, with

pt+1 (st+1 )
qt (st , st+1 ) = (4.36)
pt (st )
38 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

Proof. See Appendix 4.5

This proposition shows that each Arrow-Debreu equilibrium can be im-
plemented as a sequential markets equilibrium with a full set of one-period
Arrow securities (and vice versa, each sequential markets equilibrium is im-
plementable as an Arrow Debreu equilibrium).

Recursive Formulation of Household Problem

In order to facilitate comparisons with the standard incomplete markets

model (and for the purpose of the numerical solutions of equilibria of com-
plete markets models without use of the social planner problem) it is useful
to present a recursive formulation of the household problem from a sequential
markets equilibrium.
Suppose the stochastic process driving the economy {st } is a Markov
chain, and let π(s0 |s) denote the transition probabilities of the chain. Then
the consumer problem, under the assumption that Arrow security prices take
the form8 qt (st , st+1 ) = q(st+1 |st ), can be written recursively as

( ! )
X X
v(a, s) = max U y(s) + a − a0 (s0 )q(s0 |s) +β π(s0 |s)v(a0 (s0 ), s0 )
{a0 (s0 )}s0 ∈S
s0 s0
(4.37)

8
Under assumption 2 and the assumption that {st } is Markov, this assumption is valid.
Using (4.36) and (4.22) we find

pt+1 (st+1 )
q(st , st+1 ) =
pt (st )
−σ
πt+1 (st+1 ) ct+1 (st+1 )

= β
πt (st ) ct (st )
 −σ
yt+1 (st+1 )
= βπ(st+1 |st )
yt (st )
= q(st+1 |st )
4.2. EMPIRICAL IMPLICATIONS FOR ASSET PRICING 39

The first order and envelope condition are

!
X
q(s0 |s)U 0 y + a − a0 (s0 )q(s0 |s) = βπ(s0 |s)va (a0 (s0 ), s0 )
s0
!
X
va (a, s) = U 0 y + a − a0 (s0 )q(s0 |s)
s0
(4.38)

Combining yields a recursive version of the Euler equation, to be solved for

the policy function a0 (a, s; s0 )
!
X
q(s0 |s)U 0 y(s) + a − a0 (a, s; s0 )q(s0 |s)
s0
!
X
= βπ(s0 |s)U 0 y(s0 ) + a0 (a, s; s0 ) − a0 (a0 (a, s; s0 ), s0 ; s00 )q(s00 |s0 )
s00

We will use these equations in our discussions of complete markets models

with enforcement frictions in later chapters of this monograph.

4.2 Empirical Implications for Asset Pricing

In this section we briefly discuss the asset pricing implications of the complete
markets model.9 The previous sections derived Arrow Debreu prices and
prices of Arrow securities as a function of aggregate consumption ct (y t ) or
endowment yt (st ) = ct (y t ). These were given as:
−σ
pt+1 (st+1 ) πt+1 (st+1 ) ct+1 (st+1 )

t
qt (s , st+1 ) = =β (4.39)
pt (st ) πt (st ) ct (st )
as long as households have CRRA utility function with common risk aver-
sion σ. That is, together with the normalization p0 (s0 ) = 1 (remember we
took the initial state s0 as given) the exogenous endowment process {yt (st )}
completely determines Arrow-Debreu prices and prices of Arrow securities.
Equipped with these we can price any other asset, an asset j being defined by
9
Parts of this section are based on Kocherlakota (1996) and Campbell (1999, 2003) who
provide an extensive review of the Consumption Based Asset Pricing Model (CCAPM).
40 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

the stream of dividends dj = {djt (st )} where djt (st ) is the amount of consump-
tion goods asset j delivers at node st of the event tree. The period zero (cum
dividend) price of an asset is then simply the value of all the consumption
goods it delivers, that is
∞ X
X
P0j (d) = pt (st )djt (st )
t=0 st

and the ex-dividend price of an asset at node st , in terms of the st consump-

tion good, that pays a remaining dividend stream {dτ (sτ )}τ ≥t+1 is given by
P∞ P τ j τ
j t τ =t+1 sτ |st pτ (s )dτ (s )
Pt (d; s ) = .
pt (st )
Finally, define the one-period gross realized real return of such an asset be-
tween st and st+1 as
j
j Pt+1 (d; st+1 ) + djt+1 (st+1 )
Rt+1 (st+1 ) =
Ptj (d; st )
Example 12 The simplest examples are one-period assets such as the Arrow
securities discussed in the previous section. Such an asset, purchased at st
pays a dividend of one unit of consumption in a particular state ŝt+1 . Thus
its price at time t is given by
pt+1 (ŝt+1 )
PtA (d; st ) =
pt (st )
which is noting else but qt (st , ŝt+1 ). The associated gross realized return be-
tween st and ŝt+1 = (st , ŝt+1 ) is given by

A 0+1 pt (st ) 1
Rt+1 (ŝt+1 ) = t+1 t
= t+1
= t
pt+1 (ŝ )/pt (s ) pt+1 (ŝ ) qt (s , ŝt+1 )
j
and Rt+1 (st+1 ) = 0 for all st+1 6= st+1 .

Example 13 Now consider a one-period risk free bond, that is, a promise
to pay one unit of consumption in every node st+1 tomorrow that can follow
a given node st . The price of such an asset is given by
t+1
P
B t st+1 pt+1 (s ) X
Pt (d; s ) = t
= qt (st , st+1 ).
pt (s ) t+1 s
4.2. EMPIRICAL IMPLICATIONS FOR ASSET PRICING 41

The associated realized return for all nodes st+1 is given by

B 1 1
Rt+1 (st+1 ) = =P B
= Rt+1 (st )
PtB (d; st ) s t+1 q t (s t, s
t+1 )

and thus is nonstochastic, conditional on st . This justifies calling this asset

a risk-free bond.

Example 14 A stock (sometimes called a Lucas tree) that pays as dividend

the aggregate endowment in each period has a price (per share, if the total
number of shares outstanding is one)
P∞ P τ τ
τ =t+1 sτ |st pτ (s )yτ (s )
PtS (d; st ) =
pt (st )

Example 15 An option to buy one share of the Lucas tree at time T (at all
nodes) for a price K has a price Ptcall (st ) at node st given by
X pT (sT )
Ptcall (st ) =
 S T

max P T (d; s ) − K, 0
T t
pt (st )
s |s

Such an option is called a call option. A put option is the option to sell the
same asset, and is given by
X pT (sT )
Ptput (st ) = S T

max K − P T (d; s ), 0 .
T t
pt (st )
s |s

The price K is called the strike price (and easily could depend on sT , too).

We will now relate returns of assets in our model to aggregate consump-

tion (equivalently, aggregate income) in our model. Since both asset returns
as well as aggregate consumption can be measured in the data, this relation
will equip us with important empirically testable implications of our model.

Definition 16 A stochastic discount factor is a stochastic process {mt+1 (st+1 )}

that satisfies
j
E mt+1 (st+1 )Rt+1 (st+1 )|st = 1 for all t, st


(4.40)

and for all assets j in the economy.

42 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

A stochastic discount factor is also often called a pricing kernel, since it

incorporates all the essential information of asset prices in the model. We now
want to derive the unique stochastic discount factor for the complete markets
model. Arguably the easiest way to do this is to employ the Euler equation
from the sequential markets household maximization problem. Since we have
already argued that for all pricing purposes we can restrict ourselves to the
representative household we ignore household subscripts i from now on, until
further notice. Generalizing the household budget constraint to allow for an
arbitrary set of assets (with a generic asset denoted by j), which includes,
but need not be restricted to, the Arrow securities, we have
X X
ct (st ) + Ptj (d; st )ajt+1 (st ) ≤ yt (st ) + ajt (st−1 ) djt (st ) + Ptj (d; st )
 
j j

Attach Lagrange multiplier λt (st ) to this constraint and λt+1 (st+1 ) to the
corresponding constraints (there are as many as there are possible shocks st+1
tomorrow). The first order conditions with respect to ct (st ), ct+1 (st+1 ) and
ajt+1 (st ) read as (remember we assume that asset choices are not constrained
beyond the non-binding No Ponzi conditions)

β t πt (st )ct (st )−σ = λt (st )

β t+1 πt+1 (st+1 )ct+1 (st+1 )−σ = λt+1 (st+1 )
X
λt (st )Ptj (d; st ) = λt+1 (st+1 ) djt+1 (st+1 ) + Pt+1
j
(d; st+1 )
 
st+1

Substituting the first two equations into the third and re-arranging yields

X β t+1 πt+1 (st+1 )ct+1 (st+1 )−σ djt+1 (st+1 ) + Pt+1

j
 
(d; st+1 )
1 =
st+1
β t πt (st )ct (st )−σ Ptj (d; st )
−σ  j j

ct+1 (st+1 ) dt+1 (st+1 ) + Pt+1 (d; st+1 )
X 
t+1 t
= βπt+1 (s |s )
st+1
ct (st ) Ptj (d; st )
−σ
ct+1 (st+1 )
X 
t+1 t j
= βπt+1 (s |s ) t)
Rt+1 (st+1 )
st+1
c t (s
−σ !
ct+1 (st+1 )

j
= E β Rt+1 (st+1 )|st (4.41)
ct (st )
4.2. EMPIRICAL IMPLICATIONS FOR ASSET PRICING 43

for all assets j. Thus the unique stochastic discount factor process {mt+1 (st+1 )}
is given by −σ
ct+1 (st+1 )

t+1
mt+1 (s ) = β .
ct (st )
Note that mt+1 (st+1 ) is simply a function of aggregate consumption growth
and the two preference parameters (σ, β). Thus the equation
−σ !
ct+1 (st+1 )

j t+1 t
E β Rt+1 (s )|s = 1
ct (st )

j
is the basic empirical restriction on asset returns {Rt+1 } and consumption
growth {ct+1 /ct } implied by the complete markets model with CRRA prefer-
ences. This equation forms the basis of a large literature that investigates the
asset pricing implications of the complete markets model (or the consumption
capital asset pricing model, CCAPM, as it is sometimes called).
Now let us use and interpret condition (4.40) in the definition of a stochas-
tic discount factor further. Note that (4.40) has to hold for all assets j traded
in this economy, and thus for the risk free bond, in particular. Thus, for the
risk-free bond we have10

E mt+1 (st+1 )Rt+1

B
(st )|st = 1

B 1
Rt+1 (st ) =
E (mt+1 (st+1 )|st )
1
PtB (d; st ) = t+1 t


B
= E m t+1 (s )|s
Rt+1 (st )

that is, the price of a risk-free bond equals the conditional expectation of
the stochastic discount factor, which perhaps justifies the alternative name
“pricing kernel” more directly.
For ease of notation for what follows, let Et = E(.|st ) denote the condi-
tional expectation. The stochastic discount factor then satisfies, for all assets
j,
j
Et (mt+1 Rt+1 )=1 (4.42)
10 B
Using the fact that Rt+1 (st ) is nonstochastic conditional on st , and employing the
relation between asset prices and returns as the fact that we are considering a one-period
bond that only pays off in period t + 1.
44 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

We can rewrite (4.42) as

j j
1 = Et (mt+1 )Et (Rt+1 ) + Covt (mt+1 , Rt+1 )
and thus
j
j 1 − Covt (mt+1 , Rt+1 )
Et (Rt+1 ) = .
Et (mt+1 )
B
Using the fact that Et (mt+1 ) = 1/Rt+1 we obtain
j
Et (Rt+1 ) j
B
= 1 − Cov(mt+1 , Rt+1 ) (4.43)
Rt+1
j j
Defining net returns as rt+1 = Rt+1 − 1 and using the approximation

j j B j B
Et (Rt+1 ) 1 + Et (rt+1 ) 1 + rt+1 + Et (rt+1 ) − rt+1
B
= B
= B
Rt+1 1 + rt+1 1 + rt+1
j B
Et (rt+1 ) − rt+1 j B
= 1+ B
≈ 1 + Et (rt+1 ) − rt+1
1 + rt+1
in (4.43) we obtain
j B j j j
Et (rt+1 ) − rt+1 ≈ −Covt (mt+1 , Rt+1 ) = −ρt (mt+1 , Rt+1 )stdt (mt+1 )stdt (Rt+1 )
(4.44)
for any asset j, the risk-free bond B, and any stochastic discount factor m.
Here ρ is the correlation coefficient and std(.) denotes the standard deviation.
That is, the (conditional) expected excess return of asset j over the risk-free
B
rate rt+1 equals the (negative of) the covariance of the (gross) return of that
asset j with the stochastic discount factor, as long as the real return on the
B
risk free bond rt+1 between period t and t + 1 is close to zero, and thus the
approximation used above is accurate.
Although (4.44) holds true for any asset j and any utility function and
implied stochastic discount factor, the main implications of this equations
j B
have been studied for asset j being stocks, and thus Et (rt+1 ) − rt+1 being
the excess return on equity, and with the utility function of time-separable
CRRA form. In this case equation (4.44) becomes:
"  −σ # −σ !
t+1

S B c t+1 S c t+1 (s ) S


Et (rt+1 )−rt+1 = −ρt β , Rt+1 ∗stdt β ∗std t R t+1 .
ct ct (st )
(4.45)
4.2. EMPIRICAL IMPLICATIONS FOR ASSET PRICING 45

One quick way to assess whether the representative agent model has a
chance to rationalize the observed excess returns on equity is the following.
From the data the annual average premium to be explained is roughly 7%.
The annual standard deviation of stock returns is roughly 15.5% in n theodata.
We also observe the time series of aggregate consumption growth ct+1 ct
and
 stock
returns Rt+1 in the data. One simple way to ask whether (or, to what
extent) the complete markets model can rationalize the equity premium is
simply to calculate the right hand side for given (σ, β), replacing theoretical
with sample moments. The results of this exercise, which are carried out by
Kocherlakota (1996) and many others since are disastrous for the model, so
let us explain why by interpreting (4.45).
 −σ 
ct+1 S
First, the is no hope in explaining any premium if ρt ct
, Rt+1 >
0, that is, if (loosely speaking) consumption growth and stock returns are
negatively correlated (since ct+1 /ct is raised to a negative power). Any asset
that pays well in t + 1 (i.e. has high returns between period t and t + 1) in
bad states of the world (low consumption growth ct+1 /ct ) is a good hedge
against consumption risk, and thus does not command any risk premium
(in fact a negative premium). Thus it is the positive correlation between
consumption growth and asset returns11 that generate the equity premium,
rather than the fact that stock returns are very risky. Of course, conditional
S
on a given positive correlation between ct+1 /ct and Rt+1 , the size of the
explained equity premium is increasing in the risk of stock returns and the
volatility of consumption growth. In addition, high values of σ will, for a
given dispersion of consumption growth, help to explain a larger share of the
equity premium, by raising ct+11 /ct to larger powers. Given the empirically
S
weak, but positive correlation between ct+1 /ct and Rt+1 and the low volatility
of ct+1 /ct , the asset pricing literature was led to conclude that the empirically
observed excess returns on stocks can only be rationalized with implausibly
large values of σ. This is a version of Mehra and Prescott’s (1985) statement
of the equity premium puzzle.12

 −σ
11 ct+1
Again, the precise statement is: the negative correlation between β ct and
S
Rt+1 .
12
Mehra and Prescott (1985) considered values for σ above 10 as implausible and showed
that for values of σ below that threshold the model-implied equity premium is smaller than
the target in the data by factor of at least 10.
46 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL
  −σ 
ct+1 S
An alternative way to state the puzzle is to realize that −ρt β ct , Rt+1 ≤
1, since correlations are bounded between −1 and 1. Then equation (4.45)
implies that

"  −σ # −σ !

S B t+1
) − rt+1

Et (rt+1 ct+1 S c t+1 (s )
S

= −ρt β , Rt+1 ∗ stdt β
stdt Rt+1 ct ct (st )
S B
Et (rt+1 ) − rt+1
S

≤ stdt (mt+1 )
stdt rt+1

The entity on the left hand side is the so-called Sharpe ratio, which gives
the excess return a given asset (here stocks) commands per unit of risk (as
measured by the standard deviation of the return). The equation gives an
upper bound for that ratio: it cannot be larger than the volatility of the
stochastic discount factor. This bound, derived in various forms by Shiller
(1982) and Hansen and Jagannathan (1991), is valid for any asset (and any
stochastic discount factor). It indicates that a model that does not generate
a volatile enough stochastic discount factor has no chance in explaining the
equity premium puzzle. For annual data, the Sharpe ratio for stocks is in
the order of about 0.3 − 0.5, whereas the standard deviation of consumption
growth is not larger than 0.035 (the exact value depends somewhat on the
sample period). Thus without large values for σ (and perhaps for all possible
σ) the Hansen-Jagannathan bound is clearly violated by the unique stochastic
discount factor implied by the complete markets model: consumption growth
is simply too smooth in the data to generate a large enough equity premium,
even if it was perfectly correlated with stock returns (which it is not, either).
But was is the problem with a large σ, absent direct empirical evidence13
on its value? First, one may argue that households endowed with such large
risk aversion would make other choices (e.g. buy large amounts of all kinds of
insurance) that are at odds with the data. what is the problem with assuming
a large value for σ? But perhaps more importantly, such high assumed values
would raise a second asset pricing puzzle. To see why, note that the risk-free

13
Work by Barsky et al. (1997) and Kimball, Sahm and Shapiro (2008, 2009) has used
survey responses from the HRS and PSID to shed some light on the empirical distribution
of risk aversion in the U.S. population.
4.2. EMPIRICAL IMPLICATIONS FOR ASSET PRICING 47

rate satisfies
B 1 1 1
Rt+1 = =   −σ  =  
Et (mt+1 ) ct+1 1
Et β ct βEt  ct+1 σ
ct

Since in the data consumption is growing over time, thus ct+1 /ct tends to
be positive and on average around 1.02 in U.S. data. Therefore making σ
B
large makes Et (mt+1 ) small, and thus Rt+1 large. Why is this the case?
1
A large σ implies small IES = σ . With a small IES households desire a
smooth consumption profile. But consumption grows in the data. Thus a
large interest rate is needed to persuade them to postpone consumption. In
the data, however, the real risk free rate is small, around 1% on average.
This is Weil’s (1989) risk free rate puzzle.
One possible resolution to this puzzle is simply to make β large (larger
than 1!).14 As long as β(1 + gc )1−σ < 1 (on average), there is no problem for
lifetime utility to converge. But β > 1 and the associated degree of patience
might be considered implausible. Alternatively one might contemplate util-
ity functions in which a households’ attitude towards risk (variation of con-
sumption across states of the world) and towards intertemporal smoothing
(variation of consumption over time) is not governed by the same parameter.
Epstein and Zin (1989. 1991) propose such a class of utility functions:

Remark 17 Epstein and Zin (1989, 1991), extending earlier work by Kreps
and Porteus (1978) propose a class of utility functions that deviate from
expected utility and that can be written as
1
 1  1
 " # 1− γ
1−σ  1− γ
 1 X
u(c, st ) = ct (st )1− γ + β πt (st+1 |st )u(c, st+1 )1−σ
 
 st+1 

where u(c, st ) is lifetime (continuation) utility from a consumption allocation

c from node st onwards, and u(c, st+1 ) has the same interpretation at node
14
Suppose that consumption grows at a constant rate of 2%, then to rationalize RB =
1.01 in the presence of large risk aversion (say σ = 10, for an IES of 0.1) we would need
1
β= RB
= 1.21.
ct+1 σ
( ct )
48 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

st+1 . With these preferences the IES is controlled by the parameter γ, and
risk aversion is controlled by the separate parameter σ. These preferences
have become extremely popular in modern consumption-based asset pricing
because they help to jointly explain low risk-free rates as well as high excess
returns on stocks.

4.2.1 An Example
In this section we present an example for which we can solve for the risk free
rate and the equity premium in closed form.15 This is useful since in the ex-
ample it becomes fully transparent which parameters determine the risk-free
rate and the equity premium. The economy is populated by representative
consumers and the income of these households comes from dividends. House-
holds trade risk-free bonds in zero net supply and shares of the stock which
entitles them to the dividends of the stock.16 Also suppose that dividends
follow the process
log(dt+1 ) = log(dt ) + g + ut+1
where ut+1 is iid normally distributed with zero mean and variance σu2 . Since
dividends are the only form of income to the representative household we have
ct = dt for all t and all states of the world. Thus consumption satisfies ct+1
ct
=
g+ut+1 j
e . We can now use this in our asset pricing equation Et (mt+1 Rt+1 ) = 1.
For the risk-free rate we obtain
"  −σ #
B c t+1
Rt+1 Et β = 1
ct
B
Et βe−σ(g+ut+1 ) = 1
 
Rt+1
B
βe−σg Et e−σut+1 = 1
 
Rt+1

Now we note that because ut+1 is normal with zero mean and variance σu2 we
1 2 2
have Et [e−σut+1 ] = e 2 σ σu and thus the risk-free rate is given by
1 σ[g− 12 σσu2 ]
RB = e
β
15
This is a simplified version of the example presented in Barro (2009) which excludes
disaster risk.
16
Whether or not households also trade a full set of Arrow securities is not crucial in
this representative agent economy since there is no trade in equilibrium.
4.2. EMPIRICAL IMPLICATIONS FOR ASSET PRICING 49

or
1
rB = log(RB ) = ρ + σg − σ 2 σu2 (4.46)
2
1
where we used the definition of the time discount rate ρ as β = 1+ρ and
the approximation ρ ≈ log(1 + ρ). We see that the more patient households
are (the larger is ρ), the higher is the risk free rate. A higher consumption
(dividend) growth rate g and a lower intertemporal elasticity of substitution
raise the risk-free rate as well, for the reason discussed above. Finally, the
last term shows that the risk-free rate falls with the amount of consumption
risk. As we will discuss below, with CRRA preferences households have a
precautionary savings motive (whose size is governed by the parameter σ
as well) that is more potent the larger is consumption risk. To insure that
households are willing to consume their dividends in the presence of the extra
incentive to save, the risk free rate has to fall. This rationalizes the negative
last term in equation (4.46).
Solving for the expected return on equity is harder since we don’t know
the endogenous price of stock PtS . But note that
S
  S
S dt+1 + Pt+1 dt+1 Pt+1
Rt+1 = S
= S
+1 .
Pt Pt+1 PtS
t+1 PS
Now conjecture that the price-dividend ratio dt+1 is a time-and state invariant
S
constant so that Pt+1 = κdt+1 . Then
  S  
S dt+1 Pt+1 1 dt+1 ct+1
Rt+1 = S
+1 S
= +1 =θ
Pt+1 Pt κ dt ct
where θ = κ1 + 1 is a yet to be determined constant. Again using the asset

pricing equation, this time for stocks

S
Et (mt+1 Rt+1 ) = 1
"  1−σ #
ct+1
Et βθ = 1
ct
βθEt e(1−σ)(g+ut+1 ) = 1
 

  1 2 2
Again using Et e(1−σ)ut+1 = e 2 (1−σ) σu we have
1 (1−σ)[−g− 12 (1−σ)σu2 ]
θ= e
β
50 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

Thus

S 1 (1−σ)[−g− 12 (1−σ)σu2 ] (g+ut+1 )

Rt+1 = e e
β
1 gσ− 1 (1−σ)2 σu2 ut+1
= e 2 e
β
S 1 gσ− 12 σu2 [(1−σ)2 −1]
Et Rt+1 = e = R̄S
β
1
r̄S = log(R̄S ) = ρ + gσ + σ(σ + 2)σu2
2

Finally, the equity premium is given by

 
S B 1 2 1 2 2
r̄ − r = ρ + gσ + σ(σ + 2)σu − ρ + σg − σ σu
2 2
1
= σ(σ + 2 − σ)σu2 = σσu2
2

and thus is simply an increasing function of risk aversion and the volatility of
consumption (dividends). For similar analyses using Epstein-Zin preferences,
see e.g. Weil (1989), Tallarini (2000) or Barro (2009).

4.3 Tests of Complete Consumption Insur-

ance
As we have seen in section 4.1 a efficient consumption allocation features
perfect consumption insurance. Since equilibrium allocations in the com-
plete markets model are efficient, they share this property as well. In this
section we discuss empirical tests of this empirical prediction of the com-
plete markets model. Important papers implementing such tests are Altug
and Miller (1990), Mace (1991), Cochrane (1991), Nelson (1994), Townsend
(1994), Attanasio and Davis (1996), Hayashi, Altonji and Kotlikoff (1996),
Schulhofer-Wohl (2011) and Mazzocco and Saini (2012). Here we restrict
attention on the paper by Mace, whose test using U.S. data follows perhaps
most directly from the theory, and to the most recent extension by Mazzocco
and Saini (2012) using Indian village data.
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 51

4.3.1 Derivation of Mace’s (1991) Empirical Specifica-

tions
The paper by Mace presents the most straightforward test of the complete
consumption insurance hypothesis. The main implication of perfect risk
sharing is that individual consumption should respond to aggregate income
(or consumption) shocks, but not to idiosyncratic income shocks, no matter
whether these idiosyncratic shocks are unanticipated or not anticipated (this
is important in discriminating the complete markets model from the standard
incomplete markets models to be discussed in latter chapters).
Recall that under assumptions 1 and 2, equation (4.19) implies that
∆ ln cit = ∆ ln ct (4.47)
that is, the growth rate of individual consumption ∆ ln cit ≡ ln cit (st ) −
ci −ci
ln cit−1 (st−1 ) ≈ tci t−1 should equal the growth rate of aggregate consumption
t−1
∆ ln ct , and be independent of individual income growth and other individual-
specific shocks. That is, in the regression
∆ ln cit = α1 ∆ ln ct + α2 ∆ ln yti + it (4.48)
under the null hypothesis of complete consumption insurance (and with
CRRA utility) α1 = 1 and α2 = 0. In this regression the error term it
captures (thus far unmodeled) individual preference shocks as well as mea-
surement error in income and or consumption growth.
We can generalize our assumption 2 and still arrive at an empirical spec-
ification similar to equation (4.48) that can be used to test for complete
consumption insurance. Suppose that household preferences are of the form
i t cit (st )1−σ − 1
U i (cit (st ), st ) = eb (s ) · (4.49)
1−σ
where bi (st ) capture preference shocks such as changes in family size (which
are taken to be exogenous). Taking first order conditions in the planners
problem of maximizing
XXX
αi β t πt (st )U i (cit (st ), st )
i t st
s.t.
X X
cit (st ) ≤ yti (st )
i i
52 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

and then taking logs of these conditions yields

λt (st )
 
1 1 1
ln(cit (st )) = bi (st ) + ln(αi ) − ln (4.50)
σ σ σ β t πt (st )

Here λt (st ) is the Lagrange multiplier on the resource constraint, as before.

By taking averages of (4.50) over all agents we obtain

λ(st )
 
1 X j t 1 X j t 1 X j 1
ln(ct (s )) = b (s ) + ln(α ) − ln (4.51)
N j σN j σN j σ β t πt (st )
 
1 λ(st )
and using this equation to substitute out ln σ β t πt (st )
in equation (4.50)
yields:
! !
1 1 X 1 1 X
ln(cit (st )) = bi (st ) − bj (st ) + ln(αi ) − ln(αj )
σ N j σ N j
1 X
+ ln(cjt (st ))
N j
1 i t
1
b (s ) − ba (st ) + ln(αi ) − ln(αa ) + ln(cat (st )) (4.52)


=
σ σ
where
1 X j t
ba (st ) ≡ b (s )
N j
1 X
ln(αa ) ≡ ln(αj )
N j
1 X
ln(cat (st )) ≡ ln(cjt (st )) (4.53)
N j

are population averages. Note that we have abused notation somewhat in the
last two expressions, which are sums of logs rather than logs of sums. Finally,
by taking first differences of (4.52) to get rid off the term σ1 (ln(αi ) − ln(αa ))
we arrive at an equation that can be brought to the data (and again sup-
pressing dependence on st ):
1
∆ ln(cit ) = ∆ ln(cat ) + ∆bit − ∆bat


(4.54)
σ
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 53

Thus, with this specification of preferences we can test the complete in-
surance hypothesis with the empirical specification

∆ ln(cit ) = α1 ∆ ln(cat ) + α2 ∆ ln(yti ) + it (4.55)

where the error it now captures individual and aggregate changes in pref-
erence shocks as well as potentially measurement error.17 The null hypoth-
esis of complete consumption insurance implies that α1 = 1 and α2 = 0.
Thus the regression equation used to test for complete consumption insur-
ance remains substantially unchanged under the more general utility specifi-
cation with preference shocks, keeping in mind that in the definition above,
ln(cat (st )) 6= ln(ct ). Note that the basic prediction of the theory is that indi-
vidual household i’s consumption should not respond to idiosyncratic shocks
once aggregate consumption is accounted for. Idiosyncratic income shocks
are just one, arguably important, but not the only source of idiosyncratic
risk. Including other measures of idiosyncratic shocks in regression (4.55)
and testing the complete consumption insurance by investigating whether
the corresponding regression coefficient(s) equals zero is equally valid, from
the perspective of the theory presented so far. See especially Cochrane (1991)
on this point.
Finally, we can derive a similar specification under the assumption that
preferences take a Constant Absolute Risk Aversion (CARA) form. Suppose
the period utility function is given by:

1
U i (cit (st ), st ) = − e−γ (ct (s )−b (s ))
i t i t
(4.56)
γ

where γ is the coefficient of absolute risk aversion, then, using the same
manipulations as before one obtains

∆cit = ∆cm i m


t + ∆bt − ∆bt (4.57)

17
Note that, defining
1
it = ∆bit − ∆bat


σ

then the cross-sectional expectation of it across households i equals zero, by definition of
bat .
54 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

where
1 X i
cm
t = c
N j t
1 X i
bm
t = b (4.58)
N j t

Thus, under exponential utility the empirical specification used to test com-
plete consumption insurance is

∆cit = α1 ∆cm i i
t + α2 ∆yt + t (4.59)

with the null hypothesis being that α1 = 1 and α2 = 0.

4.3.2 Results of the Tests

Mace uses CEX data from 1980 to 1983 to estimate equations (4.55) and
(4.59). We focus our discussion on the specifications in which individual in-
come growth is used as measure of idiosyncratic risk, rather than changes
in unemployment status. In order to implement the regressions household
level data on consumption and income are needed. For each household, since
a specification in first differences is employed (in order to remove the indi-
vidual fixed effects), we require at least two observations on both income
and consumption. Thus, although the U.S. Consumer Expenditure Survey
(CEX) has no extended panel dimension, the fact that it contains income
and consumption data for four quarters for each household makes this data
set an appropriate (if not ideal) source for Mace’s application.1819
The total number of households in the CEX sample that Mace employs
is 10, 695. Each household contributes one differenced observation; Mace
chooses to use the information from the first and the forth interview of each
household (remember that each household stays in the sample for only 4
quarters) to obtain differenced consumption and income observations. For
the growth rate specification households that report non-positive income or
18
For results using more recent CEX data that are broadly consistent with Mace’s
findings, see Krueger and Perri (2005).
19
Apart from the issue of measurement error in the income and consumption data, as
Gervais and Klein point out (and then address), the fact that income and consumption
are measured for different and only partially overlapping time intervals creates additional
problems when implementing the risk sharing regressions using CEX data.
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 55

consumption in any of the two quarters have to be excluded, so that the

sample size for this specification is smaller. Mace employs several measures
of consumption in her analysis; we will report results for total consumption
expenditures, for nondurables, for services and for food expenditures. The
income variable is disposable income, defined as before-tax income minus
income taxes, deductions for social security and pension plans as well as
occupational expenses (e.g. union dues).
In Table 1 we present the results from the estimation of equation (4.55); in
addition Mace includes a constant to the regression. The first three columns
present the OLS estimates (with standard errors in parentheses) for the con-
stant, the coefficient on aggregate consumption growth and on individual
income growth. The forth column gives the test statistic of an F-test20 of
the joint hypothesis that α1 = 1 and α2 = 0 and the last column shows the
R2 of the regression.

Table 1: Growth Rate Specification

Cons. Measure α̂0 α̂1 α̂2 F-test R2
−0.04 1.06 0.04
Total Consumption 14.12∗ 0.021
(0.01) (0.08) (0.007)
−0.02 0.97 0.04
Nondurables 22.69∗ 0.027
(0.01) (0.07) (0.006)
−0.02 0.93 0.04
Services 12.44∗ 0.011
(0.01) (0.10) (0.01)
−0.02 0.91 0.04
Food 18.67∗ 0.020
(0.01) (0.07) (0.006)

The results in Table 1 are, for the most part, inconsistent with the com-
plete insurance hypothesis. For all consumption groups the F-test is deci-
sively rejected at the 5% confidence level (in fact it would be rejected at the
1% confidence level). Loosely speaking, this is mostly due to the fact that in-
dividual income growth does help to explain individual consumption growth,
20
The test statistic for the F-test is distributed as an F (2, N − 3), where N is the
number of usable household observations in the sample. For the difference specification
N = 10, 695 whereas for the growth rate specification it varies by consumption category
because households with nonpositive consumption in either interview have to be excluded.
A ∗ after the test statistic indicates that the null hypothesis can be rejected at the 95%
confidence level.
56 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

even when aggregate consumption growth is accounted for, in contrast to

what is implied by the complete consumption insurance hypothesis. Note
that for all consumption measures, based on a simple t-test the null hypoth-
esis that α1 = 1 cannot be rejected at conventional confidence levels, but
the hypothesis that α2 = 0 can be soundly rejected for all consumption mea-
sures. Also note that the constant is on the edge of being significant for most
specifications and that the R2 is quite low for all specifications. Although
the complete consumption insurance hypothesis can be statistically rejected,
note that the response of consumption to changes in income is quantitatively
small. Thus one might be lead to conclude that the hypothesis provides a
useful approximation to the data. As we will discuss below, measurement
error in income might question this assessment. Furthermore we will argue
in chapter 5 that the standard incomplete markets model (which does not
predict perfect consumption insurance) would imply consumption responses
to income shocks of a rough magnitude reported here if most income shocks
are of transitory nature.
In Table 2 we report the results for the specification relying on CARA
utility, i.e. the estimates of equation (4.59), again including an intercept.

Table 2: First Difference Specification

Cons. Measure α̂0 α̂1 α̂2 F-test R2
−77.87 1.06 0.03
Total Consumption 1.27 0.008
(19.32) (0.11) (0.02)
−13.97 0.99 0.01
Nondurables 7.71∗ 0.023
(3.33) (0.06) (0.003)
−30.47 1.01 0.01
Services 1.14 0.009
(16.47) (0.10) (0.007)
−7.46 1.01 0.005
Food 2.52 0.020
(2.12) (0.08) (0.002)

For this specification the results are more supportive of perfect consump-
tion smoothing. In particular, the F-test is rejected only for nondurable
consumption expenditures, again primarily due to the fact that individual
consumption growth is sensitive to individual income growth. Note that here
the intercept is significant for almost all measures of consumption (in theory
it should be zero). It is these results, in addition to the findings that the es-
timated value of α2 is quantitatively small, that lead Mace to the conclusion
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 57

that perfect risk sharing is a good first approximation of the data, at least
for exponential utility.
However, as Nelson (1994) argues, if one excludes households classified as
“incomplete income reporters” by the CEX (households that either did not
respond to certain income questions or gave inconsistent answers) and uses
expenditures for the entire quarter as the relevant measure of consumption
(as opposed to Mace who used only the expenditure in the month preceding
the interview) one finds a strong rejection of complete consumption insurance
even for the CARA utility specification.

4.3.3 The Problem of Measurement Error

Mace (1991) discusses several econometric issues arising in her regression
analysis. Somewhat surprisingly the paper does not include an extended
discussion of measurement error, which is a serious issue in micro data sets
in general, and for the CEX in particular. Given the survey design of the
CEX measurement error in the income variable is most likely the biggest
concern; as we will show it may bias the test of perfect risk sharing in favor
of the null hypothesis of perfect consumption insurance.
To make things simple let us suppose we want to estimate the equation

∆cit = α2 ∆yti + it (4.60)

i.e. for the moment we ignore the presence of aggregate consumption growth.
Note that if we had a single cross-section of first differenced household obser-
vations, one would exactly run a regression of the from (4.60).21 For simplicity
also assume that in the cross-section E (∆yti ) = 0, that is, individual income
changes are measured in deviation from aggregate income changes.22 The
null hypothesis of perfect risk sharing implies α̂2 = 0. For ease of exposition
we abstract from a constant in the regression and assume α0 = 0.
Suppose we have N household observations for one period t, and suppose
that the income variable is measured with error

∆zti = ∆yti + ∆vti (4.61)

21
See Cochrane (1991) for a discussion why the pooling of cross-sectional and time series
data is not free of problems and for an implementation of a test of complete risk sharing
using purely cross-sectional data.
22
This assumption is not central for the argument but makes the algebraic expressions
slighly easier.
58 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

where ∆zti is the measured income change, ∆yti is the true income change
and ∆vti is an additive measurement error satisfying E (∆vti ) = 0, where E(.)
is the cross-sectional (across households) expectation, recalling that we hold
t fixed. Furthermore assume that V ar (∆vti ) = σv2 , that V ar (∆yti ) = σy2 and
that ∆vti , it and ∆yti are all mutually independent.23
The equation we estimate is then

∆cit = α2 ∆zti + it (4.62)

and the OLS estimate for α2 is

PN
N −1 ∆cjt ∆ztj
j=1
α̂2 = j 2
(4.63)
N −1 N
P

j=1 ∆z t

We now show that α̂2 is an inconsistent estimator of α2 , with persistent bias

towards zero. We have that, multiplying out and using Slutsky’s theorem

N −1 N j j
P
j=1 ∆ct ∆zt
plimN →∞ α̂2 = plimN →∞ j 2
N −1 N
P

j=1 ∆zt

N −1 N j j
P i i


j=1 (α2 ∆yt + t ) ∆yt + ∆vt
= plimN →∞ j 2
N −1 N j
P

j=1 ∆yt + ∆vt
i 2
plimN →∞ N −1 N
P
j=1 (∆yt )
= α2 j 2 j 2
plimN →∞ N −1 N
P

−1
PN

j=1 ∆y t + plim N →∞ N j=1 ∆vt
σy2 1
= α2 · 2 2
= α2 · 2
σy + σv 1 + σσv2
y

Thus, without measurement error (σv2 = 0) α̂2 estimates α2 consistently.

However, the higher the ratio between measurement noise and signal, σv2 /σy2
the larger is the “attenuation bias” and the more is the estimator α̂2 asymp-
totically biased towards zero. Therefore, in the presence of substantial mea-
surement error in income growth or income changes we may not reject the
null of perfect risk sharing even when it should be rejected had we observed
the data without measurement error. Furthermore we might conclude that
23
The assumption that it and ∆yti are independent is already needed to make the OLS
estimate for α2 consistent even in the absence of measurement error.
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 59

the deviations from complete consumption insurance are quantitatively small

in the presence of a small estimate α̂2 when in fact that low estimate is due to
measurement error in income. This discussion is meant to give an indication
for why even the results in Mace supporting perfect risk sharing should be
taken with caution; they are also meant to motivate similar analyses like the
one in Cochrane (1991) that focus on measures of idiosyncratic income risk
which may be less prone to measurement error and correlation with the error
term than changes in or growth rates of individual income.

4.3.4 Other Empirical Issues and Solutions

In principle all variables that capture idiosyncratic shocks to a households’
consumption possibility set could be used as regressor on the right hand side
of the consumption insurance regressions. As Cochrane (1991) points out, a
good regressor should have the property that it is uncorrelated with house-
hold preference shocks (and thus the error term) and should be measured
precisely and if it has measurement error, this measurement error should not
be correlated with measurement error in consumption. Household income is
problematic in these regards.
Cochrane therefore proposes other regressors that are less prone to mea-
surement error and can be more plausibly taken as exogenous with respect
to household preference shocks. The variables he considers are days of work
lost due to sickness, due to strike, days looking for a job etc. His empiri-
cal results, derived for food consumption from the PSID are mixed; perfect
consumption insurance is rejected for some variables, but not for others. An
alternative approach to the problem of household income growth being cor-
related with the error term is to use an IV estimator. For this strategy to be
successful one needs to find an instrument that is correlated with income but
not correlated with household preference shocks (and thus the error term in
the regressions above).

4.3.5 Self-Selection into Occupations with Differential

Aggregate Risk
If households are heterogeneous with respect to their risk aversion (hetero-
geneity that will show up in the error term) and self-select into occupation
that are differentially affected by aggregate risk, then individual income will
60 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

be correlated with the error term and the standard consumption insurance
regressions will yield inconsistent result. Future versions of these notes will
discuss the papers by Schulhofer-Wohl (2011) and Mazzocco and Saini (2012)
in greater detail.
Consider the following example based on Mazzocco and Saini (2012):
∞ X
2 X
X
max αi β t πt (st )U i (cit (st ))
{cit (st ),lti (st )}
i=1 t=0 st ∈S t

s.t.
N
X
cit (st ) = Ȳt (st )
i=1
cit (st ) ≥ 0

• Standard efficiency condition

α1 Uc1 (c1t (st )) = α2 Uc2 (c2t (st ))

which together with resource constraint determines efficient split of

Ȳt (st ).

• Let U 1 , U 2 be CRRA with σ1 < σ2 .

• Plotting αi Uci (cit (st )) against cit (st ), the curves for i = 1 and i = 2
intersect once (and only once). If σ1 = σ2 , they are parallel to each
other (and on top of each other if α1 = α2 ).

• Plotting the efficient cit (st ) against Ȳt (st ), the curves (expenditure func-
tions) intersect once (and only once) if σ1 < σ2 . Let intersection point
be denoted by Ȳ ∗ .

• Consumption of household 1 efficiently varies more with Ȳt (st ) than

consumption of household 2. Less risk averse household bears more of
the consumption risk.

• Test of whether expenditure functions of households intersect forms

basis of test of preference heterogeneity.
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 61

• Absent preference shocks, with identical CRRA preferences

∆ log cit+1 = ∆ log cat+1

where
2
1X
log cat+1 = log cit+1
2 i=1

• Now suppose Ȳt < Ȳ ∗ < Ȳt+1 , then with efficient risk sharing and
σ1 < σ2 we have

∆ log c2t+1 = ∆ log cat+1 < ∆ log c1t+1

which a researcher that presupposes σ1 = σ2 has to interpret as rejec-

tion of efficient risk sharing.
• This shows potential problem of risk sharing regressions with preference
heterogeneity.
• Signing the Bias Suppose households share risk efficiently
• Suppose households have CRRA utility with σi 6= σj .
• Suppose that
i
– if ∆Ȳt+1 ≥ 0, then ∆ log yt+1 is decreasing in σi (in good aggregate
times the least risk averse do particularly well)
i
– if ∆Ȳt+1 < 0, then ∆ log yt+1 is increasing in σi (in bad aggregate
times the most risk averse do particularly well)

• Then in the regression

∆ log cit+1 − ∆ log cat+1 = ξ∆ log yt+1

i
+ εit+1
ˆ > 0, despite the fact that households share risk efficiently.
we have E(ξ)
• Intuition
i
ˆ =ξ+ Cov(∆ log yt+1 , εit+1 )
E(ξ) i
V ar(∆ log yt+1 )
i
• Since households share risk efficiently, the true ξ = 0. But Cov(∆ log yt+1 , εit+1 ) >
0. Why?
62 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

• In bad times ∆Ȳt+1 < 0, ∆ log cit+1 is particularly low for low σ house-
holds (thus εit+1 < 0) and ∆ log yt+1
i
is low for these households. Re-
versely for high σ households

• In good times the reverse is true.

i
• Thus Cov(∆ log yt+1 , εit+1 ) > 0

• Key problem: unobserved σi is an omitted variable correlated with

i
∆ log yt+1 and with error term εit+1 . Results in biased regression result.

• IV may be hard because need an instrument that is uncorrelated with

σi and correlated with income. Better not be an endogenous choice.

• Note: Schulhofer-Wohl (2011) argues that in the data households with

low σ indeed self-select into jobs with income growth more correlated
with aggregate income.

• Tests of Efficient Risk Sharing with Heterogeneity in Risk

Preferences Social Planner Problem
∞ X
N X
X
max αi β t πt (st )U i (cit (st ), lti (st ))
{cit (st ),lti (st )}
i=1 t=0 st ∈S t

s.t.
N
X XN
cit (st ) = wti (st )lti (st ) + Yt (st )
i=1 i=1
cit (st ) ≥ 0, lti (st ) ∈ [0, 1]

where Yt (st ) is nonlabor endowment in economy. Note: relative to

Mazzocco and Saini we suppress preference shocks.

• Rewrite resource constraint as

N
X
cit (st ) + wti (st )(1 − lti (st ))


i=1
XN
= wti (st ) + Yt (st ) ≡ Ȳt (st )
i=1
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 63

• Define “expenditure” of household i as

ρit (st ) = cit (st ) + wti (st )(1 − lti (st ))

• Split social planner problem into three steps

• Step 3: Conditional on given ρit (st ), what is the efficient split between
consumption and leisure

V i (ρit (st ), wti (st )) = max U i (cit (st ), lti (st ))

cit (st ),lti (st )
s.t.
ρit (st ) = cit (st ) + wti (st )(1 − lti (st ))
cit (st ) ≥ 0, lti (st ) ∈ [0, 1]

• Familiar optimality condition

Uli (cit (st ), lti (st ))
= −wti (st )
Uci (cit (st ), lti (st ))

together with constraint determines indirect utility function V i (ρit (st ), wti (st )).

• Note: if labor supply is exogenous, this step is trivial and V i (ρit (st ), wti (st )) =
U i (cit (st ))

• Step 2: Optimal allocation of expenditures across two arbitrary house-

holds i, j. Note: this is the crucial step for the tests.

V i,j (ρi,j t i t j t
t (s ), wt (s ), wt (s ))
= maxj αi V i (ρit (st ), wti (st ))
ρit (st ),ρt (st )

+αj V j (ρjt (st ), wtj (st ))

s.t.
ρt (s ) + ρt (s ) = ρi,j
i t j t t
t (s )

• Optimal solution ρi (ρi,j t i t j t

t (s ), wt (s ), wt (s )),
ρj (ρi,j t i t j t
t (s ), wt (s ), wt (s ))

• Note: if i = 1, j = 2 and labor is exogenous, we have executed step 2

before.
64 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

• Step 1: Optimal split of overall resources Ȳt (st ) between groups of

households (if N = 2, this point is mute).

• Assume that number of households is even (if not, need to group three
households into one group).

N/2
X
max V 2i−1,2i (ρt2i−1,2i (st ), wt2i−1 (st ), wt2i (st ))
{ρt2i−1,2i (st )} i=1

N/2
X
s.t. ρ2i−1,2i
t (st ) = Ȳt (st ).
i=1

• Theorem: Take an arbitrary partition of households (say the one pro-

posed in step 1). A consumption-leisure allocation that solves step 1 -
step 3 solves the social planner problem.

• Test of Risk Tolerance Heterogeneity: Suppose two households

i, j share risk efficiently (i.e. solve subproblem 2; subproblem 3 deals
with efficient allocation of consumption and leisure within a household,
subproblem 1 with efficient allocation of resources across groups).

• Suppose there are two histories st , ŝτ such that

wi (st ) = wi (ŝτ )
wj (st ) = wj (ŝτ )

• Suppose that expenditure functions cross:

ρi (ρi,j t i t j t j i,j t i t j t
t (s ), wt (s ), wt (s )) > ρ (ρt (s ), wt (s ), wt (s ))
ρi (ρi,j t i t j t j i,j t i t j t
t (ŝ ), wt (s ), wt (s )) < ρ (ρt (ŝ ), wt (s ), wt (s ))

• Then U i 6= U j . If U i , U j are CRRA, then σ i 6= σ j

• Remarks: One maintained assumption: efficient risk sharing. Thus

tests only useful in context of risk sharing tests (or if efficient risk
sharing can be ascertained independently). Also: what if later tests of
efficient risk sharing reject that hypothesis?
4.3. TESTS OF COMPLETE CONSUMPTION INSURANCE 65

• Second maintained assumption: wages held constant across st , ŝτ . Need

to find a way to do it in the data.

• Consequence of result: if test not rejected, can proceed as before. If

homogeneity of risk preferences rejected, need to device tests that are
robust to preference heterogeneity.

• Tests of Efficient Risk Sharing: If the U i are strictly concave and

if i, j share risk efficiently, then ρi , ρj are strictly increasing in ρi,j .

• Besides wi , wj and ρi,j , no idiosyncratic variable should enter ρi , ρj .

Note: if period utility affected by observable heterogeneity (e.g. family
size and composition) or unobservable heterogeneity, this can enter
ρi , ρj as well. But not e.g. nonlabor income y i .

• Note: if consumption and leisure are separable or wages are constant

over time and across states, the above two conditions are also sufficient
for efficient risk sharing.

• Implementing the Tests: For each pair i, j, estimate the expenditure

functions ρi (ρi,j i j j i,j i j
t , wt , wt ), ρ (ρt , wt , wt ).

• Test 1: Holding wages fixed, compute

gti,j = ρit − ρjt

and test whether always positive/negative. If not, evidence for prefer-

ence heterogeneity.

• Test 2: Test whether nonlabor income y i of household i enters signifi-

cantly in ρi . If yes evidence against efficient risk sharing.

• Test 3: Test whether slope of ρi , ρj with respect to ρi,j is positive. If

not, evidence against efficient risk sharing.

• Data: INCRISAT (International Crops Research Institute for the

Semi-Arid Tropics) VLS (Village Level Studies) on Indian villages.

• From 1975 6 villages, from 1981 10 villages in rural India.

• From each village 40 households (10 landless laborers, 10 small farmers,

10 medium size farmers, 10 large farmers).
66 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

• Key: weather is very important for these villages, and has lots of annual
and seasonal variation. Life is risky there.

• 3 villages selected (Aurepale, Shirapur, Kanzara)

• Monthly data from 1975 - 1985. About 120 observations for about 30
households in each village.

• Observe: labor supply, labor income, assets (used to construct nonla-

bor income), price of goods, monetary and nonmonetary transactions
(from which consumption is constructed), demographics of household
(including caste).

• Results (Summary): Standard risk sharing tests (adding income

changes to regression of consumption change) strongly rejected efficient
risk sharing.

• Tests of preference heterogeneity: homogeneous preferences rejected for

25% of pairs in Aurepale, 16% in Shirapur, 35% in Kanzara.

