Vous êtes sur la page 1sur 133

Copyright

by
Tansel Yilmazer
2002
The Dissertation Committee for Tansel Yilmazer
certifies that this is the approved version of the following dissertation:

Household Saving Behavior, Portfolio Choice and


Children: Evidence from the Survey of Consumer
Finances

Committee:

Daniel T. Slesnick, Supervisor

Don Fullerton

Maxwell B. Stinchcombe

Peter J. Wilcoxen

Jacqueline Angel
Household Saving Behavior, Portfolio Choice and
Children: Evidence from the Survey of Consumer
Finances

by

Tansel Yilmazer, B.S., M.A.

DISSERTATION
Presented to the Faculty of the Graduate School of
The University of Texas at Austin
in Partial Fulfillment
of the Requirements
for the Degree of

DOCTOR OF PHILOSOPHY

THE UNIVERSITY OF TEXAS AT AUSTIN


December 2002
UMI Number: 3110711

________________________________________________________
UMI Microform 3110711
Copyright 2004 by ProQuest Information and Learning Company.
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.
____________________________________________________________

ProQuest Information and Learning Company


300 North Zeeb Road
PO Box 1346
Ann Arbor, MI 48106-1346
Acknowledgments

I am grateful to many people who shared the best and worst moments
of ‘my dissertation years.’ First, I would like to thank my advisor, Daniel
Slesnick, for his support, patience, guidance and encouragement. I would also
like to thank my committee members Don Fullerton, Maxwell Stinchcombe,
Peter Wilcoxen and Jacqueline Angel for their valuable feedback and com-
ments. Special thanks go to Asli Kes, Anne Golla, Angela Lyons, Anne Gorney,
Görkem Çelik, Adam Winship, Matias Fontenla, Mala Velamuri, Steve Trejo,
and Vivian Goldman-Leffler for their stimulating conversations and friendship.
I am indebted to my family for their love and believing in me over these years.
Finally, I wish to thank Fikret for always being there for me, in spite of the
thousands of miles between us.

iv
Household Saving Behavior, Portfolio Choice and
Children: Evidence from the Survey of Consumer
Finances

Publication No.

Tansel Yilmazer, Ph.D.


The University of Texas at Austin, 2002

Supervisor: Daniel T. Slesnick

Using the Survey of Consumer Finances (SCF), this dissertation ex-


amines the relationship between having children and the motives of saving: (i)
to hold assets because of the return they provide, (ii) to build up reserves as
a precaution for a ‘rainy day,’ and (iii) to accumulate for anticipated future
needs, such as educational expenses.

The first chapter examines how the number of children living in the
household affects the way households allocate their wealth across different
assets, such as owner-occupied housing, risky assets and interest-bearing ac-
counts. The portfolio allocation of homeowners is compared to that of renters
by taking into account the portfolio constraint imposed by the consumption
demand for housing. The results show that the number of children increases
the housing consumption of homeowners and the share of the portfolio al-
located to owner-occupied housing. As a result of the portfolio constraint,

v
homeowners decrease the share of the portfolio invested in retirement assets
as the number of children increases.

Using a life-cycle model that incorporates precautionary motives for


saving, the second chapter investigates the relationship between household
saving and fertility decisions. By examining the implications of income uncer-
tainty on the demand for children, this chapter extends the empirical work on
precautionary savings. The results show that households with higher income
uncertainty are less likely to have a child, and after controlling for family size,
income uncertainty has little effect on household savings. Further, having an
additional child reduces savings of households with young heads and increases
savings of those with older heads.

The third chapter examines the effect of financing children’s college ed-
ucation on household savings. Using the actual college expenditures reported
in the 1983-86 SCF, the empirical model estimates the expected expenditures
on children’s college education and investigates the effect of expected college
expenses on household savings. The results show that parents save for college
expenses of their children. Also, savings for college increase with the age of
the household head. The results are consistent with the predictions the life-
cycle theory of saving that households save in advance for expected expenses
to smooth their consumption.

vi
Table of Contents

Acknowledgments iv

Abstract v

List of Tables ix

List of Figures xi

Chapter 1. Introduction 1

Chapter 2. Do Children Affect Household Portfolio Allocation? 6


2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Empirical Model . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Estimation and Results . . . . . . . . . . . . . . . . . . . . . . 24
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

Chapter 3. The Effect of Precautionary Motives on Household


Saving and Fertility 44
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 The Relationship between Fertility and Saving . . . . . . . . . 48
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Estimation and Results . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

vii
Chapter 4. Saving for Children’s College Education 73
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 A Model of Saving for College . . . . . . . . . . . . . . . . . . 80
4.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4 Empirical Specification . . . . . . . . . . . . . . . . . . . . . . 86
4.5 Estimation and Results . . . . . . . . . . . . . . . . . . . . . . 88
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

Appendices 103

Appendix A. Appendix for Chapter 2 104


A.1 Estimating Marginal Tax Rates . . . . . . . . . . . . . . . . . 104
A.2 Definition of Variables . . . . . . . . . . . . . . . . . . . . . . 105
A.3 Estimating Permanent Income . . . . . . . . . . . . . . . . . . 106

Appendix B. Appendix for Chapter 3 109


B.1 Definition of Variables . . . . . . . . . . . . . . . . . . . . . . 109

Appendix C. Appendix for Chapter 4 111


C.1 Definition of Variables . . . . . . . . . . . . . . . . . . . . . . 111

Bibliography 113

Vita 121

viii
List of Tables

2.1 Descriptive Statistics by Year . . . . . . . . . . . . . . . . . . 33


2.2 Mean Asset Shares by Year . . . . . . . . . . . . . . . . . . . 34
2.3 Expenditure on Housing, 1998 . . . . . . . . . . . . . . . . . . 35
2.4 Mean Asset Shares, 1998 . . . . . . . . . . . . . . . . . . . . . 36
2.5 Mean Asset Shares, 1998: Continued . . . . . . . . . . . . . . 37
2.6 Results from Probit Estimation . . . . . . . . . . . . . . . . . 38
2.7 Results: Asset Shares and Housing Expenditure of Homeowners 39
2.8 Homeowners: Continued . . . . . . . . . . . . . . . . . . . . . 40
2.9 Results: Asset Shares and Housing Expenditure of Renters . . 41
2.10 Renters: Continued . . . . . . . . . . . . . . . . . . . . . . . . 42
2.11 Portfolio Shares for Assets by the Number of Children and Age 43

3.1 Saving Motives by Age Groups, 1983 . . . . . . . . . . . . . . 65


3.2 Descriptive Statistics by Household Fertility Decision . . . . . 66
3.3 Mean Income Uncertainty by Household Demographics . . . . 67
3.4 Savings, Income and Income Uncertainty by Age and Fertility 68
3.5 Probit: Fertility Decision of Fecund Households . . . . . . . . 69
3.6 Regressions of SAVE1 on Income Uncertainty with Endogenous
Fertility Decision . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.7 Regressions of SAVE2 on Income Uncertainty with Endogenous
Fertility Decision . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.8 The Effect of a Change in the Fertility Decision on SAVE1 . . 72

4.1 Saving Motives By the Number of Children . . . . . . . . . . 95


4.2 Descriptive Summary of Variables . . . . . . . . . . . . . . . 96
4.3 Savings and College Expenses by the Number of Children in
College . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4 Poisson Regression: Number of Children . . . . . . . . . . . . 98
4.5 Tobit Estimates of College Expenditure Equation . . . . . . . 99

ix
4.6 College Expenditures and Savings by the Number of Children
in College . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.7 Effect of Anticipated College Expenses on Savings . . . . . . . 101

x
List of Figures

4.1 The Importance of Educational Expenses on Savings . . . . . 102

xi
Chapter 1

Introduction

Raising children is costly with their housing, educational and other ex-
penses. To meet the costs of raising their children, parents use both current
income and intertemporal transfers. Children living in the household, there-
fore, are likely to affect the level of household savings, portfolio composition
and the life-cycle profile of savings. Using data from the Survey of Consumer
Finances (SCF), this dissertation examines the relationship between children
and the motives of saving: (i) to hold assets because of the return they pro-
vide, (ii) to build up reserves as a precaution for a ‘rainy day,’ and (iii) to
accumulate for anticipated future needs, such as educational expenses.