• Efficient risk sharing tests: results fairly uniform across the two tests.
For few (less than 5%) of pairs efficient risk sharing rejected. But
enough to still formally reject efficient risk sharing on village level.

• Very few rejections across pairs of efficient risk sharing on village-caste

level.
4.4. APPENDIX A: DISCUSSION OF THE NO PONZI CONDITION 67

4.4 Appendix A: Discussion of the No Ponzi

Condition
In this appendix we further discuss the exact form of the no Ponzi condition,
following Wright (1987). To rule out Ponzi schemes he imposes a condition
of the form

−v0 (sT )aiT (sT −1 , sT ) ≤ lim inf SN (sT ) (4.64)

N →∞

where
N X
X
T
v0 (st ) yti (st ) − cit (st )


SN (s ) = (4.65)
t=T st |sT

is the time zero value of future income net of consumption (that is, the time
zero value of period savings) from node sT up to date N. Note that for a
fixed sT , the sequence {SN (sT )} need not converge24 , but if it does, then
∞ X
X
T
v0 (st ) yti (st ) − cit (st )


lim inf SN (s ) = (4.66)
N →∞
t=T st |sT

Intuitively, the no Ponzi scheme condition

ait+1 (st , st+1 ) ≥ −Āi (st+1 , q, ci ) (4.67)

lim inf SN (st+1 )
= −
v0 (st+1 )
24
Remember that for any sequence {xn }∞
n=0 the number

x∗ = lim inf xn
n→∞

is the smallest cluster point of the sequence, i.e.

x∗ = sup{w|xn < w for at most finite number of n’s}

Note that if {xn } converges, then

x∗ = lim xn
n→∞

and if {xn } is unbounded below, then

x∗ = −∞
68 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

rules out a sequence of asset holdings and consumption for which, at any
node st+1 , the beginning of the period debt ait+1 (st , st+1 ) exceeds the node
SN (st+1 )
st+1 of future savings, − lim inf
v0 (st+1 )
.
Let us interpret the No Ponzi scheme condition a bit further. Writing out
(4.64) explicitly using (4.65) we obtain:
N X
X
−v0 (sT )aiT (sT −1 , sT ) ≤ lim inf v0 (st ) yti (st ) − cit (st )


(4.68)
N →∞
t=T st |sT

Now suppose that the sequential budget constraint holds with equality (which
it will as an optimum as long as the utility function is strictly increasing).
Then
X
yti (st ) − cit (st ) = qt (st , st+1 )ait+1 (st , st+1 ) − ait (st ) (4.69)
st+1

≡ qt (st ) · ait+1 (st ) − ait (st )

where we define (in slight abuse of notation):
X
qt (st ) · ait+1 (st ) = qt (st , st+1 )ait+1 (st , st+1 ) (4.70)
st+1

as the value of the asset portfolio bought in node st . Using equation (4.70),
equation (4.68) becomes
N X
X
−v0 (sT )aiT (sT −1 , sT ) ≤ v0 (st ) qt (st ) · ait+1 (st ) − ait (st )
 
lim inf
N →∞
t=T st |sT

lim inf v0 (sT )qT (sT ) · aiT +1 (sT ) − v0 (sT )aiT (sT )

=
N →∞
X
+ v0 (sT +1 )qT +1 (sT +1 ) · aiT +2 (sT +1 ) − v0 (sT +1 )aiT +1 (sT +1 )
sT +1 |sT

X 
+... + v0 (sN )qN (sN ) · aiN +1 (sN ) − v0 (sN )aiN (sN )

sN |sT
 
 X 
= lim inf −v0 (sT )aiT (sT ) + v0 (sN )qN (sN ) · aiN +1 (sN )
N →∞  
sN |s T

(4.71)
4.5. APPENDIX B: PROOFS 69

The last equality is due to the fact that all intermediate terms cancel out.
We demonstrate this for the first terms
X
v0 (sT )qT (sT ) · aiT +1 (sT ) = v0 (sT ) qT (sT , sT +1 )aiT +1 (sT , sT +1 )
sT +1

(4.72)
X X
v0 (sT +1 )aiT +1 (sT +1 ) = v0 (sT )qT (sT , sT +1 )aiT +1 (sT +1 )
sT +1 |sT sT +1 |sT
X
= v0 (sT ) qT (sT , sT +1 )aiT +1 (sT , sT +1 )
sT +1

(4.73)
Simplifying equation (4.71) we can therefore conclude that the no Ponzi
scheme condition, as stated in (4.64), can equivalently be written as
 
X 
lim inf v0 (sN )qN (sN ) · aiN +1 (sN ) ≥ 0 (4.74)
N →∞ N T 
s |s

and thus rules out asset sequences for which the present value of future debt
is bounded above zero. Using this No Ponzi condition Wright (1987) then
states and proves proposition 11 in the main text.

4.5 Appendix B: Proofs

In this appendix we provide a sketch of a proof of proposition 11 in the main
text. See Wright (1987) for a complete proof.
Proof. It is sufficient to show that the budget sets described by the
Arrow Debreu budget constraint (4.7) and the sequential markets budget
constraint (4.28) plus the no Ponzi condition (4.64) (or alternatively (4.74))
contain the same possible consumption choices.
(1) Suppose (ci , ai ) satisfies (4.28) and (4.64). Then
N X
X
−v0 (sT )aiT (sT −1 , sT ) ≤ lim inf v0 (st ) yti (st ) − cit (st )


(4.75)
N →∞
t=T st |sT

Using the definition of Arrow Debreu prices and setting T = 0 yields

N X
X
−p0 (s0 )ai0 (s0 ) pt (st ) yti (st ) − cit (st )


≤ lim inf (4.76)
N →∞
t=0 st |s 0
70 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

Noting that ai0 (s0 ) = 0 for all s0 we find that

N X
X
pt (st ) yti (st ) − cit (st )


0 ≤ lim inf
N →∞
t=T st |s0
N X
X
pt (st ) yti (st ) − cit (st )


= lim inf
N →∞
t=0 st
∞ X
X
pt (st ) yti (st ) − cit (st )


= (4.77)
t=0 st

and thus ci satisfies the Arrow-Debreu budget constraint.25

(2) Suppose ci satisfies the Arrow Debreu budget constraint (4.7). We
want to construct asset holdings a such that (ci , a) satisfy (4.28) and (4.64).
Fix N and define

aN (sN ) = 0 for all sN (4.78)

and recursively

X
aN (st ) = cit (st ) − yti (st ) + qt (st , st+1 )aN (st , st+1 ) (4.79)
st+1

25
The last equality is valid whenever the series converges. One can circumvent the prob-
lem of possible nonconvergence by adjusting the definition of Arrow Debreu equilibrium:
one defines the Arrow debreu budget constraint as

p0 (ci − y i ) ≤ 0

where
T X
X
p0 c = lim inf pt (st )ct (st )
T →∞
t=0 st

Then all arguments go through even if, for a given price process p, the infinite sum does
not converge for some process c. For details see Wright (1987), in particular footnote 6.
4.6. APPENDIX C: PROPERTIES OF CRRA UTILITY 71

Multiplying by v0 (st ) yields

 X
v0 (st )aN (st ) = v0 (st ) cit (st ) − yti (st ) + v0 (st )qt (st , st+1 )aN (st , st+1 )

st+1
X
= v0 (s ) cit (st ) − yti (st ) +
t
v0 (st+1 )aN (st , st+1 )
 

st+1 |st

= v0 (s ) cit (st ) − yti (st )

t
 
X
v0 (st+1 ) cit+1 (st+1 ) − yt+1
i
(st+1 ) + v0 (st+1 )qt (st ) · aN (st )
 
+
st+1 |st

(4.80)
With aN (sN ) = 0 for all sN we have, via continuing substitution
N X
X
T N T
v0 (st ) cit (st ) − yti (st )
 
v0 (s )a (s ) = (4.81)
t=T st |sT

But for each st , aN (st ) is a sequence in N. Now define

at (st ) = lim inf aN (st ) (4.82)
N →∞

Taking lim inf 0 s on both sides of (4.79) yields

X
at (st ) = cit (st ) − yti (st ) + qt (st , st+1 )at+1 (st , st+1 ) (4.83)
st+1

and hence the sequential budget constraint. Taking lim inf 0 s on both sides of
(4.81) yields
N X
X
T T −1
v0 (st ) yti (st ) − cit (st )
 
−v0 (s )aT (s , sT ) = lim inf (4.84)
N →∞
t=T st |sT

and hence (ci , a) as constructed above satisfies the no Ponzi scheme condition.

4.6 Appendix C: Properties of CRRA Utility

In this appendix we discuss the basic properties of the CRRA utility function,
given by  c1−σ −1
1−σ
if σ 6= 1
U (c) =
ln(c) if σ = 1
72 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

1−σ
with σ > 0. Note that limσ→1 c 1−σ−1 = ln(c) which justifies the second line
in the definition of the CRRA utility function above. First we note that
U satisfies the “usual” properties: U is continuous, three times continu-
ously differentiable, strictly increasing (i.e. U 0 (c) > 0), strictly concave (i.e.
U 00 (c) < 0) and satisfies the Inada conditions

lim U 0 (c) = +∞
c&0

lim U 0 (c) = 0
c%+∞

4.6.1 Constant Relative Risk Aversion

Define as
U 00 (c)c
σ(c) = − 0
U (c)
the Arrow-Pratt coefficient of relative risk aversion. Hence σ(c) indicates a
household’s attitude towards risk, with higher σ(c) representing higher risk
aversion. The (relative) risk premium measures the household’s willingness
to pay (and thus reduce safe consumption c̄) to avoid a proportional con-
sumption gamble in which a household can win, but also lose, a fraction of
c̄. See figure ?? for a depiction of the risk premium.
Arrow-Pratt’s theorem states that this risk premium is proportional (up
to a first order approximation) to the coefficient of relative risk aversion σ(c̄).
This coefficient is thus a quantitative measure of the willingness to pay to
avoid consumption gambles. Typically this willingness depends on the level
of consumption c̄, but for a CRRA utility function it does not, in that the
entity σ(c) is constant for all c, and equal to the parameter σ. This explains
the name of this class of period utility functions.26
Add CRRA.wmf pic
26
The CRRA utility function belongs to the more general class of hyperbolic absolute
risk aversion utility functions, given by the general form
 µ
1−µ αc
U (c) = +ω
µ 1−µ

where µ, α, ω are parameters. It is easy to show that, up to irrelavant constants, CRRA

utility is a special case of this general form (as are the CARA and quadratic utility
functions).
4.6. APPENDIX C: PROPERTIES OF CRRA UTILITY 73

4.6.2 Constant Intertemporal Elasticity of Substitu-

tion
Now consider the lifetime utility function (for ease of exposition, in the ab-
sence of risk)
X∞
u(c) = β t U (ct ).
t=0

Define the intertemporal elasticity of substitution (IES) as iest (ct+1 , ct ) as

 
∂u(c)
 −1
∂ct+1
 c  d ∂u(c) 
t+1
d ct

 ∂ct


ct+1 c 
ct
 d t+1ct

iest (ct+1 , ct ) = −   ∂u(c)   = − 
 
∂u(c) 
∂ct+1  ∂ct+1 
d ∂u(c)   ∂u(c) 
∂ct ∂ct
   
 ∂u(c)  ct+1
 ∂ct+1  ct
∂u(c)
∂ct

that is, as the inverse of the percentage change in the marginal rate of substi-
tution MRS between consumption at t and t + 1 in response to a percentage
change in the consumption ratio ct+1 ct
. For the CRRA utility function note
that
∂u(c)  −σ
∂ct+1 ct+1
∂u(c)
= M RS(ct+1 , ct ) = β
ct
∂ct

and thus
 −1
 −σ−1
 −σβ ct+1
ct
 1
iest (ct+1 , ct ) = −  =
 
c −σ
σ
t+1

 β c 
t
ct+1
ct

and the intertemporal elasticity of substitution is constant, independent of

the level or growth rate of consumption, and equal to 1/σ. Graphically, the
IES measures the curvature of the utility function. If σ = 0 consumption
in two adjacent periods are perfect substitutes and the IES equals ies =
∞. If σ → ∞ the utility function converges to a Leontieff utility function,
consumption in adjacent periods are prefect complements and ies = 0.
74 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

The IES has a useful interpretation in terms of observable behavior. From

the first order conditions of the household problem we obtain
∂u(c)
∂ct+1 pt+1 1
∂u(c)
= = qt+1 = (4.85)
pt 1 + rt+1
∂ct

where rt+1 is the real interest rate between periods t and t + 1. Thus, for any
model in which the intertemporal Euler equation holds with equality the IES
can alternatively be written as:
  ct+1     ct+1  
d ct
d ct
ct+1 ct+1
ct ct
iest (ct+1 , ct ) = −  
∂u(c)
 = −  
∂ct+1 d 1+r1
d  t+1
∂u(c) 1
∂ct
  1+rt+1
 ∂u(c)

 ∂ct+1

∂u(c)
∂ct

that is, the IES measures the percentage change in the consumption growth
rate in response to a percentage change in the gross real interest rate, the
intertemporal price of consumption.
Note that for the CRRA utility function the Euler equation reads as
 −σ
ct+1
(1 + rt+1 )β = 1.
ct
Taking logs on both sides and rearranging one obtains

ln(1 + rt+1 ) + log(β) = σ [ln(ct+1 ) − ln(ct )]

or
1 1
ln(ct+1 ) − ln(ct ) =
ln(β) + ln(1 + rt+1 ). (4.86)
σ σ
This equation can then be used to obtain empirical estimates of the IES.
With time series data on consumption growth and real interest rates the IES
1
σ
can be estimated from a regression of the former on the later.27
27
Note that in order to interpret (4.86) as a regression one needs a theory where the
error term comes from. In models with risk this error term can be linked to expectational
errors, and (4.86) with error term arises as a first order approximation to the stochastic
version of the Euler equation. We will discuss the stochastic Euler equation at length in
part II of this monograph.
4.6. APPENDIX C: PROPERTIES OF CRRA UTILITY 75

4.6.3 Homotheticity and Balanced Growth

Finally, the lifetime utility function u is said to be homothetic if M RS(ct+s , ct ) =
M RS(λct+s , λct ) for all λ > 0 and c, that is, if scaling consumption in all pe-
riods up leaves the marginal rate of substitution of consumption between any
two periods unaffected. It is easy to verify that for a period utility function
U of CRRA variety the lifetime utility function u is homothetic, since

β t+s (ct+s )−σ β t+s (λct+s )−σ

M RS(ct+s , ct ) = = = M RS(λct+s , λct ) (4.87)
β t (ct )−σ β t (λct )−σ
Homotheticity of the lifetime utility function in turn is a desirable property
in many macroeconomic applications. It implies that, if an agent’s lifetime
income doubles, optimal consumption choices will double in each period (in-
come expansion paths are linear).28 It also means that consumption allo-
cations are independent of the units in which income and consumption are
measured. This property of the utility function is also crucial for the exis-
tence of a balanced growth in models with growth in income or endowments
(or in models with technological progress in the production function). De-
fine a balanced growth path as a situation in which consumption grows at a
constant rate, ct = (1 + g)t c0 and the real interest rate is constant over time,
rt+1 = r for all t.
Plugging in for a balanced growth path, equation (4.85) yields, for all t
∂u(c)
∂ct+1 1
∂u(c)
= M RS(ct+1 , ct ) = .
1+r
∂ct

But for this equation to hold for all t we require that

M RS(ct+1 , ct ) = M RS((1 + g)t c1 , (1 + g)t c0 ) = M RS(c1 , c0 )

and thus requires that u is homothetic (where λ = (1+g)t in equation (4.87)).

Thus homothetic lifetime utility is a necessary condition for the existence of
a balanced growth path in growth models. Above we showed that CRRA
period utility implies homotheticity of lifetime utility u. Without proof here
we state that CRRA utility is the only period utility function such that
lifetime utility is homothetic. Thus (at least in the class of time separable
28
In the absense of borrowing constraints and other frictions that will be discussed later
in this monograph.
76 CHAPTER 4. THE STANDARD COMPLETE MARKETS MODEL

lifetime utility functions) CRRA period utility is a necessary condition for

the existence of a balanced growth path, which in part explains why this
utility function is used in a wide range of macroeconomic applications.29

29
Note that the class of (non-time separable) Epstein-Zin utility functions discussed in
remark 17 above inherit the homotheticity property from the time separable lifetime utility
with CRRA period utility.
 Part II

The Standard Incomplete

Markets Model (SIM)

77
 79

In this part of the monograph we discuss the standard incomplete mar-

kets model in which households can lend (and perhaps borrow) at a risk-free
rate, with the goal of insulating consumption against fluctuations in labor
income. This literature starts with the famous work by Friedman (1957) on
the permanent income hypothesis and the life cycle hypothesis of Modigliani
and Brumberg (1954). Common to both studies is the focus on a single
consumer (with interest rates are exogenously given and not derived endoge-
nously in general equilibrium) and the absence of explicit insurance arrange-
ments against individual income risk (these are ruled out in ad hoc fashion).
These early and seminal contributions sparked important theoretical work in
the 1970’s on what became known as the “income fluctuation problem”.
In chapter 5, after a brief introduction by means of a two period toy
model, we then present a formal SIM model (with quadratic period utility and
nonbinding borrowing constraints) that implies both certainty equivalence
(households make consumption savings decisions that are -ex ante- identical
to those that are optimal in a world without income risk) as well as results
in Friedman’s permanent income hypothesis. We then discuss extensions
of the basic theory that relax the assumption of quadratic utility and/or
slack borrowing constraints and show how this can lead to precautionary
saving behavior (these concepts will be made precise below). We also discuss
how versions of the model can be solved numerically for which no analytical
solution exists (as is the norm).
The key quantitative ingredient into SIM is the stochastic labor income
or wage process to be fed into the household consumption-saving problem.
Thus, in chapter 6 we digress from the main theme of this part of the mono-
graph and briefly discuss the empirical properties of these process, with spe-
cific focus on the recent discussion in the literature about how persistent
income shocks are that hit individuals or households.
Finally, in chapter 7 we present general equilibrium versions of the SIM
model. We first discuss stationary equilibria, characterized by an endogenous
cross-sectional distribution over household characteristics (the relevant aggre-
gate state variable in this class of models) that is time invariant. although
individual households move up and down in that distribution. We will then
consider deterministic transitions in which the cross-sectional distribution of
household characteristics evolves over time in a deterministic fashion, before
concluding with a business cycle version of the model in which the economy
is hit by recurrent aggregate shocks and thus the cross-sectional distribution
follows a stochastic law of motion.
80
Chapter 5

The SIM in Partial Equilibrium

Permanent income type models assume that agents do not have access to a
complete set of contingent consumption claims. The main difference between
this type of models and the complete markets model thus manifests itself in
the budget constraints. Whereas the complete markets model in sequential
formulation has budget constraints of the form (4.28), for the SIM model we
have1
ct (st ) + qt (st )at+1 (st ) = yt (st ) + at (st−1 ) (5.1)

Here qt (st ) is the price at date t, event history st , of one unit of consumption
delivered in period t + 1 regardless of what event st+1 is realized. Notation
is crucial here: at+1 is the quantity of risk-free one period bonds being pur-
chased in period t and paying off in period t + 1. Thus at+1 (st ) is a function
only of st and not of st+1 whereas in the complete markets model with a full
set of Arrow securities these assets are indexed by (st , st+1 ) = st+1 and each
asset pays off only at a particular event history st+1 . Obviously each of these
Arrow securities has a potentially different price qt (st , st+1 ), as discussed in
the previous chapter.
The basic income fluctuation problem is to maximize

T X
X
u(c) = β t πt (st )U (ct (st ), st ) (5.2)
t=0 st ∈S t

1
In principle the budget constraints have to hold only as inequalities. At the optimal
allocations they will always hold as strict equality as long as household utility is strictly
increasing in consumption, an assumption we maintain throughout.

81
82 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

subject to the sequence of period budget constraints given in (5.1) and a

short-sale constraint on assets at+1 (st ) ≥ −Ā(st ; q) which rules out Ponzi
schemes2 , but is assumed to be sufficiently wide so as to never bind at the
optimal allocation chosen by the household. In later parts of this chapter
we will consider borrowing limits that are more stringent and potentially
binding. Let the initial wealth of an agent be denoted by a0 (s−1 ) = a0 . At
this point it is not required to make further assumptions on a0 . Unless noted
explicitly we assume that U satisfies Assumption 1 of the previous chapter (U
is twice continuously differentiable, strictly increasing, strictly concave and
satisfies the Inada conditions). Note that for the discussion in this chapter
time and household age can be used interchangeably since one household
is considered in isolation. Thus the index t will be used for a generic time
period or age of the household.

5.1 A Simple 2 Period Toy Model

To develop intuition for the solution of the general model we first look at the
simplest example of a SIM with two periods and without risk.

5.1.1 Household Decision Problem

The problem of the household then becomes

max U (c0 ) + βU (c1 ) (5.3)

cc ,c1 ,a1
s.t.
c0 + qa1 = y0 + a0 (5.4)
c1 = y1 + a1 (5.5)
c0 , c1 ≥ 0 (5.6)

One can consolidate the budget constraints to a lifetime budget constraint

which reads as:
c1 y1
c0 + = a0 + y0 + (5.7)
1+r 1+r
2
In the case of an infinite planning horizon; for the finite horizon case we continue to
impose aT +1 (sT ) ≥ 0.
5.1. A SIMPLE 2 PERIOD TOY MODEL 83

where 1 + r = 1q is the gross real interest rate. Equation (5.7) together with
the Euler equation
U 0 (c0 ) = β(1 + r)U 0 (c1 ) (5.8)
uniquely determine the optimal consumption allocation (c0 , c1 ) over the life
cycle of the household, as a function of the model parameters (y0 , y1 , r, β) and
any parameters characterizing the period utility function (e.g. the parameter
σ in the case of CRRA utility).

5.1.2 Comparative Statics

The comparative statics with respect to most model parameters can be signed
easily. Consumption in both periods depends positively on the present dis-
counted value of lifetime income (including initial wealth a0 )
y1
W0 (y0 , y1 , r, a0 ) = a0 + y0 + (5.9)
1+r
but is independent of the timing of income and thus the income profile
(y0 , y1 ). An increase in the time discount factor β (turning the household
more patient) induces a shift from early consumption c0 to late consumption
c1 . However, even in this very simple model signing the effect of changes in
the interest rate r on consumption allocations is not straightforward. An
increase in the interest rate r has three effects on consumption allocations:

1. An increase in r reduces the relative price of consumption in period 1,

relative to consumption in period 0, and thus reduces current consump-
tion and relatively less expensive compared to period 1 consumption
and increases saving and future consumption. This is the (intertempo-
ral) substitution effect.

2. An increase in r reduces the price of period 1 consumption in absolute

terms, acting like an increase in income. This is the income effect
which, ceteris paribus, increases consumption in both periods.

3. Provided that the household has income in the second period, an in-
crease in r reduces the present value of lifetime income W0 (y0 , y1 , r, a0 )
and hence reduces current and future consumption. This is the human
capital or wealth effect.
84 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Which effect dominates depends on the particular form of the utility func-
tion and the income profile. We can at least partially rank the magnitudes
of these effects for the case in which households have CRRA period utility,
discussed in the previous chapter. Recall that the parameter 1/σ measures
the intertemporal elasticity of substitution as discussed in the appendix of
the previous chapter and hence the potency of the substitution effect.3
In the case of CRRA utility the explicit solution of the model is given by:

W0 (y0 , y1 , r, a0 )
c0 = 1 1 (5.10)
β σ (1+r) σ
1+ 1+r
1 1
β σ (1 + r) σ W0 (y0 , y1 , r, a0 )
c1 = 1 1 (5.11)
β σ (1+r) σ
1+ 1+r

1
The dependence of W0 on 1 + r captures the wealth effect, the term (1 + r) σ
captures the substitution effect (absent if 1/σ = 0) and the term 1 + r in the
second denominator encompasses the income effect. Since

∂W0 (y0 , y1 , r, a0 ) y1
=− ≤0
∂(1 + r) (1 + r)2

the human wealth effect is negative for consumption in both periods, and is
arbitrarily small or large depending on the size of future income y1 , and thus
can either dominate, or is being dominated by the other two effects.
Now suppose the wealth effect is absent (e.g. because y1 = 0). Then
−2 1  !
∂c0  1 1
−1 βσ 1 1 1
= W0 (y0 , y1 , r, a0 ) ∗ 1 + β (1 + r)
σ σ ∗ (1 + r) − (1 + r) σ
σ
∂(1 + r) (1 + r)2 σ
1 1  !
 1 1
−2 β σ (1 + r) σ 1
−1
= W0 (y0 , y1 , r, a0 ) ∗ 1 + β (1 + r)
σ σ ∗ 2
1− (5.12)
(1 + r) σ

The first term in the last bracket is the positive income effect, the second term
the negative intertemporal substitution effect. expression the human capital
effect, the second term is the combination between income and substitution
effect. For CRRA utility the relative magnitude of these two effects only
3
In the absence of risk, risk aversion (as measured by σ) of households is irrelevant for
the interpretation of the solution.
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 85

depends on the size of the intertemporal elasticity of substitution, as given

by the parameter σ1 . It is immediate that we can distinguish the following
three cases:

1. If 1/σ = 1 (log-case, unit intertemporal elasticity of substitution), in-

come and substitution effect exactly cancel out and current consumption
does not respond to changes in interest rates, absent a negative human
wealth effect.

2. If 1/σ > 1 (high intertemporal elasticity of substitution), then the sub-

stitution effect dominates the income effect and hence current consump-
tion declines as reaction to an increase in the interest rate.

3. If 1/σ < 1 (low intertemporal elasticity of substitution), then the

income effect dominates the substitution effect and, absent a human
wealth effect, current consumption increases (and current saving de-
creases) with the interest rate.

Tedious but straightforward algebra also reveals that for consumption in

period 1 both the income and the substitution effect from an increase in the
interest rate are positive, and thus, absent the negative wealth effect, future
consumption rises with an increase in the interest rate. To summarize, in
the absence of income risk the only motive when making the intertemporal
consumption decision is to insulate the timing of consumption from income
variation over time, by means of (dis-saving) using the one-period risk free
bonds. In the next sections we will discuss how this behavior is affected by
the presence of income risk and the potential desire to hedge against this
income risk through precautionary saving.

5.2 The General Model with Certainty Equiv-

alence
Now we turn attention to the study of the general model in which households
live for multiple periods, face a potentially stochastic income process and
maximize (5.2) subject to (5.1). In the next section we will study versions
of the model that give rise to certainty equivalence: households will make
consumption savings decision in the presence of income risk that are (ex
ante, prior to the resolution of income risk) identical to the choices they
86 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

would make in the absence of risk. The following section will then relax the
assumptions leading to certainty equivalence and study versions of the model
in which households engage in precautionary saving behavior.4
For the rest of this chapter we impose the following assumption on the
exogenous prices (interest rates) of the one period risk-free bond.

Assumption 18 The price of the one period bond is nonstochastic and con-
stant over time,
1
qt (st ) = q = .
1+r

5.2.1 Nonstochastic Income

In order to set the stage for a comparison with the situation in which income is
stochastic, first consider a version of the model in which household income is
given by the deterministic sequence y = {yt }Tt=0 . Also assume that the process
{st } that potentially drives household preference shifts is a deterministic
sequence. Define as
T
X yt
W0 = W0 (a0 , y, r) = a0 + (5.13)
t=0
(1 + r)t

the present value of all future labor income, including initial wealth. We
assume that W0 is finite (otherwise the household maximization problem does
not have a solution).5 Forming the Lagrangian, taking first order conditions
and combining yields the standard Euler equation

Uc (ct , st ) = β(1 + r)Uc (ct+1 , st+1 ) (5.15)

4
These terms will be defined more formally below.
5
When T = ∞ we impose a short sale constraint on the bond that prevents Ponzi
schemes, but is loose enough to allow optimal consumption smoothing. Defining as the
present discounted value of remaining lifetime income
∞
X yτ
Wt+1 =
τ =t+1
(1 + r)τ −t

and assuming that the sum is finite for all t, we require that for all t bond holdings adhere
to the “natural borrowing limit”
at+1 ≥ −Wt+1 (5.14)
Note that Wt+1 is finite if r > 0 and the sequence {yt } is bounded from above.
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 87

and hence, for all t,

 t
t 1+ρ
Uc (ct , s ) = Uc (c0 , s0 ) (5.16)
1+r
where time discount factor β and time discount rate ρ are related through
1
the definition β = 1+ρ . As before in the two-period model, the sequence of
Euler equations (5.16) and the intertemporal budget constraint
T
X ct
= W0 (5.17)
t=0
(1 + r)t
jointly determine the optimal consumption choice of the household. Inspect-
ing equations (5.16) we obtain the following predictions from the determinis-
tic version of the income fluctuation problem. These predictions are sharpest
under the assumption that the rate r at which the market discounts the fu-
ture equals the subjective time discount rate ρ of households.

Specific Case ρ = r
If ρ = r then the right hand side of (5.16) is constant over time and thus
Uc (ct , st ) is time-invariant. In periods in which the preference shifter st makes
marginal utility high, consumption also has to be high, since marginal utility
is decreasing in consumption. This may provide us with a nontrivial theory
of life-cycle consumption profiles even in the absence of risk, and even under
the assumption ρ = r.
Example 19 If U is separable between consumption and preference shifters
st , then we immediately obtain that household consumption is constant over
time (over the life cycle)
ct = ct+1 (5.18)
something that seems counterfactual6 in light of the discussion in chapter
3. We can combine () with the intertemporal budget constraint () to solve
explicitly for the level of consumption. This yields, for all t,
 −1 rW0
θ 1+r if T < ∞
c0 = ct = ct = r (5.19)
1+r
W0 if T = ∞
6
Note that under the separability assumption, even if ρ 6= r, consumption does not
display a life-cycle hump shape, but rather monotonically trends upwards (if r > ρ, i.e.
if incentives to postpone consumption dominate impatience) or monotonically downwards
(if ρ > r, i.e. if impatience dominates incentives to postpone consumption).
88 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM
 
where θ = 1 − (1+r)1 T +1 is an adjustment factor needed if the lifetime hori-
zon T of the household is finite. Thus consumption at each date is equal to
“permanent income”
rW0
ct =
θ (1 + r)
At each date households consume annuity value of their lifetime wealth W0 ,
composed of initial financial wealth and the present discounted value of life-
time labor income. This is the purest version of the permanent income hy-
pothesis: consumption at each date should equal permanent income.

Example 20 For more realistic assumptions on the time (age) profile of

preference shifters it need not be the case that household consumption is con-
stant or trending monotonically upwards or downwards over time. Suppose
that st stands in for the (exogenous) number of individual members present in
the household. In the data, household size first increases and then decreases
with the age t of the household head (since this person first tends to get mar-
ried and have children, then the children leave the household and finally the
household head or spouse dies). Now suppose that the period utility function
takes the form
c1−σ
U (ct , st ) = exp(st ) ∗ t (5.20)
1−σ
so that marginal utility from household consumption expenditure ct is higher if
there are more members of the households that have to share the benefits from
a given expenditure ct . With this preference specification the Euler equation
reads as
ct 1
= exp(st − s0 ) σ (5.21)
c0
As family size increases (i.e. st > s0 ), so will optimal household consump-
tion. Thus a hump-shaped family size profile, at least qualitatively, can ac-
count for a hump-shaped household consumption expenditure profile in the
data, even if households follow the simple consumption-savings model out-
lined in this section.7

Example 21 A similar argument can be made if consumption and leisure

are non-separable in the utility function. Suppose st reflects the amount
7
Attanasio et. al. (1999) and Fernandez-Villaverde and Krueger (2007) investigate this
point from a quantitative perspective.
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 89

worked (as a share of total available non-sleeping hours) in period t, again

taken as exogenous for now. Let the period utility function be given by
1−σ
(cγt (1 − st )1−γ ) −1
U (ct , st ) = with σ > 1 and γ ∈ (0, 1) (5.22)
1−σ
where γ is a share parameter measuring the importance of consumption rel-
ative to leisure 1 − st in the utility function. The Euler equation becomes
  (1−γ)(1−σ)
ct 1 − st 1−γ+γσ
= (5.23)
c0 1 − s0
Thus, in periods where labor supply st is high, so should consumption. There-
fore a hump-shaped life cycle profile of hours worked can explain, at least
qualitatively, a hump-shaped life cycle consumption profile.8 Also note that
since, at retirement, the number of hours worked decreases sharply, consump-
tion should fall at retirement as it does in the data. Thus this argument is
the starting point for a potential resolution of the consumption retirement
puzzle.9

General Case ρ 6= r
For a fixed sequence of preference shifters {st }, a decline in the interest rate
r makes the life cycle consumption profile less steep, as future consumption
is substituted in favor of current consumption, ceteris paribus. The reverse
is true for an increase in the interest rate. Note that this is not a statement
about the level of the consumption profile but rather its slope, because the
effect of interest rate changes on consumption levels also depends on the
size of the income and human wealth effects, as discussed in the previous
section. Thus the Euler equations only capture the substitution effect and
make predictions about the shape of the consumption profile over the life
8
See Heckman (1974) for the original source of this point. A hump-shaped profile
of hours worked in turn would be part of the optimal household labor supply choice in
the presence of a hump-shaped life cycle wage profile (at least under the appropriate
assumptions on household preferences).
9
Aguiar and Hurst (2005, 2007) argue that older households (especially those in retire-
ment) pay lower prices for the same consumption goods because they spend more time
shopping for good deals. This results in a decline in life cycle consumption expenditures to-
wards the end of life, without necessarily implying a fall in utility-generating consumption
services that enter the utility function.
90 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

cycle. To determine consumption levels, in addition the intertemporal budget

constraint 5.17 has to be invoked.
Before turning to the model with a stochastic income process, in the
example below we provide an explicit solution to the household problem
when households have CRRA utility.

Example 22 Suppose households have preferences represented by (4.13). Then

the Euler equation reads as
 
−σ 1+ρ
(ct+1 ) = (ct )−σ .
1+r

Taking logs on both sides and rearranging yields

1
∆ ln ct+1 = [ln(1 + r) − ln(1 + ρ)] (5.24)
σ
Using the Euler equations and the intertemporal budget constraint one can
solve for the optimal consumption allocation, which, for all t ≥ 0, is given
by:
 t  
1+r σ 1−γ
ct = W0 (5.25)
1+ρ 1 − γ T +1
≡ M P C(t, T ) ∗ W0 (5.26)
  σ1
10 [1+r]1−σ
where γ is defined as γ = . Holding lifetime income W0 fixed we
1+ρ
find that an increase in the lifetime horizon T reduces the marginal propen-
sities to consume out of lifetime wealth M P C(t, T ) at all ages t.

5.2.2 Stochastic Income and Quadratic Preferences

Now let us consider maximizing (5.2) subject to (5.1), but now permit the
income process to be stochastic.11 Attaching Lagrange multiplier λt (st ) to
10
In order to insure that lifetime utility is finite at the optimal allocation, in addition
to assuming W0 < ∞, we require that γ < 1.
11
Again, to make the problem well-defined we need a No-Ponzi condition. For finite T
we require
aT +1 (sT ) ≥ 0 (5.27)
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 91

the event history st budget constraint and taking first order conditions with
respect to ct (st ) and ct+1 (st+1 ) yields
β t πt (st )Uc (ct (st ), st ) = λt (st ) (5.30)
β t+1 πt+1 (st+1 )Uc (ct+1 (st+1 ), st+1 ) = λt+1 (st+1 ) (5.31)
Taking first order conditions with respect to at+1 (st ) yields
X
λt (st )q = λt+1 (st+1 ) (5.32)
st+1 |st

Combining yields
1 X t+1
β t πt (st )Uc (ct (st ), st ) = β πt+1 (st+1 )Uc (ct+1 (st+1 ), st+1 ) (5.33)
q t+1 t
s |s

for all sT .
For the infinite horizon case, T = ∞, we impose
∞
X yτ +1 (sτ +1 )
at+1 (st ) ≥ − inf (5.28)
{sτ +1 }|st
τ =t+1
(1 + r)τ −(t+1)
≡ − inf Wt+1 (st ) ≡ −Āt+1 (st ) (5.29)
{sτ +1 }|st

where the right hand side is the minimum (infimum, if the minimum does not exist)
realized present discount value of future income over all infinite histories {sτ +1 } that can
follow node st . We assume that the process {Āt+1 (st )} is bounded above, which is true
as long as r > 0 and current income only depends on the current shock: yτ (sτ ) = y(sτ ).
In that case define
ymin = min y(st )
st ∈S

and
∞
X yτ +1 (sτ +1 )
Āt+1 (st ) = inf
{s τ +1
}|st
τ =t+1
(1 + r)τ −(t+1)
∞
X 1
= ymin τ −(t+1)
τ =t+1
(1 + r)
ymin (1 + r)ymin
= 1 =
1 − 1+r r
This No Ponzi constraint, the natural extension of the concept of the “natural borrowing
constraint” to a stochastic income process, insures that households having incurred the
maximal amount of debt can repay it with probability 1 (by setting consumption to zero
in all future periods). As long as the utility function satisfies the Inada conditions this
constraint will never be binding.
92 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

or
  X
t t 1+r
Uc (ct (s ), s ) = πt+1 (st+1 |st )Uc (ct+1 (st+1 ), st+1 )
1+ρ
st+1 |st
 
1+r
E Uc (ct+1 (st+1 ), st+1 )|st


= (5.34)
1+ρ
or, more compactly,
 
t 1+r
Et Uc (ct+1 , st+1 )


Uc (ct , s ) = (5.35)
1+ρ
where Et is the expectation of st+1 conditional on st . This is the standard
stochastic Euler equation for the standard incomplete markets (SIM) model.
We immediately see that if ρ = r, then the marginal utility of consumption
process {Uc (ct , st )} follows a martingale, i.e. its period t expectation of the
t + 1 variable equals the t variable.12

Remark 23 Recall that for the standard complete markets (SCM) model,
t+1 |st )
under the assumption13 that Arrow securities prices satisfy qt (st+1 ) = πt+1 (s
1+r
12
A martingale is a stochastic process {xt } that satisfies, with probability 1,

Et (xt+1 ) = xt .

A process that satisfies, with probability 1,

Et (xt+1 ) ≥ xt

is called a submartingale, and a process for which

Et (xt+1 ) ≤ xt

holds with probability 1 is called a supermartingale. Thus the relationship between ρ

and r determines whether marginal utility in the SIM model without binding borrowing
constraints is a sub-, super-, or standard martingale. The same statement applies to
consumption itself, for the case of quadratic utility discussed below.
13
¿From equation (4.39) we observe that this assumption is warranted as long as ag-
gregate income (endowment) and thus aggregate consumption is constant over time. In
chapter 7 we will construct general equilibrium models with a continuum of households in
which this will be true. Note that for the SCM model
1
= β.
1+r
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 93

delivers a set of Euler equations of the form

 
t t 1+r
Uc (ct (s ), s ) = Uc (ct+1 (st+1 ), st+1 ) for all st+1 (5.36)
1+ρ
Thus the key difference between the SCM and in the SIM models is that in the
former the Euler equation holds state by state st+1 , whereas in the SIM model
it only holds in (conditional) expectation. This is due to the fact that in the
SCM model households have access to a full set of Arrow securities, which
allows households to smooth marginal utility over time state by state, whereas
the best a household in the SIM model can do is to use the uncontingent bond
to smooth marginal utility over time on average (in expectation).

As in the deterministic case, the Euler equations (5.35) determine con-

sumption profiles over time and will also allow us to make statements about
how consumption responds to income shocks, but they are, by themselves,
insufficient to deduce the level of consumption in every period. For that we
need, in addition, the sequence14 of period budget constraints (5.1). In gen-
eral, no closed form solution for the stochastic consumption process exists,
and one has to resort to the numerical techniques discussed in section 5.6.
However, if households have quadratic period utility, then a sharp char-
acterization of the optimal consumption choices can be obtained, and these
choices obey “Certainty Equivalence”. To demonstrate this, we now make

Assumption 24 The period utility function is quadratic and separable in

consumption
1
2
U (ct (st ), st ) = − ct (st ) − c̄ + v(st ) (5.37)
2
where c̄ is the bliss level of consumption, assumed to be large relative to an
agents’ stochastic income process.15
14
As we will see below, under the assumption that interest rates are nonstochastic and
households have access only to an uncontingent short-term bond, we can consolidate the
sequence of budget constraints into an intertemporal budget constraint, exactly as in the
case without risk. This is by no means true in incomplete markets models with more
complex asset market structure and asset price processes.
15
We assume that c̄ is so large that the agent cannot afford, given his income process,
to obtaim ct (st ) = c̄ for all st . Note that it is easy to pick c̄ large enough for T finite,
since aT +1 (sT ) = 0 is required. For T = ∞ it must be chosen large enough so that the
consumption allocation ct (st ) = c̄ leads to required asset holdings that eventually violate
the no Ponzi scheme condition.
94 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Under assumptions 18 and 24 the Euler equations in (5.35) become

Et ct+1 = α1 + α2 ct (5.38)
1+ρ 1+ρ


with α1 = c̄ 1 − 1+r
and α2 = 1+r
. If furthermore ρ = r, then equation
(5.38) turns into:
Et ct+1 = ct (5.39)
i.e. not only marginal utility but consumption itself follows a martingale.
We summarize the most important implications of (5.38) or (5.39) as:

1. The agent’s optimal consumption decisions obey certainty equivalence;

comparing the rules for optimally allocating consumption over time
in the certainty and the case with income risk, (5.18) and (5.39), we
see that they are identical.16 Of course the realized consumption path
under income risk will deviate, ex post, from the path chosen by the
household in the absence of risk.

2. From equation (5.38) it follows that the optimal stochastic consumption

process chosen by the household therefore obeys the regression equation

ct+1 = α1 + α2 ct + ut+1 (5.40)

where ut+1 is a random variable satisfying Et ut+1 = 0. The main empiri-

cal implication of the SIM model in its most basic, certainty equivalence
form is that period t + 1 consumption ct+1 is perfectly predicted by pe-
riod t consumption ct and no other variables that are in the households’
information set at period t should help predict it. In particular, once ct
is included in the regression, current income yt , current assets at , past
consumption ct−1 or any other variable should not enter with a signif-
icant coefficient if included in regression 5.40. Following the literature
we call this the consumption “martingale” hypothesis.

3. Note, however, that the income realization in period yt+1 can affect
the choice ct+1 , but only through that part that constitutes a deviation
16
Of course, the assumption ρ = r is not required for the result, as it is straightforward
to show that, with quadratic utility, the consumption dynamics in the no-risk case is given
by
ct+1 = α1 + α2 ct
compared to equation (5.38) in the main text.
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 95

(a “shock”) from its period t expectation yt+1 − Et (yt+1 ). The com-

ponent of income yt+1 that is already predictable in period t (that is,
whatever of yt+1 is in the information set at time t) should not affect
consumption in period t. The fact that consumption responds to income
shocks strongly sets the SIM apart from the complete markets model
where even unexpected changes in (individual) income do not affect
consumption, since agents can perfectly insure (using the purchase of
Arrow securities) against these unexpected changes.17

Hall’s (1978) Empirical Tests of the Martingale Hypothesis Using

Macroeconomic Data
The implications discussed under point 2. above are empirically tested by
a large literature18 , starting from the seminal paper by Robert Hall (1978).
He used aggregate consumption and income data to run the two regressions,
motivated by equation (5.40)

ct+1 = α1 + α2 ct + α3 ct−1 + α4 ct−2 + α5 ct−3 + ut+1 (5.41)

ct+1 = β1 + β2 ct + β3 yt + vt+1 (5.42)

Under the null of the martingale hypothesis, α3 = α4 = α5 = 0 and β3 = 0.

Hall (1978) used per capita nondurable consumption expenditures (including
services) in constant 1972 dollars from the NIPA as his measure of ct . His
income measure for yt is total nominal disposable income per capita from
the NIPA, divided by the implicit deflator for nondurable consumption and
services.19 Hall’s basic results from regression for quarterly data from 1948-
1977 are summarized in Table 3
17
Recall that in our discussion of perfect consumption insurance in chapter 4 we never
made a distinction between predicted and unpredicted changes in income; such distinction
is unnecessary in the complete markets model.
18
Most of this literature starts from the Euler equation with CRRA utility, obtains a first
order approximation and then implements a regression similar to equation (5.42), but in
logs, on household level consumption and income data. We will derive the approximation
and discuss the relevant literature in the next section.
19
To stress this point again, Hall’s (1978) original test used aggregate data, although
the theory envisions a single household as its unit of analysis. As long as households
have linear marginal utility, equation (5.38) can be aggregated perfectly across households
and the theory makes the same predictions for aggregate consumption and income data.
Such aggregation fails for the SIM whenever we deviate from the quadratic utility case, as
demonstrated below.
96 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Table 3: Test of Martingale Hypothesis

Regression Parameter Estimates and Standard Errors (in Parenthesis) R2

α1 = 8.2 α2 = 1.13 α3 = −.04 α4 = .03 α5 = −.113
(5.41) .9988
(8.3) (.092) (.142) (.142) (.093)
β1 = −16 β2 = 1.024 β3 = −.01
(5.42) .9988
(11) (.044) (.032)

The basic message of Table 3 is that the martingale hypothesis cannot be

rejected. The coefficients on lagged consumption and current income are not
significantly different from 0, and F -tests indicate that the hypothesis that
they are jointly equal to zero cannot be rejected at standard confidence levels.
The inclusion of further lags in consumption or income does not change the
result. This is strong evidence in favor of the martingale hypothesis.
However, when Hall current stock prices (the real value of the S&P 500)
are included into the regression, as a proxy for current wealth at , the coeffi-
cient is highly significant, formally rejecting the martingale hypothesis. Thus
the empirical evidence using aggregate data is mixed, but with perhaps sur-
prisingly large support for the hypothesis from regressions using consumption
and income data.20 The more recent literature using household level data
tends to find stronger re

The Consumption Function under Certainty Equivalence

The previous analysis only used the Euler equation to derive the empirical
test of the model. Equations (5.38) and (5.39) do not constitute close-from
solutions of the model, as the right hand side of both equations contains the
endogenous (and stochastic) variable ct . We now assume ρ = r and combine
(5.39) with the household budget constraints (5.1) to solve for the optimal
consumption allocation in closed form, and to investigate how consumption
responds to income shocks, under the assumption of quadratic utility.21
In appendix 5.8 we show how to consolidate the sequence of constraints
(5.1) over time, in order to arrive at an intertemporal budget constraint
20
Note that Hall’s study was conducted prior to the development of the co-integration
literature in time series econometrics (see e.g. Engle and Granger, 1987) that raised
concerns about the interpretation of regressions using aggregate nonstationary data.
21
We could use (5.38) and derive similar results for the case ρ 6= r, but at the expense
of substantial algebra, and without obtaining additional insights.
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 97

similar to equation (5.17). For a fixed node st , this intertemporal budget

constraint reads as

T −t X T −t X
X πt+τ (st+τ |st )ct+τ (st+τ ) t−1
X πt+τ (st+τ |st )yt+τ (st+τ )
≤ a t (s ) +
τ =0 t+τ t
(1 + r)τ τ =0 t+τ t
(1 + r)τ
s |s s |s

(5.43)
T −t
 ! T −t
 !
X ct+τ  t X yt+τ  t

E s ≤ at (st−1 ) + E  s ≡ Wt (st )
τ =0
(1 + r)τ  τ =0
(1 + r)τ 
(5.44)

or, in short,

T −t T −t
X ct+τ X yt+τ
Et ≤ Et + at ≡ W t (5.45)
τ =0
(1 + r)τ τ =0
(1 + r)τ

where the dependence of all variables on st is understood and the expectation

is conditional on st .
Now, using equation (5.39) and the law of iterated expectations, we have22

Et ct+τ = ct for all τ ≥ 0.

Using this result in the intertemporal budget constraint (5.45), which at the
optimal allocation holds with equality, we can solve for consumption in closed

22
Since

Et ct+1 = ct
Et ct+2 = Et Et+1 ct+2 = Et ct+1 = ct
98 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

form23 :
θt−1 rW

1+r
t
if T < ∞
ct = r (5.46)
1+r
W t if T = ∞
 
1
where θt = 1 − (1+r)T −t+1
.

Remark 25 More explicitly,

( t
t θt−1 rW1+r
t (s )
if T < ∞
ct (s ) = r t (5.47)
1+r
Wt (s ) if T = ∞

Note that this equation holds for the fixed initial node s0 as well, so that (say,
for T = ∞),
∞
!
r X yt+τ
c0 (s0 ) = a0 + E 0
1+r τ =0
(1 + r)τ

where a0 is the exogenous initial asset position the household starts her life
with (whereas for future periods Wt (st ) is a function of the endogenous choice
at (st−1 )). Compare this to the consumption function under certain income
in (5.19): in period 0 households make identical consumption (and thus sav-
ing) choices in the presence and absence of income risk: the SIM model
with quadratic utility (and nonbinding borrowing constraints) exhibits cer-
tainty equivalence. Of course, in the stochastic case future consumption and
asset holdings respond to future income shocks (which are absent in the de-
terministic case). However, for a given amount of assets carried into period
t, households at that time period would make identical consumption-savings
choices for period t in the deterministic and the stochastic income case.
23
More explicitly,
( t
t θt−1 rW1+r
t (s )
if T < ∞
ct (s ) = r t
1+r Wt (s ) if T = ∞

Note that this equation holds for the fixed initial node s0 as well, so that (say, for T = ∞),
∞
!
r X yt+τ
c0 (s0 ) = a0 + E0
1+r τ =0
(1 + r)τ

where a0 is the exogenous initial asset position the household starts her life with (whereas
for future periods Wt (st ) is a function of the endogenous choice at (st−1 )).
5.2. THE GENERAL MODEL WITH CERTAINTY EQUIVALENCE 99

Consumption Responses to Income Shocks24

Given the consumption function in equation (5.46) we can also determine how
realized consumption changes in response to income shocks. We have already
argued that expected changes in income have no effect on consumption; here
we are interested in the impact of income shocks. Defining ∆ct = ct − ct−1
to be the realized change in consumption between period t − 1 and period t,
in Appendix 5.9 we show that
T −t
r X (Et − Et−1 )yt+s
θt ∆ct = ≡ ηt (5.48)
1 + r s=0 (1 + r)s

Here (Et − Et−1 )yt+s = Et yt+s − Et−1 yt+s is the revision of the expectation
about period t + s income between period t − 1 and period t, and thus ηt is
the annuity value of the revisions of expectations, between
 t − 1 and
 t, of all
1
future incomes. Again, the adjustment factor θt = 1 − (1+r)T −t+1 controls
for the length of remaining lifetime if T is finite. Thus the realized change
in consumption between two periods equals the annuity value of the revision
in expectations about the present discounted value of future income, ηt . Its
exact value depends on the specifics of the stochastic income process that
households face. We now consider several concrete examples.