Most U.S. households hold a large portion of their wealth in the form of
owner-occupied housing. According to the 1995 SCF, 65 percent of households
are homeowners, and the value of an average homeowner’s property is 60
percent of its total assets. Owner-occupied housing differs from other types of
wealth in its dual role as both a consumption good and an investment good.
Since households cannot separate the level of consumption of housing services
from investment in housing as an asset, the optimal level of owner-occupied
housing may be higher than the optimal level for households only interested

1
in long run returns. The demand for housing services is likely to increase with
the number of children living in the household. Therefore, the consumption
constraint can be even more binding for households with children.

Chapter 2 uses the 1989, 1992, 1995 and 1998 SCF to investigate how
the number of children living in the household affect the portfolio choice be-
tween housing and other assets. The portfolio allocation of homeowners is
compared to that of renters by taking into account the portfolio constraint
imposed by the consumption demand for housing. The empirical model also
examines the effect of children on the demand for housing services and home-
ownership decision. The results show that the number of children increases
the housing consumption of homeowners as well as the share of the portfolio
allocated to owner-occupied housing. As a result of the portfolio constraint,
homeowners decrease the portfolio share of retirement assets as the number of
children increases.

Low levels of retirement savings of U.S. households have generated sig-


nificant concern in the last twenty years. The findings of Chapter 2 show that
households with children decrease the portfolio share for retirement savings
considerably while they increase the portfolio share for housing. If the return
on housing is less than the return on retirement accounts, there is a hidden
cost of children. Explaining the size of the portfolio effect allows a better un-
derstanding of the cost of children. Also, changes in housing programs or tax
deduction rules for mortgage interest payments influence the portfolio alloca-
tion of households with children considerably by increasing or decreasing the

2
cost of homeownership.

The data on U.S. household saving show that saving rates are higher
for married couples with no children and lower for those with children. Using a
life cycle model that incorporates precautionary motives for saving, Chapter 3
investigates the relation between household saving and fertility decisions. Pre-
cautionary saving models predict that uncertainty about future income may
cause households to reduce their current consumption in order to raise their
stock of precautionary saving. By examining the implications of uncertainty
on the fertility decisions of households and incorporating fertility decisions as a
motive for household saving behavior, this chapter extends the empirical work
on precautionary saving. The 1983-89 panel of the SCF is used to examine
the interaction of income uncertainty and changes in the number of children
on the saving behavior of households at different stages of the life cycle.

The results of the empirical model in Chapter 3 show that households


with higher income uncertainty are less likely to have a child at a point in
time. The findings, however, are not consistent with the predictions of the
precautionary saving model that suggests agents faced with uncertainty about
future income increase their savings. Income uncertainty actually reduces
savings of the households with low or very high wealth holdings and does not
affect the saving behavior of other households. Also, having an additional child
decreases savings of households with young heads and increases savings of those
with older heads. This finding is consistent with the life-cycle theory of saving
and consumption and shows that household composition is an important factor

3
of life-cycle savings.

Chapter 4 examines the effect of financing children’s college education


on household savings. Understanding the effect of financing children’s college
education on household saving behavior is important for at least three rea-
sons. First, parents contribute a significant amount to their children’s college
expenses. According to the 1996 National Postsecondary Student Aid Survey,
90 percent of dependent undergraduate’s parents contributed financially to the
costs of their children’s education. Of those contributing to their children’s
college costs in 1987, about 65 percent reported using some previous savings.
Second, families who save for college reduce their eligibility for financial aid.
The college financial aid system imposes an implicit tax on the savings of
households that are potentially eligible for financial assistance. Third, the
quality-quantity model of fertility behavior assumes that parents have pref-
erences both for the expenditure per child and the number of children. This
chapter uses the amount of parental expenditure on children’s college educa-
tion as a measure for child quality. Given the rapidly rising cost of college
tuition, an analysis of financing college education and family size highlights
an important aspect of the quality-quantity model.

Using the actual college expenditures reported in the 1983-86 SCF,


Chapter 4 estimates the household’s expected expenditures on children’s col-
lege education and investigates the effect of expected college expenses on
household savings. The results show that parents save for college expenses
of their children. Also, savings for college education increases with the age of

4
the household head. These results are consistent with the predictions of the
life-cycle theory of saving and consumption that households save in advance
for expected expenses to smooth their consumption.

5
Chapter 2

Do Children Affect Household Portfolio


Allocation?

2.1 Introduction

Empirical studies of household portfolio composition have identified


large differences in portfolio allocation choices of different demographic groups.
So far, the literature has focused on the impact of demographic variables such
as the effect of age, race and gender of the household head on the portfolio
composition.1 The influence of children living in the household on the portfolio
composition has not been yet discussed.

It is likely that children living in the household affect the way a house-
hold allocates its wealth across different assets such as owner-occupied hous-
ing, risky assets, and interest-bearing accounts. For example, households with
children may purchase more housing than households with no children or they
may have a higher probability of owning a home. Parents may choose to invest
part of their household portfolio in stocks to meet the rising costs of a college
education. Conversely, they may hold most of their financial assets in riskless

1
See Poterba and Samwick [46], King and Leape [41], and Ioannides [34] for age effect;
Chiteji and Stafford [13] for race; Jianakoplas and Bernasek [35], and Sundèn and Surette [52]
for gender effects.

6
form to decrease their families’ exposure to risk.

Understanding the size of the impact of children on household portfolio


allocation is intrinsically interesting. It has also important policy implications.
If households with children allocate a larger share of their portfolio to owner-
occupied housing, then changes in housing programs or tax deduction rules for
mortgage interest payments influence their portfolio allocation by increasing or
decreasing the cost of homeownership. Also, as the result of higher consump-
tion demand for housing, households with children may decrease the portfolio
share for other assets considerably while they increase the portfolio share for
housing. Low levels of retirement savings of U.S. households have generated
significant concern in the last twenty years. The failure of households with
children to invest sufficient assets in retirement accounts may lead to a lower
retirement wealth.

Using data from the 1989, 1992, 1995 and 1998 SCF, this chapter in-
vestigates the effect of children on household portfolio composition, paying
particular attention to the impact of children on the demand for housing ser-
vices and homeownership decision. Specifically, I analyze a model in which
households decide on portfolio shares for different assets jointly with the tenure
choice (the decision of owning or renting) and the consumption demand for
housing services. I focus on how the number and age of children living in the
household affect (i) the homeownership decision, (ii) the portfolio shares for
housing and the other assets that homeowners and renters hold, and (iii) the
housing expenditure of homeowners and renters.

7
Most U.S. households hold a large portion of their wealth in the form
of owner-occupied housing. Wolff [56] uses the 1983, 1989, 1992 1995 SCF,
and King and Leape [41] examine the 1960-62 Michigan Surveys of Consumer
Finances, and both report that owner-occupied housing accounts for about 30
percent of household assets. According to the 1995 SCF, 65 percent of house-
holds are homeowners, and the value of an average homeowner’s property is 60
percent of its total assets. Owner-occupied housing differs from other types of
wealth in its dual role as both a consumption good and an investment good.2
In the presence of tax distortions and transaction costs, households cannot
separate the level of consumption of housing services from investment in hous-
ing as an asset, and the ownership of their principal residence determines the
level of consumption of housing services. The optimal level of owner-occupied
housing for households may be higher than the optimal level for households
that are only interested in long run returns. Households with children are
likely to have a higher demand for housing services and the consumption con-
straint can be even more binding. Explaining the size of the portfolio effect
allows a better understanding of the cost of children.