Example 26 Suppose the income process of a household is specified as

yt = ytp + ut (5.49)
ytp = yt−1
p
+ vt (5.50)

where ytp is the “permanent” part of current income, ut is the transitory part
and vt is the innovation to the permanent part of income.25 We assume that
ut and vt are uncorrelated iid random variables with Et ut+s = Et vt+s = 0 for
s > 0, where we adopt the timing convention that ut , vt are known when Et
is taken. This process can be rewritten as

yt = yt−1 + ut − ut−1 + vt (5.51)

24
This subsection is based on Blundell and Preston (1998).
25
Income processes with persistent (often permanent) and transitory shocks are very
commonly estimated by labor economists and used by quantitative macroeconomists in
their heterogeneous household models. Predominantly the permanent-transitory distinc-
tion is applied to log-income, however, see chapter 5.9 for a more detailed discussion.
100 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

and thus
t+s
X
yt+s = yt−1 + ut+s − ut−1 + vτ (5.52)
τ =t

We now demonstrate that agents behaving according to the SIM model with
quadratic utility react quite differently to permanent income shocks vt and
transitory income shocks ut . We then have that

ut if s = 0
Et yt+s = yt−1 − ut−1 + vt + (5.53)
0 if s > 0
Et−1 yt+s = yt−1 − ut−1 + 0 + 0 (5.54)

and hence 
ut + vt if s = 0
(Et − Et−1 ) yt+s = (5.55)
vt if s > 0
Thus we can calculate ηt in equation (5.48) as
r
ηt = ut + θt vt
1+r

and thus we find that, for the specific income process with permanent and
transitory shocks, the realized change in consumption is given by
r
θt ∆ct = ut + θt vt (5.56)
1+r
rθt−1
∆ct = ut + vt . (5.57)
1+r

Thus households adjust their consumption one for one with a permanent in-
rθt−1
come shock vt , but only change their consumption mildly, by 1+r in response
to a purely temporary shock ut .

Example 27 The last example assumes that income shocks are either per-
manent or transitory. Now suppose that households are only subject to one
income shock, and that a fraction γ of the shock is mean-reverting whereas a
fraction 1 − γ is permanent. Thus the income process takes the form

yt = yt−1 + εt − γεt−1 (5.58)

5.3. PRUDENCE 101

with γ ∈ [0, 1]. Comparing (5.58) to (5.51) we see that the case of γ = 0 cor-
responds to a process with only permanent shocks, whereas γ = 1 corresponds
to a process with only transitory shocks. Under such a process we find that

εt if s = 0
(Et − Et−1 ) yt+s = (5.59)
(1 − γ)εt if s > 0
and thus
rγ
θt ∆ct = εt + (1 − γ)θt εt (5.60)
1+r
Therefore the household’s consumption response to the εt shock is a convex
combination (with weights γ and 1 − γ) of the response to a transitory and
to a permanent shock.
Example 28 Finally, for a simple AR(1) income process of the form
yt = δyt−1 + εt (5.61)
with 0 < δ < 1 we find that
T −t  s
rεt X δ
θt ∆ct = . (5.62)
1 + r s=0 1+r

Comparing this result to example 26 above we see that the change in con-
sumption in reaction to a persistent (but not permanent) shock εt falls quan-
titatively in between that induced by a purely transitory shock (which equals
rθt−1 P −t δ
s
1+r
), since Ts=0 1+r
> 1 and that of a permanent shock (which equals
1), since δ < 1.

5.3 Prudence
So far we have analyzed the SIM model in partial equilibrium, under the
assumptions that household have quadratic utility and that constraints on
borrowing were not binding. We showed that household behavior exhibit
certainty equivalence: ex ante, prior to the realizations of risk, consumption
and savings choices are identical without and with risk. We also studied how
consumption and thus saving responds to income shocks ex post.
In the next two sections we will in turn relax both key assumptions that
gave risk to certainty equivalence, and show that the relaxation of either as-
sumption can give rise to precautionary saving behavior such that an increase
102 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

in income risk induces households to save more and consume less. In this
section26 we demonstrate that households engage in precautionary saving if
they have convex marginal utility, U 000 (c) > 0. In section 5.4 it is shown that
in the presence of liquidity constraints agents may exhibit the same behavior
even under quadratic utility.27
To set the stage for this section, recall that with quadratic utility and
absent binding liquidity constraints the consumption function was given by
(5.46) and exhibited certainty equivalence. Specifically, the optimal con-
r
sumption rule was only a function of an annuity factor 1+r (adjusted by an
additional term if T is finite) and otherwise only depends on the expected
present discounted value of lifetime income Wt . From the definition of Wt
we see that only the conditional (on period t information) first moment of
future incomes Et yt+s matters for the consumption choice, but not the extent
of income risk (the conditional variance of future labor income) or higher mo-
ments of the random variables {yt+s }s≥1 . Thus a change in the riskiness of
future labor income (with unchanged mean) or any other higher moment of
the distribution of future labor income leaves the current consumption (and
thus saving) choice completely unaltered.
The key steps in deriving this result were to exploit quadratic utility
to obtain equation (5.38) as a special case of the general stochastic Euler
equation (5.35) for the SIM model, reproduced here for completeness, but
with preference shocks suppressed:

 
1+r
Et Uc (ct+1 , st+1 )


Uc (ct ) = (5.63)
1+ρ

and then to combine this Euler equation with the intertemporal budget con-
straint (5.45), which was derived from the sequential budget constraints, a
derivation that in turn required the absence of binding borrowing constraints.
In this section we will continue to use the intertemporal budget constraint,
but will investigate the implications of the intertemporal Euler equation of
utility is not quadratic (and thus marginal utility is not linear in consump-
tion, which was required for obtaining (5.38).

26
For this part the key references include Kimball (1990), Barsky, Mankiw and Zeldes
(1986), Deaton (1991) and Carroll (1997).
27
See Zeldes (1989a, 1989b) and again Deaton (1991) for the classic references.
5.3. PRUDENCE 103

5.3.1 A Simple Model and a General Result

In order to derive the result that relates the sign of the third derivative
of the utility function to the presence of precautionary saving in the most
transparent form we first consider a simple two period example, similar in
spirit to the one studied for the certainty case in section 5.1 above.28 We
assume that income in period 0 is known and equal to y0 , but period 1 income
y1 is stochastic. For convenience29 we also shall assume that ρ = r = a0 = 0.
It is convenient to decompose period 1 income into a deterministic and a
stochastic component:
y1 = ȳ1 + ỹ1 (5.64)
where ȳ1 = E0 y1 is the (conditional) expectation of y1 at period 0 and ỹ1 is
a random variable with E0 (ỹ1 ) = 0. As before we define

W0 = y0 + ȳ1 (5.65)

as expected present discounted value of lifetime labor income. Under these

assumptions the budget constrains in both periods read as

c 0 + a1 = y 0 (5.66)
c1 = a1 + ȳ1 + ỹ1 (5.67)

and substituting out a1 in the second equation yields

c1 = y0 − c0 + ȳ1 + ỹ1
= W0 − c0 + ỹ1 (5.68)

Note that c1 is a random variable that varies with the realization of the
stochastic component of period 1 income ỹ1 .
Finally define

s = W0 − c0 (5.69)
= a1 + ȳ1 = E0 (c1 )

as saving out of lifetime wealth (which equals expected consumption in period

1). This concept of saving, a simple shift of a1 by the constant ȳ1 , will be
28
The discussion of the simple model follows Barsky, Mankiw and Zeldes (1986).
29
The results go through unchanged with ρ 6= 0 and/or r 6= 0 as well as a0 6= 0, but the
algebra becomes substantially more messy.
104 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

useful later on when stating the general precautionary result. With these
definitions the stochastic Euler equation (5.63), ignoring preference shocks,
becomes
Uc (c0 ) = E0 Uc (W0 − c0 + ỹ1 ) (5.70)
Note that the only endogenous choice in this equation is the deterministic
number c0 , the consumption choice for period 0, since W0 is an exogenous
constant and ỹ1 is an exogenous random variable. We now want to analyze
how the optimal choice of c0 depends on the parameters of the model, and
especially, how it varies with the characteristics of income risk, measured by
the variance of labor income in period 1, i.e. with σy2 = V ar0 (ỹ1 ).

Intuition for the General Result

Before turning to the general result that shows under which conditions con-
sumption c0 is decreasing (and thus saving s and a1 are increasing) in income
risk σy2 we consider a special income process where this result can be derived
using simple algebra. Thus suppose that the stochastic part of income in
period 1 can only take two values:

−ε with prob. 21

ỹ1 = (5.71)
ε with prob. 12

with 0 < ε < ȳ1 . Therefore σy2 = ε2 and ε measures the amount of income
risk the household faces. We wish to determine under what condition c0 (ε)
is a strictly decreasing function. Writing out equation (5.70) and using the
assumption that ỹ1 follows a two-point distribution yields:
1
Uc (c0 ) = [Uc (W0 − c0 + ε) + Uc (W0 − c0 − ε)] . (5.72)
2
Totally differentiating (5.72) with respect to ε delivers:
    
dc0 1 dc0 dc0
Ucc (c0 ) = Ucc (W0 − c0 + ε)) − + 1 + Ucc (W0 − c0 − ε)) − −1
dε 2 dε dε
(5.73a)

and thus
1
dc0 (ε) 2
[Ucc (W0 − c0 + ε) − Ucc (W0 − c0 − ε)]
= (5.74)
dε Ucc (c0 ) + 12 [Ucc (W0 − c0 + ε)) + Ucc (W0 − c0 − ε))]
5.3. PRUDENCE 105

The denominator of this expression is unambiguously negative (since we as-

sume that U is strictly concave). The nominator is positive if and only if

Ucc (W0 − c0 + ε)) − Ucc (W0 − c0 − ε)) (5.75)

is positive.
But this is true for arbitrary ε > 0 if and only if Uccc (c) > 0. Hence
consumption in period 0 strictly declines in reaction to a marginal increase in
period 1 income risk σy2 if and only if the third derivative of the utility function
is positive. Using equation (5.66) and (5.69) it then follows immediately that
a1 and s are strictly increasing functions of σy2 for all levels of σy2 if and only
if Uccc (c) > 0 for all c. Thus a sufficient (and necessary) condition for the
household to exhibit precautionary saving behavior (i.e. to increase saving
in response to increased income risk) in the absence of binding borrowing
constraints is strictly convex marginal utility Uc . Of course it is a simple
corollary of this result that under the assumption Uccc (c) > 0 households
deviate from certainty equivalence behavior; simply realize that c0 (ε = 0) >
c0 (ε > 0).

An Application
The model above has been used by Barsky, Mankiw and Zeldes (1986) to
show that the presence of income risk alone can invalidate the Ricardian
equivalence hypothesis. Recall that this hypothesis states that for a given
process of government spending a change in the timing of taxation does not
affect household behavior and macroeconomic aggregates. Thus if Ricardian
equivalence holds it should not matter whether current government expendi-
tures are financed via current taxes of by government debt that is redeemed
later (using future taxes).
A simple example is sufficient to make this argument. Suppose that
government spending satisfies G0 = G1 = 0 and that the benchmark policy
is one of no taxation in either period (and thus the situation analyzed in the
previous subsection). Now consider the Ricardian experiment of lowering
taxes in period 0, say, in order to stimulate consumption in the economy
in a Keynesian-style fiscal expansion. Thus in period 0 households receive
a lump-sum subsidy of size t, and in period 1 have to pay positive taxes in
order to finance the repayment of the debt the government incurred for paying
the transfers. Let τ denote the tax rate on labor income in period 1. We
assume that the government deals with a continuum population of measure
106 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

1, each member of which faces the income process specified in (5.71). Thus
government tax revenues from the income tax defined by τ are given by30

τ E0 (y1 ) = τ ȳ1

and the size of the period 0 transfers thus has to satisfy (recall that we
assumed r = 0)
t = τ ȳ1
in order for the intertemporal government budget constraint to hold.
Therefore we note that, after the policy innovation, the expected present
discounted value of lifetime labor income for each household equals

W0 (t, τ ) = y0 +t+(1−τ )E0 (ȳ1 + ỹ1 ) = y0 +t+(1−τ )ȳ1 = y0 +ȳ1 = W0 (t = τ = 0),

and the fiscal policy reform leaves the expected present discounted value of
lifetime income unchanged. If households were certainty equivalence con-
sumers, their consumption behavior would therefore be unaffected by the
Ricardian tax policy experiment. We note, however, that after-tax income
in the second period now equals (1 − τ )(ȳ1 + ỹ1 ), with associated variance
(1 − τ )2 σy2 < σy2 . Therefore, from the result in the previous section (as along
as the size of τ, t is small), the household consumes more and saves less in re-
sponse to the small temporary, debt-financed tax cut. Ricardian equivalence
fails, despite the fact that the household behaves fully rational and forward
looking, faces no borrowing constraints, and the tax is (or at least looks
like) a lump-sum tax. The tax change does affect the extent of income risk,
however, and thus induces a decline in precautionary saving. With Barsky,
Mankiw and Zeldes’ (1986) words, the household is a “Ricardian consumer
with Keynesian propensities” to consume out of current income.

The General Theory

Kimball (1990) has defined the term “prudence” to mean the “propensity
to prepare and forearm oneself in the face of uncertainty” and hence the
intensity of a precautionary saving motive induced by the preferences of
30
We invoke a law of large numbers here that guarantees that the deterministic tax
revenues from the entire population of identical households equals the expected tax revenue
from each individual household. We will discuss potenial conerns with the applicability of
the law of large numbers in chapter 7 and refer the reader to Feldman and Gilles (1985)
and Judd (1985) for the mathematical details of the problem and one potential solution.
5.3. PRUDENCE 107

households (although the concept of prudence applies to other decisions in

uncertain environments as well, and not just to saving behavior). From this
definition it is clear that the term prudence characterizes preferences, whereas
precautionary saving characterizes behavior: prudence leads to precautionary
saving, but both concepts should be kept separately.
Also note that, as should be clear already from the results in the previous
sections, that prudence and risk aversion are very distinct concepts (although
they both describe preferences). Risk aversion is controlled by the concav-
ity of the utility function (Ucc (c) < 0) whereas prudence (the precautionary
savings motive) is controlled by the convexity of the marginal utility func-
tion (Uccc (c) > 0). This distinction is most transparent for households with
quadratic utility: these households dislike the additional consumption risk
induced by additional income risk (in the SIM model) and would be willing
to pay a positive insurance premium to get rid off the additional risk, but
what they won’t do is to increase their savings in order to hedge against that
additional income risk (as we have demonstrated in section 5.2.2 that house-
holds with quadratic utility, absent binding borrowing constraints display
certainty equivalence behavior).
Now turning to the general theory of prudence and precautionary saving,
the result derived in section 5.3.1 merely demonstrates, via an example, that
Uccc (c) > 0 implies precautionary saving behavior, but it makes no statement
about the magnitude of the effect of income risk on optimal household con-
sumption and saving choices. The ingenious insight of Kimball (1990) is that
a quantitatively meaningful theory of precautionary saving based on prefer-
ences that exhibit prudence can be derived as a mathematically straightfor-
ward extension of the quantitative theory of risk aversion developed by Pratt
(1964) and Arrow (1965).
Pratt (1964) shows that the (equivalent31 ) risk premium θ(c, ỹ1 ), for a
small, zero mean absolute risk (that is, a gamble that adds or takes away the
random (zero mean) quantity ỹ1 for c), which is defined implicitly from the

31
The compensating risk premium θ∗ (c, ỹ1 ) is defined as

E0 U (c + ỹ1 + θ∗ (c, ỹ1 )) = U (c). (5.76)

108 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

equation32
E0 U (c + ỹ1 ) = U (c − θ(c, ỹ1 )) (5.77)
is explicitly characterized by
1 Ucc (c)
θ(c, ỹ1 ) = − σy2 + o(σy2 ), (5.78)
2 Uc
where o(σy2 ) is a term that converges to zero faster than does σy2 . This result
justifies the coefficient of absolute risk aversion
Ucc (c)
r(c) = − (5.79)
Uc (c)
as an appropriate measure to quantify the willingness of a household to pay
to avoid a small risk. Correspondingly, the coefficient of relative risk aversion
cUcc (c)
σ(c) = − (5.80)
Uc (c)
measures the willingness to pay to avoid small relative gambles in which
percentages of c are at risk. Of course, for the special cases of CRRA utility
1−σ
U (c) = c1−σ and CARA utility U (c) = − γ1 e−γc these entities are given,
correspondingly, by
σ
r(c) = and σ(c) = σ
c
r(c) = γ and σ(c) = γc.

The key insight of Kimball (1990), building on earlier work by Leland

(1968), Sandmo (1970) and Rothschild and Stiglitz (1970) was to notice that
if one simply adds one derivative to all expressions obtained from equation
(5.77) one immediately obtains a quantitative theory of precautionary saving.
To see this, recall from equation (5.70) that the optimal consumption-saving
choice (c0 , a1 ), or equivalently, (c0 , s) is determined by the Euler equation

E0 Uc (W0 − c0 + ỹ1 ) = Uc (c0 )

or, by the definition of s from equation (5.69):

E0 Uc (s + ỹ1 ) = Uc (W0 − s).

32
In general this premium is a function of the starting point c and the characteristics of
the random variable ỹ1 . The notations θ(c, ỹ1 ) and θ∗ (c, ỹ1 ) are meant to convey this.
5.3. PRUDENCE 109

Now define the equivalent precautionary saving premium ψ(s, ỹ1 ) implicitly
by the equation
E0 Uc (s + ỹ1 ) = Uc (s − ψ(s, ỹ1 )) (5.81)
Comparing equations (5.77) and (5.81) it follows immediately that all results
(and especially the result summarized in equation (5.78) above) derived by
Pratt (1964) for the equivalent risk premium θ(c, ỹ1 ) immediately apply to
the equivalent precautionary saving premium ψ(s, ỹ1 ), with all expressions
in the original Pratt (1964) results replaced by one higher derivative.
Similarly, Kimball (1990) defines the compensating precautionary saving
premium by ψ ∗ (s, ỹ1 ) as

E0 Uc (s + ỹ1 + ψ ∗ (s, ỹ1 )) = Uc (s), (5.82)

that is, the amount of extra saving (starting from the saving level s without
risk) the household would find optimal in the presence of risk, relative to
the risk-free case. Since the compensating precautionary saving premium is
easier to interpret (and, at least locally, equal to the equivalent premium)
we will focus on this measure of precautionary saving from now on, referring
the reader to Kimball (1990) for a complete treatment. Again comparing
equations (5.82) and (5.76) allows for a straightforward application of the
results in Pratt (1964) to the precautionary saving premium. Consequently,
we have the following

Proposition 29 (Kimball 1990). The compensating precautionary saving

premium is given by

1 Uccc (s)
ψ ∗ (s, ỹ1 ) = − σy2 + o(σy2 )
2 Ucc (s)

The residual term o(σy2 ) satisfies

o(σy2 )
lim = 0. (5.83)
σy →0 σy2
2

This result shows that the compensating precautionary saving premium

is positive if and only if Uccc (s) > 0, since Ucc < 0 by assumption. Also
note that Uccc is evaluated at s = E0 (c1 ). Perhaps more importantly, the
proposition states that the magnitude of the precautionary saving premium
110 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

is exclusively determined, to a first order, by the amount of income risk σy2

and the index of absolute prudence

Uccc (s)
p(s) = −
Ucc (s)

which justifies this index as a quantitative measure of prudence (and thus

as an index of the intensity of the precautionary savings motive). Also note
that for the special case of CRRA and CARA utility functions we find that:
σ+1
p(s) = (5.84)
c
p(s) = γ (5.85)

and in both cases risk aversion and prudence is controlled by a single param-
eter (σ and γ, respectively).33
Now that we have fully characterized the precautionary saving premium
through proposition 29 we want to further interpret what this entity ψ ∗ (s, ỹ1 )
actually measures, in terms of observable behavior. To do so, we first recall
that the Euler equation for the case without risk reads as

Uc (W0 − s) = Uc (s) (5.87)

Combining equations (5.82) and (5.87) yields

EUc (s + ỹ1 + ψ ∗ (s, ỹ1 )) = Uc (W0 − s) (5.88)

For future reference define c0 (W0 , ỹ1 ) as the optimal consumption choice as-
sociated with lifetime wealth W0 and income in the second period determined
33
As for risk aversion and the risk premium, one can also define the index of relative
prudence as
cUccc (c)
pr(c) = − (5.86)
Ucc (c)
and establish the same results as above if income risk is proportional to wealth W0 . For
CRRA and CARA utility the coefficient of relative prudence is given, respectively, by

pr(c) = σ+1
pr(c) = γc.
5.3. PRUDENCE 111

by the random variable ỹ1 , and as s(W0 , ỹ1 ) the optimal saving (out of life-
time wealth) function. The corresponding optimal policy functions in the
absence of income risk are defined as c0 (W0 , 0) and s(W0 , 0), which are of
course related by
c0 (W0 , 0) + s(W0 , 0) = W0 .
Furthermore define the lifetime wealth W0 needed to make a given consump-
tion c0 optimal as W0 (c0 , ỹ1 ). Note that W0 (c, ỹ1 ) is defined as the quantity
x that satisfies
W0 (c0 (x, ỹ1 ), ỹ1 ) = x.
It is then easy to show34 that ψ ∗ (s, ỹ1 ) equals the additional wealth required
to keep consumption at a given level c0 if a small income risk is introduced,
that is
ψ ∗ (W0 (c0 , 0) − c0 , ỹ1 ) = W0 (c0 , ỹ1 ) − W0 (c0 , 0) (5.89)
In other words, ψ ∗ (s, ỹ1 ) can also be interpreted as the magnitude of the
rightward shift of the consumption function c0 (W0 , 0) with wealth W0 on the
x-axis, at a particular level of consumption c0 , as income risk is turned on.
We now collect further important results about precautionary saving con-
tained in Kimball (1990).

Remark 30 The previous results were local in the sense that we considered
a specific value for c0 or s, and small risks. However, Kimball (1990) also
34
By the definition of W0 (c0 , ỹ1 ) we have

Uc (c0 ) = E0 Uc (W0 (c0 , ỹ1 ) − c0 + ỹ1 ) = Uc (W0 (c0 , 0) − c0 ).

where the last equality follows from equation (5.88). By definition of ψ ∗ (W0 (c0 , 0) − c0 , ỹ1 )

E0 Uc (W0 (c0 , 0) − c0 + ỹ1 + ψ ∗ (W0 (c0 , 0) − c0 , ỹ1 )) = Uc (W0 (c0 , 0) − c0 ).

Equating the left hand sides of the previous two equations yields

E0 Uc (W0 (c0 , 0) − c0 + ỹ1 + ψ ∗ (W0 (c0 , 0) − c0 , ỹ1 ))

= E0 Uc (W0 (c0 , ỹ1 ) − c0 + ỹ1 ).

But since Ucc < 0 for all c, this last equation implies that, since both W0 (c0 , 0) +
ψ ∗ (W0 (c0 , 0) − c0 , ỹ1 ) and W0 (c0 , ỹ1 ) are deterministic numbers:

W0 (c0 , 0) + ψ ∗ (W0 (c0 , 0) − c0 , ỹ1 ) = W0 (c0 , ỹ1 ).

112 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

provides a global when he shows that, for any two utility functions U, V, if
Uccc (s) Vccc (s)
− >−
Ucc (s) Vcc (s)
for all s then
ψU∗ (s, ỹ1 ) > ψV∗ (s, ỹ1 )
for all s and all nondegenerate random variables ỹ1 . This result is particularly
useful for two utility functions U, V that yield the same consumption and
savings functions in the absence of risk (e.g. if both U and V are of CRRA
form with σU > σV ; recall that ρ = r = 0). Then this result implies that
for the same introduction of risk ỹ1 households with utility function U shift
the consumption function to the right uniformly more than households with
V when income risk is introduced.

Remark 31 Often the situation of interest is one where the initial choice
situation already contains some risk, and then additional risk is added, and
the question arises what happens to consumption and saving in response to
this additional risk. As long as the new risk is independently distributed to
the already present risk, the results go through unchanged. However, if the
additional risk is, for example, a mean-preserving spread of the old risk (and
thus the additional risk is not independent of the initial risk), then the results
stated above do not carry over.

Remark 32 Finally it is possible to determine how, at a given initial level of

consumption c0 = c0 (W0 , 0) the propensity to consume out of wealth ∂c0∂W
(W0 ,0)
0
is affected by the introduction of income risk. From equation (5.89) we find

W0 (c0 , ỹ1 ) − W0 (c0 , 0) = ψ ∗ (s(c0 , 0), ỹ1 ).

where s(c0 , 0) ≡ W0 (c0 , 0) − c0 . Suppose all terms are differentiable35 in c0 ,

then
∂W0 (c0 , ỹ1 ) ∂W0 (c0 , 0) ∂ψ ∗ (s(c0 , 0), ỹ1 ) ∂s(c0 , 0)
− =
∂c0 ∂c0 ∂s(c0 , 0) ∂c0
Because of time separability consumption c0 in period zero and consumption
in period one (which is equal to s(c0 , 0) without risk) are both normal goods
and thus increase together as W0 increases. Thus ∂s(c 0 ,0)
∂c0
> 0 and the sign of
35
See footnote 19 of Kimball (1990) how to proceed if that assumption is violated.
5.3. PRUDENCE 113
∗
the left hand side is determined exclusively by the sign of ∂ψ ∂s(c
(s(c0 ,0),ỹ1 )
0 ,0)
, which
from proposition 29 is exclusively determined by the sign of
 
∂ − UUccc (s)
cc (s)
p0 (s) =
∂s

that is, the sign of ∂W0∂c

(c0 ,ỹ1 )
0
− ∂W∂c
0 (c0 ,0)
0
depends solely on whether absolute
prudence is strictly increasing, constant (such as for the CARA utility func-
tion) or strictly decreasing (such as for the CRRA utility function) in s.
Since the inverse of ∂W∂c 0 (c0 ,.)
0
is the marginal propensity to consume out of
lifetime wealth, we find that the marginal propensity to consume out of life-
time wealth in period zero (that is, the slope of the consumption function), at
a given consumption level c0 is strictly declining (increasing) with the intro-
duction of income risk if absolute prudence is strictly increasing (decreasing)
at the s associated with c0 .

Example 33 Assume that U (c) = log(c), which has positive and strictly
decreasing absolute prudence, and also assume that ỹ1 follows the process
from subsection 5.3.1. We can then determine the consumption function in
closed form as:
3 1
1
c0 (W0 ; ε) = w − 8ε2 + (W0 )2 2
4 4
where c0 (W0 ; ε) is well-defined for wealth levels W0 ≥ ε. We note that
1
c0 (W0 ; ε) < c0 (W0 ; ε = 0) = W0 for all W0 ≥ ε
2
and that for all ε > 0 the function c0 (W0 ; ε) is strictly increasing and strictly
concave, with
lim (c0 (W0 ; ε) − c0 (W0 ; ε = 0)) = 0
W0 →∞

for all ε > 0. Furthermore

1 ∂c0 (W0 ; ε = 0) ∂c0 (W0 ; ε) 3 1

= < = −  12
2 ∂W0 ∂W0 4 
ε 2

1
16 w
+ 16

and
∂c0 (W0 ; ε) 1 ∂c0 (W0 ; ε = 0)
lim = = .
w→∞ ∂W0 2 ∂W0
114 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Figure ?? shows the consumption function (for the case ε = 0.2) and the pre-
cautionary saving premium at c0 = 0.2. As predicted by the previous remark
the slope of the consumption function with risk is larger than the slope of the
consumption function without risk, for all wealth levels W0 .

Add Kimball.eps picture

5.3.2 A Parametric Example for the General Model

Now let us consider the general model with many periods. In general nothing
analytical can be said about the optimal consumption profile, and how it
responds to income shock. That is, a fully analytical counterpart to the case
with quadratic utility carried out in section 5.2.2 is not available if marginal
utility is not linear and thus households deviate from certainty equivalence
behavior. However, in this section we show that, using CRRA utility and
an approximation (whose accuracy can then evaluated numerically) we can
derive a closed form expression for consumption growth (log-changes) and
thus the consumption profile. In the next section we present a similar result
(but without having to resort to approximations) for changes in consumption
when households have CARA utility.
Now let us again start with the stochastic Euler equation from equation
(5.35):  
t 1+r
Et Uc (ct+1 , st+1 )


Uc (ct , s ) = (5.90)
1+ρ
Now we assume separability between consumption and preference shocks as
well as CRRA utility, and the equation becomes
  " −σ #
1+r ct+1
Et = 1 (5.91)
1+ρ ct
"  c −σ  #
1+r
ln ( 1+ρ ) t+1
ct
Et e = 1

Et e−σ ln(ct+1 ) eσ ln(ct )+ln(1+r)−ln(1+ρ)

= 1
eσ ln(ct )+ln(1+r)−ln(1+ρ) Et e−σ ln(ct+1 ) = 1.


(5.92)

We derived equation (5.91) from (5.92) since it will be useful in the analysis
below. We note that since ct is known when expectations Et are taken, the
5.3. PRUDENCE 115
  
−σ  −σ
1 1
term Et ct
= ct
does not create any problem for obtaining a
closed-form solution. However, the term Et (ct+1 )−σ does, since it is the
 

expectation of a term that is nonlinear in ct+1 .

Thus, in order to arrive at a regression equation similar to the one Hall
(1978) derives and estimates with quadratic utility (see equation (5.38)) most
of the empirical literature proceeds by taking a first order Taylor approxima-
tion of the function f (ct+1 ) = c−σ
t+1 around the point ct+1 = ct in (5.91). This
yields
c−σ −σ −σ−1
t+1 ≈ ct − σ(ct+1 − ct )ct
 
−σ ct+1 − ct
= ct 1−σ . (5.93)
ct
Using this in (5.91) delivers
   
1+r ct+1 − ct
Et 1 − σ =1
1+ρ ct
and thus  
ct+1 − ct 1 r−ρ
Et = . (5.94)
ct σ 1+r
which relates expected consumption growth to the gap between the interest
rate and the subjective time discount rate, with the size of the response
determined by the intertemporal elasticity of substitution σ1 . Although one
could easily bring (5.94) to the data, the equation most frequently estimated
is derived from using the additional approximations ct+1ct−ct ≈ ∆ ln ct+1 and
1+r
1 − 1+ρ = r−ρ
1+r
≈ ln(1 + r) − ln(1 + ρ) in (5.94) to obtain
1
Et ∆ ln ct+1 = [ln(1 + r) − ln(1 + ρ)] . (5.95)
σ
This equation can then be used to estimate, via OLS regression the intertem-
poral elasticity of substitution σ1 , and to test the hypothesis implied by the
model that any variable Xt in the information set of the household at time
t should not help predict consumption growth. Thus the literature on Euler
equation estimation (see e.g. Attanasio and Weber (1993, 1995) or Attana-
sio and Browning (1995) has used data on household consumption and real
interest rates to run the regression
∆ ln ct+1 = α0 + α1 ln(1 + rt+1 ) + βXt + εt+1
116 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

to obtain estimates of α̂1 = bσ1 and to test whether β̂ = 0, where the variables
Xt often include household income in period t.
Although estimating the approximated stochastic Euler equation has been
the predominant approach to determine the parameters and test the main
implication of the partial equilibrium SIM model, critiques of this approach
(see e.g. Carroll, 2001) question whether the approximation (5.93) used to
derive equation (5.95) is accurate. Related to this point, by using this ap-
proximation equation (5.95) ignores the importance of precautionary saving
(despite the fact that with CRRA utility the household is prudent and has
preference-induced precautionary saving motive).
To see this point, now assume that ln(ct+1 ) is normally distributed36 with
mean µ = Et ln(ct+1 ) and variance σc2 . Although it is in general not desirable
to make distributional assumptions on endogenous variables (such as ct+1 ),
we do it here to demonstrate the potential problem with the Euler equation
estimation approach, in an analytically tractable way. With this assumption
we find that

Et (c−σ −σ ln(ct+1 )


t+1 ) = Et e
(u−µ)2
Z ∞ − 2
−σu e
2σc
= e √ du
−∞ 2πσc
2

−
[u−(µ−σσc2 )]
2 )2 −µ2 Z ∞ 2
(µ−σσc
e 2σc
= e 2σc2
√ du
−∞ 2πσc
2 )2 −µ2
(µ−σσc
2
= e 2σc

1 2 2
= e 2 σ σc −µσ
1 2 2
= e 2 σ σc −σEt ln(ct+1 ) (5.96)

Inserting this expression into equation (5.92) yields

1 2 σ2
e−σEt ∆ ln(ct+1 )+ln(1+r)−ln(1+ρ)+ 2 σ c = 1. (5.97)
36
This in turn requires appropriate assumptions on the underlying stochastic income
process.For the two period model above, this requires that the random variable

Z = ln(κ + ỹ1 )

is normally distributed, where κ = w − c0 is a constant.

5.3. PRUDENCE 117

Finally, taking logs on both sides and rearranging gives

1 1
Et ∆ ln(ct+1 ) = [ln(1 + r) − ln(1 + ρ] + σσc2 (5.98)
σ 2
Remember that by linearizing the Euler equation became (see (5.95)):
1
Et ∆ ln ct+1 = [ln(1 + r) − ln(1 + ρ)] (5.99)
σ
and thus under the approximation expected consumption growth is chosen
in exactly the same way as in the absence of any risk (compare equation
(5.99) to the Euler equation in the absence of risk, equation (5.24)). From
equations (5.98) and (5.99) we observe that:
1. The “correct” consumption allocation does not obey certainty equiva-
lence. As CRRA utility exhibits prudence and thus induces a precau-
tionary savings motive to households, from (5.98) we see that future
consumption risk, captured in the last term of (5.98) tilts the consump-
tion profile upward in expectation, relative to the consumption growth
rule (5.99) derived under the linearization (which is in turn identical,
in expectation, to the certainty case). Consumption growth is higher
in the presence of risk since households find it optimal to postpone
consumption for precautionary motives.
2. The degree to which consumption is postponed (i.e. the degree to
which there is precautionary saving) in response to future consump-
tion (income) risk is determined by the parameter σ that controls the
magnitude of prudence for the CRRA utility function, as shown above
for the two period model.
3. All variables Xt that, at period t, help to predict the variability of fu-
ture consumption σc2 , will help to predict expected consumption growth,
according to the exact Euler equation 5.98 (but not according to the ap-
proximated Euler equation (5.99)). For example, agents with a higher
level of current assets or income may have lower future consumption
variability and thus, according to equation (5.98), lower consumption
growth. This point was made, among others, by Carroll (1992). Thus
it may be flawed to run the regression
ln ct+1 − ln ct = α1 + βXt + εt (5.100)
where Xt may be current wealth, and interpret a statistically significant
estimate of β as evidence against the SIM model with CRRA utility.
118 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

5.4 Liquidity Constraints

So far we have assumed that the household can borrow up to some arbitrarily
large amount, up to the no-Ponzi scheme condition which was assumed to
be generous enough never to be binding. In this section we investigate how
the analysis from the previous section (that mainly exploited the stochas-
tic Euler equation) has to be adjusted if households instead face potentially
binding borrowing constraints. These constraints seem empirically plausi-
ble and formal econometric tests seem to indicate not only their presence,
but also their effect on consumption allocations (see Zeldes (1989) and the
literature originating from that study).
Before investigating the impact of borrowing constraints on the Euler
equation, we first want to demonstrate under what condition such a con-
straint is likely binding. To do so we need a sharp characterization of the
optimal stochastic process of asset holdings. Such a characterization is typ-
ically not available for general utility functions, but with quadratic utility
and the resulting certainty equivalent behavior. We recall that for this case,
the optimal consumption rule of an infinitely lived consumer is given by
" ∞
#
r X yt+s
ct = Et + at . (5.101)
1+r s=0
(1 + r)s
From the budget constraint we find that
at+1 = (1 + r)(yt − ct ) + (1 + r)at (5.102)
and plugging in optimal consumption ct yields:
" ∞
#
X yt+s
at+1 = (1 + r)yt + (1 + r)at − r Et + at
s=0
(1 + r)s
∞
X ryt+s
= at + yt − Et
s=1
(1 + r)s
∞  
X yt+s yt+s
= at + yt − Et s−1 − (5.103)
s=1
(1 + r) (1 + r)s
as the optimal asset allocation decision. Thus the realized change in asset
holdings is given by
∞
X Et ∆yt+s
∆at+1 = at+1 − at = −
s=1
(1 + r)s−1
5.4. LIQUIDITY CONSTRAINTS 119

The exact properties of wealth {at+1 } depend on the stochastic income

process. For example, if the income process is of the form (5.49)-(5.50) with
transitory and permanent income shocks (ut and vt respectively), then

−ut if s = 1
Et ∆yt+s =
0 if s > 1

and thus
∆at+1 = ut (5.104)

i.e. assets follow a random walk. Permanent income shocks vt are fully ab-
sorbed by consumption and thus do not induce changes in the asset position
of the household. Transitory income shocks in contrast trigger a consump-
rut ut
tion response of only 1+r , and thus the remaining part of the shock 1+r is
at+1
absorbed by 1+r and thus assets change by the whole transitory shock ut .
This discussion shows that a certainty equivalent consumer will violate
any fixed borrowing limit at+1 ≥ −Ā with probability 1 (since assets follow
a random walk). The asset process might not may not violate the no Ponzi
condition if we specify it carefully37 , but calls into question the plausibility
of the assumption that liquidity constraints are never binding, a maintained
assumption so far.

5.4.1 The Euler Equation with Liquidity Constraints

Now let us assume that there exist borrowing constraints, and for simplicity
let us follow Deaton (1991), Schechtman (1976), Aiyagari (1994) and many
others and assume that agents cannot borrow at all38 , i.e. face the constraint
at+1 (st ) ≥ 0 for all st . These constraints may or may not be binding, depend-
ing on the realizations of the labor income shock, but we have to take these
constraints into account explicitly when deriving the stochastic Euler equa-
tion. Let us attach Lagrange multiplier µt (st ) to the borrowing constraint
at+1 (st ) ≥ 0 at event history st . The first order conditions with respect to
consumption, (5.30) and (5.31) remain unchanged. The first order condition
37
For quadratic utility the natural borrowing constraint, assumed so far, is still too tight
since negative consumption is permissible with this utility specification.
38
If there is a positive probability of zero income in every period then the constraint
at+1 (st ) ≥ 0 is the natural borrowing limit.
120 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

with respect to assets at+1 (st ) now becomes

λt (st ) X
− µt (st ) = λt+1 (st+1 ) (5.105)
1+r t+1 t s |s

with complementary slackness conditions

at+1 (st ), µt (st ) ≥ 0
at+1 (st )µt (st ) = 0 (5.106)
Combining the first order condition yields
µt (st )(1 + r) X
Uc (ct (st )) − = (1 + r)β πt+1 (st+1 |st )Uc (ct+1 (st+1 ), st+1 )
β t πt (st ) t+1 t
s |s
(5.107)
or in short
1+r
Uc (ct ) ≥ Et Uc (ct+1 ) (5.108)
1+ρ
1+r
= Et Uc (ct+1 ) if at+1 > 0
1+ρ
For the analysis below it is useful to further rewrite the Euler equation
(5.108). Note from the budget constraint that
at+1
c t = y t + at −
1+r
≤ y t + at (5.109)
where the last inequality follows from the liquidity constraint at+1 ≥ 0. Thus
either at+1 = 0 and therefore ct = yt + at and
1+r
Uc (yt + at ) = Uc (ct ) ≥ Et Uc (ct+1 )
1+ρ
or at+1 > 0, therefore ct < at + yt and thus (using strict concavity of the
utility function)
1+r
Uc (yt + at ) < Uc (ct ) = Et Uc (ct+1 )
1+ρ
Hence the Euler equation can be compactly written as
 
1+r
Uc (ct ) = max Uc (yt + at ), Et Uc (ct+1 ) (5.110)
1+ρ
5.4. LIQUIDITY CONSTRAINTS 121

5.4.2 Precautionary Saving Due to Liquidity Constraints

In section 5.2.2 we showed that in the absence of binding liquidity constraints
households with quadratic utility obey certainty equivalence and thus do not
engage in precautionary saving behavior. Now we demonstrate that the
presence of potentially binding liquidity constraints induces precautionary
saving behavior even with quadratic utility.
With quadratic preferences equation (5.110) becomes
 
1+r
−(ct − c̄) = max −(yt + at − c̄), − (Et ct+1 − c̄) (5.111)
1+ρ
To simplify the algebra assume ρ = r. Then (5.111) can be rewritten as39
ct = min{yt + at , Et ct+1 } (5.112)
= min{yt + at , Et min{yt+1 + at+1 , Et+1 ct+2 }} (5.113)
From equation (5.113) we make the following observations:
1. If the certainty equivalence solution (5.101) (or its finite horizon coun-
terpart) has associated asset holdings (5.103) that satisfy at+1 ≥ 0
with probability 1, then it is the optimal consumption allocation even
in the presence of borrowing constraints. Whether or not this is the
case obviously depends crucially on the stochastic income process. If
the income process takes the form (5.49)-(5.50) and thus asset holdings
follow a random walk, then the constraint at+1 ≥ 0 is violated with
probability one.
2. In the absence of borrowing constraints we know from our discussion
of certainty equivalence that
ct = Et (ct+s )
for all s > 0. Suppose now there exists an s > 0 (e.g. s = 1) such that
for some realization of the income shock yt+s (with positive probability)
the household is borrowing constrained, and thus yt+s +at+s < Et+s ct+s .
Then equation (5.112) implies that (using the law of iterated expecta-
tions) ct < Et (ct+s ). Thus, even if the liquidity constraint is not binding
in period t, future binding constraints affect the current consumption
choice ct .
39
Remember that max{−x, −y} = min{x, y}.
122 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

3. The previous argument also implies that any change in the environ-
ment that affects the incidence of future binding borrowing constraints
affects current consumption. Importantly, suppose the variance of fu-
ture income increases (say, for yt+1 ), making lower realizations of yt+1
possible or more likely. If consequently the set of yt+1 values for which
the borrowing constraint binds becomes larger, then

Et min{yt+1 + at+1 , Et ct+2 } (5.114)

declines and so does ct , since in more instances the minimum is the

first of the two objects in equation (5.114). Saving increases in reac-
tion to increases in future income risk, because households, afraid of
future contingencies of low consumption (agents with quadratic util-
ity are risk-averse) and aware of their inability to smooth low income
shocks via borrowing, increase their precautionary savings. Note that
we obtain precautionary savings behavior without prudence (i.e. with-
out a precautionary saving motive induced by preference), and purely
from the existence of liquidity constraints (and risk aversion). Hence
when observing increases in saving in response to increased income
risk, this may have a preference-based explanation (agents are pru-
dent: Uccc > 0) or an (incomplete) financial market-based interpreta-
tion (credit markets prevent or limit borrowing).

5.4.3 Empirical Tests of Liquidity Constraints

[TBC Discuss the empirical literature on liquidity constraints]
This Euler equation, which sometimes holds with inequality, depending
on whether the liquidity constraint is binding, is the basis for empirical tests
for liquidity constraints. In particular, Zeldes (1989) subdivides the sample
of households into two groups, depending on their current wealth positions,
with one group composed of households whose liquidity constraint is most
likely not binding (high wealth households) and the other group composed of
households whose constraint is most likely binding (low wealth households).
The empirical test then consists of running a regression like Hall (1978), but
for both groups separately (and with micro data rather than macro data)
and ask whether current income helps forecast future consumption growth.
for both groups As evidence for the presence of liquidity constraints would
be interpreted as a finding which shows that current income helps predict
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 123

consumption growth for the low-wealth group, but not for the high-wealth
group. Zeldes (broadly) finds such evidence and concludes that borrowing
constraints are important in shaping consumption choices, at least for the
currently wealth-poor. In evaluating his result, however, you should keep in
mind that we argued above that with preferences that exhibit prudence, cur-
rent income may help predict consumption growth, if it contains information
about the extent of future income uncertainty. While it is not entirely obvi-
ous why this should be the case for low-wealth households and not for high-
wealth households, note that wealth accumulation is an endogenous choice,
so the low wealth-group sample is not simply a random sample. Therefore it
is not inconceivable that low-wealth households have income processes with
very different stochastic properties that high-wealth households, including
predictability of future income risk by current income. [TBC]

5.5 Prudence and Liquidity Constraints: The-

ory
Finally we will discuss optimal consumption choices with both a precaution-
ary saving motive and binding borrowing constraints. As we will see in the
next section, most of the analysis relies on numerical approximations of the
consumption or the savings function, as analytical solutions for this problem,
in contrast to the certainty equivalence case, are usually not available. How-
ever, the theoretical literature has provided qualitative characterizations of
the optimal consumption-savings choices, under varying assumptions about
the stochastic income process and the relative size of the interest rate r the
and time discount rate ρ. As it will turn out, the theoretical results depend
crucially on the assumptions about the relative magnitudes of r and ρ; this
in turn have important implications for the relationship between r and ρ in
the general equilibrium models in the next chapter, in which the interest rate
is determined endogenously. The important references from this literature
include Schechtman (1976), Schechtman and Escudero (1977), Yaari (1977),
Sotomayor (1984), Clarida (1987), Caballero (1990, 1991), Deaton (1991),
Huggett (1993) and Chamberlain and Wilson (2000), Carroll and Kimball
(1996, 2001).
In the class of SIM models we study consumption-savings decisions are
determined by two main motives. First, the degree of impatience ρ of
124 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

households, relative to the market interest rate, determines whether, ceteris

paribus, consumption is rising, falling and constant over time. Second, the
precautionary savings motive calls for the postponement of consumption in
favor of saving, and thus, again ceteris paribus, an upward sloping consump-
tion profile. Below we will show, roughly speaking, that with ρ ≤ r the
(im)patience motive and the precautionary savings motive reinforce each
other and consumption as well as assets rise over time without bounds, at
least as long as households are infinitely lived.40 In contrast, if ρ > r then
the precautionary and the impatience motives compete, and under appro-
priate conditions asset holdings and consumption remain bounded even if
the household lives forever, and consumption and asset possess an ergodic
distribution. Thus we need to consider the cases ρ > r, ρ = r and ρ < r
separately. We briefly discuss the situation where the planning horizon of the
household is finite and thus follows a meaningful life cycle. However, sharp
theoretical results are only available for an infinite planning horizon when
households face a time (age) invariant income process.

5.5.1 Finite Lifetime

If T is finite, obviously the stochastic processes for consumption and assets
ct and at+1 remain bounded, independent of the relationship between r and
ρ (recall that the income process was assumed to be bounded).41 The life
cycle profile of consumption is determined by the relative strength of the
impatience and the precautionary savings motive. Attanasio et al. (1999)
and Gourinchas and Parker (2003), and many others since, have shown that
as long as the interest rate and the time discount factor satisfy r < ρ and
are of “appropriate” size, then for realistic income processes (and thus re-
alistic income risk) estimated from micro data such as the PSID, early in
the life cycle remaining lifetime income risk is high and thus the precau-
40
This result will also imply that, since aggregate asset demand is infinite with ρ < r in
the long, no steady state general equilibrium (to be defined in chapter 7) will exist with
r > ρ.
41
This also implies that aggregate asset demand is finite even if r > ρ and thus this case
cannot be ruled out to occur (and does occur frequently in applications) in general equi-
librium life cycle models. General equilibrium models with many overlapping generations,
each of which faces a income fluctuation problem with finite horizon, have become popular
tools to analyze policy reforms, from social security reform to fundamental tax reform.
See e.g. Conesa and Krueger (1999, 2006) and Conesa et al. (2009) for representative
examples.
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 125

tionary saving motive dominates and consumption initially grows. Later in

life the consumer has accumulated an asset buffer against that risk; in ad-
dition remaining lifetime income risk declines towards retirement. Thus the
impatience motive starts to dominate and consumption declines (especially
if mortality risk that increases the effective time rate is factored in). Com-
bining both observations delivers a model-generate (expected) consumption
profile that is hump-shaped as in the data. The asset life cycle profile is
dominated by the precautionary saving early in life; in addition the life cy-
cle household saves for retirement so that assets increase over the life cycle
towards retirement and are then decumulated in order to finance old age
consumption. The quantitative magnitude of saving for retirement is in turn
crucially affected by the generosity of the social security system. These re-
sults are derived from numerical simulations of the model since in general no
analytical solutions or qualitative characterizations of the finite horizon life
cycle model with uninsurable income risk exist.

5.5.2 Infinite Horizon

If households live forever and the stochastic income process they face is time-
invariant as well, the only two forces governing the both the consumption
and savings behavior are the impatience motive and the precautionary sav-
ings motive as the life cycle savings motive is absent without a time (age)
dependent income process. Since the quantitative importance of the impa-
tience motive is determined by the size of the interest rate r, relative to the
time discount factor ρ, it is intuitive that the theoretical properties of the
optimal consumption-savings allocation depend crucially on the relative size
of r and ρ. We therefore distinguish three cases.

Case ρ < r
In this case both the (im-)patience motive and the precautionary saving
motive point towards postponing consumption in favor of saving. In fact, it
is optimal to accumulate assets without bounds as we will demonstrate next.
Proposition 34 [Sotomayor (1984), Chamberlain and Wilson (200)]: Let
r > ρ > 0 and assume that either U is bounded or that the income process is
iid with support yt ∈ [a, A] where a ≥ 0 and A < ∞. Then
lim ct = ∞ almost surely (5.115)
t→∞
126 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

and
lim at = ∞ almost surely (5.116)
t→∞

The basic intuition for this result can be derived from the Euler equation
with liquidity constraints, equation (5.108). Under the assumption that ρ < r
we obtain that
1+r
Uc (ct ) ≥ Et Uc (ct+1 ) (5.117)
1+ρ
> Et Uc (ct+1 ).

Hence the stochastic process for marginal utility {Uc (ct )}, which is strictly
positive, follows a supermartingale.42 We now can invoke the martingale
convergence theorem (which is reviewed in remark 35 below) it follows that
the sequence of random variables {Uc (ct )} converges almost surely to some
limit random variable Uc (c). Iterating on the Euler equation, for all t
 t+1
1+ρ
E0 Uc (ct+1 ) ≤ Uc (c0 ) (5.118)
1+r

and thus E0 Uc (ct+1 ) → 0. But since Uc (.) is strictly positive, it must be the
case that not only {Uc (ct )} converges almost surely to some limit random
variable Uc (c), which is what the martingale convergence theorem guaran-
tees, but it converges almost surely to Uc (c) = 0. From the strict concavity
and the Inada conditions of the utility function it follows that consumption
converges to ∞ with probability one, that is, equation (5.115) is satisfied.
With a bounded labor income process, to finance diverging consumption re-
quires diverging asset income and thus equation (5.116) follows. will exist.
Although we refer the reader to Sotomayor (1984) and Chamberlain and Wil-
son (2000) for proofs of this result, note that the assumptions in the previous
proposition are used to insure that lifetime utility of the consumer remains
finite under the optimal consumption allocation.
42
In fact, for the seqence of random variables {Uc (ct )} to be a supermartingale it is
required that
E |Uc (ct )| < ∞
which is assured if the endowment process {yt } is such that positive consumption is is
possible with probability 1 (or alternatively, if Uc (0) < ∞, but this would violate the
Inada condition).
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 127

Remark 35 The martingale convergence theorem is due to the American

Mathematician Joseph Leo Doob. In order to formally state the theorem
in our context, it is required to describe the probability space on which our
stochastic consumption processes are defined. Let F denote the set of all
infinite sequences43 s = (s0 , s1 . . . , st , . . .), let F denote a σ-algebra on F and
π : F →[0, 1] denote a probability measure defined over the measurable space
(F, F). A random variable X : F → R is a measurable function with respect
to F, i.e. a function X such that for all a ∈ R the set

{s ∈ F : X(s) ≤ a} ∈ F.

A filtration is a sequence of σ-algebras {Ft } with Ft ⊂ F that satisfy

Ft ⊂ Ft+1

for all t. For our purposes we can interpret Ft as capturing the information
the household has at time t. Imagine that at time zero nature draws an
infinite sequence s ∈ F, but which s was drawn is unknown to the household
(she just knows the probability measure π over all possible sequences). As
time unfolds the household learns more and more about what infinite history
s was drawn. For example, suppose that st ∈ {1, 2} for all t. Then s ∈ F is
an infinite history of 1’s and 2’s. Then

F0 = {, F, F01 , F02 }

where
F0i = {ŝ ∈ F : ŝ0 = i},
that is, at time zero the household can distinguish between infinite histories
whose first entry differs (but not among histories with identical first entries).
Similarly
F1 = {, F, F111 , F112 , F121 , F122 , . . .}
where the set
F1ij = {ŝ ∈ F : ŝ0 = i, ŝ1 = j}.
Now households can distinguish between infinite histories that differ in one of
the first two entries. A stochastic process is a sequence {Xt , Ft }∞
t=0 such that

43
For ease of notation we treat s0 as random here too, rather than as the fixed intial
node (as we did so far).
128 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

{Ft }∞ ∞
t=0 is a filtration and {Xt }t=0 is a sequence of random variables such that
Xt is measurable with respect to Ft for all t. We say that {Xt }∞ t=0 is adapted
∞ t t
to the filtration {Ft }t=0 . Our notation {ct (s ), πt (s )} can now be related to
this rigorous way of describing the stochastic structure of the economy. First,
we can construct the probability πt (st ) of event history st from the probability
measure π over infinite histories as
Z
t
πt (s ) = 1A(st ) dπ

where
A(st ) = {ŝ ∈ F : ŝt = st }.
Note that formally πt (.) is a probability measure on the measurable space
(S t , Ft ). Finally, although consumption, defined as a stochastic process, is a
sequence of functions ct : F → R mapping infinite histories into the real
numbers, the measurability of ct with respect to Ft requires that for any two
infinite histories s, ŝ ∈ F with st = ŝt we have

ct (s) = ct (ŝ).

Thus we can write, in short, ct (st ) for all infinite histories for which the
finite history until period t is given by st . Finally, with our definition of what
exactly a stochastic process is we can now define a martingale as a stochastic
process {Xt , Ft }∞
t=0 such that for all t

Xt = E [Xt+1 |Ft ]

almost surely.44 Similar a submartingale is defined as a stochastic process

such that for all t
Xt < E [Xt+1 |Ft ]
almost surely, and a a supermartingale is defined as a stochastic process such
that for all t
Xt > E [Xt+1 |Ft ]
44
In addition the stochastic process has to satisfy the regularity condition

E (|Xt |) < ∞

for all t.
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 129

almost surely. The martingale convergence theorem then states that a super-
martingale {Xt , Ft }∞
t=0 that satisfies

K = sup E(|Xt |) < ∞

t

converges to a limit random variable X ∗ almost surely and that the limit
random variable X ∗ satisfies E(|X ∗ |) ≤ K. Note that X ∗ in general is a
nondegenerate random variable, rather than a fixed number.

The result in proposition 34 also implies that, in a model with many

infinitely lived households each of which solving an income fluctuation, there
will not be a long-run stationary asset distribution if r > ρ, since the asset
holdings of all households diverge with probability one. Consequently no
stationary general equilibrium exists45 in such a model (which we will study
in chapter 7) that satisfies r > ρ. We therefore now turn to the cases r ≤ ρ.

Case ρ = r
Sotomayor (1984) and Chamberlain and Wilson (2000) show that the pre-
vious result (that consumption and assets diverge to ∞ almost surely) goes
through even for ρ = r, if the income process is sufficiently stochastic, in a
sense to be made precise below.

Deterministic Labor Income That a sufficiently risky income process is

required for the result can easily be seen by considering the limiting case in
which labor income follows a deterministic process. Therefore, now assume
that the endowment process is a deterministic sequence {yt }∞ t=0 . Absent the
tight borrowing constraint this is of course a trivial case where consumption is
constant over time, see subsection 5.2.1. However, with the tight borrowing
constraints at+1 ≥ 0 even the deterministic case is not straightforward to
tackle. If income is falling over time, then the household can realize constant
consumption at the optimal level by saving early and dissaving once income
falls. But if labor income tends to grow with time, in order to implement a
constant consumption stream the household would need to borrow against
future higher labor income which the tight borrowing constraint prevents.
45
One possibility to circumvent this problem yet remain within the class of models
in which household age is not a state variable (and thus life cycle considerations are
abstracted from) is to introduce a constant and sufficiently large probability of death of
each household, as in Yaari (1965).
130 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Then we have the following proposition, due to Chamberlain and Wilson

(2000):

Proposition 36 Define
∞
r X yτ
xt = (5.119)
1 + r τ =t (1 + r)τ −t

Then
c̄ := lim ct = sup xt =: x̄ (5.120)
t→∞ t

Proof. See Sargent and Ljungquist (2nd edition), chapter 16.3.1 Note
that for this result U need not be bounded, a maintained assumption of
Chamberlain and Wilson (2000) who of course are mainly interested in the
stochastic labor income case.
The number xt is the period t annuity value of future labor income. Since
the income process {yt } is bounded from above (because yt can take only
finitely many values) the sequence {xt } is bounded from above and thus x̄ is
a finite number which gives the maximal (over time) annuity value of income.
The intuition for this result is easiest to see if there is a finite date T at
which the borrowing constraint binds for the last time. In that case aT = 0
and aT +τ > 0 for τ > 0. Since ρ = r and there is no income risk the Euler
equation then implies (see equation (5.108)) that:
cT = cT +1 = cT +τ = ĉ for all τ > 0 (5.121)
ct ≤ cT for all t < T (5.122)
The budget constraints from period T onwards read as
aT +1
cT + = yT
1+r
aT +τ +1
cT +τ + = yT +τ + aT +τ
1+r
Consolidating these into a lifetime budget constraint (from T onwards) yields46
∞ ∞
X cT +τ X yT +τ
=
τ =0
(1 + r)τ τ =0
(1 + r)τ
46
This follows as long as
aT +τ +1
lim = 0.
τ 7→∞ (1 + r)τ +1
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 131

and thus
∞ ∞ ∞
X cT +τ X 1 (1 + r)ĉ X yT +τ
= ĉ = =
τ =0
(1 + r)τ τ =0
(1 + r)τ r τ =0
(1 + r)τ

and thus, for all t and τ,

∞
r X yT +τ
ct ≤ cT = cT +τ = ĉ = = xT .
1 + r τ =0 (1 + r)τ

The fact that

aT +τ +1
lim ≥0
τ 7→∞ (1 + r)τ +1
is implied by the borrowing constraint aT +τ +1 ≥ 0, and a consumption-savings plan that
has
aT +τ +1
lim >0
τ 7→∞ (1 + r)τ +1

cannot be optimal, intuitively, because the household “leaves resources on the table” rather
than consuming it. The condition
aT +τ +1
lim ≤0
τ 7→∞ (1 + r)τ +1

is called the transversality condition. It is an optimality condition (like the Euler equa-
tion), that, under appropriate conditions on the utility function, the endowment process
and the interest process, is jointly sufficient, together with the Euler equations, for an
optimal consumption-savings plan. Under alternative conditions on the fundamentals the
transversality condition is a necessary condition for an optimal allocation.
Note the fundamental distinction between a No-Ponzi condition and the transversality
condition. The No-Ponzi condition is required for the household maximization problem to
have a solution and in the current context would read as
aT +τ +1
lim ≥ 0.
τ 7→∞ (1 + r)τ +1

The transversality condition is an optimality condition and states

aT +τ +1
lim ≤ 0.
τ 7→∞ (1 + r)τ +1

Taking the No Ponzi condition and the transversality condition (and asserting its neces-
sity) together, any candidate for an optimal consumption-savings plan of a well defined
deterministic consumption-savings problem has to satisfy:
aT +τ +1
lim = 0.
τ 7→∞ (1 + r)τ +1
132 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

That is, in this case consumption is weakly rising over time until the date the
borrowing constraint binds for the last time, and then constant and equal
to the annuity value of income from that date onwards. The proposition
generalizes this result to the case in which the borrowing constraint binds
infinitely often and thus consumption ceases to rise only in the time limit.

Stochastic Labor Income Now let’s turn to the more interesting case in
which the endowment process is stochastic. Here the result that consumption
and asset holdings diverge as time extends to infinity is restored, provided
that, for any event history st the present discounted value of future income is
sufficiently stochastic. The next proposition, again proved under various as-
sumptions by Sotomayor (1984) and Chamberlain and Wilson (2000), states
this formally.

Proposition 37 Suppose that either (a) U is bounded and {yt } is a stochas-

tic process that satisfies the following condition: there exists ε > 0 such that
for all α ∈ R and all st
( ∞
)
τ
X y(s ) 
 t <1−ε
prob α ≤ τ −t ≤ α + ε s (5.123)
τ =t
(1 + r)

or (b) {yt } is a sequence of nondegenerate iid random variables with support

yt ∈ [a, A] with a ≥ 0 and A < ∞. Then, almost surely,

lim ct = ∞ (5.124)
t→∞
lim at = ∞ (5.125)
t→∞

Note from (b) that Sotomayor (1984) needs no assumptions on the bound-
edness of the utility and the level of variability of the income process (other
that it has to be nondegenerate) to prove the result, but needs the iid assump-
tion. Dispensing with the iid comes at the cost of having to make U bounded
and the income process “sufficiently” stochastic so that present discounted
value of future income leaves set [α, α + ε] with probability of at least ε. So
unfortunately the theorem does not apply to an economy with both serially
correlated shocks and standard utility functions (CRRA or CARA).
Note that one can show that condition (5.123) holds for income following a
finite-state Markov chain with Π(y) > 0 for all y and π(y 0 |y) > 0 for all y, y 0 .
The previous theorem implies that, under the appropriate conditions, the
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 133

state space for assets at is unbounded. In particular, since with probability 1

asset holdings explode, the demand for assets in a steady state of an economy
composed of agents with ρ = r (and facing the income fluctuation problem
discussed here) is infinite and thus no steady state equilibrium can exist in
which the interest rate, endogenously determined in general equilibrium, will
satisfy ρ = r.

Case ρ > r
At this stage it will prove helpful to formulate the income fluctuation problem
with borrowing constraints recursively. For the general problem in which
income follows a stationary, finite state Markov chain47 the Bellman equation
is
( )
1 X
v(a, y) = max U (c) + π(y 0 |y)v(a0 , y 0 ) (5.126)
a0 ,c≥0 1 + ρ y0
a0
s.t. c + = y+a (5.127)
1+r
As first order condition we obtain
1+r X
Uc (c) ≥ π(y 0 |y)v(a0 , y 0 ) (5.128)
1 + ρ y0
= if a0 > 0

The envelope condition reads as

v 0 (a, y) = Uc (c) (5.129)

so that the Euler equation once again becomes

1+r X
Uc (c) ≥ π(y 0 |y)Uc (c0 ) (5.130)
1 + ρ y0
= if a0 > 0 (5.131)

or ( )
1+r X
Uc (c) = max Uc (y + a), π(y 0 |y)Uc (c0 ) (5.132)
1 + ρ y0
47
We will deal with the case of nonstationary income below.
134 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Both equations (5.126) and (5.132) can be used to compute optimal policy
functions a0 (a, y) and c(a, y) as well as the value function v(a, y) using exactly
the same iterative procedures as described in the last section, just with the
modification that the Bellman equation has the additional constraint a0 ≥ 0
and the Euler equation has two parts now.

Income Process IID If the income process is iid we can reduce the state
space from two to one dimension by introducing the variable cash at hand
x = a + y. The Bellman equation becomes
(  )
a0

1 X
v(x) = max u x− + π(y 0 )v(a0 + y 0 ) (5.133)
0≤a0 ≤(1+r)x 1+r 1 + ρ y0

with Euler equation

( )
a0 (x) 0 0
  
1+r X a (x )
Uc x − = max Uc (x), π(y 0 )Uc a0 (x) + y 0 − (5.134)
1+r 1 + ρ y0 1+r
( )
0 0 0

1+r X a [a (x) + y ]
= max Uc (x), π(y 0 )Uc a0 (x) + y 0 −
1 + ρ y0 1+r

Schechtman and Escudero (1977) provide the following characterizations of

the optimal policy function a0 (x) for next periods’ asset holdings and for
consumption c(x). We first consider the deterministic case.

Proposition 38 Let T = ∞ and Π(y) = 1 (no uncertainty). Then there

exists an x̃ such that if x ≥ x̃, then

x0 = a0 (x) + y < x (5.135)

A liquidity constrained agent remains so forever, i.e. a0 (y) = 0 and c(y) = y.

Consumption is strictly increasing in cash at hand, or

dc(x)
>0 (5.136)
dx
There exists an x̄ > y such that a0 (x) = 0 for all x ≤ x̄ and a0 (x) > 0 for all
0
x > x̄. Finally dc(x)
dx
≤ 1 and dadx(x) < 1 + r.
5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 135

Proof. If a0 (x) > 0 then from envelope and FOC

1+r 0 0
v 0 (x) = v (x )
1+ρ
< v 0 (x0 ) (5.137)
1+r
since 1+ρ < 1 by our maintained assumption. Since v is strictly concave (v
inherits all the properties of U ) we have x > x0 . Obviously, if a0 (x) = 0 and
x > y, then x0 = y < x. Thus pick any x̃ > y.
For second part, suppose that a0 (y) > 0. Then from the first order condi-
tion and strict concavity of the value function
1+r 0 0
v 0 (y) = v (a (y) + y)
1+ρ
< v 0 (a0 (y) + y)
< v 0 (y) (5.138)
a contradiction. Hence a0 (y) = 0 and c(y) = y.
The last part we also prove by contradiction. Suppose a0 (x) > 0 for all
x > y. Pick arbitrary such x and define the sequence {xt }∞
t=0 recursively by

x0 = x (5.139)
xt = a0 (xt−1 ) + y ≥ y (5.140)
If there exists a smallest T such that xT = y then we found a contradiction,
since then a0 (xT −1 ) = 0 and xT −1 > 0. So suppose that xt > y for all t. But
then a0 (xt ) > 0 by assumption. Hence
1+r 0
v 0 (x0 ) = v (x1 )
1+ρ
 t
1+r
= v 0 (xt )
1+ρ
 t
1+r
< v 0 (y)
1+ρ
 t
1+r
= u0 (y) (5.141)
1+ρ
where the inequality follows from the fact that xt > y and the strict concavity
of v. the last equality follows from the envelope theorem and the fact that
a0 (y) = 0 so that c(y) = y.
136 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

But since v 0 (x0 ) > 0 and u0 (y) > 0 and 1+ρ1+r

< 1, we have that there exists
 t
finite t such that v 0 (x0 ) > 1+ρ
1+r
u0 (y), a contradiction.
The proof that the derivatives of the policy functions have the asserted
properties follow from the strict concavity of the value function and the first
order condition.
This last result bounds the optimal asset holdings (and hence cash at
hand) from above for T = ∞. Since computational techniques usually rely
on the finiteness of the state space we want to make sure that for our theory
the state space can be bounded from above. For the finite lifetime case there
is no problem. The most an agent can save is by consuming 0 in each period
and hence
t
X
t+1
at+1 (xt ) ≤ xt ≤ (1 + r) a0 + (1 + r)j ymax (5.142)
j=0

which is bounded for any finite lifetime horizon T < ∞.

The last theorem says that cash at hand declines over time or is constant
at y, in the case the borrowing constraint binds. The theorem also shows
that the agent eventually becomes credit-constrained: there exists a finite τ
such that the agent consumes his endowment in all periods following τ. This
follows from the fact that marginal utility of consumption has to increase
1+r
at geometric rate 1+ρ if the agent is unconstrained (and thus consumption
to decline) and from the fact that once he is credit-constrained, he remains
credit constrained forever. This can be seen as follows. First x ≥ y by
the credit constraint. Suppose that a0 (x) = 0 but a0 (x0 ) > 0. Since x0 =
a0 (x) + y = y we have that x0 ≤ x. Thus from the previous proposition
a0 (x0 ) ≤ a0 (x) = 0 and hence the agent remains credit-constrained forever.
For the infinite lifetime horizon, under deterministic and constant income
we have a full qualitative characterization of the allocation: If a0 = 0 then
the consumer consumes his income forever from time 0. If a0 > 0, then cash
at hand and hence consumption is declining over time, and there exists a
time τ (a0 ) such that for all t > τ (a0 ) the consumer consumes his income
forever from thereon, and consequently does not save anything.
We now consider the stochastic case with income being iid over time.
The proof of the following proposition is identical to the deterministic case.
Remember the assumption that the minimum income level y1 ≥ 0.

Proposition 39 Consumption is strictly increasing in cash at hand, i.e.

5.5. PRUDENCE AND LIQUIDITY CONSTRAINTS: THEORY 137

c0 (x) ∈ (0, 1]. Optimal asset holdings are either constant at the borrowing
0
limit or strictly increasing in cash at hand, i.e. a0 (x) = 0 or dadx(x) ∈ (0, 1 + r)

It is obvious that a0 (x) ≥ 0 and hence x0 (x, y 0 ) = a0 (x) + y 0 ≥ y1 so we

have y1 > 0 as a lower bound on the state space for x. We now show that
there is a level x̄ > y1 for cash at hand such that for all x ≤ x̄ we have that
c(x) = x and a0 (x) = 0

Proposition 40 There exists x̄ > y1 such that for all x ≤ x̄ we have c(x) =
x and a0 (x) = 0

Proof. Suppose, to the contrary, that a0 (x) > 0 for all x ≥ y1 . Then,
using the first order condition and the envelope condition we have for all
x ≥ y1
1+r 0 0 1+r 0
v(x) = Ev (x ) ≤ v (y1 ) < v 0 (y1 ) (5.143)
1+ρ 1+ρ
Picking x = y1 yields a contradiction.
Hence there is a cutoff level for cash at hand below which the consumer
consumes all cash at hand and above which he consumes less than cash at
hand and saves a0 (x) > 0. So far the results are strikingly similar to the
deterministic case. Unfortunately here it basically ends, and therefore our
analytical ability to characterize the optimal policies. In particular, the very
important proposition showing that there exists x̃ such that if x ≥ x̃ then
x0 < x̃ does not go through anymore, which is obviously quite problematic
for computational considerations. In fact we state, without a proof, a result
due to Schechtman and Escudero (1977).

Proposition 41 Suppose the period utility function is of constant absolute

risk aversion form u(c) = −e−c , then for the infinite life income fluctuation
problem, if Π(y = 0) > 0 we have xt → +∞ almost surely.

Proof. See Schechtman and Escudero (1977), Lemma 3.6 and Theorem
3.7
Fortunately there are fairly general conditions under which one can, in
fact, prove the existence of an upper bound for the state space. Again we
will refer to Schechtman and Escudero for the proof of the following results.
Intuitively why would cash at hand go off to infinity even if the agents are
impatient relative to the market interest rate, i.e. even if β(1 + r) < 1? If
138 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

agents are very risk averse, face borrowing constraints and a positive proba-
bility of having very low income for a long time, they may find it optimal to
accumulated unbounded funds over time to self-insure against the eventual-
ity of this unlikely, but very bad event to happen. It turns out that if one
assumes that the risk aversion of the agent is sufficiently bounded, then one
can rule this out.

Proposition 42 Suppose that the marginal utility function has the property
that there exist finite eu0 such that

lim (logc u0 (c)) = eu0 (5.144)

c→∞

Then there exists a x̃ such that x0 = a0 (x) + yN ≤ x for all x ≥ x̃.

Proof. See Schechtman and Escudero (1977), Theorems 3.8 and 3.9
The number eu0 is called the asymptotic exponent of u0 . Note that if the
utility function is of CRRA form with risk aversion parameter σ, then since

logc c−σ = −σ logc c = −σ (5.145)

we have eu0 = −σ and hence for these utility function the previous proposition
applies. Also note that for CARA utility function
c
logc e−c = −c logc e = − (5.146)
ln(c)
c
− lim = −∞ (5.147)
c→∞ ln(c)

and hence the proposition does not apply.

So under the proposition of the previous theorem we have the result that
cash at hand stays in the bounded set X = [y1 , x̃].48 Consumption equals cash
at hand for x ≤ x̄ and is lower than x for x > x̄, with the rest being spent on
capital accumulation a0 (x) > 0. Figure ?? below shows the situation for the
case in which income can take only two possible realizations Y = {y1 , yN }.
Add Bewley1.wmf picture
48
If x0 = a0 + y0 happens to be bigger than x̃, then pick x̃0 = x0 .
5.6. PRUDENCE AND LIQUIDITY CONSTRAINTS: COMPUTATION139

Income Serially Correlated, but Stationary Now consider the case

where income is correlated over time and follows a Markov chain with tran-
sition π. Now the trick of reducing the state to the single variable cash at
hand does not work anymore. This was only possible since current income
y and past saving a entered additively in the constraint set of the Bellman
equation, but neither variable appeared separately. With serially correlated
income, however, current income influences the probability distribution of
future income. There are several possibilities of choosing the state space
for the Bellman equation. One can use cash at hand and current income,
(x, y), or asset holdings and current income (a, y). Obviously both ways are
equivalent and I opted for the later variant, which leads to the functional
equation (5.133). What can we say in general about the properties of the
optimal policy functions a0 (a, y) and c(a, y)? Huggett (1993) proves a propo-
sition similar to the ones above showing that c(a, y) is strictly increasing in
a and that a0 (a, y) is constant at the borrowing limit or strictly increasing
(which implies a cutoff ā(y) as before, which now will depend on current
income y). What turns out to be very difficult to prove is the existence of an
upper bound of the state space, ã such that a0 (a, y) ≤ a if a ≥ ã. Huggett
proves this result for the special case that income can only take two states
y ∈ {yl , yh } with 0 < yl < yh and π(yh |yh ) ≥ π(yh |yl ), and CRRA utility. See
his Lemmata 1-3 in the appendix. I am not aware of any more general result
for the non-iid case. With respect to computation in more general cases, we
have to cross our fingers and hope that a0 (a, y) eventually (i.e. for finite a)
crosses the 450 -line for all y.

5.6 Prudence and Liquidity Constraints: Com-

putation
The analysis in the previous section is as far as one can push the model
analytically. Apart from very special cases (see Caballero (1990) for an ex-
plicit solution when utility is CARA, agents life for finite time, the interest
rate is zero and income follows a random walk with normally distributed iid
disturbances) the model cannot be solved analytically.
For computational purposes we want to make the problem a consumer
faces recursive. From now on, for the rest of this chapter, unless oth-
erwise noted, we assume that the stochastic process governing labor in-
140 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

come is described by a finite state, stationary Markov process with domain

y ∈ Y = {y1 , . . . , yN } and transition probabilities π(y 0 |y), where we assume
that y1 ≥ 0 and yi+1 > yi . As state variables we choose current asset hold-
ings and the current labor income shock (a, y). In order to make the problem
well-behaved we have to make sure that agents don’t go into debt so much
that they can’t pay at least the interest on that debt and still have non-
negative consumption. Let Ā be the maximum amount an agent is allowed
a0
to borrow. Since consumption equals c = y + a − 1+r , we have, for an agent
that borrowed to the maximum amount a = −Ā, received the worst income
shock y1 and just repays interest (i.e. a0 = −Ā):
Ā
c = y1 + Ā −≥0 (5.148)
1+r
Non-negativity of consumption implies the borrowing limit
1+r
Ā = y1 (5.149)
r
which we impose on the consumer. Since Ā also equals the present discounted
value of future labor income in the worst possible scenario of always obtain-
ing the lowest income realization y1 , we may call this borrowing limit the
“natural debt limit” (see Aiyagari (1994)). Note that, since borrowing up to
the borrowing limit implies a positive probability of zero consumption next
period (if π(y1 |y) > 0 for all y ∈ Y ), this borrowing constraint is not going
to be binding, as long as the utility function satisfies the Inada conditions.
For any a ∈ [−Ā, ∞) and any y ∈ Y we then can write Bellman’s equation
as
(  )
a0
 X
0 0 0
vt (a, y) = max U y+a− +β π(y |y)vt+1 (a , y )
−Ā≤a0 ≤(1+r)(a+y) 1+r y0
(5.150)
The first order conditions to this problem is
a0
 
1 0
X
U y+a− =β π(y 0 |y)vt+1
0
(a0 , y 0 ) (5.151)
1+r 1+r y0
0
where vt+1 is the derivative of vt+1 with respect to its first argument. The
envelope condition is
a0
 
0 0
vt (a, y) = U y + a − (5.152)
1+r
5.6. PRUDENCE AND LIQUIDITY CONSTRAINTS: COMPUTATION141

Combining both conditions yields

a0 a00
  X  
0 0 0 0 0
U y+a− = β(1 + r) π(y |y)U y + a − (5.153)
1+r y 0
1+r

Defining
a0
ct = y + a − (5.154)
1+r
a00
ct+1 = y 0 + a0 − (5.155)
1+r
we obtain back our stochastic Euler equation (5.35). There are several im-
portant special cases that merit a brief discussion.

5.6.1 IID Income shocks

If labor income shocks are iid over time with probability density π, then
one can reduce the state space to a single variable, so called cash at hand
x = a + y. The budget constraint then becomes
a0
c+ =x (5.156)
1+r
and since
x 0 = a0 + y 0 (5.157)
we have the associated Bellman equation
(  )
a0
 X
vt (x) = max U x− +β π(y 0 )vt+1 (a0 + y 0 )
−Ā≤a0 ≤(1+r)x 1+r y0

5.6.2 Serially Correlated, Mean-Reverting Income Shocks

With serially correlated income shocks the state space consists of (a, y), asset
holdings and the current income shock. Independent of whether the state
space is (a, y) or only cash at hand x, we can either work on the value
function directly or use the Euler equation to solve for the optimal policies.
Depending on whether the time horizon of the household is finite or infinite,
different iterative procedures can be applied:
142 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Finite Time Horizon T < ∞

1. Value function iteration: We want to find sequences of value func-

tions {vt (a, y)}Tt=0 and associated policy functions {ct (a, y), a0t (a, y)}∞t=0 .
Given that the agent dies at period T we can normalize vT +1 (a, y) ≡ 0.
Then we can iterate backwards on the equation (5.150): at period t we
know the function vt+1 (., .), hence can solve the maximization problem
to find functions a0t (., .), vt (., .) and ct (., .). Note that for each t there
are as many maximization problems to solve as there are admissible
(a, y)-pairs.

2. Policy function iteration on the Euler equation. We are looking for

sequences of policy functions {ct (a, y), a0t (a, y)}∞
t=0 . Again, given that
the agent dies at period T + 1 we know that at period T all income will
be consumed and nothing we be saved, thus

cT (a, y) = a + y (5.158)
a0T (a, y) = 0 (5.159)

The Euler equation between period t and t + 1 reads as

a0t (a, y) a0t+1 (at (a, y), y 0 )

  X  
0 0 0 0 0
U y+a− = β(1+r) π(y |y)U y + at (a, y) −
1+r y 0
1+r
(5.160)
where a0t+1 (., .)
is a known function from the previous step and we
want to solve for the function a0t (., .). Note that for a given (a, y) no
maximization is needed to find a0t (a, y) as in the previous procedure,
one just has to find a solution to a (potentially highly nonlinear) single
equation.

Infinite Time Horizon T = ∞

1. Value function iteration: now we look for a time invariant value func-
tion v(a, y) and associated policy functions a0 (a, y) and c(a, y). We need
to find a fixed point to Bellman’s equation, since there is no final period
to start from. Thus we make an initial guess for the value function,
5.6. PRUDENCE AND LIQUIDITY CONSTRAINTS: COMPUTATION143

v 0 (a, y) and then iterate on the functional equation

(  )
0

a X
v n (a, y) = max U y+a− +β π(y 0 |y)v n−1 (a0 , y 0 )
0
−Ā≤a ≤(1+r)(y+a) 1+r y0
(5.161)
until convergence, i.e. until
n
v − v n−1 ≤ ε (5.162)

where ε is the desired precision of the approximation. Note that at

each iteration we generate policy functions a0n (a, y) and cn (a, y). We
know that under appropriate assumptions (β < 1, U bounded) the
operator defined by Bellman’s equation is a contraction mapping and
hence convergence of the iterative procedure to a unique fixed point is
guaranteed.

2. Policy function iteration on the Euler equation. Again we have no final

time period, so we guess an initial policy a00 (a, y) or c0 (a, y) and then
iterate on
a0n (a, y) a0n−1 (a0n (a, y), y 0 )
  X  
0 0 0 0 0n
U y+a− = β(1+r) π(y |y)U y + a (a, y) −
1+r y0
1+r
(5.163)
until 0n
a − a0n−1 ≤ εa

(5.164)
Deaton (1992) argues that under the assumption β(1 + r) < 1 the
operator defined by the Euler equation is a contraction mapping, so
that convergence to a unique fixed point is guaranteed.

5.6.3 Permanent Shocks

Deaton (1991) and Carroll (1997), among many others, demonstrate how to
tackle problems in which the level of income of households follow a random
walk. For agents living forever, there is no hope to obtain finite bounds
on the amount of assets that are potentially accumulated, hence the state
space is unbounded, which poses problems for the computation of optimal
policies (and the value function). It should also be noted that in such an
economy there is in general no hope for a stationary general equilibrium,
144 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

since the income distribution is nonstationary and thus in general the wealth
distribution will be for any fixed interest rate r. Hence when dealing with
general equilibrium models with infinitely lived agents, one of the maintained
assumptions is stationarity of the individual income processes.49
The partial equilibrium problem with nonstationary income process and
infinite horizon can be tackled, however, through an appropriate change of
variables. For this trick to work it is crucial to assume that households
have CRRA period utility. First assume that, following Deaton (1991) that
the labor income process of the household features iid growth rates, that is,
assume that
yt+1
zt+1 = (5.165)
yt

is a sequence of iid random variables. This implies that

log(yt+1 ) − log(yt ) (5.166)

is equal to an iid random variable, and thus implies that the natural loga-
rithm of income follows a random walk, potentially with drift

log(yt+1 ) = µ + log(yt ) + εt (5.167)

where {εt } is a sequence of iid random variables with E(εt ) = 0. Note that
this specification implies positive income in all periods, with probability one.
In order to overcome the problem that the state space for assets is unbounded,
which makes the dynamic programming problem of the household difficult
to tackle, normalize all variables by the current level of income

ct
θt = (5.168)
yt
xt at + y t
wt = = (5.169)
yt yt

49
Note that the same comment does not apply for economies in which agents die with
probability 1 in finite time in which case typically a stationary general equilibrium can be
found.
5.6. PRUDENCE AND LIQUIDITY CONSTRAINTS: COMPUTATION145

The budget constraint then becomes

at+1
ct + = y t + at
1+r
xt+1 − yt+1
ct + = xt
1+r
wt+1 − 1
θt + zt+1 = wt
1+r
(1 + r)(wt − θt )
wt+1 = +1 (5.170)
zt+1
and, by dividing both sides of the Euler equation by yt−σ we obtain
(   " −σ #)
1 + r c t+1
θt−σ = max wt−σ , Et
1+ρ yt
   
−σ 1+r  −σ 
= max wt , Et (θt+1 zt+1 ) (5.171)
1+ρ
We are searching for a time-invariant policy function θ(w) solving this Euler
equation. Note that, written in recursive formulation θt corresponds to θ(w)
and θt+1 to θ(w0 ) where w0 = (1+r)(w−θ(w))
z0
. Assuming that z 0 = zt+1 can take
on only a finite number of values we obtain
(  X    −σ )
1 + r (1 + r)(w − θ(w))
θt−σ = max w−σ , π(z 0 ) θ + 1 z0
1 + ρ z0 z0
(5.172)
This Euler equation can be solved numerically for the optimal policy function
θ(w) of consumption per current income (the average propensity to consume)
by doing policy iteration on50
(  X   −σ )
1 + r (1 + r)(w − θ(w))
θ(w)−σ = max w−σ , θ + 1 z0
1 + ρ z0 z0
(5.174)
50
Deaton and Laroque (1992) show that with nonstationary income the operator asso-
ciated with the Euler equation is a contraction if
1+r
E(z −σ ) < 1 (5.173)
1+ρ
146 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

Note that Deaton (1991) is not able to develop a condition under which the
state space for w is bounded above. It is obviously bounded below by 1
since assets at are restricted to be nonnegative. In accordance to the model
with stationary income he establishes a threshold w̄ such that θ(w) = w and
w0 (w) = 1 for all w ≤ w̄; i.e. for low levels of normalized cash at hand it is
optimal for the consumer to eat all cash at hand today and save nothing for
tomorrow.
As discussed in the next chapter, a very popular specification of the house-
hold labor income process postulates that log-labor income is composed of
a permanent component that follows a random walk and a transitory shock.
That is, suppose labor income follows the process51

ln(yt+1 ) = pt+1 + εt+1

pt+1 = pt + ηt+1

Here ηt+1 and εt+1 are permanent and transitory income shocks that are
independently (over time and of each other) and normally distributed with
2 σ2
variances (σε2 , ση2 ) and means (− σ2ε , − 2η ). Denote the cumulative distribution
function of the Normal distribution by N.
As before one can make the problem stationary by expressing all variables
relative to the permanent component of income, pt , see e.g. Carroll (1997).
Now the state space consists of assets a, the permanent component of log-
income p and the transitory shock e. The dynamic programming problem
can be written as
 1−σ Z Z 
c 0 0 0 0 0
v(a, p, ε) = max +β v(a , p + η , ε )dN (ε )dN (η )
c,a0 ≥0 1−σ ε0 η 0
s.t.
a0
c+ = exp(p + ε) + a
1+r
With the usual definition of cash at hand

x = y+a
= exp(p + ε) + a
51
In example 26 we studied a similar process, but there the level of labor income (rather
than its log) was subject to permanent and transitory shocks (and the period utility
function was quadratic rather than of CRRA form).
5.6. PRUDENCE AND LIQUIDITY CONSTRAINTS: COMPUTATION147

we can reduce the state space and re-write the recursive problem as
 1−σ Z Z 
c 0 0 0 0 0
v(x, p) = max +β v(exp(p + ε ) + (1 + r)(x − c), p + η )dN (ε )dN (η )
0≤c≤x 1−σ ε0 η 0
(5.175)
0
where we have used the fact that by definition of x and the budget constraint

x0 = exp(p0 + ε0 ) + a0
= exp(p0 + ε0 ) + (1 + r)(x − c)

Note, however, that since p follows a random walk, the problem above is not
stationary. Therefore we attempt to make it stationary by defining variables
relative to the permanent level of income:
c x
c̃ = and w =
exp(p) exp(p)

We now conjecture that, due to our assumption of CRRA utility (which

guarantees homotheticity of the lifetime utility function), the value function
takes the form
v(x, p) = ṽ(w) exp(p)1−σ
and that we can find associated policy functions c̃(w) and w0 (w) that solve the
dynamic programming problem associated with the “deflated” value function
ṽ(w). Dividing (5.175) by exp(p)1−σ and using the conjectured form of the
value function yields
 1−σ Z Z 
c̃ 0 1−σ 0 0 0
ṽ(w) = max +β exp(p − p) ṽ(w )dN (ε )dN (η )
0≤c̃≤w 1−σ ε0 η 0

with
x0 exp(p0 + ε0 ) + (1 + r)(x − c)
w0 = =
exp(p0 ) exp(p0 )
(1 + r)(x − c) (1 + r)(w − c̃)
= exp(ε0 ) + 0
= exp(ε0 ) +
exp(p + η ) exp(η 0 )
and thus the Bellman equation becomes
 1−σ   
(1 + r)(w − c̃)
Z Z
c̃ 0 1−σ 0 0 0
ṽ(w) = max +β exp(η ) ṽ exp(ε ) + dN (ε )dN (η )
0≤c̃≤w 1−σ ε0 η 0 exp(η 0 )
148 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

which we need to solve numerically for the value function ṽ(w) and the as-
sociated policy function c̃(w). Once we have determined those, consumption
and asset levels are given by
 
x
c(x, p) = exp(p)c̃
exp(p)
0
a (x, p) = (x − c(x, p))(1 + r).

Alternatively one can use the Euler equation (and policy function iter-
ation) to determine the optimal consumption policy. In original form, the
Euler equation associated with Bellman equation (5.175) reads as
 Z Z 
−σ −σ 0 0 0 0 0
c(a, p, ε) = max (exp(p + ε) + a) , β(1 + r) c(a , p + η , ε )dN (ε )dN (η )
ε0 η0

or, in terms of cash at hand

 Z Z 
−σ −σ 0 0 0 0
c(x, p) = max x , β(1 + r) c(x , p + η )dN (ε )dN (η )
ε0 η0
0 0 0
with x = exp(p + η + ε ) + (1 + r)(x − c(x, p))

As before, conjecturing that c(x, p) = exp(p)c̃ (w) with w = x/ exp(p) and

dividing the Euler equation by exp(p)−σ yields
( Z Z   −σ
−σ −σ 0 0 (1 + r)(w − c̃(w))
c̃ (w) = max w , β(1 + r) exp(η )c̃ exp(ε ) + 0
dN (ε0 )dN (η
ε0 η 0 exp(η )

which is a nonlinear functional equation in the unknown function c̃ with

the one state variable w. Obviously very similar transformations yield the
appropriate Bellman equations and Euler equations for the finite horizon
case.

5.6.4 From Policy Functions to Simulated Time Series

Once one has solved for the policy functions one can simulate time paths for
consumption and asset holdings for a particular agent. Start at some asset
level a0 , possibly equal to zero, then draw a sequence of random income
numbers {yi }M t=0 according to the specified income process. This will usually
require first drawing random numbers from a uniform distribution and then
5.6. PRUDENCE AND LIQUIDITY CONSTRAINTS: COMPUTATION149

mapping these numbers into the appropriate labor income realizations. Now
one can generate a time series for consumption and asset holdings (for the
infinite horizon case) by

c0 = c(a0 , y0 ) (5.176)
a1 = a0 (a0 , y0 ) (5.177)

and recursively

ci = c(ai , yi ) (5.178)
ai+1 = a0 (ai , yi ) (5.179)

The same applies to the finite time horizon case, but here the policy functions
being applied also vary with time.

5.6.5 Implementation of Specific Algorithms

All algorithms for the different cases require the approximation of policy or
value functions over the state space of potential incomes and assets, (a, y),
the latter in principle being a continuous state variable.52 There is a good
number of alternative approaches of how to approximate these functions,
ranging from a full discretization of the problem, to approximations that
divide the state space into smaller pieces and then approximate the functions
under consideration by simple functions (these are so-called finite element
methods) to approaches that approximate the unknown function over the
entire state space by a weighted sum of simple functions (e.g. polynomials),
so called projection methods. For a complete treatment of these methods,
see chapters 10-12 of Judd (1998), McGrattan (1999) or Herr and Maussner
(2009).53

52
In the canonical consumption-savings problem the sate space is typically small. For
problems with multiple continuous state variables and smooth functions to be approx-
imated Krueger and Kubler (2004) propose a sparse grid method for approximation of
multidimensional (but smooth) functions. Recently, Judd, Maliar and Maliar (2012) de-
veloped a related algorithm that endogenizes the choice of the sparse grid.
53
Recently Carroll (2006) proposed an endogenous grid method that in practice speeds
up computation of the class of problems discussed in this section tremendously.
150 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

5.7 Prudence and Liquidity Constraints: Em-

pirical Implications
In this section we discuss the implications of the SIM with respect to two
important empirical dimensions and review the relevant empirical and quan-
titative literature. First, in the next subsection we investigate how strongly
(if at all) consumption is expected to respond to income shocks, or reversely,
how well households that only have access to a simple state-uncontingent
asset can self-insure. Second, we derive the implications for the evolution of
the cross-sectional consumption distribution of a life-cycle version of the SIM
and confront it with the data.

5.7.1 Time Series Implications: How Does Consump-

tion Respond to Income Shocks
Recall from section 5.2.2 that the certainty equivalence version of the SIM
with β(1 + r) = 1 and an income process with transitory and permanent
shocks implies that the realized change in household consumption is given
by
rθ−1
∆ct = t ut + vt
1+r
where vt is a permanent
 income shock, ut is a transitory income shock and
1
θt = 1 − (1+r)T −t+1 is an adjustment factor that depends on how close the
household is towards the end of life T. If T = ∞ the household is infintely
lived and θt = 1. Thus, according to this version of the model, consumption
responds one for one to permanent income shocks and almost not at all to
transitory income shocks if the household is both young (t is small relative
to T ) and the interest rate r is small.54

Closed form Solution for CARA Utility

Wang (2003) for a closed form solution with CARA utility. Same upshot as
with quadratic utility, but with constant precautionary savings term
54 rθ −1
Of course, if t = T we find 1+r
t
= 1 and household consumption responds strongly
to ”temporary” income shocks since at the end of life a temporary income shock is a
permanent shock.
5.7. PRUDENCE AND LIQUIDITY CONSTRAINTS: EMPIRICAL IMPLICATIONS151

Empirical Estimates of Insurance Coefficients

With certainty equivalence and a specific income process the question how
strongly consumption responds to income shocks with different persistence
has a precise answer, as demonstrated above. More broadly, Blundell, Pista-
ferri and Preston (2008) propose as measure for how strongly consumption
responds (or better, does not respond) the following concept of consumption
insurance coefficients. Suppose that the stochastic part of log-labor earnings
can be written as
t X
X N
log(yit ) = ατn xnit−τ (5.180)
τ =0 n=1

where the (x1it , x2it , . . . , xN

it )
are N income shocks (think of them as shocks
with different persistence), assumed to be independently distributed across
time, households and across each other (cov(xnit , xm
jτ ) = 0 for all i, j, all t, τ,
all m, n).

Example 43 One popular stochastic earnings process (see section 5.6.3 and
also the next chapter) models the logarithm of earnings as the sum of a
permanent and transitory shock, as did equations (5.49) and (5.50) for the
level of earnings:

log(yit ) = pit + εit (5.181)

pit = pit−1 + ηit (5.182)

where εit is the transitory shock and ηit is the permanent shock. This process
fits into the general structure of equation (5.180) by letting N = 2, (x1it , x2it ) =
(εit , ηit ) and α01 = 1, ατ1 = 0 for all τ > 0, and ατ2 = 1 for all τ ≥ 0. Recall
that this process also implies

∆ log(yit ) = ηit + ∆εit (5.183)

The consumption insurance coefficient for income shock n is then defined

as
Covi (∆ log(cit ), xnit )
φnt = 1 − (5.184)
V ari (xnit )
where Covi and V ari are taken with respect to the cross-section of house-
hold observations indexed by i observed at time (or age) t. The number
φnt measures what share of stochastic labor income variability due to shock
152 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

n does not translate into corresponding consumption changes,that is, how

well household consumption is insured about the n-th component of income
risk. Thus these measures of consumption insurance have become important
“summary statistics”.
Equipped with this definition we can now ask how much consumption
insurance is there in the data, and how well do our theoretical models cap-
ture this degree of consumption insurance. The model statistics are easy to
compute, either by hand (if the model has an analytical solution) or through
straightforward simulations. To estimate the empirical counterpart is signif-
icantly harder since the income shocks {xnit } are not directly observed in the
data (but they are observed in simulations of the theoretical models), even
in the ideal situation where we have panel data on income and consumption,
{cit , yit }.
To estimate the denominator V ari (xnit ) of equation (5.184) one needs to
make specific assumption of the form of the income process. To estimate
the numerator Covi (∆ log(cit ), xnit ) one in addition needs to know (or at
least make assumptions about) the stochastic process household consumption
growth follows. Blundell et al. (2008) assume that log-earnings follow the
process in equations (5.181)-(5.182), and they postulate that unexplained55
consumption growth is determined by the process56

∆ log(cit ) = πtη ηit + πtε εit + ξit

where the remaining error term ξit is uncorrelated with the permanent and
transitory income shock at all leads and lags. One can loosely interpret πtη
as the (potentially time- or age-varying) marginal propensity to consume out
of a permanent income shock, with πtε possessing the same interpretation
for a transitory shock, but should keep in mind that ∆ log(cit ) is the change
in log-consumption (i.e. consumption growth) rather than the change in
consumption levels. In fact, ignoring this distinction (both for income and
consumption), the canonical PIH model with certainty equivalence and in-
finite horizon has the implication that πtη = 1 and πtε = 1+rr
, as shown in
55
That is, after consumption growth has been purged from its predictable components
(that are due to, e.g. life cycle effects and family size and composition as well as other
observable household characteristics).
56
Blundell et al. (2008) provide an appendix where they argue that this consumption
growth process is consistent with a first order approximation of the Euler equation (of
course, with non-binding borrowing constraints) in a model where income follows the
permanent-transitory process stipulated and households have CRRA period utility.
5.7. PRUDENCE AND LIQUIDITY CONSTRAINTS: EMPIRICAL IMPLICATIONS153

section 5.6.3.
Note that with this postulated consumption process
Covi (πtη ηit + πtε εit + ξit , εit )
φεt = 1− = 1 − πtε
V ari (εit )
Covi (πtη ηit + πtε εit + ξit , ηit )
φηt = 1− = 1 − πtη
V ari (ηit )
and thus the insurance coefficients are just one minus the marginal propen-
sities to consume out of the various income shocks. But again, as long as
(εit , ηit ) are not directly observable, neither the φxt nor the πtx can be empir-
ically estimated.
However, given the assumed income and consumption growth processes,
recognize that, using equation (5.183),

Covi (∆ log(yit ), ∆ log(yit+1 )) = Covi (ηit + εit − εit−1 , ηit+1 + εit+1 − εit ) = −V ari (εit )
Covi (∆ log(cit ), ∆ log(yit+1 )) = Covi (πtη ηit + πtε εit + ξit , ηit+1 + εit+1 − εit )
= −πtε V ari (εit ) = −Covi (∆ log(cit ), εit ).

Therefore we can estimate

Covi (∆ log(cit ), εit ) Covi (∆ log(cit ), ∆ log(yit+1 ))
φεt = 1 − =1−
V ari (εit ) Covi (∆ log(yit ), ∆ log(yit+1 ))
which only requires household consumption and income data with small panel
dimension (two periods for consumption, three for income). Note that the
last ratio has a nice interpretation, it is the regression coefficient of a cross-
sectional regression of consumption growth on income growth, using one
period ahead income growth as an instrument.
A similar, but more data-demanding calculation yields

Covi (∆ log(yit ), ∆ log(yit−1 ) + ∆ log(yit ) + ∆ log(yit+1 )) = V ari (ηit )

Covi (∆ log(cit ), ∆ log(yit−1 ) + ∆ log(yit ) + ∆ log(yit+1 )) = Covi (∆ log(cit ), ηit )

and thus
Covi (∆ log(cit ), ηit )
φηt = 1 −
V ari (ηit )
Covi (∆ log(cit ), ∆ log(yit−1 ) + ∆ log(yit ) + ∆ log(yit+1 ))
= 1−
Covi (∆ log(yit ), ∆ log(yit−1 ) + ∆ log(yit ) + ∆ log(yit+1 ))
154 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

This expression has the same instrumental variable interpretation as above,

but now with ∆ log(yit−1 )+∆ log(yit )+∆ log(yit+1 ) as instrument for ∆ log(yit ).
Also note that now we need for subsequent income observations per house-
hold to estimate the consumption insurance coefficient for permanent shocks.
Next discuss empirical findings, also Blundell, Pistaferri and Saporta-
Eksten (2013)
Next discuss Kaplan and Violante (2011) for quantitative results in OLG
model

5.7.2 Cross-Sectional Implications: How Does Con-

sumption Inequality Evolve over the Life Cycle
Review Deaton-Paxson and Storesletten et al. (2004) and Blundell and Pre-
ston (1996)
rθt−1
ct = ct−1 + ut + vt
1+r
and thus
rθt−1 2
V ari (ct ) = V ari (ct−1 ) + σu + σv2
1+r
since by assumption Corri (ct−1 , ut ) = Corri (ct−1 , vt ) = Corri (vt , ut ) = 0.
With only permanent shocks consumption inequality should increase linearly
with the age of the cohort (and at the same rate as income inequality).