While the dual role of housing has been recognized, its impact on the
portfolio choice between housing and other assets has not been discussed much.
Exceptions are the theoretical model of Brueckner [7], the general equilibrium
model of Berkovec and Fullerton [4] and the numerical analysis of Flavin and
Yamashita [20]. Brueckner analyzes the behavior of homeowners. In his model,

2
See Henderson and Ioannides [27] and Berkovec and Fullerton [4]

8
an investment constraint requires that the quantity of housing owned is at
least as large as the quantity of housing consumed. His model analyzes the
resulting distortion of the effect of this investment constraint on the portfolio
choice of homeowners. The results of his model show that when the constraint
imposed by housing is binding, the homeowner’s optimal portfolio is inefficient
in a mean-variance framework. In Berkovec and Fullerton, households decide
on tenure and quantity of housing taking both consumption and investment
motives into account. Their simulation concentrates on the effect of taxes
on the tenure choice and owner-occupied housing. Flavin and Yamashita use
numerical methods to calculate the mean-variance efficient frontier. Their
results show that the portfolio constraint imposed by the consumption demand
for housing causes a life-cycle pattern in the portfolio shares for stocks and
bonds such that the ratio of stocks to net worth increases as the household
head gets older. Neither of these studies explicitly analyzes the determinants
of the consumption demand for housing and the portfolio share for housing.
This chapter extends the previous studies of portfolio choice by examining the
effect of both consumption and investment motives on the portfolio share for
housing and other assets.

The literature on housing demand has recognized the role of children


on the tenure choice and the demand for housing services. For example, Harun
et al. [26] treat the presence of children in the household as endogenous and
find that a 10 percent increase in the probability of having a child raises
the likelihood of homeownership by 2.5 percent. Robst et al. [36] show that

9
an additional child increases the probability of owning a home by around 8
percent. Goodman and Kawai [25] find that larger households prefer more
housing. After controlling for the household size, however, their results show
that the presence of children in school has either an insignificant or a negative
effect on the demand for housing. Ihlanfeldt [33] reports housing demand
estimates obtained separately from two samples-recent movers and nonmovers.
Among recent movers, the importance of the current and expected family size
differs between owners and renters: while renters demand more housing with
an increase in family size and expectation of an additional child within the next
nine months, these variables do not affect the housing demand of homeowners.
The results of the previous studies show that dependent children have some
impact on the demand for housing. However, as noted in Goodman [24], little
systematic treatment of children has appeared in the estimation of tenure
choice and housing demand.

Besides housing, U.S. households typically invest in only a few of the


assets available in the economy. For example, according to the 1995 SCF, only
41 percent of households held stocks directly or indirectly in IRAs, 401(k)s,
defined benefit pensions and mutual funds. Many studies have investigated
the reasons that most households choose to hold incomplete portfolios. The
information cost of monitoring and managing a portfolio is suggested as an im-
portant reason for holding riskless assets. Demographic characteristics such as
age, marital status, and race of the household head are shown to be significant
factors that reduce the level of information cost that would be sufficient to

10
discourage households from investing in risky assets. For example, Bertaut [5]
uses the 1983-89 SCF to analyze the effect of household characteristics on port-
folio allocation. His results show that household characteristics such as age and
education of the household head are significant in explaining the probability of
owning stocks. King and Leape [41] analyze a model in which investors choose
to hold incomplete portfolios, and they estimate equations for both the prob-
ability of owning an asset and its demand conditional upon ownership. Their
findings show that age and marital status of the household head significantly
affect the probability of asset ownership. In the conditional demand equations,
however, the effect of age and marital status appears to be significant only for
some of the assets. Using the Panel Study of Income Dynamics, Chiteji and
Stafford [13] link independent young African-American adults back to their
parents. Their finding is that parents who held stocks are more likely to have
children who hold stocks as young adults.

Children living in the household have not been the focus of any study
examining the portfolio choice of households. This chapter aims to do so by
examining the effect of the number and the age of children on household port-
folio choice. The theoretical model developed in the chapter shows how the
portfolio constraint imposed by the consumption demand for housing affects
the portfolio shares for housing and other assets. The empirical model com-
pares the portfolio allocation of homeowners to that of renters, taking into
account the effect of children on the consumption demand for housing. The
results show that the number of children has a positive and significant effect

11
on the probability of owning a home. The number of children also increases
the housing demand of homeowners. As a result of the portfolio constraint im-
posed by the housing demand of children, homeowners decrease the portfolio
share in retirement accounts while they increase the portfolio share in hous-
ing. Controlling for the number of children and other variables, homeowners
with all children older than age 13 invest a greater share of their portfolio in
vehicles and other real estate and a smaller share of their portfolio in housing.
Children living in the household also affects the portfolio choice of renters.
Renters invest a smaller share of their portfolio in interest-bearing accounts
with an increase in the number of children. The main conclusion of the chapter
is that homeowners shift their resources from retirement accounts to housing
with an increase in the number of children.

The remainder of this chapter is organized as follows. Section 2.2 in-


troduces the theoretical model and discusses the empirical specification of the
model. Section 2.3 describes the data set and the variables used in the empir-
ical work. The estimation results are reported in Section 2.4. A summary of
the findings and concluding remarks are presented in Section 2.5.

2.2 The Model


2.2.1 Theory

This section examines the behavior of a consumer deciding whether to


rent or own a home, and how much to allocate to other risky assets. The con-
sumer maximizes a multiperiod utility function. Following Brueckner [7] and

12
Henderson and Ioannides [27], I assume that third and subsequent periods are
buried in the indirect utility function given remaining wealth at the beginning
of the second period. A consumer in this economy is assumed to obtain utility
from the current consumption of a single nondurable good (c), housing services
(hc ), and consumption in future periods that depends on the random total re-
turn R from the investment portfolio. The consumer’s objective function can
be written as follows:
U (c, hc ) + δE[V (R + y)], (2.1)

where y is future labor income, U gives the utility from the current consump-
tion, V is an indirect utility function, E gives the expected utility, and δ is
the discount factor.

The dollar amount of asset j purchased is denoted aj , j = 0, 1, .., J,


with a0 being the riskless asset. The only source of uncertainty is assumed to
be from returns on J + 1 assets and owner-occupied housing (h). Short selling
is ruled out for all assets including housing, so that aj ≥ 0, j = 0, 1, .., J, and
h ≥ 0. The j th asset earns a gross return of rj , and owner-occupied housing
earns rh .

If the consumer purchases a house, then she holds owner-occupied hous-


ing (h > 0) and is constrained to consume the same amount of owner-occupied
housing in her portfolio (hc = h). The first period budget constraint is given
by
J
X
c=w− poh hc − aj , (2.2)
j=0

13
where w is her initial wealth and poh is the current price of a unit of housing.
The total return of the portfolio is given by
J
X
R = rh h + rj aj . (2.3)
j=0

If the consumer rents a house, then the first period budget constraint
is given by
J
X
c = w − por hc − aj , (2.4)
j=0

where por is the price of a unit of housing for renters. The total return of the
portfolio is given by
J
X
R= rj aj , (2.5)
j=0

since h is equal to zero for renters.

In the model, the return on housing and the return on other assets
are assumed to be normal variables with the expected values rh and rj , j =
1, 2, .., J, respectively. For homeowners, the total portfolio return R is a normal
random variable with the expected value
J
X
R = rh h + r0 a0 + r j aj (2.6)
j=1

and the standard deviation


J
X J X
X K
2
σ = (θhh h + 2 haj θhj + aj ak θjk )1/2 , (2.7)
j=1 j=1 k=1

where θhh and θjj , j = 1, ...J, are the variances of rh and rk , θjk is the co-
variance of returns between asset j and k, and θjh is the covariance of returns
between asset j and housing. For renters, h = 0 in equations (2.6) and (2.7).

14
Following Fama and Miller [21] and Brueckner [7], I rewrite the objec-
tive function (2.1) in terms of R, σ, and the standard normal variable z as
follows:
Z
U (c, hc ) + δ V (R + σz + y)φ(z)dz, (2.8)

where φ(.) is the standard normal density function. The consumer’s problem
is to choose c∗ , h∗c and a∗j , j = 0, 1, .., J, that maximize (2.1) subject to (2.2),
(2.3), (2.6) and (2.7). The consumer also decides on c∗ , h∗c , and a∗j , j =
0, 1, .., J, that maximize (2.8) subject to (2.4), (2.5), (2.6) and (2.7) and decides
to own or rent a house comparing the utilities in two outcomes.