5.8 Appendix A: The Intertemporal Budget

Constraint
In this appendix we derive the intertemporal budget constraint (5.43) from
the sequence of budget constraints (5.17). For a specific node st the latter
reads as
at+1 (st )
ct (st ) + = yt (st ) + at (st−1 ). (5.185)
1+r
In the next period, for node st+1 it reads as

at+2 (st+1 )
ct+1 (st+1 ) + = yt+1 (st+1 ) + at+1 (st ) (5.186)
1+r
5.8. APPENDIX A: THE INTERTEMPORAL BUDGET CONSTRAINT155

πt+1 (st+1 |st )

Solving (5.186) for at+1 (st ) and multiplying by 1+r
yields:

at+2 (st+1 )
!
πt+1 (st+1 |st )at+1 (st ) ct+1 (st+1 ) + 1+r
− yt+1 (st+1 )
= πt+1 (st+1 |st )
1+r 1+r
(5.187)
Since at+1 (st ) is only a function of st (but not st+1 ), equation (5.187) holds
for every node st+1 following st . Thus we can sum (5.187) over all nodes
st+1 |st to obtain
at+2 (st+1 )
!
at+1 (st ) X πt+1 (st+1 |st ) X ct+1 (st+1 ) + 1+r
− yt+1 (st+1 )
= at+1 (st ) = πt+1 (st+1 |st )
1+r t+1 t
1 + r t+1 t
1+r
s |s s |s
(5.188)
Substituting (5.188) back into (5.185) yields
X πt+1 (st+1 |st )ct+1 (st+1 ) X πt+1 (st+1 |st )at+2 (st+1 )
ct (st ) + +
t+1 t
1+r t+1 t
(1 + r)2
s |s s |s
t+1 t t+1
X πt+1 (s |s )yt+1 (s )
= yt (st ) + + at (st−1 ) (5.189)
1+r
st+1 |st

Repeating this procedure until period T yields

T −t X
X πt+τ (st+τ |st )ct+1 (st+1 ) X πT (sT |st )aT +1 (sT )
ct (st ) + +
τ =1 t+τ t
(1 + r) τ
T t
(1 + r)T −t+1
s |s s |s
T −t X
X πt+τ (st+τ |st )yt+τ (st+τ )
= yt (st ) + τ
+ at (st−1 ). (5.190)
τ =1 t+τ t
(1 + r)
s |s

Thus as long as the term

X πT (sT |st )aT +1 (sT )
≥0 (5.191)
T t
(1 + r)T −t+1
s |s

we obtain equation (5.43). But equation (5.191) follows from the No-Ponzi
conditions imposed in (5.27) and (5.28). For T < ∞ this is trivial since
aT +1 (sT ) ≥ 0 for all sT immediately implies (5.191). For T = ∞ we note
that the constraint
156 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM

at+1 (st ) ≥ −Āt+1 (st )

together with the assumption that the process {Āt+1 (st )} is bounded above
implies that
X πT (sT |st )aT +1 (sT )
lim = 0.
T →∞
T t
(1 + r)T −t+1
s |s

5.9 Appendix B: Derivation of Consumption

Response to Income Shocks
Our goal in this appendix is to derive equation (5.48). Rewrite (5.46) as

rWt
θt ct = (5.192)
1+r
rWt−1
θt−1 ct−1 = . (5.193)
1+r

Expanding equation (5.192) by using the definition for Wt yields

T −t
rWt rat r X yt+s
θt ct = = + Et (5.194)
1+r 1 + r 1 + r s=0 (1 + r)s

Plugging in for at from the period t − 1 budget constraint

at = (1 + r) (at−1 + yt−1 − ct−1 ) (5.195)

yields
T −t
r X yt+s
θt ct = r (at−1 + yt−1 − ct−1 ) + Et (5.196)
1 + r s=0 (1 + r)s

Bringing the term rct−1 to the left and dividing both sides by (1 + r) delivers

T −t
θt ct + rct−1 r r X yt+s
= (at−1 + yt−1 ) + E t s (5.197)
1+r 1+r (1 + r)2 s=0 (1 + r)
5.9. APPENDIX B: DERIVATION OF CONSUMPTION RESPONSE TO INCOME SHOCKS 157

Subtracting θt−1 ct−1 from both sides, using (5.193) and the definition of Wt−1
one gets
T −t T −t+1
θt ct + rct−1 r X yt+s r r X yt−1+s
−θt−1 ct−1 = E t s + yt−1 − E t−1
1+r (1 + r)2 s=0 (1 + r) 1 + r 1+r s=0
(1 + r)s
(5.198)
But now we note that the left hand side of equation (5.198) equals:
 
θt ct + rct−1 θt ct 1 r
− θt−1 ct−1 = − 1− − ct−1
1+r 1+r (1 + r)T −t+2 1 + r
 
θt ct 1 1
= − 1+r− − r ct−1
1+r 1+r (1 + r)T −t+1
θt ∆ct
= (5.199)
1+r
Plugging (5.199) back into equation (5.198) and multiplying both sides by
(1 + r) finally yields
T −t −t+1
TX
r X yt+s yt−1+s
θt ∆ct = Et s + ry t−1 − rE t−1
1 + r s=0 (1 + r) s=0
(1 + r)s
T −t −t+1
TX
r X yt+s yt−1+s
= Et s − rE t−1
1 + r s=0 (1 + r) s=1
(1 + r)s
T −t −t+1
TX
r X yt+s r yt−1+s
= Et s − Et−1
1 + r s=0 (1 + r) 1+r s=1
(1 + r)s−1
T −t T −t
r X yt+s r X yt+s
= Et s − Et−1
1 + r s=0 (1 + r) 1+r s=0
(1 + r)s
T −t
r X (Et − Et−1 )yt+s
= = ηt
1 + r s=0 (1 + r)s

which is equation (5.48) in the main text.

158 CHAPTER 5. THE SIM IN PARTIAL EQUILIBRIUM
Chapter 6

A Digression: Stochastic
Earnings or Wage Processes

Before turning to general equilibrium versions of standard incomplete mar-

kets models we briefly want to discuss the specification of the key quantitative
ingredient into the model, the stochastic labor income process (or wage pro-
cess, if labor supply is a choice variable). The discussion in this section will
follow Guvenen (2007) who in turn builds upon papers by Liliard and Weiss
(1979), MaCurdy (1982), Abowd and Card (1989), Baker (1997), Haider
(2001), Meghir and Pistaferri (2004) and many others.
In most of the literature labor earnings or income eiht of individual or
household i with actual or potential labor market experience (or age) h at
time t is assumed to follow a process of the form

log(eiht ) = g(θt , Xit ) + f (γi , Xit ) + log(yit ) (6.1)

where θt , γi are parameter vectors, Xit are observable characteristics of house-

hold (individual) i at time t and log(yit ) is the stochastic part of earnings.
We will now discuss the exact specification of each of these components.

1. Deterministic determinants of earnings whose effects are not allowed to

vary across households, g(θt , Xit ): the parameter vector can potentially
change over time, but is restricted to be identical across households.
Typical observables Xit that are used in the literature are age and (po-
tential or actual) labor market experience of the individual/household,

159
160CHAPTER 6. A DIGRESSION: STOCHASTIC EARNINGS OR WAGE PROCESSES

dummies for family composition or education1 or birth cohort, or a

unit vector (so that this component includes time dummies). Guvenen
(2007) specifies this term as a cubic polynomial in potential experience
hit of the household

g(θt , Xit ) = θ0t + θ1t hit + θ2t h2it + θ3t h3it .

The key restriction is that the parameter vector θ is allowed to vary

with time t, but not with household identity i. Time dependence of
the constant θ0t allows for time fixed effects impacting earnings (e.g.
business cycle effects). The time varying coefficients on the experience
(or age) polynomial reflect potentially time varying returns to labor
market experience.

2. Household-specific deterministic determinants of the life cycle earnings

profile, f (γi , Xit ): the parameter vector γi is assumed to be constant
over time, but can potentially vary across households. There are two
main contenders for the form of f, nested in the general formulation
specified by Guvenen (2007):

f (γi , Xit ) = αi + βi hit (6.2)

where (αi , βi ) are household-specific parameters, drawn from a popu-

lation distribution that is normal with zero mean, variances σα2 , σβ2 and
covariance σαβ . There is almost no disagreement that one should allow
for σα2 > 0, that is, there are initial and permanent differences in the
levels of household lifetime earnings processes. Note, however, that
the more heterogeneity is already taken out of the data through part
1. (either education, family composition dummies, or separate analysis
for different group), the lower is the estimate of σα2 going to be. The
real controversy (and one whose resolution turns out to matter a lot for
the estimates of the entire earnings process, and thus the quantitative
properties of our models) is about whether one allows heterogeneity
across individuals/household in the slope of the life cycle earnings pro-
file, that is, whether one allows for σβ2 > 0. Early papers allowed for
σβ2 > 0 and estimated this variance to be significantly different from
zero. In a very influential paper MaCurdy (1982) argued for σβ2 = 0 and
1
Alternatively, the analysis is done separately for different groups, where group identity
is often determined by the level of education attained by the household.
 161

then estimated a stochastic earnings process of the form (6.1) impos-

ing this restriction. We will return to MaCurdy’s argument below, and
here just mention that most of the literature since MaCurdy followed
his lead (e.g. Abowd and Card (1989), Meghir and Pistaferri (2004),
Storesletten et al. (2004) and many others). Guvenen (2007) chal-
lenges this position and argues that the evidence that leads MaCurdy
to assume σβ2 = 0 is weak. He then allows for σβ2 > 0, estimates it to
be significantly positive and argues that omitting slope heterogeneity
may bias the estimates governing the stochastic part of the earnings
process log(yit ). So the first key question we will address is whether
one should allow for σβ2 > 0.

3. The stochastic part of the earnings process log(yit ): this is the part
we will approximate as a Markov chain below and take as input into
our quantitative models. An infinite horizon model abstracts from the
life cycle aspects of earnings as specified in part 1. and 2. Thus the
key model input in this class of models is the stochastic component of
earnings. Even in quantitative life cycle models where parts 1. and 2.
are modelled as well, the stochastic part of the earnings process is a key
ingredient of the model. Typically, the stochastic part of log-earnings
is modeled as the sum of a transitory and a persistent shock of the form

log(yit ) = zit + φt εit (6.3)

where
zit = ρzit−1 + πt ηit (6.4)
and the innovations are uncorrelated standard normal random vari-
ables, that are also uncorrelated with the (αi , βi ). Note that the persis-
tent part has to be initialized with some zi0 , often this initial condition
is simply set to zero. Thus the stochastic part of log-earnings is mod-
eled as a sum of a transitory shock with (potentially time varying)
variance φ2t and a persistent shock whose conditional variance is given
by πt2 and the persistence is given by the (time-invariant) parameter
ρ. Sometimes a low-order MA term is allowed for the transitory shock,
and some authors a priori impose ρ = 1, estimating the stochastic part
as a combination of a transitory and a fully permanent shock. The key
question for this part of the process is how persistent income shocks are,
that is, how big is ρ and how large is πt2 relative to φ2t . Since we know
162CHAPTER 6. A DIGRESSION: STOCHASTIC EARNINGS OR WAGE PROCESSES

from our previous discussion that more persistent shocks are harder to
self-insure against, the empirical answer to this question is evidently of
great importance for the quantitative properties of our models.

6.1 Estimation
The parameters to be estimated are the θt ’s, the variances and the covariance
of (αi , βi ) and the (φ2t , πt2 ) as well as ρ. In a first stage the θt ’s are estimated
by regressing individual or household log-earnings on observable household
characteristics, in Guvenen’s (2007) case a cubic polynomial in potential
experience hit . In a second stage the stochastic part of the earnings equation
is estimated (if αi , βi are random coefficients their variances are typically
estimated jointly with the remaining stochastic part of the process). While
there are various methods to do this, the most popular method is to use a
minimum distance estimator that, by choice of the parameters, minimizes
the weighted distance between specific cross-sectional moments measured in
the data and implied by the model.
The moments used in estimation are cross-sectional variances and auto-
covariances of log(eiht ), that is,

E(log(eiht )2 )
E(log(eiht ) log(eih+n,t+n ))

for n ≥ 1. For the data, E(.) denotes cross-sectional sample averages, for
the statistical model given jointly by (6.2)-(6.4) we can compute these cross-
sectional moments explicitly (alternatively one could simulate them easily).
Both in the data and in the model these moments could be computed sepa-
rately for different experience levels h, in practice small sample sizes in the
data leads most researchers to combine (i.e. take some average of) the data
for different experience groups. Guvenen (2007) provides the formulas for
these moments from the model and argues that from these moments the pa-
rameters of the statistical model are identified. In particular, the moments
from the model are given by

E(log(eiht )2 ) = σα2 + 2σαβ h + σβ2 h2 + φ2t + E(ziht

2
)
E(log(eiht ) log(eih+n,t+n )) = σα2 + 2σαβ (2h + n) + σβ2 h(h + n) + ρn E(ziht
2
)
2
where E(ziht ) in turn is a function of ρ and the πt2 .
6.2. RESULTS 163

6.2 Results
Guvenen (2007) estimates the process above on US PSID data from 1968 to
1993. The following table summarizes his results.

Estimates for Earnings Process from PSID

Sample ρ σα2 σβ2 corrαβ π̄ 2 φ̄2
0.988 0.058 0.015 0.061
All – –
(.024) (.011) (.007) (.010)
0.979 0.031 0.0099 0.047
College – –
(.055) (.021) (0.013) (.020)
0.972 0.053 0.011 0.052
High School – –
(.023) (.015) (.007) (.008)
0.821 0.022 0.00038 −0.23 0.029 0.047
All
(.030 (.074) (.00008) (.43) (.008) (.007)
0.805 0.023 0.00049 −0.70 0.025 0.032
College
(.061) (.112) (.00014) (1.22) (.015) (.017)
0.829 0.038 0.00020 −0.25 0.022 0.034
High School
(.029) (.081) (.00009) (.59) (.008) (.007)

The first three rows display the results if the stochastic earnings process is
estimated under the restriction that σβ2 = 0 (and consequently σαβ = 0). We
observe that with this restriction the persistent shock is very persistent as
the estimate of ρ is close to, and statistically indistinguishable from 1. This
result is very much consistent with the previous literature that has imposed
σβ2 = 0. There is substantial heterogeneity in the intercept of life cycle earn-
ings profiles (σ̂α2 is significantly different from zero) and this heterogeneity
is decreasing with educational attainment. Both the estimates of the size of
transitory and persistent shocks is consistent with the previous literature as
well (the table displays time averages of πt2 and φ2t ).
Once slope heterogeneity is allowed (the last three rows of the table), three
findings stand out. First, the estimated heterogeneity across households in
the slope of life cycle earnings profile is significantly different from zero in a
statistical sense, and is substantial from an economic point of view. Guvenen
(2007) calculated that of the entire cross-sectional earnings dispersion of a
cohort that has reached age 55, 75% of that variance for individuals with
a college degree and 55% of those with high school degree is accounted for
164CHAPTER 6. A DIGRESSION: STOCHASTIC EARNINGS OR WAGE PROCESSES

by heterogeneity in slopes of life cycle earnings profiles (rather than shocks

or initial differences in levels). Not surprisingly this share is higher for col-
lege graduates since they, according to the estimates from the table, face
significantly higher profile heterogeneity. Even though the estimates for σβ2
are small in magnitude, as long as earnings of households grow at perma-
nently different rates across households, by later ages in the life cycle these
different slopes generate substantial dispersion across household in earnings.
Second, if we allow for slope heterogeneity, the estimated degree of hetero-
geneity in earnings levels σα2 declines substantially and becomes statistically
insignificant. Third, the estimated persistence of the income process declines
massively, from a process that is essentially a random walk to a strongly
mean-reverting process.2 Again these findings are typical for the literature
that allows for slope heterogeneity in earnings.

Thus the results in the table essentially contrast two views of how the
stochastic earnings process of individuals or households look like. According
to what Guvenen (2007) calls the Restricted Income Profile (RIP) view, life
cycle earnings are characterized by substantial heterogeneity in initial levels,
no heterogeneity in slopes and highly persistent shocks over the life cycle. In
contrast, according to the Heterogeneous Income Profiles (HIP) view, the key
heterogeneity across households is in the slopes of life cycle earnings profiles
which are realized at labor market entry. Subsequent earnings shocks are
not very persistent (and thus, as we have seen above, should be quite easy to
self-insure against). We will now discuss further where the initial evidence
(mainly by MaCurdy, 1982) in favor of RIP came from and why σβ2 = 0 and
ρ ≈ 1 seem to go hand in hand, and σβ2  0 and ρ  1 seem to go hand in
hand.

2
The variances of the transitory component declines somewhat in size, while that of
the persistent component increases (but remember, the persistent component is not nearly
as persistent anymore). Also note that the correlation between αi and βi is estimated to
be negative (albeit insignificantly so). This may suggest that households face trade-offs
between careers with low initial earnings levels but subsequently higher earnings growth,
and those with higher levels and lower growth.
6.3. INTERPRETATION 165

6.3 Interpretation

6.3.1 RIP vs. HIP

Why did MaCurdy (1982) and many other following his lead conclude that
σβ2 = 0? Suppose for simplicity3 that the true process has φt = 0 (only persis-
tent shocks) and that the variance of the persistent shock is not time varying,
πt = π. Then the stochastic part of the earnings process becomes

log(êiht ) = αi + βi hit + zit

zit = ρzit−1 + πηit .

Now let us consider year-by-year earnings growth rates

∆ log(êiht ) = log(êiht ) − log(êih−1t−1 ) = βi + ∆zit ,

according to the model. Let us also assume that the process started at
t = −∞ so that
∞
X
zit = π ρt−τ ηiτ
τ =0

and hence

∆zit = zit − zit−1

= (ρ − 1)zit−1 + πηit
" ∞
#
X
= π (ρ − 1) ρt−1−τ ηiτ + ηit
τ =0

Therefore if we calculate the so-called autocovariance function between earn-

ings growth at time t and experience h and earnings growth n ≥ 2 periods

3
Simplicity means that the basic argument does not depend on these assumptions (but
the exact algebra does).
166CHAPTER 6. A DIGRESSION: STOCHASTIC EARNINGS OR WAGE PROCESSES

from now we find

Cov [∆ log(êiht ), ∆ log(êih+nt+n )]

= E [∆ log(êiht ) · ∆ log(êih+nt+n )]
" ∞
! ∞
!#
X X
= σβ2 + π 2 E (ρ − 1) ρt−1−τ ηiτ + ηit (ρ − 1) ρt+n−1−τ ηiτ + ηit+n
τ =0 τ =0
" ∞
#
X
= σβ2 + π 2 (ρ − 1)ρn−1 + (ρ − 1)2 ρn ρ2(t−1−τ )
τ =0
" ∞
#
X
= σβ2 − π 2 ρn−1 (1 − ρ) 1 − (1 − ρ)ρ ρ2(t−1−τ )
τ =0
2 n−1
π ρ (1 − ρ)
= σβ2 − . (6.5)
1+ρ
The key observation is that the second term exponentially declines with n, so
that for large enough leads n we should see empirically Cov [∆ log(êiht ), ∆ log(êih+nt+n )] >
0 if σβ2 > 0. Intuitively, if the slopes of life cycle earnings profiles vary across
individuals, in the cross sections earnings growth rates today and in the
future (at least the far future) have to be positively correlated.
When MaCurdy (and the subsequent literature) computed the autoco-
variance function from the data, he found that they were all negative for
n ≤ 10 (he did not compute it for longer leads, partially because the data
was not available in 1982), and insignificantly different from zero for n ≥ 2.
He interpreted this as inconsistent with σβ2 > 0. Furthermore the fact that
for n ≥ 2 the autocovariances are statistically zero requires, in light of (6.5),
that ρ = 1. The subsequent literature has then proceeded as MaCurdy: com-
pute the autocovariance function, see whether it all negative, conclude that
σβ2 = 0, and from it being zero at leads n ≥ 2 conclude that ρ ≈ 1 (which
comes out of the estimation when imposing σβ2 = 0 also). Note that in the
presence of transitory shocks the autocovariance function of the model is neg-
ative rather than zero for n = 1 even if ρ = 1, so the significantly negative
autocovariance function at lead n = 1 is easily rationalized with the presence
of transitory shocks..
Guvenen’s (2007) key argument against MaCurdy’s argument is the low
power of the test. he shows that with his empirical estimates for σβ2 > 0 the
first n for which the model autocovariance is significantly positive is large,
in the order of n = 12 or above. He also shows that the model-simulated
6.3. INTERPRETATION 167

autocovariances (with the estimated parameters) are not inconsistent with

the data. While Guvenen’s evidence is likely not the last word in this debate,
it casts some doubt on the arguments against HIP advanced so far in the
literature.

6.3.2 Size of ρ
The last point we want to discuss is that why, if the truth is HIP but we
estimate the earnings process under the restriction σβ2 = 0, we may end up
with an estimate of ρ that is biased upward. Take two households with the
same initial income and deterministic but different earnings growth

log(eit ) = α + βi t (6.6)

with β1 < β2 . Suppose the econometrician estimates

log(eit ) = ρ log(eit−1 ) + ηit . (6.7)

with log(ei0 ) = α/ρ. In the first period the econometrician observes

log(e11 ) = α + β1
log(e21 ) = α + β2

and has to interpret this, from the perspective of (6.7), as a negative shock
for household 1 and a positive shock for household 2. Since under the truth
(6.6) earnings of both households deviate more and more from the common
average over time, the econometrician has to interpret these larger and larger
deviations as repeated and very persistent negative shocks for household 1
and positive shocks for household 2. Guvenen’s (2007) figure 1 provides an
instructive visualization of this point. Thus when one estimates the earnings
process without allowing for slope heterogeneity when in fact it is present one
tends to find significantly higher values for ρ than when estimating it with σβ2
permitted. Note that the difference between an estimate (on annual basis)
of ρ = 0.98 (RIP) and ρ = 0.82 (HIP) is huge: under the RIP estimate after
20 years still 2/3’s of a shock is present in current earnings, whereas under
the HIP estimate only 2% of the shock 20 years ago is present. Obviously
self-insurance will do very well dealing with the latter shock, and not well at
all dealing with the former shock.
168CHAPTER 6. A DIGRESSION: STOCHASTIC EARNINGS OR WAGE PROCESSES
Chapter 7

The SIM in General

Equilibrium

In this section we will look at general equilibrium versions of the PILCH

model discussed in the last chapter. In general equilibrium the interest rate(s)
and real wage(s) are determined endogenously within the model. Three rea-
sons for considering general equilibrium come to my mind

1. It imposes theoretical discipline. As we saw above, the behavior of

consumption and saving over time depends crucially on the relative
size of the interest rate r and the time discount factor ρ. In the partial
equilibrium analysis so far both r and ρ were exogenous parameters,
to be chosen by the model builder. In general equilibrium, once ρ is
chosen, r will be determined endogenously. The relationship between ρ
and r shifts from being arbitrarily chosen to an equilibrium relationship;
as we will see, only a stationary equilibrium with r < ρ can exist in
these models.

2. It gives rise to an endogenously determined consumption and wealth

distribution and hence provides a theory of wealth inequality. To be
successful, this model should be able to reproduce stylized facts of
empirical wealth distributions. Note that it is not a theory of the
income distribution, as the stochastic income process is an input to the
model (and needs to be chosen by the model builder), rather than a
result of the model.

3. It enables meaningful policy experiments. Partial equilibrium models,

169
170 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

by keeping interest rates and wages fixed when analyzing the change in
household behavior due to changes in particular policies (tax reform,
social security reform, welfare reform), may over- or understate the full
effects of such a reform. A claim, e.g., that a reform of the social se-
curity system towards a system that invests more funds in the stock
market is welfare improving may ignore the fact that, once large ad-
ditional funds are invested in the stock market, the excess return of
stocks over bonds (or the population growth rate plus the growth rate
of productivity), may diminish or vanish altogether. Only in cases in
which one can reasonably expect relative prices to remain uninfluenced
by policy reforms (for a small open economy, say) or one can convinc-
ingly argue that price effects are quantitatively unimportant1 would a
partial equilibrium analysis yield unbiased results.

7.1 A Model Without Aggregate Uncertainty

7.1.1 The Environment
But let us leave the realm of ideology and describe the model. The economy
is populated by a continuum of measure 1 of individuals that all face an
income fluctuation problem of the sort described above. Each individual
has the same stochastic labor endowment process {yt }∞ t=0 where yt ∈ Y =
{y1 , y2 , . . . yN }. A households’ labor income in a particular period is given
by wt yt where wt is the real wage that one unit of labor commands in the
economy, which will be constant in a stationary equilibrium. Hence yt can
be interpreted as the number of efficiency units of labor a household can
supply in a given time period. The labor endowment process is assumed
to follow a stationary Markov process. Let π(y 0 |y) denote the probability
that tomorrow’s endowment takes the value y 0 if today’s endowment takes
the value y. We assume a law of large numbers to hold: not only is π(y 0 |y)
the probability of a particular agent of a transition from y to y 0 but also the
deterministic fraction of the population that has this particular transition.2
1
It is not clear to me, though, how one would arrive at such a conclusion without
actually first doing the general equilibrium analysis.
2
The validity of such a law of large numbers what subject to some debate in the 1980’s.
See Judd (1985), Feldman and Gilles (1985) and Uhlig (1996) for representative papers. I
will briefly discuss the main result in Feldman and Gilles (1985).
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 171

Let Π denote the stationary distribution associated with π, assumed to

be unique. We assume that at period 0 the income of all agents, y0 , is given,
and that the distribution of incomes across the population is given by Π.
Given our assumptions, then, the distribution of income in all future periods
is also given by Π. In particular, the total labor endowment in the economy
(in efficiency units) is given by
X
L̄ = yΠ(y) (7.2)
y

As before, let denote by πt (y t |y0 ) the probability of event history y t , given

initial event y0 . We have
πt (y t |y0 ) = π(yt |yt−1 ) ∗ . . . ∗ π(y1 |y0 ) (7.3)
Hence, although there is substantial idiosyncratic uncertainty about a par-
ticular individual’s labor endowment and hence labor income, the aggregate
What one would think and like to be true is the following: define as set of agents
I = (0, 1] with associated measure space (I, B(I), m) where m is the Lebesgue measure
(essentially measuring length or mass of an interval of agents) and B(I) is the Borel sigma
algebra. Define a probability space on which the joint stochastic income process for the
continuum of agents is defined as (Ω, F, P ). For simplicity we only consider the case where
all agents draw income in only one period. Finally define a continuum of random variables
as y(i, .) : Ω → {0, 1} where we assume for simplicity that income can only take the values
0 and 1 and does so with equal probability for each household. What one would hope for
is that for all sets G = (c, d] ∈ I and almost all ω ∈ Ω
Z
1
y(i, ω)m(di) = m(G). (7.1)
G 2
However, if one insists that the y(i, .) are pairwise independent for any two agents, then
Feldman and Gilles (1985) and Judd (1985) show that for all ω either y(., ω) is not m-
measurable or there exists a set G such that (7.1) does not hold. However, for our purposes
pairwise independence is not necessary: we just would like all random variables y(i, .) to
have a desired distribution Π and the we would like to have a population distribution over
incomes y(i, .) that equals Π for all ω, that is, for aggregates to be nonstochastic.
Luckily Feldman and Gilles show that for a probability space (Y, B(Y ), Π) there exists
a continuum of random variables y(i, .) : Ω → Y such that for all i the random variable is
distributed according to Π and that for all ω ∈ Ω and all D ∈ B(Y )
m(i ∈ I : {y(i, ω) ∈ D}) = Π(D)
that is, the population income distribution is nonstochastic and given by Π. Evidently, in
light of the first proposition by Feldman and Gilles the continuum of random variables
cannot be pairwise independent.
172 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

labor endowment and hence labor income in the economy is constant over
time, i.e. there is no aggregate uncertainty. Without this assumption there
would be no hope for the existence of a stationary equilibrium in which wages
w and interest rates r are constant over time. In the next section we will
consider a model that is similar to this one, but will include one source of ag-
gregate uncertainty, in addition to the idiosyncratic labor income uncertainty
present in this model.
Each agent’s preferences over stochastic consumption processes are given
by
X ∞
u(c) = E0 β t U (ct ) (7.4)
t=0

1
with β = 1+ρ and ρ > 0. As before the agent can self-insure against idiosyn-
cratic labor endowment shocks by purchasing at period t uncontingent claims
to consumption at period t + 1 at a price qt = 1+r1t+1 . Again, in a stationary
equilibrium rt+1 will be constant across time. The agents budget constraint
is given by
ct + at+1 = wt yt + (1 + rt )at (7.5)
Note that we slightly change the way assets are traded: instead of zero
1
coupon bonds being traded at a discount q = 1+r we now consider a bond
that trades at price 1 today and earns gross interest rate (1+rt+1 ) tomorrow.
We do this for the following reason: the asset being traded will be physical
capital, with the interest rate being determined by the marginal product of
capital. As long as the interest rate is constant as in Aiyagari (1994) both for-
mulations are equivalent (if you derive the corresponding Euler equations for
both formulations, you’ll find that they’re identical). With aggregate uncer-
tainty as in Krusell and Smith (1998), however, it would make a substantial
difference whether agents can trade a risk free bond or, as Krusell and Smith
assume, risky capital. Thus, in order to be consistent with their formulation,
I opted for changing the budget constraint. Evidently all theoretical results
from the last section about the income fluctuation problem still apply (as
long as the interest rate is nonstochastic), because the optimality conditions
for the households are identical across both formulations.
We impose an exogenous borrowing constraint on asset holdings: at+1 ≥
0. Aiyagari (1994) considers several alternative borrowing constraints, but
uses the no-borrowing constraint in his applications. The agent starts out
with initial conditions (a0 , y0 ). His consumption at period t after endowment
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 173

shock history y t has been realized is denoted by ct (a0 , y t ) and his asset hold-
ings by at+1 (a0 , y t ). Let Φ0 (a0 , y0 ) denote the initial measure over (a0 , y0 )
across households. In accordance with our previous assumption the marginal
distribution of Φ0 with respect to y0 is assumed to be Π. At each point of
time an agent is characterized by her current asset position at and her cur-
rent income yt . These are her individual state variables. What describes the
aggregate state of the economy is the cross-sectional distribution over individ-
ual characteristics Φt (at , yt ). This concludes the description of the household
side of the economy.
On the production side we assume that competitive firms, taking as given
wages wt and interest rates rt , have access to a standard neoclassical produc-
tion technology3
Yt = F (Kt , Lt ) (7.6)

where F ∈ C 2 features constant returns to scale and positive but diminishing

marginal products with respect to both production factors. We also assume
the Inada conditions to hold. Firms choose labor inputs and capital inputs
to maximize the present discounted value of profits. As usual with constant
returns to scale and perfect competition the number of firms is indetermi-
nate and without loss of generality we can assert the existence of a single,
representative firm. In this economy the only asset is the physical capital
stock. Hence in equilibrium the aggregate capital stock Kt has to equal the
sum of asset holdings of all individuals, i.e. the integral over the at ’s of all
agents. We assume that capital depreciates at rate 0 < δ < 1. The aggregate
resource constraint is then given as

Ct + Kt+1 − (1 − δ)Kt = F (Kt , Lt ) (7.7)

where Ct is aggregate consumption at period t. We are now ready to de-

fine an equilibrium. We first define a sequential markets equilibrium, then
a recursive equilibrium and finally a stationary recursive equilibrium. The
first definition is done in order to stress similarities and differences with the
complete markets model, the third because this is what we will compute nu-
merically, and the second is presented because the third is a special case, and
with aggregate uncertainty we will need the more general definition anyway.
3
In the next section aggregate uncertainty is introduced by incorportating stochastic
productivity into the aggregate production function, in the tradition of real business cycles.
174 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Definition 44 Given Φ0 , a sequential markets competitive equilibrium is al-

locations for households {ct (a0 , y t ), at+1 (a0 , y t )}∞
t=0,y t ∈Y t , allocations for the
representative firm {Kt , Lt }t=0 , and prices {wt , rt }∞
∞
t=0 such that

1. Given prices, allocations maximize (7.4) subject to (7.5) and subject to

the nonnegativity constraints on assets and consumption and the non-
negativity condition on assets

2.

rt = Fk (Kt , Lt ) − δ (7.8)
wt = FL (Kt , Lt ) (7.9)

3. For all t
Z X
Kt+1 = at+1 (a0 , y t )π(y t |y0 )dΦ0 (a0 , y0 ) (7.10)
y t ∈Y t
Z X
Lt = L̄ = yt π(y t |y0 )dΦ0 (a0 , y0 ) (7.11)
y t ∈Y t
Z X
ct (a0 , y t )π(y t |y0 )dΦ0 (a0 , y0 ) + Kt+1
y t ∈Y t
= F (Kt , Lt ) + (1 − δ)Kt (7.12)

The last three conditions are the asset market clearing, the labor market
clearing and the goods market clearing condition, respectively
Now let us define a recursive competitive equilibrium. We have already
conjectured what the correct state space is for our economy, with (a, y) being
the individual state variables and Φ(a, y) being the aggregate state variable.
First we need to define an appropriate measurable space on which the mea-
sures Φ are defined. Define the set A = [0, ∞) of possible asset holdings
and the set Y of possible labor endowment realizations. Define by P(Y ) the
power set of Y (i.e. the set of all subsets of Y ) and by B(A) the Borel σ-
algebra of A. Let Z = A × Y and B(Z) = P(Y ) × B(A). Finally define by M
the set of all probability measures on the measurable space M = (Z, B(Z)).
Why all this? Because our measures Φ will be required to be elements of M.
Now we are ready to define a recursive competitive equilibrium. At the heart
of any RCE is the recursive formulation of the household problem. Note that
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 175

we have to include all state variables in the household problem, in particular

the aggregate state variable, since the interest rate r will depend on Φ. Hence
the household problem in recursive formulation is
X
v(a, y; Φ) = max 0
u(c) + β π(y 0 |y)v(a0 , y 0 ; Φ0 ) (7.13)
c≥0,a ≥0
y 0 ∈Y

s.t. c + a0 = w(Φ)y + (1 + r(Φ))a (7.14)

Φ0 = H(Φ) (7.15)

The function H : M → M is called the aggregate “law of motion”. Now let

us proceed to the equilibrium definition.

Definition 45 A recursive competitive equilibrium is a value function v :

Z × M → R, policy functions for the household a0 : Z × M → R and
c : Z × M → R, policy functions for the firm K : M → R and L : M → R,
pricing functions r : M → R and w : M → R and an aggregate law of
motion H : M → M such that

1. v, a0 , c are measurable with respect to B(Z), v satisfies the household’s

Bellman equation and a0 , c are the associated policy functions, given r()
and w()

2. K, L satisfy, given r() and w()

r(Φ) = Fk (K(Φ), L(Φ)) − δ (7.16)

w(Φ) = FL (K(Φ), L(Φ)) (7.17)

3. For all Φ ∈ M
Z
K (Φ ) = K(H(Φ)) = a0 (a, y; Φ)dΦ
0 0
(7.18)
Z
L(Φ) = ydΦ (7.19)
Z Z
c(a, y; Φ)dΦ + a0 (a, y; Φ)dΦ (7.20)
= F (K(Φ), L(Φ)) + (1 − δ)K(Φ)

4. The aggregate law of motion H is generated by the exogenous Markov

process π and the policy function a0 (as described below)
176 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Now let us specify what it means that H is generated by π and a0 . H

basically tells us how a current measure over (a, y) translates into a measure
Φ0 tomorrow. So H has to summarize how individuals move within the
distribution over assets and income from one period to the next. But this is
exactly what a transition function tells us. So define the transition function
QΦ : Z × B(Z) → [0, 1] by4
X  π(y 0 |y) if a0 (a, y; Φ) ∈ A
QΦ ((a, y), (A, Y)) = (7.21)
0
0 else
y ∈Y

for all (a, y) ∈ Z and all (A, Y) ∈ B(Z). QΦ ((a, y), (A, Y)) is the probability
that an agent with current assets a and current income y ends up with
assets a0 in A tomorrow and income y 0 in Y tomorrow. Suppose that Y is a
singleton, say Y = {y1 }. The probability that tomorrow’s income is y 0 = y1 ,
given today’s income is π(y 0 |y). The transition of assets is non-stochastic
as tomorrows assets are chosen today according to the function a0 (a, y). So
either a0 (a, y) falls into A or it does not. Hence the probability of transition
from (a, y) to {y1 } × A is π(y 0 |y) if a0 (a, y) falls into A and zero if it does
not fall into A. If Y contains more than one element, then one has to sum
over the appropriate π(y 0 |y).
How does the function QΦ help us to determine tomorrow’s measure over
(a, y) from today’s measure? Suppose QΦ were a Markov transition matrix
for a finite state Markov chain and Φt would be the distribution today. Then
to figure out the distribution Φt tomorrow we would just multiply Q by Φt ,
or
Φt+1 = QTΦt Φt (7.22)
where here T stands for the transpose of a matrix. But a transition function
is just a generalization of a Markov transition matrix to uncountable state
spaces. Hence we need integrals:
Z
0
Φ (A, Y) = (H(Φ)) (A, Y) = QΦ ((a, y), (A, Y))Φ(da × dy) (7.23)

The fraction of people with income in Y and assets in A is that fraction of

people today, as measured by Φ, that transit to (A, Y), as measured by QΦ .
In general there no presumption that tomorrow’s measure Φ0 equals to-
day’s measure, since we posed an arbitrary initial distribution over types,
4
Note that, since a0 is also a function of Φ, Q is implicitly a function of Φ, too.
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 177

Φ0 . If the sequence of measures {Φt } generated by Φ0 and H is not constant,

then obviously interest rates rt = r(Φt ) are not constant, decision rules vary
with Φt over time, and the computation of equilibria is difficult in general.
We would have to compute the function H explicitly, mapping measures (i.e.
infinite-dimensional objects) into measures. This is exactly what we are going
to discuss in the next section. In this section we are interested in stationary
Recursive Competitive Equilibria.

Definition 46 A stationary recursive competitive equilibrium is a value func-

tion v : Z → R, policy functions for the household a0 : Z → R and c : Z → R,
policies for the firm K, L, prices r, w and a measure Φ ∈ M such that

1. v, a0 , c are measurable with respect to B(Z), v satisfies the household’s

Bellman equation and a0 , c are the associated policy functions, given r
and w

2. K, L satisfy, given r and w

r = FK (K, L) − δ (7.24)
w = KL (K, L) (7.25)

3.
Z
K = a0 (a, y)dΦ (7.26)
Z
L = ydΦ (7.27)
Z Z
c(a, y)dΦ + a0 (a, y)dΦ = F (K, L) + (1 − δ)K (7.28)

4. For all (A, Y) ∈ B(Z)

Z
Φ(A, Y) = Q((a, y), (A, Y))dΦ (7.29)

where Q is the transition function induced by π and a0 as described

above (not indexed by Φ anymore)
178 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Note the big simplification: value functions, policy functions and prices
are not any longer indexed by measures Φ, all conditions have to be satisfied
only for the equilibrium measure Φ. The last requirement states that the
measure Φ reproduces itself: starting with measure over incomes and assets
Φ today generates the same measure tomorrow. In this sense a stationary
RCE is the equivalent of a steady state, only that the entity characterizing
the steady state is not longer a number (the aggregate capital stock, say) but
a rather complicated infinite-dimensional object, namely a measure.5

7.1.2 Theoretical Results: Existence and Uniqueness

In this section we want to summarize what we know about the existence and
uniqueness of a stationary RCE. We first argue that the problem of existence
boils down to the question of whether one equation in one unknown, namely
the interest rate, has a solution. To see this, first note that by Walras’ law
one of the market clearing conditions is redundant; so let’s ignore the goods
market equilibrium condition. The labor market equilibrium is easy and
gives us L = L̄ and L̄ is exogenously given, specified by the labor endowment
process. It remains the asset market clearing condition
Z
K = K(r) = a0 (a, y)dΦ ≡ Ea(r) (7.30)

where Ea(r) are the average asset holdings in the economy. This condition
requires equality between the demand for capital by firms and the supply of
capital by households (last period’s demand for assets, with physical capital
being the only asset in the economy).
5
If we restrict attention to a finite set of capital stocks, A = {a1 , . . . , aM } then Φ is
an M ∗ N × 1 column vector and the Markov transition function Q is an M ∗ N × M ∗ N
matrix with generic element qij,kl giving the probability of going from (a, y) = (ai , yl )
to (a0 , y 0 ) = (ak , yl ) tomorrow. Using the convention that rows index states today and
columns index states tomorrow, an invariant measure Φ has to satisfy the matrix equation

Φ = QT Φ.

That is, Φ is the eigenvector (rescaled to have the sum of its rows equal to 1) associated
with an eigenvalue λ = 1 of the matrix QT . Since QT is a stochastic matrix (every row
sums to 1 and all entires are nonnegative) it has at least one unit eigenvalue and thus at
least one stationary measure Φ. But if QT has multiple unit eigenvalues there are multiply
statioanry measures (in fact, a continuum of them, since convex combinations of stationary
measures are themselves stationary measures.
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 179

¿From (7.24) it is clear that the capital demand of the firm K(r) is a
function of r alone, defined implicitly as

r = Fk (K(r), L̄) − δ (7.31)

since labor supply L = L̄ > 0 is exogenous. Furthermore we know from the

assumptions on the production function that K(r) is a continuous, strictly
decreasing function on r ∈ (−δ, ∞) with

lim K(r) = ∞ (7.32)

r→−δ
lim K(r) = 0 (7.33)
r→∞

It is obvious that the average capital supply (asset demand) by households

Ea(r) satisfies Ea(r) ∈ [0, ∞] for all r in (−δ, ∞). It is our goal to charac-
terize Ea(r). In particular, if Ea(r) is continuous and satisfies

lim Ea(r) < ∞ (7.34)

r→−δ
lim Ea(r) > 0 (7.35)
r→∞

then a stationary recursive competitive equilibrium exists. Furthermore, if

Ea(r) is strictly increasing in r (the substitution effects outweighs the income
effect), then the stationary RCE is unique.
The first thing we have to argue is that average capital supply (asset
demand) is in fact a function of r alone and that it is well defined on the
appropriate range for r. First we note that the wage rate w is a function
solely of r via
w(r) = FL (K(r), L̄) (7.36)
As r increases, K(r) decreases, the capital-labor ratio declines and with it the
wage rate. Hence, once r is known, the recursive problem of the household
can be solved, because all prices are known.
The first step is to ask under what conditions the recursive problem of
the household has a solution and under which conditions the support of asset
holdings is bounded from above (it is bounded from below by 0). Various
assumptions give us these results, but here are the most important ones
(note that Aiyagari (1994) does not prove anything for the case of serially
correlated income and assumes bounded utility, Huggett (1993) does, but
with additional assumptions on the structure of the Markov process).
180 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Proposition 47 (Huggett 1993) For ρ > 0, r > −1 and y1 > 0 and CRRA
utility with σ > 1 the functional equation has a unique solution v which is
strictly increasing, strictly concave and continuously differentiable in its first
argument. The optimal policies are continuous functions that are strictly
increasing (for c(a, y)) or increasing or constant at zero (for a0 (a, y))

Similar results can be proved for the iid case and arbitrary bounded U
with ρ > r and ρ > 0, see Aiyagari (1994).
With respect to the boundedness of the state space, as seen in the pre-
vious section we require ρ > r and additional assumptions. Under these
assumptions there exists an ā s.t. a0 (ā, yN ) = ā and a0 (a, y) ≤ ā for all
y ∈ Y and all a ∈ [0, ā] = A. From now on we will restrict the state space to
Y × A and it will be understood that Z = A × Y. Thus, under the maintained
assumptions by Huggett or Aiyagari there exists an optimal policy a0r (a, y)
indexed by r. The next step is to ask what more is needed to make aggregate
asset demand Z
Ea(r) = a0r (a, y)dΦr (7.37)

well-defined, where Φr has to satisfy

Z
Φr (A, Y) = Qr ((a, y), (A, Y))dΦr (7.38)

and Qr is the Markov transition function defined by ar as

X  π(y 0 |y) if a0 (a, y) ∈ A
r
Qr ((a, y), (A, Y)) = (7.39)
0
0 else
y ∈Y

Hence all is needed for Ea(r) to be well-defined is to establish that the

operator Tr∗ : M → M defined by
Z
∗
(Tr (Φ)) (A, Y) = Qr ((a, y), (A, Y))dΦ (7.40)

has a unique fixed point (that Tr∗ maps M into itself follows from SLP,
Theorem 8.2). To show this Aiyagari (in the working paper version, and
quite loosely described) draws on a theorem in SLP and Huggett on a similar
theorem due to Hopenhayn and Prescott (1992). In both theorems the key
condition is a monotone mixing condition that requires a positive probability
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 181

to go from the highest asset level ā to a intermediate asset level in N periods

and an evenly high probability to go from 0 assets to an intermediate asset
level also in N periods. More precisely stated, the theorem by Hopenhayn
and Prescott states the following. Define the order “ ≥ ” on Z as

z ≥ z 0 iff [(z1 ≥ z10 and z2 = z20 ) or (z 0 = c = (0, y1 )) or (z = d = (ā, yN ))

(7.41)
Under this order it is easy to show that (Z, ≥) is an ordered space, Z together
with the Euclidean metric is a compact metric space, ≥ is a closed order,
c ∈ Z and d ∈ Z are the smallest and the largest elements in Z (under order
≥) and (Z, B(Z)) is a measurable space. Then we have (see Hopenhayn and
Prescott (1992), Theorem 2)

Proposition 48 If

1. Qr is a transition function

2. Qr is increasing

3. There exists z ∗ ∈ Z, ε > 0 and N such that

P N (d, {z : z ≤ z ∗ }) > ε and P N (c, {z : z ≥ z ∗ }) > ε (7.42)

Then the operator Tr∗ has a unique fixed point Φr and for all Φ0 ∈ M
the sequence of measures defined by

Φn = (T ∗ )n Φ0 (7.43)

converges weakly to Φr

Here P N {z, Z) is the probability of going from state z to set Z in N

steps. Instead of proving this result (which turns out to be quite tough) we
will explain the assumptions, heuristically verify them and discuss what the
theorem delivers for us. Assumption 1 requires that Qr is in fact a transition
function, i.e. Qr (z, .) is a probability measure on (Z, B(Z)) for all z ∈ Z and
Qr (., Z) is a B(Z)-measurable function for all Z ∈ B(Z). Given that a0 (a, y)
is a continuous function, the proof of this is not too hard. The assumption
that Qr is increasing means that for any nondecreasing function f : Z → R
we have that Z
(T f ) (z) = f (z 0 )Qr (z, dz 0 ) (7.44)
182 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

is also nondecreasing. The proof that Qr satisfies monotonicity is straight-

forward, given that a0 (a, y) is increasing in both its arguments6 , so that
bigger z’s make Qr (z, .) put more probability mass on bigger z 0 . Together
with f being nondecreasing the result follows. Finally, why is the monotone
mixing condition 3. satisfied? Pick (this is going to be very heuristic here)
z ∗ = ( 21 (a0 (0, yN ) + ā) , y1 ). Starting at d, with a sequence of bad shocks y1
one converges to assets 0 monotonically and from c one converges to ā with
a sequence of good shocks yN . Hence with probability bigger than zero one
goes with finite steps N1 or less from d to something below z ∗ and with finite
steps N2 or less from c to something above z ∗ . Take N = max{N1 , N2 }.
The conclusion of the theorem then assures the existence of a unique
invariant measure Φr which can be found by iterating on the operator Tr∗ .
Convergence is in the weak sense, that is, a sequence of measures {Φn } con-
verges weakly to Φr if for every continuous and bounded real-valued function
f on Z we have Z Z
lim f (z)dΦn = f (z)dΦr (7.45)
n→∞

The argument in the preceding section demonstrated that the function

Ea(r) is well-defined on r ∈ [−δ, ρ). Since a0r (a, y) is a continuous function
jointly in (r, a), see SLP, Theorem 3.8 and Φr is continuous in r (in the
sense of weak convergence), see SLP, Theorem 12.13, the function Ea(r) is
a continuous function of r on [−δ, ρ). Note that the real bottleneck in the
argument is in establishing an upper bound of the state space for assets, as
Z needs to be compact for the theorem by Hopenhayn and Prescott to work.
To bound the state space an interest rate r < ρ is needed, as the previous
sections showed.
The remaining things to be argued are that limr→−δ Ea(r) < ∞ and
limr→ρ Ea(r) > K(ρ). The first condition is obviously satisfied whenever a
bound on the state space for assets, ā, can be found. So this part of the
problem is easily resolved. We also know from the previous section that for
ρ = r all consumers accumulate infinite assets eventually, so that, loosely
speaking Ea(ρ) = ∞. What is asserted (but not proved) by Aiyagari is that
as r approaches ρ from below, Ea(r) goes to ∞ by some form of continuity
argument (but note that, strictly speaking, Ea(ρ) is not well-defined).
This is a so far missing step in the existence proof. Otherwise we have es-
6
For a0 (a, y) to be increasing in y we need to assume that the exogenous Markov chain
does not feature negatively correlated income.
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 183

tablished that both K(r) and Ea(r) are continuous functions on r ∈ (−δ, ρ),
that for low r we have that Ea(r) < K(r) and for high r < ρ we have
Ea(r) > K(r). These arguments together then guarantee the existence of
r∗ such that
K(r∗ ) = Ea(r∗ ) (7.46)
and hence the existence of a stationary recursive competitive equilibrium.
With respect to uniqueness the verdict is negative, as we can’t prove the
monotonicity of the function Ea(r) which is due to offsetting income and
substitution effects of saving with respect to the interest rate directly, as
well as the indirect effect that the interest rate has on the wage rate. Even
harder is the question about the stability of the stationary equilibrium. We
know that, provided the economy starts with initial distribution Φ0 = Φr∗ ,
then the economy remains there forever. The question arises whether, for
an arbitrary initial distribution Φ0 6= Φr∗ it is the case that Φt → Φr∗ and
rt → r∗ (either locally or globally). Note that to assess this question one has
to examine either the sequential equilibrium or the law of motion H in the
full-blown RCE. To the best of my knowledge no stability result has been
established for these types of economies as of today.