For both homeowners and renters, this problem can be solved in two
stages. In the first stage, the asset levels aj , j = 0, 1, ..., J, are chosen optimally
with hc and σ held constant. In the second stage, hc (and thus σ) is chosen
optimally. The empirical model described in the next section focuses on the
interaction between these two stages of decision making.

2.2.2 Empirical Model

The joint determination of whether own a house (H=1) or not (H=0),


how much to spend on housing (Eh ), and shares of wealth to allocate to each
asset j (sj ), and to owner-occupied housing (sh ) is modeled as follows. First,
a household determines whether to own or rent a house:

H = 1 if Xh β1 + ε1 > 0
(2.9)
= 0 otherwise,

where Xh is a vector of year dummies and characteristics that are associated

15
with the probability of owning a house, β1 is a parameter vector, and ε1 is an
error term. Second, the household decides on the share of portfolio allocated
to each asset and housing, and also the housing expenditure:


 sj = Xβoj + εoj j = 0, 1, ..., J
If owner, sh = Xβh + εh

log Eh = Xc βoc + εoc .
(2.10)

 sj = Xβrj + εrj j = 0, 1, ..., J
If renter, sh = 0 (2.11)

log Eh = Xc βrc + εrc ,

where X and Xc are vectors of household characteristics and year dummies;


βoj , βrj , j = 0, 1, ..., J, βh , βoc and βrc are the parameter vectors to be esti-
mated; and εoj , εrj , j = 0, 1, ..., J, εh , εoc , and εrc are the error terms.

Separate equations are specified for homeowners and renters, and the
error terms in equations (2.9) - (2.11) are assumed to have a joint normal
distribution. The two stage method described in Lee and Trost [42] is used to
estimate the model. In the first stage, a probit model of the tenure choice
in equation (2.9) provides an estimate of β1 . In the second stage, I use
φ(X βˆ1 )/Φ(X βˆ1 ), as a regressor in estimating (2.10) for homeowners, where
φ and Φ are probability density and cumulative distribution of the standard
normal distribution, respectively. Similarly, φ(X βˆ1 )/(1 − Φ(X βˆ1 )) is used as
a regressor for renters in estimating (2.11).

16
2.3 Data

The data for this study are taken from the 1989, 1992, 1995 and 1998
SCF, a triennial survey conducted by the Federal Reserve Board. The sur-
vey contains detailed information on household portfolios, income, and de-
mographic characteristics. Each survey consists of a representative sample of
the U.S. population and a supplement of high-wealth households drawn from
Internal Revenue Service file of high-income returns.3

Total assets are grouped into six categories: 1) ACCOUNT includes


all holdings of checking accounts, saving accounts, certificates of deposit, call
accounts, money market deposit accounts; 2) STOCK includes all assets held
in stocks, all types of bonds, and mutual funds; 3) RETIRE includes IRAs,
Keogh, 401(k)s, and other defined contribution plans; 4) HOUSE is the market
value of owner-occupied housing; 5) VEHICLE is the value of all the vehicles
the household owns; 6) RESTATE includes the market value of seasonal resi-
dences and other property; and 7) OTHER includes trusts, cash value of life
insurance, and other assets like arts and precious metals. Investments in busi-
nesses are not included in total assets because they generate an income that
is difficult to separate from earnings.

The consumption demand for housing is computed for renters and


homeowners as follows. For owners, the cost of housing services depends on

3
In the 1989 SCF, the supplement consists of 866 out of 3,143 households; in 1992, 1,480
out of 3,906; in 1995, 1,519 out of 4,299; and in 1998, 1,409 out of 4,309 households. The
SCF constructs sample weights to blend the supplements with the area-probability sample
to get a more representative sample of the U.S. population.