7.1.3 Computation of the General Equilibrium

Since finding a stationary RCE really amounts to finding an interest rate r∗
that clears the asset market, consider the following algorithm

1. Fix an r ∈ (−δ, ρ). For a fixed r we can solve the household’s recursive
problem (e.g. by value function iteration or policy function iteration).
This yields a value function vr and decision rules a0r , cr , which obviously
depend on the r we picked.

2. The policy function a0r and π induce a Markov transition function Qr .

Compute the unique stationary measure Φr associated with this tran-
sition function from (7.29)

3. Compute excess demand for capital

d(r) = K(r) − Ea(r) (7.47)

Note that d(r) is just a number. If this number happens to equal

zero, we are done and have found a stationary RCE. If not we update
184 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

our guess for r and start from 1. anew. If Ea(r) is increasing in r

(which means that the substitution effect outweighs the income effect
and potential indirect income effects from changes in wages) we know
that d(r) is a strictly decreasing function. The updating should then
be such that if d(r) > 0, increase the guess for r, if d(r) < 0 decrease
the guess for r.

7.1.4 Qualitative Results

One general qualitative result of these models is excess saving and the over-
accumulation of capital, compared to a complete markets model in which a
full array of insurance contracts against idiosyncratic income uncertainty is
at the disposal of households. Remember that equilibria in this model are
Pareto efficient, from which it readily follows that the stationary equilibrium
in the present model (or any equilibrium, for that matter) is suboptimal (in
an ex-ante welfare sense, of course). As seen in Chapter 2, without aggregate
uncertainty consumption of all agents would be constant in any steady state
of a complete markets model. From the complete markets Euler equation
it follows right away that the stationary equilibrium interest rate with com-
plete markets satisfies rCM = ρ. But this means, since r∗ < ρ that the steady
state capital stock under complete markets satisfies K CM < K ∗ , i.e. there is
overaccumulation of capital in this economy, compared to the first-best. Put
another way, since total gross investment (and total saving) in the steady
state of both economies equal

S CM = δK CM (7.48)
S ∗ = δK ∗ > δK CM = S CM (7.49)

i.e. agents in this model oversave because of precautionary reasons. Only

potentially binding liquidity constraints are needed for this result (in non of
the discussion above did we ever assume Uccc > 0), with prudence strength-
ening the quantitative importance of the result. One of the main objectives
of Aiyagari’s quantitative analysis is to assess whether this excess precau-
tionary saving is quantitatively important, i.e. whether PILCH people, in
general equilibrium, accumulate significantly too many assets compared to
the benchmark complete markets model. As metric for this the aggregate
savings rate and real interest rate is used. Note for future reference that the
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 185

aggregate saving rate in a stationary equilibrium is given by

S K
s= =δ (7.50)
Y Y

7.1.5 Numerical Results

Calibration
Before solving the model Aiyagari has to specify functional forms of prefer-
ences and technology. Period utility is assumed to be CRRA and Aiyagari
experiments with values σ = {1, 3, 5}. The model period length is taken to
be one year, and he picks ρ = 0.0416 (β = 0.96). For the complete mar-
kets model this yields an (empirically reasonable) annual real interest rate of
4.16%, and we know that in the present model the interest rate will be lower.
The aggregate production function is assumed to take Cobb-Douglas form
with share parameter α = 0.36. Since α equals the capital share of income
and capital income as share of GDP is roughly 36% on average for US postwar
data (the number is somewhat sensitive to how the public capital stock and
consumer durables are treated). The annual depreciation rate is set at δ =
8%, values usually used in these types of calibration exercises (Cooley and
Prescott (1995) provide a careful description of the calibration procedure for
a standard RBC model).
Compared to standard RBC theory the only nonstandard input the model
requires is the stochastic labor income process, where we note that labor
income is just labor endowment scaled by a constant. Aiyagari poses the
following AR(1) process for the natural logarithm of labor earnings

21
log(yt+1 ) = θ log(yt ) + σε 1 − θ2 εt+1 (7.51)

Note that for this process we find

Cov(log(yt+1 ), log(yt ))
corr(log(yt+1 ), log(yt )) = p = θ(7.52)
V ar(log(yt )) ∗ V ar(log(yt+1 ))
V ar(log(yt+1 )) = σε2 (7.53)

For this income process log(yt ) can take any value in (−∞, ∞). Our theory
developed so far has assumed a finite state space for yt ∈ Y or equiva-
lently Y log for log(yt ). The transformation of processes with continuous state
space into finite state Markov chains was pioneered in economics by George
186 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Tauchen (1986) and roughly goes like this. First we pick the finite set Y log .
Aiyagari picks the number of states to be 7. Since log(yt ) can take any value
between (−∞, ∞), first subdivide the real line into 7 intervals

5
I1 = (−∞, − σε )
2
5 3
I2 = [− σε , − σε )
2 2
3 1
I3 = [− σε , − σε )
2 2
1 1
I4 = [− σε , σε )
2 2
1 3
I5 = [ σε , σε )
2 2
3 5
I6 = [ σε , σε )
2 2
5
I7 = [ σε , ∞) (7.54)
2
and take as state space for log-income the set of “midpoints” of the intervals

Y log = {−3σε , −2σε , −σε , 0, σε , 2σε , 3σε } (7.55)

The matrix of transition probabilities is then determined as follows. Fix a

state si = log(y) ∈ Y log today. The conditional probability of a particular
state sj = log(y 0 ) ∈ Y log tomorrow is then computed by

(x−θsi )2
Z − 2
0 e 2σy

π(log(y ) = sj | log(y) = si ) = √ dx (7.56)

Ij 2πσy
1
where σy = σε (1 − θ2 ) 2 . That is, we integrate the Normal distribution with
mean log(y) and variance σε2 (1 − θ2 ) over the interval Ij . Doing this for all
states today and tomorrow (this integration has to be done either numeri-
cally or one has to use tables for the pdf of a Normal distribution) yields
the transition matrix π. Now we can find the stationary distribution of π,
hopefully unique, by solving the matrix equation

Π = πT Π (7.57)
7.1. A MODEL WITHOUT AGGREGATE UNCERTAINTY 187

for Π. Given that the states for log(y) are given by Y log , we find the state
space for levels of income as

Ỹ = {e−3σε , e−2σε , e−σε , 1, eσε , e2σε , e3σε } (7.58)

Now we compute the average labor endowment as

X
ȳ = yΠ(y) (7.59)
y∈Ỹ

and normalize all states by ȳ to arrive at

Y = {y1 , . . . , y7 } (7.60)
e−3σε e−2σε e−σε 1 eσε e2σε e3σε
{ , , , , , , }
ȳ ȳ ȳ ȳ ȳ ȳ ȳ
and now the average (and aggregate) labor endowment equals
X
yΠ(y) = 1. (7.61)
y∈Y

Of course, this normalization of aggregate labor supply to 1 only fixes the

units of account for this economy and thus is completely innocuous. How
good the approximation of the continuous process with its finite discretization
is depends on the parameter values that one chooses for the persistence θ and
dispersion σε (and of course on the number of states allowed in the discretiza-
tion). Although Aiyagari’s Table 1 seems to indicate that the approximation
works fine for the parameter values he considers, it is my experience that for
high values of θ (close to 1) the discretization has a harder and harder time
getting the persistence correct, in addition to the dispersion being somewhat
off already for θ = 0.9.
Aiyagari computes stationary equilibria for parameter values in the range

θ ∈ {0, 0.3, 0.6, 0.9} (7.62)

σε ∈ {0.2, 0.4} (7.63)

Quite a few labor economists these days believe, however, that the persistence
parameter should be higher than θ = 0.9, closer to 1 (what happens if θ = 1?).
But then there is substantial debate whether the simple AR(1) process for
the log of labor earnings is the appropriate stochastic process to model this
variable. Instead let us turn to the results.
188 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Results
First note that for the Cobb-Douglas production function and L̄ = 1 we have
Y = K α and
r + δ = αK α−1 (7.64)
Thus
αY αδ
r+δ = = (7.65)
K s
where s is the aggregate saving rate. Thus

αδ
s= (7.66)
r+δ

and there is a one-to-one mapping between equilibrium interest rates r and

equilibrium aggregate savings rates s. For the benchmark case of complete
markets we have rCM = ρ = 4.16% and thus, given the other parameters
chosen, s = 23.7%. Aiyagari’s Table II contains aggregate savings rates and
interest rates for the incomplete markets general equilibrium model under
various assumptions about risk aversion (also prudence) σ, persistence θ and
dispersion σε . The most important findings are:

• Keeping prudence and dispersion fixed, an increase in the persistence

of the income shock leads to increased precautionary saving and bigger
overaccumulation of capital, compared to the complete markets bench-
mark.

• Keeping fixed persistence and dispersion in income, an increase in pru-

dence σ leads to more precautionary saving and more severe overaccu-
mulation of capital

• Keeping prudence and income persistence constant, an increase in the

dispersion of the income process leads to more precautionary saving
and more severe overaccumulation of capital.

• Quantitatively, the biggest effect on precautionary saving seem to stem

from the persistence parameter. In particular, for high income persis-
tence (and high dispersion) the aggregate saving rate is 14 percentage
points higher than in the complete markets benchmark.
7.2. AN INCOMPLETE MARKETS MODEL WITH UNSECURED DEBT AND EQUILIBRIUM DEF

The other implications of the model that Aiyagari discusses, the welfare
cost of idiosyncratic income fluctuations (as well as the welfare benefit from
access to self-insurance), as well as the predictions of the model with respect
to the wealth and consumption distribution will be addressed once the model
is enriched by aggregate fluctuations.

7.2 An Incomplete Markets Model with Un-

secured Debt and Equilibrium Default
Basic idea: incorporate household default on uncollateralized debt into stan-
dard SIM model. Institutional motivation: chapter 7 of U.S. bankruptcy
code. Discussion based on simplified version of Chatterjee et. al., (EC,
2007). Focus is on explaining main mechanism of the model, not to present
the most general version of Chatterjee et al. (2007).

7.2.1 Model Overview

• Starting point: standard SIM GE model as in Aiyagari (1994).

• Production firms completely standard.

• Continuum (of measure 1) of infinitely lived households solve standard

income fluctuation problem, can borrow at a loan-size dependent inter-
est rate schedule, can default.

• Competitive financial intermediaries extend loans, break even after ac-

counting for losses from default.

• Loan interest schedule determined endogenously in equilibrium.

7.2.2 Institutional Details of Bankruptcy

• Bankruptcy flag h ∈ {0, 1}. Summarizes past default behavior. h is a
state variable.

• Current bankruptcy decision d ∈ {0, 1}.

• Consequence of d = 1: Can’t save (and of course can’t borrow) in

current period.
190 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

• Consequence of h = 1:

– Can’t borrow.
– Fraction γ ∈ (0, 1) of income lost.

State Transitions

• If h = 0, then household can choose d ∈ {0, 1}. If d = 0, then h0 = 0.

If d = 1, then h0 = 1.

• If h = 1, then d ∈ {0}. With probability λ, h0 = 1. With probability

1 − λ, h0 = 0.

7.2.3 Household Problem in Recursive Formulation

• Individual State variables (a, y, h).

• Assets a ∈ A = R.

• Exogenous idiosyncratic income y ∈ Y = [ymin , ymax ], iid over time,

identically distributed across households, with measure π(.). Assume
ymin > 0 and ymax < ∞ and a law of large numbers.

• Bankruptcy flag h ∈ H = {0, 1}.

• Population distribution over states (a, y, h) described by measure Φ

over measurable space (Z, B(Z)) where Z = A × Y × Z and B(Z) =
B(A) × B(Y ) × P(Z).

Household Budget Sets

• Household take as given a wage w and loan price function q(a, y, h; a0 ).

• Households choose consumption c, assets (loans if negative) a0 as well

as d.

• Budget set of admissible (c, a0 ) pairs depends on (a, y, h) and on d ∈

{0, 1}.
7.2. AN INCOMPLETE MARKETS MODEL WITH UNSECURED DEBT AND EQUILIBRIUM DEF

• If h = 0 and d = 0:
B(a, y, h = 0; d = 0) = {c ≥ 0, a0 ∈ A :
c + q(a, y, h; a0 )a0 ≤ wy + a}.
Note that for some (a, y) ∈ A × Y this set might be empty (and the
household will need to default).
• If h = 0 and d = 1:
B(a, y, h = 0; d = 1) = {c ≥ 0, a0 = 0 : c ≤ wy}.

• If h = 1:
B(a, y, h = 1) = {c ≥ 0, a0 ≥ 0 :
c + q(a, y, h; a0 )a0 ≤ (1 − γ)wy + a}

Dynamic Programming Problem of Households

• Value function v(a, y, h).
• For h = 0: If B(a, y, h = 0; d = 0) = ∅, then “involuntary default and
 Z 
0 0 0
v(a, y, h = 0) = u(wy) + β v(0, y , h = 1)dπ(y )

If B(a, y, h = 0; d = 0) 6= ∅, then decision between repayment and

“voluntary” (strategic) default:
v(a, y, h = 0) = max
d∈{0,1}

max(c,a0 )∈B(a,y,h=d=0) Ru(c) + β v(a0 , y 0 , h0 = 0)dπ(y 0 )

R

u(wy) + β v(0, y 0 , h0 = 1)dπ(y 0 )

• For h = 1:
v(a, y, h = 1) = max
(c,a0 )∈B(a,y,h=1;d=0)
0 0 0 0
  R 
λ v(a , y , h = 1)dπ(y )+
u(c) + β
(1 − λ) v(a0 , y 0 , h0 = 0)dπ(y 0 )
R

• Solution is a value function v(a, y, h) and optimal policy functions

c(a, y, h), a0 (a, y, h) and d(a, y, h).
192 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

7.2.4 Production Firms

• Measure 1 of identical, competitive firms.
• Constant returns to scale technology F (K, L)
• Profit maximization implies
w = FL (K, L)
r = FK (K, L) − δ

7.2.5 Financial Intermediaries

• Representative financial intermediary owns the capital stock K and
buys new capital K 0 . Earns net return r on capital.
• They issue loans (deposits). Perfect competition loan by loan type
(a, y, h; a0 )
• Consider a loan of size a0 < 0 to a type (a, y, h) household (if a0 ≥ 0
call it a deposit).

– Price q(a, y, h; a0 ). If a0 < 0, intermediary gives household q(a, y, h; a0 )a0

today for promise to repay a0 tomorrow.
– (Equilibrium) default probability p(a, y, h; a0 )
– Number n(a, y, h; a0 ) of loans of type (a, y, h; a0 ).

• Maximization problem
Z
max0 n(a, y, h; a0 ) ∗
n(a,y,h;a )≥0 (a,y,h;a0 )
[1 − p(a, y, h; a0 )] a0
 
0 0
q(a, y, h; a )a −
1+r

7.2.6 Stationary Recursive Competitive Equilibrium

Definition 49 A Stationary RCE is a value function v(a, y, h) and policy
functions c(a, y, h), a0 (a, y, h), d(a, y, h) for the households, aggregate capital
and labor (K, L), quantities of loan/deposit contracts n(a, y, h; a0 ), default
frequencies p(a, y, h; a0 ), loan prices q(a, y, h; a0 ), factor prices (w, r) and a
probability measure Φ such that:
7.2. AN INCOMPLETE MARKETS MODEL WITH UNSECURED DEBT AND EQUILIBRIUM DEF

1. Given (w, q), v solves the household Bellman equation and c, a0 , d are
the optimal policy functions.

2. Given (w, r), (K, L) satisfy the production firms’ first order condition.

3. Given (q, p, r), n solves the financial intermediary’s problem.

4. Consistency of default probabilities: For all (a, y, h; a0 )

Z
0
p(a, y, h; a ) = d(a0 , y 0 , h0 = 0)π(y 0 |y)dy 0

5. Loan markets clearing: For all (a, y, h; a0 )

n(a, y, h; a0 ) = 1{a0 (a,y,h)=a0 } Φ(a, y, h)

6. Labor market clearing:

Z Z
L= ydΦ = ydπ

7. Capital market clearing:

Z Z
K= q(a, y, h; a0 )a0 n(a, y, h; a0 )da0 dΦ

8. Goods market clearing:

Z Z
c(a, y, h)dΦ + δK = F (K, L) − γw yΦ(da, dy, h = 1)

9. Φ is an invariant probability measure of the associated Markov transi-

tion function induced by π, a0 and d.

Remark 50 Paper proves existence of Stationary RCE by restricting a0 ∈ A,

with A a finite set.
194 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

7.2.7 Characterization of Household Default Decision

• Assume that a household indifferent between defaulting or nor decides
to default (CCNR call this the maximum default set). Recall: default
policy d(a, y, h).
• Along h-dimension: for h = 1, default is not an option. For h = 0,
there is a choice.
• Along the y-dimension: for a given loan level a, define as default set
D(a) = {y ∈ Y : d(a, y, h = 0) = 1}
Proposition: D(a) is either empty or a closed interval.
• Along the a-dimension:

– If a ≥ 0, then D(a) is empty [it is never optimal to “default” on

positive assets].
– Let ã < a < 0, then D(a) ⊆ D(ã) [default set is expanding in
indebtedness].

• Implied default probability on a loan a0 for a type (a, y, h = 0):

p(a, y, h; a0 ) = π(D(a0 )) = p(a0 )
Proposition: Since y 0 is iid, the default probability is just a function of
loan size a0 . Furthermore p(a0 ) = 0 for all a0 ≥ 0.

7.2.8 Characterization of Equilibrium Loan Interest Rate

Function
• Intermediaries’ problem
X
max0 n(a, y, h; a0 ) ∗
n(a,y,h;a )≥0
(a,y,h;a0 )

[1 − p(a, y, h; a0 )] a0
 
0 0
q(a, y, h; a )a −
1+r
X
= max0 n(a, y, h; a0 )a0 ∗
n(a,y,h;a )≥0
(a,y,h;a0 )

[1 − p(a, y, h; a0 )]
 
0
q(a, y, h; a ) −
1+r
7.2. AN INCOMPLETE MARKETS MODEL WITH UNSECURED DEBT AND EQUILIBRIUM DEF

• Thus if n(a, y, h; a0 ) > 0 then

1 − p(a, y, h; a0 )
q(a, y, h; a0 ) =
1+r
and we assume that this is also the price for contracts not traded in
equilibrium (for n(a, y, h; a0 ) = 0).
The previous characterization of the household default set implies that (see
their theorem 6):
1. Since p(a, y, h; a0 ) = p(a0 ) we have q(a, y, h; a0 ) = q(a0 ).
2. Since for all a0 ≥ 0 we have p(a0 ) = 0, for those a0
1
q(a0 ) =
1+r
3. Since for ã0 < a0 < 0, p(ã0 ) > p(a0 ) we have
q(ã0 ) ≤ q(a0 )
that is, loan interest rates are increasing in loan size.
4. There exists ā0 < 0 small enough such that
q(ā0 ) = 0.
The loan size ā0 is an effective borrowing limit.
• Augmenting the Model: Constant probability of death ρ.
• Introduce persistent type s of households. Assume that s follows Markov
chain with transitions π(s0 |s).
• Make income persistent by assuming and y ∼ πs (y). Interpret s as
partially capturing socioeconomic characteristics: model blue collar vs.
white collar households.
• In model households default because of bad earnings shocks. In data
they also default because of

– Large medical bills: introduce nondiscretionary health spending

shocks ζ(s).
– Divorce: introduce preference shocks η(s).
196 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

7.2.9 Bringing the Model to the Data: Calibration and

Estimation
• Challenge is to simultaneously account for high debt levels and high
default probabilities. Medical and divorce shocks necessary.

• Calibration to

– Aggregate statistics: standard since production side is neoclassical

growth model.
– Earnings and wealth distributions: choose the appropriate earn-
ings process.
– Debt and bankruptcy statistics: choose medical and divorce shocks
appropriately.

7.2.10 Quantitative Predictions of the Model

• Table 1: Reasons for filing for bankruptcy in data.

• Table 2: Extended version of the model is consistent with incidence of

bankruptcy filings, average debt of those in debt.

• Figure 3: Model matches U.S. wealth distribution well.

• Figure 5, table 5: who defaults?

• Figure 6: implied loan prices (interest rates).

Add private information (Narajabad, Athreya et al.)

Long term contracts (Drozd and Nosal)
Collateralized debt (mortgages) and default (foreclosures): Jeske and
Krueger, Garriga and Schlagenhauf

7.3 Unexpected Aggregate Shocks and Tran-

sition Dynamics
In this section we consider hypothetical thought experiments of the following
form. Suppose the economy is in a stationary equilibrium, with a given
7.3. UNEXPECTED AGGREGATE SHOCKS AND TRANSITION DYNAMICS197

government policy and all other exogenous elements that define preferences,
endowments and technology. Now, unexpectedly, either government policy
or some exogenous elements of the economy (such as the labor productivity
process) change. This change was completely unexpected by all agents of the
economy (a zero probability event), so that no anticipation actions were taken
by any agent. The exogenous change may be either transitory or permanent;
for the general discussion to follow this does not make a difference. We
want to study the transition path induced by the exogenous change, from
the old stationary equilibrium to a new stationary equilibrium (which may
coincide with the old stationary equilibrium in case the exogenous change is
of transitory nature, or may differ from it in case the exogenous change is
permanent.
For concreteness, suppose that the economy to start with is the standard
Aiyagari economy we studied in the previous section. As an example, we
consider as exogenous unexpected change the sudden permanent introduc-
tion of a capital income tax at rate τ. The receipts are rebated lump-sum
to households as government transfers T. The initial policy is characterized
by τ = T = 0. Obviously, since a tax on capital income changes households’
savings decisions, we expect that, due to the policy change, individual be-
havior and thus aggregate variables such as the interest rate, wage rate and
the capital stock change. We would also hope that, over time, the economy
settles down to its new stationary equilibrium associated with the capital
income tax. But since the economy starts, pre-reform, with an aggregate
state (aggregate capital, wealth distribution) not equal to the final station-
ary equilibrium, one would expect that it requires time for the economy to
settle down to its new stationary equilibrium. In other words, there will be
a nontrivial transition path induced by the reform.

7.3.1 Definition of Equilibrium

We now want to define an equilibrium that allows for such transition path
and then outline how one would possibly compute such a transition path. As
should be clear from the previous discussion that no analytical characteriza-
tion of the transition path is available in general, so that we have to rely on
computational analysis.
Since, on the aggregate level, the transition path is characterized by a
deterministic sequence of prices, quantities and distributions we will cast
the definition and solution of the model in sequential notation, with the
198 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

household decision problem still being formulated recursively. Let Z = Y ×

R+ be the set of all possible (yt , at ). Let B(R+ ) be the Borel σ-algebra of
R+ and P(Y ) be the power set of Y . Finally let B(Z) = P(Y ) × B(R+ ) and
M be the set of all finite measures on the measurable space (Z, B(Z)).
First let’s write down the household problem
X
vt (a, y) = max 0
u(c) + β π(y 0 |y)vt+1 (a0 , y 0 )
c≥0,a ≥0
y 0 ∈Y
0
s.t. c + a = wt y + (1 + (1 − τt )rt )a + Tt

Note that value functions are now functions of time also, since aggregate
prices and policies may change over time.

Definition 51 Given the initial capital stock K0 and initial distribution Φ0 ,

and fiscal legislation {τt }∞
t=0 a competitive equilibrium is a sequence of in-
dividual functions for the household {vt , ct , at+1 : Z → R}∞ t=0 , sequences of
production plans for the firm {Lt , Kt }t=0 , factor prices {wt , rt }∞
∞
t=0 , govern-
ment transfers {Tt }∞t=0 , and a sequence of measures {Φ} ∞
t=1 such that, for all
t,

1. (Maximization of Households): Given {wt , rt } and {Tt , τt } the func-

tions {vt } solve Bellman’s equation for period t and {ct , at+1 } are the
associated policy functions

2. (Marginal Product Pricing): The prices wt and rt satisfy

wt = FL (Kt , Lt ) (7.67)

rt = FK (Kt , Lt ) − δ. (7.68)

3. (Government Budget Constraint):

Tt = τt rt Kt (7.69)

for all t ≥ 0.

4. (Market Clearing):
Z
ct (at , yt )dΦt + Kt+1 = F (Kt , Lt ) + (1 − δ)Kt (7.70)
7.3. UNEXPECTED AGGREGATE SHOCKS AND TRANSITION DYNAMICS199
Z
Lt = yt dΦt (7.71)
Z
Kt+1 = at+1 (at , yt )dΦt (7.72)

5. (Aggregate Law of Motion):7

Φt+1 = Γt (Φt ) (7.75)

Definition 52 A stationary equilibrium is an equilibrium such that all ele-

ments of the equilibrium that are indexed by t are constant over time.

7.3.2 Computation of the Transition Path

We are interested in computing the following equilibrium. At time 0 the
economy is in the stationary equilibrium associated with τ0 = 0 and asso-
ciated distribution Φ0 (from this distribution all other aggregate variables,
such as Kt , rt , wt can easily be derived) and associated value function v0 and
policy functions c0 , a1 . Now at time t = 1 the policy changes permanently,
τt = τ > 0 for all t ≥ 1. Let the new stationary equilibrium associated with
τ be denoted by Φ∞ ,with associated value function v∞ and policy functions
c∞ , a∞ In the previous section we discussed how to compute such stationary
equilibria. Now we want to compute the entire transition path and trace out
the welfare consequences of such a policy innovation.
The key idea is to assume that after T periods the transition from the old
to the new stationary equilibrium is completed. We will discuss below how
to choose T, but heuristically it should be large enough so that the economy
7
The functions Γt can be written explicitly as follows. Define Markov transition func-
tions Qt : Z × B(Z) → [0, 1] induced by the transition probabilities π and the optimal
policy at+1 (a, y) as
X  π(y 0 |y) if at+1 (a, y) ∈ A
Qt ((a, y), (A, Y)) = (7.73)
0
0 else
y ∈Y

for all (y, a) ∈ Z and all (A, Y) ∈ B(Z). Then

Z
Φt+1 (A, Y) = [Γt (Φt )] (A, Y) = Qt ((a, y), (A, Y))dΦt (7.74)

for all (A, Y) ∈ B(Z).

200 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

had enough time to settle down to the new stationary equilibrium. The key
insight then is to realize that if vT = v∞ , then for a given sequence of prices
{rt , wt }Tt=1 the household problem can be solved backwards (note that this
is true independent of wether people live forever or not). This suggests the
following algorithm.

Algorithm 53 1. Fix T
2. Compute the stationary equilibrium Φ0 , v0 , c0 , a0 , r0 , w0 , K0 associated
with τ = τ0 = 0
3. Compute the stationary equilibrium Φ∞ , v∞ , c∞ , a∞ , r∞ , w∞ , K∞ asso-
ciated with τ∞ = τ. Assume that
ΦT , vT , cT , aT , rT , wT , KT = Φ∞ , v∞ , c∞ , a∞ , r∞ , w∞ , K∞
−1
4. Guess a sequence of capital stocks {K̂t }Tt=1 Note that since the capital
stock at time t = 1 is determined by decisions at time 0, K̂1 = K0 . Also
note that Lt = L0 = L̄ is fixed. Thus with the guesses on the capital
−1
stock we also obtain {r̂t , ŵt }Tt=1 determined by
ŵt = FL (K̂t , L̄)
r̂t = FK (K̂t , L̄) − δ
T̂t = τt r̂t K̂t .
−1 −1
5. Since we know vT (a, y) and {r̂t , ŵt , T̂t }Tt=1 we can solve for {v̂t , ĉt , ât+1 }Tt=1
backwards.
−1
6. With the policy functions {ât+1 } we can define the transition laws {Γ̂t }Tt=1 .
But since we know Φ0 = Φ1 from the initial stationary equilibrium, we
can iterate the distributions forward
Φ̂t+1 = Γ̂t (Φ̂t )
for t = 1, . . . , T − 1.
7. With {Φ̂t }Tt=1 we can compute
Z
Ât = adΦ̂t

for t = 1, . . . , T.
7.3. UNEXPECTED AGGREGATE SHOCKS AND TRANSITION DYNAMICS201

8. Check whether  
max Ât − K̂t  < ε
 
1≤t<T

−1
If yes, go to 9. If not, adjust your guesses for {K̂t }Tt=1 in 4.

9. Check whether Φ̂T − ΦT < ε. If yes, the transition converges smoothly

into the new steady state and we are done and should save {v̂t , ât+1 , ĉt , Φ̂t , r̂t , ŵt , K̂t }.
If not, go to 1. and increase T.

This procedure determines all the variables we are interested in along

the transition path: aggregate variables such as rt , wt , Φt , Kt and individual
decision rules cr , at+1 . It turns out that the value functions vt enable us to
make statements about the welfare consequences of a tax reform for which
we just have computed the transition path.

7.3.3 Welfare Consequences of the Policy Reform

¿From our previous algorithm we obtain the sequence of value functions
{vt }Tt=0 . Remember the interpretation of the value functions: v0 (a, y) is the
expected lifetime utility of an agent with assets a and productivity shock y at
time 0 in the initial stationary equilibrium, that is, for a person that thinks he
will live in the stationary equilibrium with τ = 0 forever. Similarly v1 (a, y)
is the expected lifetime utility of an agent with assets a and productivity
y that has just been informed that there is a permanent tax change. This
lifetime utility takes into account all the transition dynamics through which
the agent is going to live. Finally vT (a, y) = v∞ (a, y) is the lifetime utility
of an agent with characteristics (a, y) born in the new stationary equilibrium
(i.e. of an agent that does not live through the transition).
So in principle we could use v0 , v1 and vT to determine the welfare con-
sequences from the reform. The problem, of course, is that utility is an
ordinal concept, so that comparing v0 (a, y) with v1 (a, y) we can only de-
termine whether agent (a, y) gains or loses from the reform, but we cannot
meaningfully discuss how big these gains or losses are.
Now suppose that the period utility function is of CRRA form

c1−σ
U (c) =
1−σ
202 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

Consider the optimal consumption allocation in the initial stationary equi-

librium, in sequential formulation, {cs }∞
s=0 , where for simplicity we have sup-
pressed the explicit dependence of the history of productivity shocks. Then
∞
X ct1−σ
v0 (a, y) = E0 βt
s=0
1−σ

Now suppose we increase consumption in each date, in each state, in the

old stationary equilibrium, by a fraction g, so that the new allocation is
{(1 + g)cs }∞
s=0 . The lifetime utility from that consumption allocation is
∞ ∞
X
t [(1 + g)ct ]1−σ X ct1−σ
v0 (a, y; g) = E0 β = (1+g)1−σ E0 βt = (1+g)1−σ v0 (a, y)
s=0
1−σ s=0
1 − σ

Obviously v0 (a, y; g = 0) = v0 (a, y). If we want to quantify the welfare con-

sequences of the policy reform for an agent of type (a, y) we can ask the
following question: by what percent g do we have to increase consumption in
the old stationary equilibrium, in each date and state, for the agent to be in-
different between living in the old stationary equilibrium and living through
the transition induced by the policy reform.8 This percent g solves

v0 (a, y; g) = v1 (a, y)

or

(1 + g)1−σ v0 (a, y) = v1 (a, y)

  1
v1 (a, y) 1−σ
g(a, y) = −1
v0 (a, y)
Evidently, this number is bigger than zero if and only if v1 (a, y) > v0 (a, y),
in which case the agent benefits from the reform, and g(a, y) measures by
how much, in consumption terms. Note that the number g(a, y) depends
on an agents’ characteristics (one would expect households with a lot of
assets to lose badly, households with little assets to lose not much or to
even gain -remember that taxes are lump-sum redistributed). Also note that
we only need to know v0 (a, y) and v1 (a, y) to compute this number, but our
computation of the transition path gives us the value functions v0 , v1 anyhow.
8
Lucas (1978) proposed this consumption equivalent variation measure in order to
assess the welfare costs of business cycles.
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES203

The computation of consumption equivalent variation, explicitly taking

into account the transition path, is the theoretically correct way to assess
the welfare consequences of a reform. Often studies try to assess the steady
state welfare consequences of a policy reform, in particular if these studies
do not explicitly compute transition paths. Even though I think that these
welfare numbers are often not very meaningful, let us quickly describe the
procedure. Obviously one can compute
  1
vT (a, y) 1−σ
gss (a, y) = −1
v0 (a, y)
with the interpretation that this number is the welfare gain of an agent being
born with characteristics.(a, y) into the new as opposed to the old stationary
equilibrium. If we want to study welfare consequences of a policy reform for
households before their identity is revealed, we may define as
R  1
vT (a, y)dΦT 1−σ
gss = R −1
v0 (a, y)dΦ0
R
the expected steady state welfare gain. Here vT (a, y)dΦT is the expected
lifetime utility of an agent in the new stationary equilibrium, before the
agent knows with what characteristics
R he or she will be born (behind the veil
of ignorance). The quantity v0 (a, y)dΦ0 is defined correspondingly. Even
though these measures are not hard to compute (in fact, one does not need to
compute the transition path to compute these numbers), they do ignore the
fact that it takes time to go from the old stationary equilibrium to the new
one. For the policy example at hand, the increase of the capital tax is likely
to induce a decline in the capital stock, thus lower aggregate consumption
and thus steady state welfare losses. What these losses ignore is that along
the transition path part of the capital stock is being eaten, with associated
consumption and welfare derived from it. Therefore welfare measures based
on steady state comparisons may give only fairly limited information about
the true welfare consequences of policy reforms.

7.4 Aggregate Uncertainty and Distributions

as State Variables
In section 7.1 we discussed a model where agents faced significant amounts of
idiosyncratic uncertainty, which they could, by assumption, only self-insure
204 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

against by precautionary saving via risk-less bonds. In the aggregate, the

economy was stationary in that output, investment and aggregate consump-
tion were constant over time. Thus by construction the model could not
speak to how aggregate consumption, saving and investment fluctuates over
the business cycle.
In this section we introduce aggregate uncertainty, in addition to idiosyn-
cratic uncertainty, into the model. In contrast to idiosyncratic uncertainty,
which is in principle insurable (but insurance against which we ruled out),
there is no mutual insurance against aggregate shock, unless one analyzes a
two-country world where the two countries have aggregate shocks which are
imperfectly correlated.

7.4.1 The Model

In the spirit of real business cycle theory aggregate shocks take the form of
productivity shocks to the aggregate production function

Yt = st F (Kt , Lt ) (7.76)

where {st } is a sequence of random variables that follows a finite state Markov
chain with transition matrix π. We will follow Krusell and Smith (1998) and
assume that the aggregate productivity shock can take only two values

st ∈ {sb , sg } = S (7.77)

with sb < sg and denote by π(s0 |s) the conditional probability of the aggre-
gate state transiting from s today to s0 tomorrow. Krusell and Smith (1998)
interpret sb as an economic recession and sg as an expansion.
The idiosyncratic labor productivity yt is assumed to take only two values

yt ∈ {yu , ye } = Y (7.78)

where yu < ye stands for the agent being unemployed (having low labor pro-
ductivity) and ye stands for the agent being employed. Krusell and Smith
attach the employed-unemployed interpretation to the idiosyncratic labor
productivity variable, an interpretation which will be important in the cal-
ibration section. Obviously the probability of being unemployed is higher
during recessions than during expansions, and the Markov chain governing
idiosyncratic labor productivity should reflect this dependence of idiosyn-
cratic uncertainty on the aggregate state of the economy.
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES205

In particular, let denote by

π(y 0 , s0 |y, s) ≥ 0 (7.79)

the conditional probability of an agent having individual productivity to-

morrow of y 0 and the aggregate state being s0 tomorrow, conditional on the
individual and aggregate state y and s today. For example π(yu , sb |ye , sg ) is
the probability of getting laid off tomorrow in a recession if the economy is
booming today and the individual is employed. Consistency with aggregate
transition probabilities requires that
X
π(y 0 , s0 |y, s) = π(s0 |s) for all y ∈ Y and all s, s0 ∈ S (7.80)
y 0 ∈Y

i.e. the probability y0 ∈Y π(y 0 , s0 |y, s) of going from a particular individual

P
state y and aggregate state s to some individual state y 0 and a particular
aggregate state s0 equals the aggregate transition probability π(s0 |s).
We again assume a law of large numbers, so that idiosyncratic uncer-
tainty averages out, and only aggregate uncertainty determines the number
of agents in states y ∈ Y. Obviously this number will vary with today’s aggre-
gate state s. Krusell and Smith assume that the share of unemployed people
only depends on s (but not on past aggregate states). Then let by Πs (y)
denote the fraction of the population in idiosyncratic state y if aggregate
state is s; e.g. Πsb (yu ) is the deterministic number of unemployed people if
the economy is in a recession. The assumption that Πs (y) only depends on
s imposes further restriction on the Markov transition matrix π:
X π(y 0 , s0 |y, s)
Πs0 (y 0 ) = Πs (y) for all s, s0 ∈ S (7.81)
y∈Y
π(s0 |s)

Suppose the economy is in a boom today, s = sg . The fraction of people

unemployed in a recession tomorrow, Πs0 =sb (y 0 = yu ) equals the fraction
of employed people today Πs=sg (y = ye ) who get laid off in the recession,
u ,sb |ye ,sg )
π(yu |sb , ye , sg ) = π(yπ(s b |sg )
, plus the fraction of unemployed people today
u ,sb |yu ,sg )
Πs=sg (y = yu ) who remain unemployed, π(yu |sb , yu , sg ) = π(yπ(s b |sg )
. The
same restriction applies for all other states.
The exogenous Markov chain driving the economy is therefore described
by the two sets of cardinality 2, S and Y and the joint 4 × 4 transition matrix
206 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

for idiosyncratic and aggregate productivity π

 
π(yu , sb |yu , sb ) π(yu , sb |ye , sb ) π(yu , sb |yu , sg )
π(yu , sb |ye , sg )
 π(ye , sb |yu , sb ) π(ye , sb |ye , sb ) π(ye , sb |yu , sg )
π(ye , sb |ye , sg ) 
π=  π(yu , sg |yu , sb ) π(yu , sg |ye , sb ) π(yu , sg |yu , sg )

π(yu , sg |ye , sg ) 
π(ye , sg |yu , sb ) π(ye , sg |ye , sb ) π(ye , sg |yu , sg )
π(ye , sg |ye , sg )
(7.82)
In the previous section we defined, proved existence of and computed a sta-
tionary equilibrium, in which prices and the cross-sectional distribution of
assets (i.e. all macroeconomic variables of interest) were constant over time.
Obviously, with aggregate shocks there is no hope of such an equilibrium
to exist. This makes our extension interesting in the first place, because it
allows business cycle fluctuations, but also will impose crucial complications
for the computation of the economy.
Since the cross-sectional distribution Φ of assets will vary with the aggre-
gate shock, we have to include it as a state variable in our formulation of a
recursive equilibrium. The aggregate state variables also include the aggre-
gate productivity shock s, since this shock determines aggregate productivity
and hence aggregate wages and interest rates.
For the recursive formulation of the household problem we note that the
individual state variable is composed of individual asset holdings and income
shocks (a, y), whereas the aggregate state variables include the aggregate
shock and the distribution of assets (s, Φ). Hence the households’ problem in
recursive formulation is
( )
XX
v(a, y, s, Φ) = max 0
U (c) + β π(y 0 , s0 |y, s)v(a0 , y 0 , s0 , Φ0 )
c,a ≥0
y 0 ∈Y s0 ∈S
s.t. (7.83)
0
c+a = w(s, Φ)y + (1 + r(s, Φ))a (7.84)
Φ0 = H(s, Φ, s0 ) (7.85)

Note that s0 is a determinant of Φ0 since it specifies how many agents have

idiosyncratic shock y 0 = yu and how many agents have y 0 = ye . We have the
following definition of a recursive competitive equilibrium

Definition 54 A recursive competitive equilibrium is a value function v :

Z × S × M → R, policy functions for the household a0 : Z × S × M → R
and c : Z × S × M → R, policy functions for the firm K : S × M → R and
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES207

L : S × M → R, pricing functions r : S × M → R and w : S × M → R and

an aggregate law of motion H : S × M × S→ M such that
1. v, a0 , c are measurable with respect to B(S), v satisfies the household’s
Bellman equation and a0 , c are the associated policy functions, given r()
and w()
2. K, L satisfy, given r() and w()
r(s, Φ) = FK (K(s, Φ), L(s, Φ)) − δ (7.86)
w(s, Φ) = FL (K(s, Φ), L(s, Φ)) (7.87)

3. For all Φ ∈ M and all s ∈ S

Z
K(H(s, Φ)) = a0 (a, y, s, Φ)dΦ (7.88)
Z
L(s, Φ) = ydΦ (7.89)
Z Z
c(a, y, s, Φ)dΦ + a0 (a, y, s, Φ)dΦ (7.90)
= F (K(s, Φ), L(s, Φ)) + (1 − δ)K(s, Φ) (7.91)

4. The aggregate law of motion H is generated by the exogenous Markov

process π and the policy function a0 (as described below)
Again define the transition function QΦ,s,s0 : Z × B(Z) → [0, 1] by
X  π(y 0 , s0 |y, s) if a0 (a, y, s, Φ) ∈ A
QΦ,s,s0 ((a, y), (A, Y)) = (7.92)
0
0 else
y ∈Y

The aggregate law of motion is then given by

Z
0 0
Φ (A, Y) = (H(s, Φ, s )) (A, Y) = QΦ,s,s0 ((a, y), (A, Y))Φ(da × dy) (7.93)

Before describing the computational strategy for the recursive equilibrium

some comments are in order:
1. One should first define a sequential markets equilibrium. Such an equi-
librium will in general exist, but it is impossible to compute such an
equilibrium directly. No claim of uniqueness of a sequential markets
equilibrium can be made.
208 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

2. For the recursive equilibrium defined above Krusell and Smith assert
that the current cross-sectional asset distribution and the current shock
are sufficient aggregate state variables. This is not formally proved,
and such a proof would be very hard. What one has to prove is that a
recursive equilibrium generates a sequential equilibrium in the following
sense: Cross-sectional distributions are generated, starting from the
initial condition (s0 , Φ0 ) as follows
Φ1 (s1 ) = H(s0 , Φ0 , s1 ) (7.94)
Φt+1 (st+1 ) = H(st , Φt (st ), st+1 ) (7.95)
prices are generated by
rt (st ) = r(st , Φt (st )) (7.96)
wt (st ) = w(st , Φt (st )) (7.97)
and, starting from initial conditions (a0 , y0 ), an individual households’
allocation is generated by
c0 (a0 , y0 , s0 ) = c(a0 , y0 , s0 , Φ0 ) (7.98)
a1 (a0 , y0 , s0 ) = a0 (a0 , y0 , s0 , Φ0 ) (7.99)
and in general recursively
ct (a0 , y t , st ) = c(at (a0 , y t−1 , st−1 ), yt , st , Φt ) (7.100)
at+1 (a0 , y t , st ) = a0 (at (a0 , y t−1 , st−1 ), yt , st , Φt ) (7.101)
Similarly optimal choices of the firm can be generated. Thus, a given
recursive equilibrium generates sequential allocations and prices; it re-
mains to be verified that these prices are in fact a sequential equilib-
rium.
3. Finally, the issue of existence of a recursive equilibrium arises. We
know that, since a sequential equilibrium in general exists, there is a
state space large enough such that a recursive equilibrium (recursive
in that state space) exists. So the issue is whether a recursive equilib-
rium in which the aggregate state only contains the current shock and
the current wealth distribution does exist. Although this state space
seems “natural” in some sense, there is no guarantee of existence of
such a recursive equilibrium. The analysis of this economy is purely
computational in spirit as neither the existence, uniqueness, stability or
qualitative features of the equilibrium can be theoretically established.
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES209

7.4.2 Computation of the Recursive Equilibrium

The key computational problem is the size of the state space. The aggregate
wealth distribution Φ is an infinite-dimensional object, the aggregate law
of motion therefore maps an infinite-dimensional space into itself. What
we look for is a low-dimensional approximation of the wealth distribution.
Why do agents need to keep track of the aggregate wealth distribution? In
order to figure out today’s interest and wage rate the aggregate (average)
capital stock (i.e. the first moment of the wealth distribution) is sufficient.
Thus, to forecast tomorrow’s factor prices, all the agent needs to forecast is
tomorrow’s aggregate capital stock. The need to keep track of the current
wealth distribution stems from the fact that the entire wealth distribution Φ
today, not only its first moment K, determines tomorrow’s aggregate capital
stock via Z
K = a0 (a, y, s, Φ)dΦ
0
(7.102)

Another way of putting it, the average capital stock tomorrow is not equal to
the saving function of some average, representative consumer, evaluated at
today’s average capital stock. If the optimal policy function for tomorrow’s
assets, would feature a constant propensity to save out of current assets and
income (a0 being linear in a, with same slope for all y ∈ Y ), then exact
aggregation would occur and in fact the average capital stock today would
be a sufficient statistic for the average capital stock tomorrow. This insight
is important for the quantitative results to come.
The computational strategy that Krusell and Smith (and many others
since) follow is to approximate the distribution Φ with a finite set of mo-
ments. Remember that Φ is the distribution over (a, y). Obviously, since y
can only take two values, the second dimension is not the problem, so in what
follows we focus on the discussion of the distribution over assets a. Let the
n-dimensional vector m denote the first n moments of the asset distribution
(i.e. the marginal distribution of Φ with respect to its first argument).
We now posit that the agents use an approximate law of motion

m0 = Hn (s, m) (7.103)

mapping the first n moments of the asset distribution today, m, into the first
n moments of the asset distribution tomorrow, m0 . Note that by doing so
agents are boundedly rational in the sense that moments of higher order than
n of the current wealth distribution may help to more accurately forecast the
210 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

first n moments tomorrow. It is the hope that, according to some metric to

be discussed later, agents don’t make severe forecasting errors of tomorrow’s
average capital stock (the only variable they care about for forecasting future
prices), and that the resulting approximate equilibrium is in some sense close
to the rational expectations equilibrium in which agents use the entire wealth
distribution to forecast tomorrow’s prices.
The next step in the computation of an approximate equilibrium is to
choose the number of moments and the functional form of the function Hn .
Since the agents only need to know next periods average capital stock, Krusell
and Smith first pick n = 1 and pose the following log-linear law of motion
(remember the law of motion for the stochastic neoclassical growth model
with log-utility and Cobb-Douglas production)

log(K 0 ) = as + bs log(K) (7.104)

for s ∈ {sb , sg }. Here (as , bs ) are parameters that need to be determined. The
recursive problem of the household then becomes
( )
XX
v(a, y, s, K) = max 0
U (c) + β π(y 0 , s0 |y, s)v(a0 , y 0 , s0 , K 0 )
c,a ≥0
y 0 ∈Y s0 ∈S
s.t. (7.105)
0
c + a = w(s, K)y + (1 + r(s, K))a (7.106)
log(K 0 ) = as + bs log(K) (7.107)

Note that we have reduced the state space from something that includes the
infinite-dimensional space of measures to a four dimensional space (a, y, s, K) ∈
R × Y × S × R. The algorithm for computing an approximate recursive equi-
librium is then as follows:

1. Guess (as , bs )

2. Solve the households problem to obtain decision rules a0 (a, y, s, K)

3. Simulate the economy for a large number of T periods for a large

number N of households (in their exercises Krusell and Smith pick
N = 5000 and T = 11000: start with initial conditions for the econ-
omy (s0 , K0 ) and for each household (ai0 , y0i ). Draw random sequences
{st }Tt=1 and {yti }T,N 0
t=1,i=1 and use the decision rule a (a, y, s, K) and the
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES211

perceived law of motion for K to generate sequences of {ait }T,N

t=1,i=1 .
Aggregate to find the implied sequence of aggregate capital stocks
N
1 X i
Kt = a (7.108)
N i=1 t

Discard the first τ periods, because of dependence on initial conditions.

4. With the remaining observations run the regressions

log(K 0 ) = αs + βs log(K) (7.109)

to estimate (αs , βs ) for s ∈ S. If the R2 for this regression is high and

(αs , βs ) ≈ (as , bs ) stop. An approximate equilibrium is found. Other-
wise update guess for (as , bs ). If guesses for (as , bs ) converge, but R2
remains low, try to add higher moments to the aggregate law of motion
and/or experiment with a different functional form for the aggregate
law of motion.

7.4.3 Calibration
In this economy no analytical results can be proved and one has to resort
to numerical analysis. Krusell and Smith take the model period to be one
quarter (note that this is a business cycle model). As preferences they as-
sume CRRA utility with rather low risk aversion of σ = 1 (i.e. log-utility).
The time discount factor is chosen to be β = 0.994 = 0.96 on an annual
basis, which implies a yearly subjective time discount rate of ρ = 4.1%
as in Aiyagari. Also similar to Aiyagari they take the capital share to be
α = 0.36. As annual depreciation rate they choose δ = (1−0.025)4 −1 = 9.6%,
slightly higher than Aiyagari, but within the range of values commonly used
in real business cycle studies. The remaining parameters describe the joint
aggregate-idiosyncratic labor productivity process. Unfortunately the paper
itself does not contain a precise discussion of the parameterization, so that
the discussion here relies partly on Krusell and Smith’s information from
the paper, partly on Imrohoroglu’s (1989) paper to which they refer to and
partly on the FORTRAN code posted on Tony Smith’s web site.
The calibration strategy is to first calibrate the aggregate component of
the productivity process, i.e. the set S and the 2×2 matrix π(s0 |s). Remember
212 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

that the aggregate state represents recessions and expansions. With respect
to S they choose
S = {0.99, 1.01} (7.110)
I would think that the standard deviation of the technology shock is a bit on
the small side with 0.01. In fact, using Cooley and Prescott’s (1995) contin-
uous state process and discretizing into a two state chain yields a standard
deviation of the shock of about 0.02. Krusell and Smith claim that with their
aggregate process they are able to generate aggregate fluctuations of output
similar to US data, which, given the information in the paper I wasn’t able
to verify. Given that Cooley and Prescott need sufficiently more variance
in the technology shock to generate business cycles of reasonable size, this
must mean that the economy with heterogenous agents and uninsurable id-
iosyncratic risk, for a given aggregate shock variance, is more volatile than
its representative agent counterpart.
As for the transition matrix for the aggregate shock the first assumption
made is symmetry, so that π(sg |sg ) = π(sb |sb ). Krusell and Smith choose
π(sg |sg ) such that, conditional on being in the good state today, the expected
time in the good state are 8 quarters, or

8 = [1 − π(sg |sg )] 1 + 2π(sg |sg ) + 3π(sg |sg )2 + . . .

 
(7.111)

or
1
8 =
1 − π(sg |sg )
7
π(sg |sg ) = (7.112)
8
so that
7 1
 
0
π(s |s) = 8
1
8
7 (7.113)
8 8

It follows that, conditional on being in the bad state today, the expected
time of staying there is also 8 quarters.
With respect to idiosyncratic labor productivity, the state space is chosen
as
Y = {0.25, 1} (7.114)
Remember that the idiosyncratic states are meant to represent employment
and unemployment, so that it is assumed that an unemployed person makes
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES213

25% of the labor income of an employed person. Imrohoroglu (1989) argues

that, although the average level of unemployment compensation is higher, a
large fraction of the unemployed do not receive benefits (Imrohoroglu quotes
a number of 61%), so that the 25% figure is argued as reasonable. Krusell and
Smith adopt this value. The level of productivity for an employed person is
free for normalization, and for the sake of the discussion I set it to 1 (doubling
both states in Y obviously only changes units, but does not affect the results).
With respect to the transition probabilities for the idiosyncratic produc-
tivity we note that
π(y 0 , s0 |y, s)
π(y 0 |s0 , y, s) = (7.115)
π(s0 |s)
or
π(y 0 , s0 |y, s) = π(y 0 |s0 , y, s) ∗ π(s0 |s) (7.116)
So in order to specify π(y 0 , s0 |y, s), given our previous work we have to specify
the four (for each pair (s0 , s)) 2 × 2 matrices π(y 0 |s0 , y, s) indicating, condi-
tional on an aggregate transition from s to s0 , what the individual’s proba-
bilities of transition from employment to unemployment are. These should
vary with the aggregate transition (s0 , s).
The calibration of these matrices are governed by the following observa-
tions. In an expansion the average time of unemployment, conditional on
being unemployed today, is equal to 1.5 quarters. This implies that
∞
X
1.5 = [1 − π(y 0 = yu |s0 = sg , y = yu , s = sg )] i ∗ π(y 0 = yu |s0 = sg , y = yu , s = sg )i−1
i=1
1
= 0 0
1 − π(y = yu |s = sg , y = yu , s = sg )
1 1
π(y 0 = yu |s0 = sg , y = yu , s = sg ) = 1 − = (7.117)
1.5 3
and hence π(y 0 = ye |s0 = sg , y = yu , s = sg ) = 32 . Similarly, to match the
fact that in bad times the average time of unemployment, conditional on
unemployed, equals 2.5 quarters we find π(y 0 = yu |s0 = sb , y = yu , s = sb ) =
0.6 and hence π(y 0 = ye |s0 = sb , y = yu , s = sb ) = 0.4. The authors then
assume that the probability of remaining unemployed when times switch
from expansion to recession is 1.25 times the same probability when the
economy was already in a recession, implying π(y 0 = yu |s0 = sb , y = yu , s =
sg ) = 0.75 and thus π(y 0 = ye |s0 = sb , y = yu , s = sg ) = 0.25. Finally, the
214 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

probability of remaining unemployed when times switch from recession to

expansion is assumed to be 0.75 times the same probability when times were
already good. This gives π(y 0 = yu |s0 = sg , y = yu , s = sb ) = 0.25 and
π(y 0 = ye |s0 = sg , y = yu , s = sb ) = 0.75.
The final transition probabilities are derived from the facts that in re-
cessions the unemployment rate is Πsb (yu ) = 10% and in expansions it is
Πsg (yu ) = 4%. With the probabilities already derived and equation (7.81)
this uniquely pins down the remaining probabilities

π(y 0 = yu |s0 = sg , y = ye , s = sg ) = 0.028

π(y 0 = ye |s0 = sg , y = ye , s = sg ) = 0.972
π(y 0 = yu |s0 = sb , y = ye , s = sb ) = 0.04
π(y 0 = ye |s0 = sb , y = ye , s = sb ) = 0.96
π(y 0 = yu |s0 = sb , y = ye , s = sg ) = 0.079
π(y 0 = ye |s0 = sb , y = ye , s = sg ) = 0.921
π(y 0 = yu |s0 = sg , y = ye , s = sb ) = 0.02
π(y 0 = ye |s0 = sg , y = ye , s = sb ) = 0.98 (7.118)

In short, the best times for finding a job are when the economy moves from
a recession to an expansion, the worst chances are when the economy moves
from a boom into a recession. Combining the aggregate transition prob-
abilities with the idiosyncratic probabilities, conditional on the aggregate
transitions, finally yields as transition matrix (7.82)
 
0.525 0.035 0.09375 0.0099
 0.35 0.84 0.03125 0.1151 
π=  0.03125 0.0025 0.292 0.0245 
 (7.119)
0.09375 0.1225 0.583 0.8505

With this parameterization we are ready to report results on computed ap-

proximate equilibria.

7.4.4 Numerical Results

The model generates three basic entities of interest: an aggregate law of
motion
m0 = Hn (s, m), (7.120)
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES215

individual decision rules a0 (a, y, s, m) and time-varying cross-sectional wealth

distributions Φ(a, y). First let’s focus on the aggregate law of motion. Re-
member that we allowed agents only to use the first n moments of the current
wealth distribution to forecast the first n moments of tomorrow’s wealth dis-
tribution, i.e. agents are boundedly rational. In particular, the aggregate
law of motion perceived by agents may not coincide with the actual law of
motion, which is just another way of saying that we are NOT computing a
rational expectations equilibrium.
Since agents only have to forecast tomorrow’s aggregate (average) capital
stock (i.e. the first moment of Φ0 ), Krusell and Smith start out with n = 1.
After repeated updating of the coefficients in (7.109) the coefficients converge
and one obtains as perceived law of motion of agents
log(K 0 ) = 0.095 + 0.962 log(K) for s = sg (7.121)
log(K 0 ) = 0.085 + 0.965 log(K) for s = sb (7.122)
How irrational are agents? We would have computed a perfectly legitimate
rational expectations equilibrium if the actual law of motion for the capital
stock follows (7.121) to (7.122) exactly. Remember that at each step, to
update the coefficients in (7.121) to (7.122), a simulated time series for the
aggregate capital stock in conjunction with a sequence of aggregate shocks
{(st , Kt }Tt=0 is generated. By dividing the sample into periods with st = sb
and st = sg one then can run the regressions
log(Kt+1 ) = αg + βg log(Kt ) + εgt+1 for periods with st = sg (7.123)
log(Kt+1 ) = αb + βb log(Kt ) + εbt+1 for periods with st = sb (7.124)
In fact, given the computational algorithm, for the last step of the iteration
on the perceived law of motion they did obtain α̂g = 0.095, α̂b = 0.085 and
β̂g = 0.962, β̂b = 0.965. From econometrics we remember that with regression
errors
ε̂jt+1 = log(Kt+1 ) − α̂j − β̂j log(Kt ) for j = g, b (7.125)
we define the standard deviation of the regression error as
s
1 X j
2
σj = ε̂ (7.126)
Tj t∈τ t
j

where τj is the set of time indices for which st = sj and Tj is the cardinality
of that set. Note that, since the regression is on log-income, σj can be inter-
preted as the average percentage error for the prediction of the capital stock
216 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

made by the regression. The R2 of the regression, as alternative measure of

fit, is defined as that fraction of the variation in tomorrow’s log-capital stock
that is explained by the variation of today’s log-capital stock, or

2
ε̂jt
P
t∈τj
Rj2 = 1 − P
2 (7.127)
t∈τj log Kt+1 − log K

If σj = 0 for j = g, b or equivalently, if Rj2 = 1 for j = g, b then households’

perceived aggregate law of motion is exactly correct for all periods, i.e. agents
do not make forecasting errors and are perfectly rational. Low σj and high Rj2
are taken as evidence that the actual law of motion generated by individual
behaviors and aggregation does not depart much from the perceived law of
motion, i.e. that agents make only small forecasting errors.
Krusell and Smith obtain Rj2 = 0.999998 for j = b, g and σg = 0.0028,
σb = 0.0036, i.e. extremely low average forecasting errors made by agents.
They calculate maximal forecasting errors for interest rates 25 years into
the future of 0.1% for their simulations. Unfortunately they do not report
magnitudes of corresponding utility losses from forecasting errors (say, in
consumption equivalent variation), but assert that given the small magnitude
of forecasting errors even for the far future these utility losses from bounded
rationality are negligible.
Before giving intuition for the results a word of caution is in order. In
all of computational economics, since computer precision is limited, the best
one can achieve is to compute an approximate equilibrium, i.e. allocations
and prices in which markets almost clear (excess demand is ε away from
zero). This approximate equilibrium may be arbitrarily far away from a
true equilibrium (i.e. prices and allocations for which markets clear exactly).
The study by Krusell and Smith is not unusual in this respect. But even
if this general problem were absent, Krusell and Smith’s equilibrium is at
best an approximation to a rational expectations equilibrium and the true
rational expectations equilibrium may be arbitrarily far away from the com-
puted equilibrium. Krusell and Smith’s agents make small forecasting errors
(according to their σj , Rj2 metric), but the aggregate law of motion at which
agents make no forecasting errors (i.e. a rational expectations equilibrium)
will in general involve higher moments and may look very different from the
one they found.
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES217

7.4.5 Why Quasi-Aggregation?

Suppose all agents, for all interest rates r(K) and wages w(K) have linear
savings functions with the same marginal propensity to save, so that
a0 (a, y, s, K) = as + bs a + cs y (7.128)
Then
Z Z
0 0
K = a (a, y, s, K)dΦ = as + bs adΦ + cs L̄
= ãs + bs K (7.129)
R
since K = adΦ. Therefore exact aggregation obtains and the first moment
of the current wealth distribution (which equals the current capital stock) is
in fact a sufficient statistic for Φ for forecasting the aggregate capital stock
tomorrow.
In their Figure 2, Krusell and Smith plot savings functions (note: for a
particular current aggregate stock and a particular shock s). We see that
they are almost linear with same slope for y = yu and y = ye . The only
exceptions are unlucky agents (y = yu ) with little assets which are liquidity
constrained and hence have a low (zero) marginal propensity to save. How-
ever, since these (few) agents hold a negligible fraction of aggregate wealth,
they don’t matter for the aggregate capital dynamics. All other agents have
almost identical propensities to save, thus individual saving decisions almost
exactly aggregate, and the current aggregate capital stock is almost a suffi-
cient statistic when forecasting tomorrow’s capital stock: quasi-aggregation
obtains.
The key question is why individual savings functions a0 are almost linear
in current assets a at just about all current asset levels. ¿From Figure 2 of
Krusell and Smith we see that the slope of a0 when plotted against a is roughly
equal to 1 for all but very low asset levels. Remember from the PILCH model
with certainty equivalence that optimal consumption was given by
T −t
!
r X yt+s
ct = Et + at (7.130)
1+r s=0
(1 + r)s
r
and hence agents consume a fraction 1+r of current assets. That is, they save
out of current assets for tomorrow
 
at+1 r
= 1− at + G(y) (7.131)
1+r 1+r
218 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

where G(y) is a function of the stochastic income process. Thus under cer-
tainty equivalence
at+1 = at + H(y) (7.132)
and thus the saving function a0 has slope 1 under certainty equivalence (and
ρ = r). In Krusell and Smith’s economy agents are prudent and face liquidity
constraints, but almost act as if they are certainty equivalence consumers.
Why? A few reasons come to my mind:

1. Agents are prudent, but not all that much. A σ = 1 is at the lower end
of the empirical estimates for risk aversion. As we saw from Aiyagari’s
(1994) paper, the amount of precautionary saving increases significantly
with increases in σ, and so should the deviation of agents decision rules
from certainty equivalence.

2. The unconditional standard deviation of individual income is roughly

0.2, at the lower end of the estimates used by Aiyagari. As we saw,
higher variability, induces more quantitatively important precautionary
saving.

3. Probably most important, negative income shocks (unemployment) are

infrequent and not very persistent, so that they don’t force a large
departure in behavior from certainty equivalence. Krusell and Smith’s
agents never become permanently disabled and don’t face permanent
wage declines after being laid off (once they find a new job, their relative
wage is as high as before they were laid off)

Krusell and Smith consider various sensitivity tests with respect to these
and other dimensions (change in time discount rates, endogenous labor-
leisure choice etc.), and claim that the result of quasi-aggregation (only the
mean capital stock matters for forecasting tomorrows mean capital stock) is
robust to changes in the model parameterization.

7.4.6 Rich People are Not Rich Enough

The model generates an endogenous consumption and wealth distribution,
something that all of representative agents macroeconomics is silent about.
Note that the income distribution is, by specifying the income process that
households face, an input into the model. To the extent that the income
7.4. AGGREGATE UNCERTAINTY AND DISTRIBUTIONS AS STATE VARIABLES219

process and hence the cross-sectional income distribution is realistic, one

would hope that the resulting wealth distributions (remember that this dis-
tribution changes over time) is on average consistent with the cross-sectional
wealth distribution in the data. Unfortunately, the model does a fairly bad
job reproducing the US wealth distribution, in particular it fails to generate
the high concentration of wealth at the upper end of the distribution. In
the data, the richest 1% of the US population holds 30% of all household
wealth, the top 5% hold 51% of all wealth. For the model described above
the corresponding numbers are 3% and 11%, correspondingly. In the model
people save to buffer their consumption against unemployment shocks, but
since these shocks are infrequent and of short duration, they don’t save all
that much.
There are several ways of making a small fraction of the population
save a lot, and hence making them accumulating a large fraction of over-
all wealth. The repair job that Krusell and Smith propose is to make some
agents (stochastically) more patient than others. In particular, they assume
that the time discount factor of agents β is stochastic and follows a three
state Markov chain with β ∈ B and transition probabilities γ(β 0 |β). The
dynamic programming problem of agents then becomes
v(a, y, β, s, K) (7.133)
( )
XXX
= max
0
U (c) + β π(y 0 , s0 |y, s)γ(β 0 |β)v(a0 , y 0 , β 0 , s0 , K 0 )
c,a ≥0
y 0 ∈Y s0 ∈S β 0 ∈B

s.t.
0
c + a = w(s, K)y + (1 + r(s, K))a (7.134)
log(K 0 ) = as + bs log(K) (7.135)
They pick B = {0.9858, 0.9894, 0.993}. Hence annual discount rates differ
between 2.8% for the most patient agents and 5.6% for the most impatient
agents. As transition matrix Krusell and Smith propose (remember that the
model period is one quarter)
 
0.995 0.005 0
γ =  0.000625 0.99875 0.000625  (7.136)
0 0.005 0.995
which implies that 80% of the population has discount factor in the middle
and 10% are on either of the two extremes (in the stationary distribution).
220 CHAPTER 7. THE SIM IN GENERAL EQUILIBRIUM

It also implies that patient and impatient agents remain so for an expected
period of 50 years. De facto this parameterization creates three deterministic
types of agents. The patient agents are the ones that will accumulate most
of the wealth in this economy. With this trick of stochastic discount factors
Krusell and Smith are able to approximate the US wealth distribution to a
reasonably accurate degree (see their Figure 3): in the modified economy the
richest 1% of the population holds 24% of all wealth, the richest 5% hold 55%
percent of all wealth. Also note that the quasi-aggregation result extends to
the economy with stochastic discount factors.
 Part III

Complete Market Models with

Frictions

221
 223

In this chapter we will consider models that, in spirit, build on the com-
plete markets model considered in Chapter 3, in the same fashion the models
in Chapter 4 and 5 built on the simple Pilch model in Chapter 3. Most em-
pirical studies reject the complete insurance hypothesis and thus cast doubt
upon the complete markets model as a reasonable description of reality.
Therefore models are desired that predict some, but not perfect risk shar-
ing. The Pilch model (in its general equilibrium form) generates some “risk
sharing” via self-insurance: agents smooth part of their income fluctuation
by asset accumulation and decumulation, with the part being determined
by preferences and the nature of the income shocks. But remember that
there are no formal risk-sharing arrangements in the Pilch model, as explicit
contingent insurance contracts, which agents in the model would have an
incentive to trade and financial intermediaries would have an incentive to
offer, are ruled out by assumption, without any good reason from within the
model.9
The models we study in this chapter will allow a full set of contingent
consumption claims being traded (in the decentralized version) or being allo-
cated by the social planner (in the centralized version of the model). That full
insurance does not arise as optimal and equilibrium allocation is due to infor-
mational and/or enforcement frictions, which are explicitly modeled. These
models adhere to the principle, most forcefully articulated by Townsend that
the world is constrained efficient, and that is up to the researcher/modeler
to find the right set of constraints that give rise to model allocations which
are in line with empirical consumption allocations.
We will study two such sets of constraints: the first stemming from the
fact that in actual economies financial contracts are only imperfectly enforce-
able (because there exist explicit bankruptcy provisions in the legal code or
it is too costly to always enforce repayment), the second deriving from the
fact that individual incomes and/or actions are imperfectly observable, so
that contingent claims payments cannot be directly conditioned on these.
Both types of models will deliver allocations characterized by some, but (de-
pending on parameterizations) imperfect insurance, which is due explicitly
to these frictions.

9
Townsend (1985) and Cole and Kocherlakota (1997) study environments with private
information and show that the optimal contract between agents and financial intermedi-
aries is a simple uncontingent debt contract, closely resembling the one-period uncontin-
gent bonds that we let agents trade in the Pilch model.
224
Chapter 8

Limited Enforceability of
Contracts

We start with models in which perfect insurance is prevented because at each

point of time every agent can default on her financial obligations, with the
ensuing punishment being modeled as exclusion from future credit market
participation. The seminal contributions to this literature include Kehoe and
Levine (1993), Kocherlakota (1996) and Alvarez and Jermann (2000). Kehoe
and Levine (1993), building on earlier work of Eaton and Gersowitz (1981),
develop a competitive equilibrium model with a small number of agents incor-
porating imperfect enforceability of contracts (in the standard Arrow Debreu
model one of the crucial implicit assumptions is perfect enforcement of con-
tracts); their 2001 Econometrica paper compares this model to a standard
Pilch model in a fairly accessible way. Kocherlakota studies the same en-
vironment, but lets agents interact strategically; the no-default constraints
become restrictions derived from the requirement of subgame perfection and
constrained efficiency. Finally, Alvarez and Jermann (2000) show how to, un-
der weak assumptions, decentralize constrained-efficient allocations arising in
Kehoe and Levine and Kocherlakota as sequential markets equilibrium with
a full set of Arrow securities and state-contingent borrowing constraints that
are not “too tight”, in the sense that agents can borrow up to the amount
in which they are indifferent between repaying and defaulting (and being
excluded from future intertemporal trade).
Lately some authors have extended these models to settings with a con-
tinuum of agents (the counterpart of Aiyagari’s model), see Krueger (1999)
and used them to address empirical questions (Krueger and Perri, 2005,

225
226 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

2006a,b). Lustig (2001) develops a method to handle aggregate uncertainty

in these models (as in Krusell and Smith), and uses it to study asset prices.1
In this chapter we will first study how to characterize (and compute) con-
strained efficient allocations for a model with 2 agents, how to decentralize it
as a competitive Arrow-Debreu, a sequential markets and a subgame perfect
equilibrium. We then use it to study the relationship between the variability
of income and the resulting consumption distribution. We then will discuss
how to deal with a continuum of agents, both in terms of theory and in terms
of computation, and discuss other applications.

8.1 The Model

The model is populated by 2 (types of) infinitely lived households, i = 1, 2,
that, in each period, consume a single perishable consumption good. Let by
eit = eit (st ) > 0 denote the stochastic symmetric endowments governed by
the random variable st ∈ S, where the set S is finite. As before let denote
by st = (st , . . . s1 ) ∈ S t denote a history of shocks. The stochastic process is
assumed to be Markov with transition probabilities π(st+1 |st ) and invariant
distribution Π. For given initial shock s0 , whose distribution is given by Π,
the probabilities of endowment shock histories are given by
π(st ) = π(st |st−1 ) ∗ . . . π(s1 |s0 ) (8.1)
Assume that the endowment processes are symmetric in that if e1t (st ) =
e1t (st ) = e, then there exists a ŝt with π(st ) = π(ŝt ) and e2t (st ) = e2t (st ) = e.
That is, both agents face identical stochastic endowment processes.
A consumption allocation is denoted by (c1 , c2 ) = {c1t (st ), c2t (st )}∞ t=1,st ∈S t
and agents are assumed to have preferences over consumption streams given
by
X∞ X
U (c) = (1 − β) β t π(st )u(cit (st )) (8.2)
t=0 st ∈S t

where β ∈ (0, 1) and u is strictly increasing, strictly concave and C 2 and

satisfies INADA conditions.
1
An OLG-version of the Kehoe and Levine model is studied by Azariadis and Lamber-
tini (2001) and applied to the study of social security by Andolfatto and Gervais (2000).
Development economists applied these class of models to study formal and informal risk-
sharing arrangements in rural Africa (see Udry 1995) and India (see Ligon, Thomas and
Worrall 2000).
8.1. THE MODEL 227

The definition of a Pareto efficient allocation is standard, but we repeat

it here for the sake of refreshing memory.
Definition 55 An allocation (c1 , c2 ) is Pareto efficient if it is resource fea-
sible
c1 + c2 = e1 + e2 (8.3)
for all t, all st , and there is no other feasible allocation (ĉ1 , ĉ2 ) such that
U (ĉi ) ≥ U (ci ), with at least one inequality strict.
Note from Chapter 2 that an allocation is Pareto efficient if and only if
it is resource feasible and satisfies
u0 (c1t (st ))
= γ for all t, all st , (8.4)
u0 (c2t (st ))
for some γ > 0.
In this economy there are strong incentives to share the endowment risk
both agents face. The first question we want to pose and answer is what
is the extent of risk sharing possible if agents cannot commit to long term
contracts that are not individually rational? In order to do so we will com-
pute constrained efficient allocations using recursive contract techniques, in
which “promised utility” acts as a state variable. Then we decentralize these
allocation within a
1. Dynamic transfer game: individual rationality translates into subgame
perfection (Kocherlakota 1996)
2. Arrow Debreu competitive market. Individual rationality constraints
enter individual consumption sets (Kehoe and Levine 1993, 2001)
3. Sequential market where agents enter (long-term) relationship with fi-
nancial intermediaries that set appropriate borrowing constraints for
Arrow securities (Alvarez and Jermann, 2000).
Now let us formalize the individual rationality constraints that agents
face in the light of their inability to commit to repaying their debt. First
define the continuation lifetime expected utility from allocation (c1 , c2 ) for
agent i in node st of the event tree as
U (ci , st ) = (1 − β)u(cit (st )) +
X∞ X
(1 − β) β τ −t π(sτ |st )u(ciτ (sτ )) (8.5)
τ =t+1 sτ |st
228 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

We now impose the following individual rationality constraints on alloca-

tions
U (ci , st ) ≥ U (ei , st ) ≡ U i,Aut (st ) (8.6)

These constraints on allocations say that, at no point of time, no contingency

any agent would prefer to walk away from the allocation (c1 , c2 ), with the
consequence of living in financial autarky and consuming her own endowment
from that node onward.

8.2 Constrained Efficient Allocations

Let Γ denote the set of all allocations that satisfy the resource constraints
(8.3) and the individual rationality constraints (8.6). Also define Ū i (s0 ) =
max(c1 ,c2 )∈Γ U (ci , s0 ) and U (s0 ) = U i,Aut (s0 ) as the upper and lower bounds
of lifetime utility agent i can obtain from any allocation that is feasible and
satisfies the individual rationality constraints (remember that endowments
are symmetric). We have the following definition.

Definition 56 An allocation (c1 , c2 ) ∈ Γ is constrained efficient if there is

no other feasible allocation (ĉ1 , ĉ2 ) ∈ Γ such that U (ĉi ) ≥ U (ci ), with at least
one inequality strict.

For given initial s0 , the constrained Pareto Frontier Ws0 : [U 1 (s0 ), Ū 1 (s0 )] →
[U 2 (s0 ), Ū 2 (s0 )] is defined as

Ws0 (U ) = max U (c2 , s0 )

(c1 ,c2 )∈Γ

s.t. U (c1 , s0 ) ≥ U (8.7)

The number Ws0 (U ) is interpreted as the maximal lifetime utility (conditional

on s0 having been realized) agent 2 can obtain from any constrained-feasible
allocation if agent 1 is guaranteed at least lifetime utility [U 1 (s0 ), Ū 1 (s0 )].
We immediately have the following

Proposition 57 An allocation (c1 , c2 ) is constrained efficient if and only if

it solves the above maximization problem, for some U ∈ [U 1 (s0 ), Ū 1 (s0 )].
8.3. RECURSIVE FORMULATION OF THE PROBLEM 229

8.3 Recursive Formulation of the Problem

We now want to construct the constrained efficient consumption allocation
(sometimes called a recursive contract, because it will, as we will see, have the
nature of a long-term risk sharing contract. The basic idea for doing so goes
back to Spear and Srivastava (1987) and Abreu (1988): since the individual
rationality constraints constrain continuation utilities of allocations, make
continuation utility a state variable in the recursive problem.
The constrained Pareto frontier was defined as maximizing the lifetime
utility of the second agent, subject to guaranteeing the first agent at least
a certain lifetime utility U . For the recursive problem this lifetime utility
becomes a state variable (together with the current exogenous state of the
world s). In order to notationally distinguish between the recursive and the
sequential problem let the lifetime utility of agent 1 in the recursive formu-
lation be denoted by w (obviously we can switch the roles of agent 1 and
2).
Thus state variables for the recursive problem are (w, s) and the Bellman
equation reads as2
( )
X
V (w, s) = max (1 − β)u(c2 ) + β π(s0 |s)V (w0 (s0 ), s0 )
(c1 ,c2 ,{w (s )}s0 ∈S )
0 0
s0 ∈S
s.t. (8.8)
c1 + c2 = e1 (st ) + e2 (st ) (8.9)
0 0 0 2,Aut 0
V (w (s ), s ) ≥ U (s ) (8.10)
0 0 1,Aut 0
w (s ) ≥ U (s ) (8.11)
w0 (s0 ) ≤ Ū 1 (s0 ) (8.12)
X
w = (1 − β)u(c1 ) + β π(s0 |s)w0 (s0 ) (8.13)
s0 ∈S

The state space is (w, s) ∈ [U 1 (s0 ), Ū 1 (s0 )]×S. This is a standard dynamic
programming problem that can be solved with standard techniques. Note
that first the values of autarky have to be determined, which is done by
2
I deviate from the timing convention in Kocherlakota, who lets the economy start after
an initial s0 has been drawn, and analyzes allocations from period 1 onward. I formulate
the problem so that consumption also takes place in period 0, after the period 0 shock has
been realized. This is more consistent with my discussion in earlier chapters and will help
to simplify the discussion of the continuum economy.
230 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

solving the equations

X
U i,Aut (s) = (1 − β)u(ei (s)) + β π(s0 |s)U i,Aut (s0 ) (8.14)
s0

These are N =card(S) equations in N unknowns U i,Aut (s), s ∈ S. Once we

have the values of autarky, we can solve the dynamic programming problem3
for policy functions ci (w, s) and w0 (w, s; s0 ).
Sequential consumption allocations are derived from the recursive policy
functions ci (w, s) and w0 (w, s; s0 ) as usual. The initial conditions for the
sequential problem are (U, s0 ). Then we derive the consumption allocation
as follows:

ci0 (s0 ) = ci (U, s0 )

w11 (s1 ) = w0 (U, s0 ; s1 )
w12 (s1 ) = V (w11 (s1 ), s1 ) (8.15)
and recursively
wt1 (st ) = w0 (wt−11
(st−1 ), st−1 ; st )
wt2 (st ) = V (wt1 (st ), st )
cit (st ) = ci (wti (st ), st ) (8.16)
The remaining question is whether these allocations in fact solve the se-
quential planning problem and whether the function Ws0 satisfies Ws0 (U ) =
V (U, s0 ) for all s0 ∈ S. But this is nothing else but proving Bellman’s princi-
ple of optimality, which in fact has been done by Thomas and Worrall (1988).
Kocherlakota provides a partial characterization of constrained efficient allo-
cations; I will defer the discussion of this to the paper presentation and will
discuss a simple example below.

8.4 Decentralization
Since both the papers by Kocherlakota and by Kehoe and Levine will be
presented in class, I will skip the discussion of how constrained efficient al-
locations can be decentralized within a subgame perfect equilibrium of a
3
Strictly speaking one also needs the upper bound on utilities Ū (s). One guesses these
bounds and then iterates between guesses and solutions of the Bellman equation, until one
attains V (Ū (s), s) = U 2,Aut (s) for all s.
8.4. DECENTRALIZATION 231

transfer game or a competitive equilibrium with enforcement constraints. I

will instead present a decentralization of constrained-efficient allocations as a
sequential markets equilibrium with state-contingent borrowing constraints
that is due to Alvarez and Jermann (2000).
This sequential markets equilibrium is almost identical to the one dis-
cussed in Chapter 3; there, however, we picked borrowing constraints that
were very “loose” in that they just prevented Ponzi schemes, but did al-
low first-best consumption smoothing. Now we are looking for borrowing
constraints that have more bite, in that they resemble exactly the incentive
constraints from the social planners problem.
Let by qt (st , st+1 ) denote the price of the Arrow security at node st that
pays out one unit of the consumption good if tomorrow’s shock is st+1 . Also
denote P by ai0 the initial conditions of asset holdings for agent i, where of
course i ai0 = 0. Note that there exists a one-to-one mapping between the
(a10 , a20 ) and the position on the Pareto frontier U. By Ṽ (ait (st ), st ) let denote
the continuation utility of agent i when entering node st with assets ait (st ).
It satisfies
Ṽ (ait (st ), st ) = max
i i
U (ci , st ) (8.17)
c ,a
s.t.
X
cit (st ) + qt (s t
, st+1 )ait+1 (st , st+1 ) ≤ eit (st ) + ait (st ) (8.18)
st+1

ait+1 (st+1 ) ≥ Āi (st+1 ) (8.19)

Evidently Ṽ (ai0 , s0 ) is the lifetime utility of agent i with initial conditions

(ai0 , s0 ) which faces the borrowing constraints Āi (st+1 ). Alvarez and Jermann
define borrowing constraints that are not too tight as satisfying
V (Āit (st ), st ) = U i,Aut (st ) (8.20)
That is, the borrowing limits are such that agents that have borrowed
up the maximum are exactly indifferent between repaying their debt and de-
faulting, with the punishment of default being specified as financial autarky.
The definition of a sequential market equilibrium with borrowing constraints
that are not too tight is standard and hence omitted.
Define Arrow securities prices as
u0 (cit+1 (st , st+1 ))
 
t 0
qt (s , st+1 ) = max βπ(z |z) (8.21)
i=1,2 u0 (cit (st ))
232 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

and implied Arrow-Debreu prices as

Q(st |s0 ) = q0 (s0 , s1 ) ∗ q1 (s1 , s2 ) ∗ . . . ∗ qt−1 (st−1 , st ) (8.22)

For any allocation (c1 , c2 ), the implied interest rate are said to be high if
XX
Q(st |s0 ) ∗ c1t (st ) + c2t (st ) < ∞


(8.23)
t≥0 st

Alvarez and Jermann then prove the following results:

Proposition 58 Suppose a constrained efficient consumption allocation (c1 , c2 )

has high implied interest rates. Then it can be decentralized as sequential mar-
kets equilibrium with some initial conditions and borrowing constraints that
are not too tight.

Proposition 59 Suppose that a constrained efficient allocation (c1 , c2 ) has

some risk sharing, that is, for all t and all st there exists an i such that

U (ci , st ) > U (ei , st ) (8.24)

Then the implied interest rates are high.

Combining these two results we have the following

Corollary 60 Any nonautarkic constrained efficient consumption allocation

can be decentralized as a sequential markets equilibrium with borrowing con-
straints that are not too tight.

The main implication of this result is that one can solve for the entire set
of potential equilibrium allocations by solving the recursive planning problem
and then find Arrow securities prices and borrowing constraints from (8.20)
and (8.21) that, together with the consumption allocation, form a sequential
markets equilibrium.

8.5 An Application: Income and Consump-

tion Inequality
In this section we will consider a simple example of the model below and
use it to study how, in the model, an increase in income inequality affects
8.5. AN APPLICATION: INCOME AND CONSUMPTION INEQUALITY233

the consumption distribution. This example follows the one used in Krueger
and Perri’s (2006) application of limited commitment models to income and
consumption inequality.
Let us assume that the aggregate state of the world can take two values
st ∈ S = {1, 2}. Let individual endowments be given by the functions

1 1 + ε if st = 1
e (st ) = (8.25)
1 − ε if st = 2

2 1 − ε if st = 1
e (st ) = (8.26)
1 + ε if st = 2
That is, if st = 1 then agent 1 has currently high endowment 1 + ε, if st = 2
then agent 2 has currently high endowment. Here ε measures the variability
of individual endowment and also turns out be the (unconditional) standard
deviation of the cross-sectional income distribution.
Which agent is rich follows a stochastic process that is assumed to be
Markov with transition probabilities
 
δ 1−δ
π= (8.27)
1−δ δ
where δ ∈ (0, 1) denotes the conditional probability of agent 1 being rich
tomorrow, conditional on being rich today (and symmetrically for agent 2).
Thus δ is a measure of persistence of the income process, with δ = 21 denoting
the iid case and δ = 1 reflecting a deterministic income process (that is,
permanent shocks). The stationary distribution associated with π, for any
given δ ∈ (0, 1) is given by Π(s) = 21 for all s ∈ S, and we assume that
Π(s0 ) = 21 for all s0 ∈ S. This assumption makes agents ex ante identical and
allows us to restrict attention to symmetric equilibrium allocations.
The good thing about this model is that we can solve for the continuation
expected discounted utility from autarky analytically. Since there are only
two aggregate states at each point of time, the continuation value from the
autarkic allocation can take two values, one for the currently rich agent,
denoted by U (1+ε), and one for the currently poor agent, denoted by U (1−ε).
These two values solve the following recursion

U (1 + ε) = (1 − β)u(1 + ε) + δU (1 + ε) + (1 − δ)U (1 − ε)
U (1 − ε) = (1 − β)u(1 − ε) + δU (1 − ε) + (1 − δ)U (1 + ε) (8.28)
234 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

Solving these two equations in two unknowns yields

1
U (1 + ε) = {(1 − β) u(1 + ε) + β(1 − δ) [u(1 + ε) + u(1 − ε)]}
D
1
U (1 − ε) = {(1 − β) u(1 − ε) + β(1 − δ) [u(1 − ε) + u(1 + ε)]}
(8.29)
D
where
(1 − βδ)2 − (β − βδ)2
D= >0 (8.30)
1−β
The utility from autarky is the weighted sum of the utility from consuming
the endowment today, u(1 + ε) (or u(1 − ε) for the currently poor agent)
and the expected future utility, which is proportional to u(1 + ε) + u(1 −
ε). A change in ε thus changes current consumption and the risk of future
consumption.
We are interested in how the variability of the income process ε affects
the consumption distribution. Since the only reason perfect insurance is not
possible is the presence of the individual rationality constraints, we first want
to characterize the right hand side of these constraints U (1 + ε) and U (1 − ε),
as functions of ε.
Let by U F B = u(1) denote the lifetime utility from the first best, per-
fect risk sharing allocation. This lifetime utility would be obtained by both
agents if the individual rationality constraints were absent, i.e. with complete
markets and no commitment problem.
By simple inspection of (8.29) we can prove the following

Lemma 61 The continuation utilities from the autarkic allocation satisfy

the following properties

1. U (1 + ε)|ε=0 = U (1 − ε)|ε=0 = u(1) ≡ U F B

2. U (1 + ε) and U (1 − ε) are strictly concave and differentiable for all

ε ∈ [0, 1)
dU (1−ε)
3. dε
< 0 for all ε ∈ [0, 1)

4. dU (1+ε) > 0 and limε→1 dU (1+ε) <0

dε  dε
ε=0

5. There exists a unique ε1 = arg maxε∈[0,1) U (1 + ε). For all ε ∈ [0, ε1 ) we

have dU (1+ε)
dε
> 0 and for all ε ∈ (ε1 , 1) we have dU (1+ε)
dε
<0
8.5. AN APPLICATION: INCOME AND CONSUMPTION INEQUALITY235

6. There exists at most one ε2 ∈ (0, 1) such that U (1 + ε2 ) = u(1). If ε2

exists, it satisfies ε2 > ε1 .

The proof of this lemma is straightforward and hence omitted. The in-
terpretation is also fairly simple: without income fluctuations the autarkic
allocation is first-best; since an increase in income variability reduces current
consumption and increases future consumption risk, it reduces the continua-
tion utility for the currently poor agent; for the rich agent initially the direct
effect of currently higher consumption dominates the risk effect, but as ε be-
comes large the risk effect becomes dominant; the last two properties follow
from strict concavity of U (1 + ε) and the signs of the derivatives at ε = 0
and ε → 1. The first figure at the end of this chapter graphically summarizes
the lemma.
We now want to characterize constrained-efficient consumption alloca-
tions in this model and analyze how they change with changes in ε. For this
we note that in any insurance mechanism as the one described in this model it
is efficient to transfer resources from the currently rich to the currently poor
agent. Therefore the constrained-efficient consumption distribution features
maximal insurance, subject to delivering at least the continuation utility of
autarky to the currently rich agent. This argument motivates the following
proposition, which is due to Kehoe and Levine (2001):4

Proposition 62 The constrained-efficient consumption distribution is com-

pletely characterized by a number εc (ε) ≥ 0. The agent with income 1 + ε
consumes 1 + εc (ε) and the agent with income 1 − ε consumes 1 − εc (ε) re-
gardless of her past history. The number εc (ε) is the smallest non-negative
solution of the following equation

max(U F B , U (1 + ε)) = U (1 + εc (ε)) (8.31)

The interpretation of this result goes as follows: if max(U F B , U (1 + ε)) =

FB
U , then εc (ε) = 0 and the resulting allocation is the first-best, complete
risk sharing allocation. From the previous lemma we know that this case
applies to all ε ≥ ε2 , if such ε2 exists. Now suppose that max(U F B , U (1 +
4
This proposition is less general than Kehoe and Levine’s original proposition, because
their environment also contains a long-lived asset. It is more general since, making agents
ex ante identical enables us to establish the characterization of the constrained-efficient
consumption distribution from the initial model period onward, and not just after a tran-
sition period after which initial conditions cease to matter, as in Kehoe and Levine.
236 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

ε)) = U (1 + ε). Obviously one solution for εc (ε) = ε, i.e. the autarkic
consumption allocation. We note from the previous lemma that if ε ≤ ε1 ,
then this is in fact the resulting consumption allocation. In this case, for
small ε, there is no risk sharing possible whatsoever, since at any level of
risk sharing the rich agent would have an incentive to default. However, if
ε ∈ (ε1 , ε2 ) then there exists an εc (ε) < ε1 such that U (1 + ε) = U (1 + εc (ε)).
In this case the resulting consumption allocation features some risk sharing,
but not complete risk sharing. The second figure at the end of this chapter
shows a sample path for the income process and the resulting constrained-
efficient sample path for the consumption allocation.
How does the dispersion of the consumption distribution εc (ε) vary with
an increase in income dispersion. From the previous discussion we immedi-
ately obtain the following proposition (see Krueger and Perri, 2006a):

Proposition 63 For given δ, starting from a given income dispersion ε a

marginal increase in ε leads to a decrease in consumption inequality if and
only if εc (ε) < ε (in the initial equilibrium there is positive risk sharing).
The decrease is strict if and only if 0 < εc (ε) < ε0 (in the initial equilibrium
there is positive, but not complete risk sharing).

This discussion is summarized in the third figure at the end of this chap-
ter. Similarly one can establish how the consumption distribution varies
with changes in the persistence of the income process. See Kehoe and Levine
(2001) or Krueger and Perri (2006a) for a proof.

Proposition 64 For a given income dispersion ε a marginal increase in

persistence δ leads to an increase in consumption inequality. The increase
is strict if and only if 0 < εc (ε, δ) < ε (in the initial equilibrium there is
positive, but not complete risk sharing).

So for this simple example we have a complete characterization of the

cross-sectional consumption distribution and how it varies with the parame-
ters of the model. However, the resulting consumption distribution is some-
what trivial. since the model consists of only two agents. For serious quanti-
tative work we therefore would like a similar model, but with a large number
of agents, in some sense a counterpart of Aiyagari’s (1994) or Huggett’s (1993)
model for a world with limited commitment. The formulation and numerical
solution of such model is discussed next.
8.6. A CONTINUUM ECONOMY 237

8.6 A Continuum Economy

In this section we formulate a version of the above model with a continuum
of infinitely lived agents. Let {yt } denote an agent’s idiosyncratic income
process, which is assumed to be a finite state Markov chain with transi-
tion probabilities π and associated invariant distribution Π. As before an
endowment shock history is denoted by y t , with probability of that history
occurring, conditional on the initial income shock y0 , equal to π(y t |y0 ). We
again assume a law of large numbers so that π(y t |y0 ) is also the deterministic
fraction of the population that started with y0 and has experienced history
y t . It is also assumed that the initial income distribution (and hence the
income distribution at any future date) is given by Π. Agents in this con-
tinuum economy face the same commitment problem as the agents in the
simple economy before.
Let Φ(a0 , y0 ) denote the initial distribution over assets and income. We
first want to compute and characterize constrained-efficient allocations and
then decentralize them as equilibria with borrowing constraints that are not
too tight, in the spirit of Alvarez and Jermann (2000). Note that an agents’
initial wealth level (and initial income realization) determines how much
expected discounted lifetime utility this agent can obtain. Let this utility
level be given by w0 ; we will formally establish the mapping between (a0 , y0 )
and (w0 , y0 ) below. Let the initial distribution of utility promises and income
be denoted by Ψ(w0 , y0 ).
Previously we solved for the constrained-efficient consumption allocation
by maximizing one agent’s expected utility, subject to the promise-keeping
constraint of delivering at least a certain minimum utility level to the other
agent (and, of course, subject to the individual rationality constraints). With
a continuum of agents this is evidently impossible, since there would be a
continuum of promise-keeping constraints. Atkeson and Lucas (1992, 1995)
propose a dual approach to overcome this problem: instead of maximizing
someone’s utility, subject to enforcement constraints, the resource constraint
and individual rationality constraints one minimizes the cost of delivering a
given level of promised utility to a particular agent; the distribution of utility
promises is then adjusted so as to preserve resource feasibility.
Atkeson and Lucas first formally define constrained efficiency (in their
dual sense), then formulate a sequential social planners problem that solves
for constrained-efficient allocations and then make this problem recursive
and finally prove that the policy functions from the recursive problem induce
238 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

constrained-efficient sequential consumption allocations {ct (w0 , y t )}. We will

directly go to the recursive formulation; readers interested in the details may
consult Krueger (1999).
In order to make the cost minimization operative we first have to endow
the social planner with a time discount factor that measures the relative
“price” of resources today versus tomorrow. Let this shadow interest rate be
denoted by R ∈ (1, β1 ). We will see later that this R will correspond to the
risk-free interest rate in the decentralization of the allocations. Also define
the function C ≡ u−1 denote the inverse of the period utility function u. The
entity C(u) is interpreted as the amount of current consumption needed to
generate period utility u. The key idea of Atkeson and Lucas is to realize that,
since the individual rationality constraints involve continuation utilities, one
may let the planner allocate utilities instead of consumption and use promises
of expected discounted utility as state variables.5 The function C may then
be used to translate utility allocations back into consumption allocations.
So let the individual state variables in the dynamic programming problem
be (w, y), that is, the current promise of expected discounted future utility
an agent enters the period with and the current income realization y. In
order to satisfy this utility promise the planner can either give the agent
current utility h or expected utility from tomorrow onward, conditional on
tomorrow’s income shock y 0 , g(y 0 ). The dynamic programming problem the
planner solves is then given by
 
1 1 X
V (w, y) = min0 1 − C(h) + π(y 0 |y)V (g(y 0 ), y 0 ) (8.32)
h,g(y ) R R y0 ∈Y
s.t.
X
w = (1 − β)h + β π(y 0 |y)g(y 0 ) (8.33)
y 0 ∈Y

g(y 0 ) ≥ U Aut (y 0 ) (8.34)

Here V (w, y) is the total (normalized) resource cost the planner has to
minimally spend in order to fulfill his utility promises w, without violating
the individual rationality constraint of the agent: he can’t promise less from
tomorrow onwards, in any state of the world, than the agent would obtain
from the autarkic allocation. As before, multiplying the current cost by the
5
Of course this is not their idea; it goes back at least to Abreu (1986) and Spear and
Srivastava (1987).
8.6. A CONTINUUM ECONOMY 239

factor 1 − R1 expresses total cost in the same units as current cost; it is

an innocuous normalization that just changes the units in which V (w, y) is
measured. Note that if the income process is iid, the variable y does not
appear in the minimization problem and hence the value function V does
not depend on y.
It is a standard exercise to show that the operator induced by this func-
tional equation is a contraction, that the resulting unique fixed point V is
differentiable and strictly convex (strict convexity is not that straightforward,
but note that since u is strictly concave, C is strictly convex) and that the re-
sulting policies h(w, y) and g(w, y; y 0 ) are single-valued continuous functions.
Furthermore h is strictly increasing in w and g is either constant at U Aut (y 0 )
or strictly increasing in w as well.
Attach Lagrange multiplier λ to the promise-keeping constraint (8.33)
and multipliers π(y 0 |y)µ(y 0 ) to the individual rationality constraints (8.34).
The first order conditions and envelope conditions read as
 
1
1− C 0 (h) = λ(1 − β) (8.35)
R
1 0
V (g(y 0 ), y 0 ) = βλ − µ(y 0 ) (8.36)
R
V 0 (w, y) = λ (8.37)
Combining equations (8.36) and (8.37) and the complementary slackness
conditions yields
V 0 (g(y 0 ), y 0 ) = RβV 0 (w, y) − µ(y 0 ) (8.38)
or
V 0 (g(y 0 ), y 0 ) ≤ RβV 0 (w, y)
= if g(y 0 ) > U Aut (y 0 ) (8.39)
Now suppose that income is iid and that Rβ < 1, something that was true
in general equilibrium in the standard incomplete markets model in the last
chapter and is true (under weak conditions) in this model as well (see again
Krueger (1999) or Krueger and Perri (2006b)). Then the Euler equation
(8.39) becomes
V 0 (g(y 0 )) < RβV 0 (w)
= if g(y 0 ) > U Aut (y 0 ) (8.40)
240 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

Since V is strictly convex, this implies that either g(w; y 0 ) < w or g(w; y 0 ) =
U Aut (y 0 ). That is, for states in which the individual rationality constraints are
not binding utility promises for tomorrow are lower than for today; for states
in which the constraint is binding utility promises are raised sufficiently in
order to prevent the agent from reneging. Define w = miny∈Y U Aut (y) and
w̄ = maxy∈Y U Aut (y). It is clear that for all w we must have g(w; y 0 ) ≥
w. On the other hand we have that for ymax = arg maxy∈Y U Aut (y) we
have g(U Aut (ymax ), ymax ) = U Aut (ymax ) and for all w > U Aut (ymax ) we have
g(w, y 0 ) ≤ g(w, ymax ) ≤ U Aut (y 0 ). Thus for iid income shocks the state space
for promised utility w, denoted by W, is bounded: W = [w, w̄]. For 2 income
shocks the fourth figure at the end of this chapter demonstrates the situation;
for a formal proof of the assertions above see Krueger and Perri (2006b). You
may want it instructive to follow an agent with given initial utility promise
w0 through his life with a sequence of income shocks: one positive income
shock brings utility to w = w, with a sequence of bad income shock making
the agent move down in promised utility, until he hits w = w, with a single
good shock putting him back at w = w̄.

So far resource feasibility and the determination of the distribution of

promised utilities has not been discussed; in fact, the beauty of Atkeson and
Lucas’ (1992, 1995) approach is that it separates the dynamic programming
problem the planner has to solve for a given individual (w, y) from the prob-
lems solved for other individuals in society. We want to find a gross real
shadow interest rate and associated invariant measure over utility promises
such that the resources available to the planner equals the resources given
out by the planner.

The remaining discussion of this model has strong similarities with the
discussion in Aiyagari (1994) and Huggett (1993) for standard incomplete
markets models. First, for a given interest rate R, the dynamic programming
problem (??) delivers value function V (w, y) and policy functions h(w, y) and
g(w, y; y 0 ). In order to find the associated stationary distribution over utility
promises and income shocks ΨR we first determine the Markov transition
function QR induced by π and g(w, y; y 0 ). First one has to prove that there is
an upper bound for utility promises w, denoted by w̄ (this is straightforward
for π being iid, but not for the general case). Then denote the state space for
utility promises by W = [w, w̄], define Z = W ×Y and B(Z) = B(W )×P(Y ).
8.6. A CONTINUUM ECONOMY 241

Then the transition function QR : Z × B(Z) → [0, 1] is given by

X  π(y 0 |y) if g(w, y; y 0 ) ∈ A
QR ((w, y), (A, Y)) = (8.41)
0
0 else
y ∈Y

The invariant measure ΨR over (w, y) then satisfies

Z
ΨR (A, Y) = QR ((w, y), (A, Y))dΨR (8.42)

Of course one has to prove that such invariant measure exists and is unique,
the proof of which for the iid case is contained in Krueger (1999). For the
general, non-iid case I’m not aware of any such result, which is similar to the
status of theoretical results available for the standard incomplete markets
model discussed in the previous chapter.
The last figure at the end of this chapter shows the stationary consump-
tion distribution associated with ΨR (w, y), again for the iid case with only
two possible income shocks. Since consumption C(h(w, y)) is strictly increas-
ing in utility promises w, the consumption distribution mimics the utility dis-
tribution ΨR . In particular, all agents with currently high shock consume the
same, and the maximum in the consumption distribution, with agents with
a sequence of bad shocks working themselves down through the consumption
distribution until they have hit rock bottom.
So for a given intertemporal shadow interest rate R we now know how de-
termine the associated stationary utility promise distribution. So far nothing
assures that the social planner can satisfy this distribution of utility promises
with the aggregate resources available to society. Total resources available
are Z X
ydΨR = yΠ(y) (8.43)
y

How about total resources needed by the planner? An agent that enters
this period with utility promises w and current income y is awarded current
utility h(w, y). Thus requires, in terms of resources, C(h(w, y)). Thus total
resources required to satisfy utility distribution ΨR in the current period (and
by stationarity in each period) are
Z
C(h(w, y))dΨR (8.44)
242 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

and therefore the excess resource requirements, as a function of the gross

shadow interest rate R, are given by
Z
d(R) = (C(h(w, y)) − y) dΨR (8.45)

To finish the determination of a stationary constrained-efficient allocation

requires to find an R∗ such that d(R∗ ) = 0. Computationally this can be
done by searching over R ∈ (1, β1 ) for such R∗ . Theoretically, for the iid case
one can show that d(R) is a continuous and increasing function in R on (1, β1 ).
With additional assumptions on u and the income process one can show that
limR→ 1 d(R) > 0 and that limR→1 d(R) ≤ 0, proving existence of a stationary
β
constrained-efficient consumption allocation. Since it is hard to prove that
d(R) is strictly increasing, uniqueness is hard to establish (obviously the set
of efficient R0 s is convex). Again see Krueger (1999) or Krueger and Perri
(2006b) for the details.
Finally we can decentralize a constrained efficient stationary consumption
allocation as a sequential markets equilibrium with borrowing constraints
that are not too tight, in the spirit of Alvarez and Jermann. Agents trade
a full set of Arrow securities; in this model these are securities that pay
out conditional on individual income realizations, rather than aggregate re-
alizations. The determination of the borrowing constraints that are not too
tight is as before and hence omitted. Finally, the price of an Arrow security
bought/sold today by an agent with current individual income shock y that
pays out one unit of consumption if tomorrow’s income shock for that agent
is y 0 turns out to be

π(y 0 |y)
q(y 0 |y) = qπ(y 0 |y) = (8.46)
R
which justifies that R is in fact called a shadow interest rate: it turns out
to be the equilibrium risk free interest rate in the corresponding equilibrium
with borrowing constraints that are not too tight.
Finally we want to discuss the relationship between initial promised utili-
ties w0 and initial assets a0 . So suppose we have found a stationary constrained-
efficient utility distribution Ψ(w, y), a corresponding R∗ and associated value
and policy functions V (w, y), h(w, y) and g(w, y; y 0 ). First, analogously to
(8.15) and (8.16) from the recursive policy functions we can construct sequen-
tial constrained-efficient consumption allocations {ct (w0 , y0 )} for an agent
8.6. A CONTINUUM ECONOMY 243

with initial conditions (w0 , y0 ). What is missing is the connection between

initial conditions (w0 , y0 ) and (a0 , y0 ) (and the corresponding relationship be-
tween Ψ(w0 , y0 ) and Φ(a0 , y0 )). Define Arrow-Debreu prices associated with
(8.46) as

Q(y0 ) = Π(y0 )
Q(y t ) = q(yt |yt−1 ) ∗ . . . q(y1 |y0 )Π(y0 ) (8.47)

Then initial assets associated with (w0 , y0 ), denoted by a0 (w0 , y0 ), are

given as
∞ X
X
Q(y t ) ct (w0 , y t ) − yt


a0 (w0 , y0 ) = (8.48)
t=0 y t |y0

and the associated equilibrium consumption allocations are given by

ct (a0 , y t ) = ct (a−1 t
0 (a0 , y0 ), y ) (8.49)

where w0 = a−1 0 (a0 , y0 ) is the inverse function of (8.48) with respect to the
first argument (note that this function is well-defined because a0 (w0 , y0 ) is
strictly increasing in w0 ). The distribution Φ(a0 , y0 ) is then determined as

Φ(a0 , y0 ) = Ψ(a−1
0 (a0 , y0 ), y0 ) (8.50)

What one then can prove (see Krueger (1999) is that for an initial dis-
tribution Φ(a0 , y0 ) given by (8.50), the allocation determined by (8.49) and
prices (8.46) are a competitive equilibrium with borrowing constraints that
are not too tight (or alternatively, form a competitive equilibrium in the spirit
of Kehoe and Levine (1993), where equilibrium prices are given by (8.47)).
A few final comments: the method of using utility promises as state
variables to make dynamic contracting problems recursive has been used
in a number of applications, for example in the study of optimal unemploy-
ment insurance (Shavell and Weiss 1979, Hopenhayn and Nicolini 1997, Zhao
2000), asset pricing and the equity premium (Alvarez and Jermann 2000,
Lustig 2001), sovereign debt (Atkeson 1991, Kletzer and Wright 2000), redis-
tributive taxation (Krueger and Perri 2006b, Attanasio and Rios-Rull 2001),
time-consistent monetary policy (Chang 1998), time-consistent fiscal policy
(Phelan and Stacchetti 1999, Sleet 1998) and risk sharing in village economies
(Udry 1994, Ligon, Thomas and Worrall 2001).
Also, instead of using utility promises Marcet and Marimon (1999) have
developed a recursive method that use (cumulative) Lagrange multipliers as
244 CHAPTER 8. LIMITED ENFORCEABILITY OF CONTRACTS

state variables. As before, the same problems with the curse of dimensionality
when the number of heterogeneous agents becomes large arises.
So what next: on the methodological side one would like to figure out
how to handle models with a large number of agents and aggregate uncer-
tainty, both theoretically and numerically. Lustig (2001) takes a big step in
that direction. On the substantial side a careful study of redistribution and
insurance over the business cycle, of optimal insurance of large income risks
in the future, both by government policies and private arrangements, and of
the dynamics of the income, consumption and wealth distribution seems to
be fruitful avenues for future research.
Chapter 9

Private Information

In this section we consider another friction that may prevent perfect risk
sharing from occurring as a result of efficient or equilibrium consumption
allocations. As in standard Arrow Debreu theory now households and fi-
nancial intermediaries can write legally binding and enforceable contracts.
However, we now assume that individual income realizations (and individual
consumption) is private information of the agent. Financial intermediaries or
the social planner have to rely on reports of income by agents. Consumption
allocations therefore have to structured in such a way that households find
it optimal to tell the truth about their income realization, rather than to
lie about it. We will first consider the problem of a financial intermediary
dealing with a single agent in isolation, before discussing a model with many
agents, an aggregate resource constraint and an endogenous interest rate

9.1 Partial Equilibrium

Our treatment of the partial equilibrium case is motivated by the seminal
papers by Green (1987) and Thomas and Worrall (1990), which in turn is
nicely discussed in Ljungqvist and Sargent’s book. Consider a risk-neutral
financial intermediary that lives forever and discounts the future at time
discount factor β ∈ (0, 1). This financial intermediary (sometimes called
principal) engages in a long-term relationship with a risk averse household
(sometimes called agent) that also discounts the future at factor β. The agent
faces a stochastic income process {yt }, assumed to be iid with finite support
Y = {y1 , y2 , . . . yN } and probabilities (π1 , π2 , . . . , πN ); and seeks insurance

245
246 CHAPTER 9. PRIVATE INFORMATION

from the risk-neutral principal. Both parties can commit to long-term con-
tracts, so that in the absence of private information the optimal consumption
allocation for the agent, subject to the financial intermediary breaking even,
is
ct (y t ) = E(yt ) = E(y) = 1
where the last equality is by assumption (that is, we normalize mean income
to 1). Thus the agent hands over his realized income in every period and
receives constant consumption equal to mean income back from the financial
intermediary. His lifetime utility equals
∞ X
X
U (c) = (1 − β) πt (y t )β t u(ct (y t )) = u(1)
t=0 yt

and expected utility of the financial intermediary (or profits) equal

∞ X
X
πt (y t )β t yt − ct (y t )


W (c) = (1 − β)
t=0 yt
∞
X X
= (1 − β) βt πt (y t ) (yt − E(yt ))
t=0 yt
= 0

The big problem with this consumption allocation is that the agent’s
income realizations are private information. If the agent is promised constant
consumption independent of his income report, then he would always report
yt = y1 , keep the difference between his true income and the report, receive
1 from the financial intermediary and do strictly better. Then, however, the
principal would lose money, since his profits would be
∞ X
X
W = (1 − β) πt (y t )β t (y1 − 1)
t=0 yt
= y1 − 1 < 0.

Thus the tension in the current environment is between the provision of

insurance (without the informational frictions it is efficient for the financial
intermediary to insure the agent, since he is better taking risk because of his
risk-neutrality) and the provision of incentives to tell the truth (and we saw
9.1. PARTIAL EQUILIBRIUM 247

above that without providing these incentives the principal is going to lose
money).
What we want to construct in the following is the efficient long-term
insurance contract between the two parties, explicitly taking into account
the informational frictions in this environment. We will again immediately
proceed to the recursive formulation of the problem, understanding that in
principle one should first write down the sequential problem, then the re-
cursive problem and then prove the principle of optimality.1 The recursive
formulation of the contracting problem again makes use of promised lifetime
utility w as a state variable. Let us first pose the dynamic programming
problem and then explain it:
X
V (w) = min πs [(1 − β)ts + βV (ws )] (9.1)
{ts ,ws }N
s=1

s.t.
X
w = πs [(1 − β)u(ts + ys ) + βws ] (9.2)
(1 − β)u(ts + ys ) + βws ≥ (1 − β)u(tk + ys ) + βwk ∀s, k (9.3)
ts ∈ [−ys , ∞) (9.4)
ws ∈ [w, w̄] (9.5)

We cast the optimal contracting problem as the problem of minimizing cost

of the contract for the principal, subject to delivering lifetime utility w to
the agent (as assured by the promise-keeping constraint (9.2)), subject to
the agent having the correct incentives to report income truthfully (see con-
straints (9.3)), and subject to the domain restrictions (9.4) and (9.5) which
we will discuss below.2 The choice variables are the state-dependent transfers
the principal makes to the agent today, ts , and the state-dependent continu-
ation utility promises ws from tomorrow onwards.
1
In fact, the incentive constraints in the sequential formulation are much more complex.
That they can be simplyfied to period-by-period incentive constraints (so-called temporary
incentive compatibility constraints, in Green’s (1987) language) is one of the major results
proved in Green’s paper. Only because of this result do we obtain a simple recursive
formulation of the problem.
2
We can restrict ourselves to contracts that induce truthtelling behavior because the
revelation principle assures that whatever insurance can be provided with an arbitrary
mechanism can also be provided with a mechanism (contract) in which agents report their
income directly (a direct mechanism) and are provided with the appropriate incentives to
report income truthfully.
248 CHAPTER 9. PRIVATE INFORMATION

The constraints (9.3) say the following: suppose state s with associated
income realization ys is realized. If the agent reports the truth, he receives
transfers ts and continuation utility ws , for total lifetime utility

u(ts + ys ) + βws .

If, on the other hand, he falsely reports yk , he receives transfers tk and

continuation utility wk , for total lifetime utility

u(tk + ys ) + βwk .

Note that even when mis-reporting income, his true lifetime utility depends
on his true, rather than his reported income ys . The constraints (9.3) simply
state that it has to be in the agents’ own interest to truthfully reveal his
income. The domain restriction on transfers is self-explanatory: the principal
can make arbitrarily large transfers, but cannot force the agent to make
payments bigger than their current (truthfully reported) income ys . With
respect to the bounds on promised utility we note the following. Let w̄ =
supc u(c) and w = inf c u(c); evidently it is not possible to deliver more lifetime
utility that w̄ even with infinite resources and it is not possible to deliver less
lifetime utility than w even without giving the agent any consumption even.
We explicitly allow w = −∞ and w̄ = +∞, but will make more restrictive
assumptions later. For now we assume that
Assumption: The utility function u : [0, ∞) → R is strictly increasing,
strictly concave, at least twice differentiable and satisfies the Inada condi-
tions.
So far is has been left unclear what determines the w the agent will start
the contract with. A higher initial w means more lifetime utility for the
agent, but less lifetime utility for the principal. The expected lifetime utility
for the principal, conditional on delivering w to the agent, is given by

W (w) = 1 − V (w)

since V (w) measures the lifetime expected costs from the contract and 1
measures the expected revenues from the agent, equal to his expected per
period and (by normalization of 1 − β) lifetime revenue. Thus varying the w
traces out the constrained utility possibility frontier between principal and
agent. One particularly important w is that w that solves V (w) = 1, that is,
that lifetime utility of the agent that yields 0 profits for the principal. If V
is strictly increasing in w, such w is necessarily unique.
9.1. PARTIAL EQUILIBRIUM 249

9.1.1 Properties of the Recursive Problem

By the standard contraction mapping arguments a unique solution V to the
functional equation exists Let us first bound V (w). The cheapest way to
provide lifetime utility w is by delivering constant consumption. This may
not be incentive compatible, in fact it will not be, but surely provides a lower
bound on V (w). So define cf b (w) as

u(c) = w

or cf b (w) = u−1 (w). The value of the transfers needed to deliver the constant
consumption stream cf b (w) are given by

V (w) = cf b (w) − 1.

Obviously
V (w) ≤ V (w)
with strict inequality if any of the incentive constraints is binding. On the
other hand, the principal can decide to give constant transfers ts = t (and
constant continuation utility ws = w) Surely, since transfers are independent
of reported income, this induces truth-telling. In order to obtain the utility
promise w, the transfer has to satisfy
X
πs u(ys + t) = w
s

Denote this transfer by t̄(w). The cost of this policy is given by

V̄ (w) = t̄(w)

and evidently
V (w) ≤ V̄ (w)
with strict inequality whenever the distribution of income shocks is not de-
generate.
Observe the following facts. Since cf b (w) is strictly increasing and strictly
convex, so is V (w). Furthermore, as long as u satisfies the Inada conditions we
have V 0 (w) = 0 and V 0 (w̄) = ∞, since u(.) and cf b (.) are inverse functions
of each other. Finally we have V (w) = −1 and V (w̄) = ∞. Similarly,
the function t̄(w) and thus V̄ (w) are strictly increasing and strictly convex,
250 CHAPTER 9. PRIVATE INFORMATION

with V̄ (w̄) = ∞. While this does not necessarily mean that V (w) is strictly
increasing (it is straightforward to show that it is) and strictly convex (it is
harder to show that it is), this discussion gives us a fairly tight bound on
V (w).
The fact that V (w) is strictly increasing in w is intuitive: delivering
higher lifetime utility to the agent requires higher transfers on average by
the principal. Since marginal utility is declining, a given additional unit of
transfers increases utility by less and less, thus one would expect the cost
function to be strictly convex. We will go ahead and assert this without
proof.
Now we discuss further properties of the optimal contract. Consider the
constraints

(1 − β)u(ts + ys ) + βws ≥ (1 − β)u(ts−1 + ys ) + βws−1 (9.6)

(1 − β)u(ts−1 + ys−1 ) + βws−1 ≥ (1 − β)u(ts + ys−1 ) + βws (9.7)

Adding them results in

u(ts + ys ) + u(ts−1 + ys−1 ) ≥ u(ts−1 + ys ) + u(ts + ys−1 )

or
u(ts + ys ) − u(ts−1 + ys ) ≥ u(ts + ys−1 ) − u(ts−1 + ys−1 )
Since u is strictly concave and ys > ys−1 we have ts−1 ≥ ts (it helps to draw a
picture to convince you of this). From (9.6) it then easily follows that ws ≥
ws−1 . That is, it is efficient to give households with lower income realizations
higher transfers. But in order to provide the incentives of not always claiming
to have had low income realizations, these households are “punished” with
lower continuation utilities. We summarize this in the following

Proposition 65 In an efficient contract, lower income reports are rewarded

with higher current transfers, but lower continuation utilities: ts−1 ≥ ts and
ws ≥ ws−1 for all s ∈ S

Second we claim that the principal only has to worry about the agent
lying locally, that is, if income is ys , the principal has to only worry about
the agent reporting ys−1 or ys+1 . As long as the agent does not have an
incentive to lie locally, he does not want to lie and report ys−2 or ys+2 or
even more extreme values. Formally
9.1. PARTIAL EQUILIBRIUM 251

Proposition 66 Suppose that the true income state is s and that all local
constraints of the form

(1 − β)u(ts + ys ) + βws ≥ (1 − β)u(ts+1 + ys ) + βws+1 (9.8)

(1 − β)u(ts + ys ) + βws ≥ (1 − β)u(ts−1 + ys ) + βws−1 (9.9)

hold. Then all other incentive constraints are satisfied.

Proof. We will show that, if these constraints are satisfied, then a house-
hold would not want to report s − 1 when the true state is s + 1. A simple
repetition of the argument will then show that the household would not want
to report s − 1 for any state k ≥ s. The case of mis-reporting upwards is, of
course, symmetric and hence omitted.
Since s > s − 1, we have, from the previous result, ts ≤ ts−1 and thus
(again by concavity of u)

u(ts + ys+1 ) − u(ts + ys ) ≥ u(ts−1 + ys+1 ) − u(ts−1 + ys ) (9.10)

Multiplying (9.10) by (1 − β) and adding it to (9.9) yields

(1 − β)u(ts + ys+1 ) + βws ≥ (1 − β)u(ts−1 + ys+1 ) + βws−1

But
(1 − β)u(ts+1 + ys+1 ) + βws+1 ≥ (1 − β)u(ts + ys+1 ) + βws
(use (9.9), but for s + 1) and thus

(1 − β)u(ts+1 + ys+1 ) + βws+1 ≥ (1 − β)u(ts + ys+1 ) + βws

≥ (1 − β)u(ts−1 + ys+1 ) + βws−1

That is, if the household does not want to lie one state down from s to s − 1,
he does not want to lie two states down, from s + 1 to s − 1 either.
This result reduces the set of incentive constraints substantially. It turns
out that we can reduce it even further. Without proof we state the following
proposition (if you are interested in the proof, consult Thomas and Worrall
(1990) or the discussion of the same paper in Sargent and Ljungqvist’s book;
note that the proof requires strict convexity of the cost function V ).

Proposition 67 The local downward constraints (9.9) are always binding,

the local upward constraints (9.8) are never binding.
252 CHAPTER 9. PRIVATE INFORMATION

Proof. Omitted
Intuitively, this result makes a lot of sense. Remember, the purpose of the
contract is for the principal to insure the agent against bad income shocks,
with high transfers in states with bad income realizations. This, of course,
triggers the incentive to report lower income than actually realized, rather
than higher income. Thus the lying-down constraints are the crucial con-
straints. With this result the dynamic programming problem becomes more
manageable. In particular, if N = 2, there is only one incentive constraint
and the promise keeping constraint. It is then quite feasible to characterize
the qualitative properties of the optimal contract in more detail, which we
leave to the discussion of Atkeson and Lucas (1992) in class.
Let us consider instead a simple example with N = 2 income states. From
our characterization of the binding pattern of the constraints the dynamic
programming problem becomes
V (w) = min π [(1 − β)t1 + βV (w1 )] + (1 − π) [(1 − β)t2 + βV (w2 )]
{t1 ,t2 ,w1 ,w2 }
s.t.
w = π [(1 − β)u(t1 + y1 ) + βw1 ] + (1 − π) [(1 − β)u(t2 + y2 ) + βw2 ]
(1 − β)u(t2 + y2 ) + βw2 ≥ (1 − β)u(t1 + y2 ) + βw1
where π is the probability of low income y1 . Let us proceed under the as-
sumption that V is differentiable (which is not straightforward to prove).
The the first order and envelope conditions read as
π(1 − β) − λπ(1 − β)u0 (t1 + y1 ) + µ(1 − β)u0 (t1 + y2 ) = 0
(9.11)
0 0
(1 − π)(1 − β) − λ(1 − π)(1 − β)u (t2 + y2 ) − µ(1 − β)u (t2 + y2 ) = 0
(9.12)
0
πβV (w1 ) − λπβ + µβ = 0
(9.13)
0
(1 − π)βV (w2 ) − λ(1 − π)β − µβ = 0
(9.14)
0
V (w) = λ
(9.15)
Rewriting equations (9.13) and (9.14) and using (9.15) yields
µ µ
λ = V 0 (w) = V 0 (w1 ) + = V 0 (w2 ) −
π 1−π
9.2. ENDOGENOUS INTEREST RATES IN GENERAL EQUILIBRIUM253

Thus, since µ > 0, we have (by strict convexity of the cost function V ) that
w1 < w < w2 , that is, utility promises spread out over time. Rewriting (9.11)
and (9.12) yields
µ
1 = λu0 (t1 + y1 ) − u0 (t1 + y2 )
π
 µ 0 µ  µ 0
= λ− u (t1 + y1 ) + [u0 (t1 + y1 ) − u0 (t1 + y2 )] > λ − u (t1 + y1 )
 π  π π
µ
1 = λ+ u0 (t2 + y2 )
1−π

and thus
µ 0
1 = V 0 (w2 )u0 (c2 ) = V 0 (w1 )u0 (c1 ) + [u (c1 ) − u0 (t1 + y2 )]
π
> V 0 (w1 )u0 (c1 )

The first equality (and again using strict convexity of the cost function) show
that if future promises w2 are increased (in response to, say, an increase in w),
then it is necessarily optimal to also increase c2 . That is, the principal should
spread costs over time, increasing both current costs and well as future costs.
[To be completed]

9.2 Endogenous Interest Rates in General Equi-

librium
See the papers by Phelan and Townsend (1991) and Atkeson and Lucas 1992,
1995 should be discussed

9.3 Applications
9.3.1 New Dynamic Public Finance
Paper that makes the closest connection to the traditional Ramsey tax lit-
erature is Werning (2007). Key paper starting this literature is Golosov,
Kocherlakota and Tsyvinski (2003) and accessible summaries are Golosov,
Tsyvinski and Werning (2006) as well as Kocherlakota (20xx).
254 CHAPTER 9. PRIVATE INFORMATION
 Part IV

Conclusions

255
Bibliography

[1] Aguiar, M. and E. Hurst (2005), “Consumption vs. Expenditure,”

Journal of Political Economy, 113, 919-948.
[2] Aguiar, M. and E. Hurst (2007), “Lifecycle Prices and Production,”
American Economic Review, 97, 1533-1559.
[3] Aiyagari, R. (1994), “Uninsured Risk and Aggregate Saving,” Quar-
terly Journal of Economics, 109, 659-684.
[4] Aiyagari, R. (1995), “Optimal Capital Income Taxation with Incom-
plete Markets, Borrowing Constraints, and Constant Discounting,”
Journal of Political Economy, 103, 1158-1175.
[5] Aiyagari, R. and McGrattan, E. (1998), “The Optimum Quantity of
Debt,” Journal of Monetary Economics, 42, 447-469.
[6] Ales, L. and P. Maziero (2007), “Accounting for Private Information,”
Working Paper, University of Minnesota.
[7] Altug, S. and Miller, R. (1990), “Household Choices in Equilibrium,”
Econometrica, 58, 543-70.
[8] Atkeson, A. and Lucas, R. (1992), “On Efficient Distribution with Pri-
vate Information,” Review of Economic Studies, 59, 427-453.
[9] Atkeson, A. and Lucas, R. (1995), “Efficiency and Equality in a Sim-
ple Model of Efficient Unemployment Insurance,” Journal of Economic
Theory, 66, 64-88.
[10] Attanasio, O. (1992), “Intertemporal Substitution in Consumption:
Theory and Evidence,” Carnegie Rochester Series on Public Policy,
supplement to Journal of Monetary Economics.

257
258 BIBLIOGRAPHY

[11] Attanasio, O. (1999), “Consumption”, in Handbook of Macroeconomics,

J. Taylor and M. Woodford (eds.), Elsevier Science.

[12] Attanasio, O. and M. Browing (1995), “Consumption over the Life

Cycle and over the Business Cycle,” American Economic Review, 85,
1118-1137.

[13] Attanasio, O. and S. Davis (1996), “Relative Wage Movements and

the Distribution of Consumption,” Journal of Political Economy , 104,
1227-1262.

[14] Attanasio, O. and De Leire, T. (1994), “IRAs and Household Saving

Revisited: Some New Evidence,” NBER Working Paper 4900.

[15] Attanasio, O. and Weber, G. (1993), “Consumption Growth, the Inter-

est Rate and Aggregation,” Review of Economic Studies, 60, 631-649.

[16] Attanasio, O. and Weber, G. (1995), “Is Consumption Growth Consis-

tent with Intertemporal Optimization? Evidence from the Consumer
Expenditure Survey,” Journal of Political Economy, 103, 1121-1157.

[17] Attanasio, O., J. Banks, C. Meghir and G. Weber, (1999), “Humps and
Bumps in Lifetime Consumption”, Journal of Business and Economic
Statistics.

[18] Barro, R. (1974), “Are Government Bonds Net Wealth?,” Journal of

Political Economy, 82, 1095-1117.

[19] Barro, R. (2009), “Rare Disasters, Asset Prices, and Welfare Costs,”
American Economic Review, 99, 243–264.

[20] Barsky, R., G. Mankiw and S. Zeldes (1986), “Ricardian Consumers

with Keynesian Propensities,” American Economic Review, 76, 676-
691.

[21] Ben-Porath, Y. (1967), “The Production of Human Capital and the

Life Cycle of Earnings,” Journal of Political Economy, 75, 352-365.

[22] Bernheim, D. (1991), “How Strong are Bequest Motives? Evidence

Based on Estimates on the Demand for Annuities”, Journal of Political
Economy, 99, 899-927.
BIBLIOGRAPHY 259

[23] Bewley, T. (1986), “Stationary Monetary Equilibrium with a Con-

tinuum of Independently Fluctuating Consumers,” in Werner Hilden-
brand and Andreu Mas-Colell (eds.) Contributions to Mathematical
Economics in Honor of Gerard Debreu, North-Holland, 79-102.
[24] Blanchard, O. and Mankiw, G. (1988), “Consumption: Beyond Cer-
tainty Equivalence”, American Economic Review Papers and Proceed-
ings, 78, 173-77.
[25] Blundell, R. and I. Preston (1998), “Consumption Inequality and In-
come Uncertainty, Quarterly Journal of Economics, 113, 603-640.
[26] Blundell, R., L. Pistaferri and I. Preston (2007), “Imputing Consump-
tion in the PSID using Food Demand Estimates from the CEX,” Work-
ing Paper, Stanford University.
[27] Blundell, R., L. Pistaferri and I. Preston (2007), “Consumption In-
equality and Partial Insurance,” American Economic Review, 98, 1887-
1921.
[28] Browning, M. and T. Crossley (2003), “Shocks, Stocks and Socks,
“Smoothing Consumption over a Temporary Income Loss,” Working
Paper, McMaster University.
[29] Browning, M.. and Lusardi, L. (1996), “Household Saving: Micro The-
ories and Micro Facts”, Journal of Economic Literature, 34, 1797-855.
[30] Browning, M. and Meghir, C. (1991), “The Effects of Male and Female
Labour Supply on Commodity Demands”, Econometrica 59, 925-51.
[31] Browning, M., Deaton, A. and Irish, M. (1985), “A Profitable Ap-
proach to Labor Supply and Commodity Demands over the Life-Cycle,”
Econometrica, 53, 503-43.
[32] Caballero, R. (1990), “Consumption Puzzles and Precautionary Sav-
ings,” Journal of Monetary Economics, 25, 113-36.
[33] Caballero, R. (1991), “Earnings Uncertainty and Aggregate Wealth
Accumulation,” American Economic Review, 81, 859-571.
[34] Campbell, J. (1999), “Asset Prices, Consumption and the Business
Cycle” in Handbook of Macroeconomics, Vol. 1C
260 BIBLIOGRAPHY

[35] Campbell, J. (2003), “Consumption-Based Asset Princing,” in Hand-

book of the Economics of Finance, Vol. 1B

[36] Campbell, J. and Deaton, A. (1989), “Why is Consumption so

Smooth?,” Review of Economic Studies, 56, 357-74

[37] Campbell, J. and Mankiw, G. (1989), “Consumption, Income, and

Interest Rates: Reinterpreting the Time Series Evidence”, in NBER
Macroeconomics Annual, The MIT Press.

[38] Carroll, C. (1992), “The Buffer-Stock Theory of Saving: Some Macroe-

conomic Evidence,” Brookings Papers on Economic Activity, 61-135.

[39] Carroll, C. (1997), “Buffer-Stock Saving and the Life Cycle/Permanent

Income Hypothesis,” Quarterly Journal of Economics, 112, 1-55.

[40] Carroll, C. (2001), “Death to the Log-Linearized Consumption Euler

Equation! (And Very Poor Health to the Second-Order Approxima-
tion),” Advances in Macroeconomics, 1, Article 6.

[41] Carroll, C. (2006), “The Method of Endogenous Gridpoints for Solving

Dynamic Stochastic Optimization Problems,” Economics Letters, 312-
320.

[42] Carroll, C. and Samwick, A. (1997), “The Nature of Precautionary

Saving,” Journal of Monetary Economics, 40, 41-71.

[43] Caselli, F. and Ventura, J. (2000), “A Representative Consumer Theory

of Distribution,” American Economic Review, 90, 909-926.

[44] Chamberlain, G. (1984), “Panel Data”, in Handbook of Econometrics,

vol. 2, Zvi Griliches and Michael D. Intriligator (eds.) Elsevier Science.

[45] Chamberlain, G. and C. Wilson (2000), “Optimal Intertemporal Con-

sumption under Uncertainty,” Review of Economic Dynamics, 365-395.

[46] Chambers, M., C. Garriga and D. Schlagenhauf (2007), “Accounting for

Changes in the Homeownership Rate,” Working Paper, Florida State
University.
BIBLIOGRAPHY 261

[47] Chatterjee, S. (1994), “Transitional Dynamics and the Distribution of

Wealth in a Neoclassical Growth Model,” Journal of Public Economics,
54, 97-119.

[48] Chatterjee, S., D. Corbae, M. Nakajima and V. Rios-Rull (2007), “A

Quantitative Theory of Unsecured Consumer Credit with Risk of De-
fault,” Econometrica, 75, 1525-1590.

[49] Clarida, R. (1987), “Consumption, Liquidity Constraints, and Asset

Accumulation in the Presence of Random income Fluctuations,” Inter-
national Economic Review, 28, 339-351.

[50] Cochrane, J. (1991), “A Simple Test of Consumption Insurance”, Jour-

nal of Political Economy, 99, 957-976.

[51] Conesa, J. and Krueger, D. (1999), “Social Security Reform with Het-
erogeneous Agents,” Review of Economic Dynamics, 2, 757-795.

[52] Conesa, J. and Krueger, D. (2006), “On the Optimal Progressivity of

the Income Tax Code,” Journal of Monetary Economics, 53, 1425-1450.

[53] Cooley, T. and Soares, J. (1999), “A Positive Theory of Social Security

Based on Reputation,” Journal of Political Economy, 107, 135-160.

[54] Cox, D (1987), “Motives for Private Income Transfers,” Journal of

Political Economy, 95, 508-46.

[55] Dávila, J., J. Hong, P. Krusell and J. Rı́os-Rull (2011), “Constrained

Efficiency in the Neoclassical Growth Model with Uninsurable Idiosyn-
cratic Shocks,” Working paper.

[56] Deaton, A. (1985), “Panel Data from Time Series of Cross-Sections,”

Journal of Econometrics, 30, 109-26.

[57] Deaton, A. (1991), “Saving and Liquidity Constraints,” Econometrica,

59, 1221-1248.

[58] Deaton, A. (1992), Understanding Consumption, Oxford University

Press.

[59] Deaton, A. (1997), The Analysis of Household Surveys, The Johns

Hopkins University Press.
262 BIBLIOGRAPHY

[60] Deaton, A. and Paxson, C (1994), “Intertemporal Choice and Inequal-

ity,” Journal of Political Economy, 102, 384-94.

[61] Deaton, A. and Paxson, C. (1997), “The Effects of Economic and Pop-
ulation Growth on National Saving and Inequality,” Demography, 34,
97-114.

[62] Diamond, P. and Hausman, J. (1984), “Individual Retirement and Sav-

ing Behaviour,” Journal of Public Economics, 23, 81-114.

[63] Dynan, K. (1993), “How Prudent are Consumers?,” Journal of Political

Economy, 101, 1104-13.

[64] Engle, R. and C. Granger (1987), “Co-Integration and Error Correc-

tion: Representation, Estimation and Testing,” Econometrica, 251-276.

[65] Engen, E., Gale, W. and Scholz, J. (1996), “The Illusory Effect of
Saving Incentives on Saving”, Journal of Economic Perspectives, 10,
113-38.

[66] Farber, H. and Gibbons, R (1996), “Learning and Wage Dynamics,”

Quarterly Journal of Economics, 111, 1007-1047.

[67] Fernandes, A. and Phelan, C. (2000), ”A Recursive Formulation for Re-

peated Agency with History Dependence,” Journal of Economic The-
ory, 91, 223-247.

[68] Fernandez-Villaverde, J. and D. Krueger (2002), “Consumption and

Saving over the Life Cycle: How Important are Consumer Durables?”
Proceedings of the 2002 North American Summer Meetings of the
Econometric Society: Macroeconomic Theory.

[69] Fernandez-Villaverde, J. and D. Krueger (2007), “Consumption over

the Life Cycle: Facts from the Consumer Expenditure Survey” Review
of Economics and Statistics, 89, 552-565.

[70] Flavin, M. (1981), “The Adjustment of Consumption to Changing Ex-

pectations About Future Income,” Journal of Political Economy, 89,
974-1009.
BIBLIOGRAPHY 263

[71] Galı̀, J. (1990), “Finite Horizon, Life-Cycle Savings and Time Series
Evidence on Consumption,” Journal of Monetary Economics, 26, 433-
52.
[72] Gervais, M. and P. Klein (2010), “Measuring Consumption Smoothing
in CEX Data,” Journal of Monetary Economics, 57, 988-999.
[73] Gollier, C. (2002), “What Does the Classical Theory Have to Say About
Household Portfolios?,” in Household Portfolios, L. Guiso, M. Halias-
sos, T. Jappelli (eds.), MIT Press, Cabridge, MA.
[74] Golosov, M., N. Kocherlakota and A. Tsyvinski (2003), “Optimal Indi-
rect and Capital Taxation,” Review of Economic Studies, 70, 569-587.
[75] Golosov, M., A. Tsyvinski and I. Werning (2006), “New Dynamic Pub-
lic Finance: A User’s Guide,” NBER Macroeconomics Annual 2006,
317-363.
[76] Gourinchas, P. and Parker, J. (2002), “Consumption over the Life Cy-
cle,” Econometrica 70, 47-89.
[77] Green, E. (1987), “Lending and the Smoothing of Uninsurable In-
come,” in: Prescott, E. and Wallace, N. (eds) Contractual Arrange-
ments for Intertemporal Trade, University of Minnesota Press, 3-25.
[78] Guiso, L., Jappelli, T. and Terlizzese, D. (1992), “Earnings Uncertainty
and Precautionary Saving,” Journal of Monetary Economics 30, 307-
37.
[79] Guvenen, F. (2007a), “An Empirical Investigation of Labor Income
Processes,” Working Paper, University of Texas
[80] Guvenen, F. (2007b), “Learning your Earning: Are Labor Income
Shocks Really Very Persistent?” American Economic Review, 97, 687-
712.
[81] Haliassos, M. and Bertaut, C. (1995), “Why Do so Few Hold Stocks?,”
Economic Journal, 105, 1110-29.
[82] Hall, R. (1978), “Stochastic Implications of the Life Cycle-Permanent
Income Hypothesis: Theory and Evidence,” Journal of Political Econ-
omy, 86, 971-987.
264 BIBLIOGRAPHY

[83] Hall, R. (1988), “Intertemporal Substitution in Consumption,” Journal

of Political Economy, 86, 971-87.

[84] Hansen, L. and Singleton, K. (1983), “Consumption, Risk Aversion and

the Temporal Behavior of Stock Market Returns,” Journal of Political
Economy 91, 249-265.

[85] Hayashi, F. (1982), “The Permanent Income Hypothesis: Estimation

and Testing by Instrumental Variables,” Journal of Political Economy,
90, 895-916.

[86] Hayashi, F. (1985), “The Permanent Income Hypothesis and Consump-

tion Durability: Analysis Based on Japanese Panel Data,” Quarterly
Journal of Economics, 100, 1083-113.

[87] Hayashi, F. (1987), ”Tests for Liquidity Constraints: A Critical Survey

and Some New Observations,” in: Bewley, T. (ed.) Advances in Econo-
metrics, Vol. 2, Fifth World Congress, Cambridge University Press,
91-120.

[88] Hayashi, F. (1997) Understanding Saving, MIT Press.

[89] Hayashi, F., Altonji, J. and Kotlikoff, L. (1996), “Risk Sharing Between
and Within Families,” Econometrica, 64, 261-94.

[90] Heathcote, J. (2005), “Fiscal Policy with Heterogeneous Agents and

Incomplete Markets,” Review of Economic Studies, 72, 161-188.

[91] Heathcote, J., K. Storesletten and G. Violante (2004a), “The Macroe-

conomic Implications of Rising Wage Inequality in the United States,”
Working Paper, New York University.

[92] Heathcote, J., K. Storesletten and G. Violante (2004b), “Two Views

of Inequality Over the Life-Cycle,” Journal of the European Economic
Association, 3, 765-775.

[93] Heathcote, J., K. Storesletten and G. Violante (2007a), “Consumption

and Labor Supply with Partial Insurance: An Analytical Framework,”
Working Paper, New York University.
BIBLIOGRAPHY 265

[94] Heathcote, J., K. Storesletten and G. Violante (2007b), “Insurance and

Opportunities: The Welfare Implications of Rising Wage Dispersion,”
Working Paper, New York University.

[95] Heathcote, J., K. Storesletten and G. Violante (2009), “Quantitative

Macroeconomics with Heterogeneous Households,” Annual Review of
Economics, 1, 319-354.

[96] Heaton, J. (1995), “An Empirical Investigation of Asset Pricing with

Temporally Dependent Preference Specifications,” Econometrica, 63,
681-717.

[97] Heckman, J. (1974), “Life Cycle Consumption and Labor Supply: An

Explanation of the Relationship between Income and Consumption
Over the Life Cycle,” American Economic Review, 64, 188-194.

[98] Heer, B. and A. Maussner (2009), Dynamic General Equilibrium Mod-

eling (2nd ed.), Springer-Verlag, Berlin, Germany.

[99] Hopenhayn, H. and E. Prescott (1992), “Stochastic Monotonicity and

Stationary Distributions for Dynamic Economies,” Econometrica, 60,
1387-406.

[100] Huang, K. and J. Werner (2004), “Implementing Arrow-Debreu Equi-

libria by Trading Infinitely-Lived Securities,” Economic Theory, 24,
603-623.

[101] Hubbard, G., Skinner, J. and Zeldes, S. (1994), “The Importance of

Precautionary Motives in Explaining Individual and Aggregate Sav-
ing,” Carnegie Rochester Series on Public Policy, 40, 59-125.

[102] Hubbard, G., Skinner, J. and Zeldes, S. (1995), “Precautionary Saving

and Social Insurance,” Journal of Political Economy, 103, 360-99.

[103] Huggett, M. (1993), “The Risk-Free Rate in Heterogeneous-Agent

Incomplete-Insurance Economies,” Journal of Economic Dynamics and
Control, 17, 953-969.

[104] Huggett, M. (1997), “The One-Sector Growth Model with Idiosyn-

cratic Shocks: Steady States and Dynamics,” Journal of Monetary
Economics, 39, 385-403.
266 BIBLIOGRAPHY

[105] Huggett, M. and Ventura, (1999), “On the Distributional Effects of

Social Security Reform,” Review of Economic Dynamics, 2, 498-531.

[106] Hurd, M. (1987), “Savings of the Elderly and Desired Bequests,” Amer-
ican Economic Review, 77, 298-312.

[107] Hurd, M. (1989), ”Mortality Risk and Bequests,” Econometrica, 57,

779-813.

[108] Judd, K. (1998), Numerical Methods in Economics, MIT Press, Cam-

bridge, MA.

[109] Judd, K., L. Maliar and S. Maliar (2012), “Merging Simulation and
Projection Approaches to Solve High-Dimensional Problems,” NBER
Working Paper 18501.

[110] Kantor, S. and Fishback, P. (1996), “Precautionary Saving, Insurance,

and the Origin of Workers’ Compensation,” Journal of Political Econ-
omy, 104, 419-42.

[111] Kehoe, T. and D. Levine (2001), “Liquidity Constrained Markets versus

Debt Constrained Markets,” Econometrica, 69, 575-598.

[112] Kimball, M. (1990), “Precautionary Saving in the Small and in the

Large,” Econometrica, 58, 53-73.

[113] King, M. and Dycks Mireaux, L. (1982), “Asset Holdings and the Life-
Cycle,” Economic Journal 92, 247-67.

[114] Klein, P. and Rios-Rull, V. (1999), “Time-Consistent Optimal Fiscal

Policy,” International Economic Review, 44, 1217-1246.

[115] Kocherlakota, N. (1996a), “Implications of Efficient Risk Sharing with-

out Commitment,” Review of Economic Studies, 63, 595-609.

[116] Kocherlakota, N. (1996b), “The Equity Premium: It’s Still a Puzzle,”

Journal of Economic Literature, 34, 42-71.

[117] Kocherlakota, N. (2010), The New Dynamic Public Finance, Princeton

University Press, Princeton, NJ.
BIBLIOGRAPHY 267

[118] Kotlikoff, L. (1988), “Intergenerational Transfers and Savings,” Journal

of Economic Perspectives, 2, 41-58.
[119] Kotlikoff, L. and Summers, L. (1981), “The Role of Intergenerational
Transfers in Aggregate Capital Accumulation,” Journal of Political
Economy, 89, 706-32
[120] C. Koulovatianos (2007), “The Possibility of Demand Aggregation
when Agents are Infinitely-Lived,” Working Paper, University of Vi-
enna.
[121] Krueger, D. and F. Kubler (2004), “Computing Equilibrium in OLG
Models with Stochastic Production,” Journal of Economic Dynamics
and Control, 28, 1411-1436.
[122] Krueger, D. and H. Lustig (2007), “The Irrelevance of Market In-
completeness for the Price of Aggregate Risk,” NBER Working Paper
12634.
[123] Krueger, D., H. Lustig and F. Perri (2007), “Evaluating Asset Pric-
ing Models with Limited Commitment using Household Consumption
Data,” NBER Working Paper 13650.
[124] Krueger, D. and F. Perri (1999), “Risk Sharing: Private Insurance Mar-
kets or Redistributive Taxes?,” Federal Reserve Bank of Minneapolis
Staff Report 262.
[125] Krueger D. and F. Perri (2003), “On the Welfare Consequences of the
Increase in Inequality in the United States,” 2003 NBER Macroeco-
nomics Annual.
[126] Krueger, D. and F. Perri (2005), “Understanding Consumption
Smoothing: Evidence from the U.S. Consumer Expenditure Data,”
Journal of the European Economic Association, MIT Press, 3, 340-349.
[127] Krueger D. and F. Perri (2006), “Does Income Inequality Lead to Con-
sumption Inequality: Evidence and Theory,” Review of Economic Stud-
ies, 73, 163-193.
[128] Krusell, P., Quadrini, V. and Rios-Rull, V. (1997), “Politico-Economic
Equilibrium and Economic Growth,” Journal of Economic Dynamics
and Control, 21, 243-72.
268 BIBLIOGRAPHY

[129] Krusell, P. and A. Smith (1997), “Income and Wealth Heterogeneity,

Portfolio Choice, and Equilibrium Asset Returns,” Macroeconomic Dy-
namics, 1, 387-422.

[130] Krusell, P. and Smith, A. (1998), “Income and Wealth Heterogeneity

in the Macroeconomy,” Journal of Political Economy, 106, 867-896.

[131] Krusell, P. and Smith, A. (2003), “Consumption and Savings Decisions

with Quasi-Geometric Discounting,” Econometrica, 71, 365-375.

[132] Laibson, D. (1997), “Golden Eggs and Hyperbolic Discounting,” Quar-

terly Journal of Economics, 112, 443-77.

[133] Lambertini, L. (2003), “Endogenous Debt Constraints in Life Cycle

Economies,” Review of Economic Studies, 70, 1-27.

[134] Lettau, M. and Uhlig, H. (1999), “Rules of Thumb versus Dynamic

Programming,” American Economic Review, 89, 148-174.

[135] Li, Wenli and R. Yao (2007), “The Life-Cycle Effects of House Price
Changes,” Journal of Money, Credit, and Banking, 1375-1409.

[136] Livshits, I., J. MacGee and M. Tertilt (2007), “Accounting for the Rise
in Consumer Bankruptcies,” Working Paper, Stanford University.

[137] Mace, B. (1991), “Full Insurance in the Presence of Aggregate Uncer-

tainty,” Journal of Political Economy, 99, 928-956.

[138] MaCurdy, T. (1982), “The Use of Time Series Processes to Model the
Error Structure of Earnings in a Longitudinal Data Analysis,” Journal
of Econometrics, 18, 82-114.

[139] Mankiw, G. and Zeldes, S. (1991), “The Consumption of Stockholders

and Non-Stock-Holders,” Journal of Financial Economics, 29, 97-112.

[140] Mariger, R. (1987), “A Life-Cycle Consumption Model with Liquidity

Constraints: Theory and Empirical Results,” Econometrica, 55, 533-
57.

[141] Mazzocco, M. and S. Saini (2012), “Testing Efficient Risk Sharing

with Heterogeneous Risk Preferences,” American Economic Review,
102, 428-468.
BIBLIOGRAPHY 269

[142] McGrattan, E. (1999), “Applications of Weighted Residual Methods

to Dynamic Economic Models,” in A. Scott and R. Marimon (eds.)
Computational Methods for the Study of Dynamic Economies, Oxford
University Press, Oxford, UK.
[143] Meghir, C. and Weber, G. (1996), “Intertemporal Non-Separability or
Borrowing Restrictions? A Disaggregate Analysis Using a US Con-
sumption Panel,” Econometrica, 64, 1151-82.
[144] Mehra, R. and Prescott, E. (1985), “The Equity Premium: A Puzzle,”
Journal of Monetary Economics, 15, 145-161.
[145] Modigliani, F. (1986), “Life Cycle, Individual Thrift, and the Wealth
of Nations,” American Economic Review, 76, 297-312.
[146] Modigliani, F. (1988), “The Role of Intergenerational Transfers and
Life Cycle Saving in the Accumulation of Wealth,” Journal of Economic
Perspectives, 2, 15-40.
[147] Narajabad, B. (2007), “Information Technology and the Rise of House-
hold Bankruptcy,” Working Paper, Rice University.
[148] Nelson, J. (1994), “On Testing for Full Insurance Using Consumer Ex-
penditure Survey Data: Comment,” Journal of Political Economy, 102,
384-394.
[149] Pischke, J. (1995), “Individual Income, Incomplete Information, and
Aggregate Consumption,” Econometrica, 63, 805-39.
[150] Pistaferri, L. (2001), “Superior Information, Income Shocks and the
Permanent Income Hypothesis,” Review of Economics and Statistics,
83, 465-476.
[151] Quadrini, V. (1997), “Entrepreneurship, Saving and Social Mobility,”
Review of Economic Dynamics, 3, 1-40.
[152] Rios-Rull, V. (1999), “Computation of Equilibria in Heterogeneous
Agent Models,” in: Marimon and Scott (eds.), Computational Methods
for the Study of Dynamic Economies, Oxford University Press.
[153] Runkle, D. (1988), “Liquidity Constraints and the Permanent Income
Hypothesis,” Journal of Monetary Economics, 27, 73-98.
270 BIBLIOGRAPHY

[154] Sargent, T. (1993) Bounded Rationality in Macroeconomics, Clarendon

Press.

[155] Schechtman, J. (1976), “An Income Fluctuation Problem,” Journal of

Economic Theory, 12, 218-241.

[156] Schechtman, J. and Escudero, V. (1977), “Some Results on ”An Income

Fluctuation Problem”,” Journal of Economic Theory, 16, 151-166.

[157] Shea, J. (1995), “Union Contracts and the Life-Cycle Permanent In-
come Hypothesis,” American Economic Review, 85, 186-200.

[158] Schulhofer-Wohl, S. (2011), “Heterogeneity and Tests of Risk Sharing,”

Journal of Political Economy, 119, 925-958.

[159] Shorrocks, A. (1975), “The Age Wealth Relationship: A Cross-

Sectional and Cohort Analysis,” Review of Economics and Statistics,
57, 155-63.

[160] Stoker, T. (1993), “Empirical Approaches to the Problem of Aggrega-

tion Over Individuals,” Journal of Economic Literature, 31, 1827-1874.

[161] Storesletten, K. C. Telmer and A. Yaron (2004a), “Consumption and

Risk Sharing over the Life Cycle,” Journal of Monetary Economics,
2004, 51, 609-633

[162] Storesletten, K., C. Telmer and A. Yaron (2004b), “Cyclical Dynamics

in Idiosyncratic Labor-Market Risk,” Journal of Political Economy,
112, 695-717.

[163] Storesletten, K., C. Telmer and A. Yaron (2007), “Asset Pricing with
Idiosyncratic Risk and Overlapping Generations,” Review of Economic
Dynamics, 10, 519-548.

[164] Tallarini, T. (2000), “Risk-Sensitive Real Business Cycles,” Journal of

Monetary Economics, 45, 507-532.

[165] Townsend, R. (1994), “Risk and Insurance in Village India,” Econo-

metrica, 62, 539-91.
BIBLIOGRAPHY 271

[166] Udry, C. (1994), “Risk and Insurance in a Rural Credit Market: An

Empirical Investigation in Northern Nigeria,” Review of Economic
Studies, 61, 495-526.

[167] Wang, N. (2003), “Caballero Meets Bewley: The Permanent-Income

Hypothesis in General Equilibrium,” American Economic Review, 93,
927-936

[168] Weil, P. (1989), “The Equity Premium Puzzle and the Risk-Free Rate
Puzzle,” Journal of Monetary Economics, 24, 401-421.

[169] Werning, I. (2007), “Optimal Fiscal Policy with Redistribution,” Quar-

terly Journal of Economics, 122, 925-967.

[170] Wilson, R. (1968), “The Theory of Syndicates,” Econometrica, 36, 119-

132.

[171] Yaari, M. (1965), “Uncertain Lifetime, Life Insurance, and the Theory
of the Consumer,” Review of Economic Studies, 32, 137–150.

[172] Zeldes, S. (1989), “Consumption and Liquidity Constraints: an Em-

pirical Investigation,” Journal of Political Economy, 97, 305-46.