17
the gross value of the residence (G), maintenance and depreciation costs (d),
the property tax rate (τp ), the mortgage interest payment (m), the interest
rate (r), the income tax rate (τ ), the rate of increase in the nominal price of
housing (ρ) and the overall inflation rate (π). The housing expenditures (Eh )
of homeowners are then defined as

Eh = [(1 − τ )r + d + (1 − τ )τp − (ρ − π)]G − mτ. (2.12)

This formulation assumes that homeowners claim tax deductions for property
taxes and mortgage interest payments. For renters, the annual rental expen-
diture reported in the SCF is used as the consumption demand for housing.

To calculate the housing expenditure by using equation (2.12), I make


several assumptions. Following Henderson and Ioannides [28], I assume an
annual rate of depreciation of d=0.015 for each of the sample years. Property
tax rates and mortgage interest payments are reported in the SCF. The interest
rate, r, is assumed to be the interest rate on treasury bills, and the rate of
increase in house prices, ρ, is the rate of increase in the median sale price
of houses in that year. The inflation rate, π, is the annual inflation rate
calculated using the CPI-U deflator. Since marginal tax rates are not reported
in the SCF, I impute them using detailed account information on the sources
of income and demographics for each household. The calculation of marginal
tax rates is described in Appendix A.1.

A few restrictions are imposed on the sample. First, households that


neither rent nor own their homes are excluded for lack of information to cal-

18
culate housing expenditure.4 Second, households with the highest 0.1 percent
weighted wealth holdings in each wave of the SCF are dropped, to avoid the
influence of extreme outliers on the regression.5 The final sample consists of
13,989 observations; 2,900, 3,509, 3,773 and 3,807 households in 1989, 1992,
1995 and 1998, respectively.

Table 2.1 shows the summary statistics for all the variables used in
the estimation. The variables are described in detail in Appendix A.2. Sam-
ple demographics show the age of the household head (AGE), marital status
(MARRIED) and gender (FEMALE) of the household head and the fraction
of homeowners (HOMEOWN), most of which have not changed much over
time.6 However, both mean and median wealth (ASSET) have risen since
1992. The same pattern is true for permanent income (INCOME). As a proxy
for permanent income, I take the estimated earnings of the household head and
the spouse at the age of 45 and an individual-specific effect. The calculation
of permanent income follows King and Dicks-Mireaux [40] and is described
in Appendix A.3. The calculated expenditure of housing consumption (Eh )

4
A household is assumed to be a homeowner if (i) it owns the house/apartment that
it lives in or owns it as a part of a condo, a co-op or a townhouse association; (ii) it
owns both the mobile house and the site; or (iii) it owns part or all of the farm/ranch
on which it lives on. A household is assumed to be a renter if it rents all or part of the
farm/ranch/apartment/house/mobile home in which it lives. In 1989, 1992, 1995 and 1998,
respectively, 116, 183, 317 and 309 households were neither renters nor owners and were
dropped from the sample.
5
Of the remaining households, 127, 214, 209 and 193 were in the 0.1 percentile of the
weighted wealth distribution in the 1989, 1992, 1995, and 1998 SCF, respectively.
6
The SCF defines the head of the household to be the husband for all married households.
Therefore, households with female heads are headed by single females.

19
was higher for homeowners in 1992 than in other years due to the decline in
house prices in that year. The average number of children (NCHILD) living
in the household declined from 0.83 in 1989 to 0.75 in 1995 and stayed the
same in 1998. The percentage of households with all children older than age
13 (CHAGE13) has stayed the same since 1992.

As shown in Table 2.2, the composition of households’ portfolios reveals


the importance of housing as an asset. HOUSE is the most important asset,
representing 39.4 percent of total assets in 1998. The second largest asset
in the households’ portfolios is VEHICLE (18.6 percent in 1998), followed by
ACCOUNT. The portfolio share for ACCOUNT declined from 14.3 percent
in 1989 to 11.2 percent in 1995, but it rose to 13.2 percent in 1998 due to an
increase in the portfolio share for saving accounts.

Table 2.2 presents interesting changes in household portfolio structures


over time. First, the share for RETIRE increases sharply. Assets in these
accounts increased from 5.7 percent of total assets in 1989 to 10.5 percent in
1998. Second, there is a steady growth in the portfolio share for STOCK and
a steady decline in the portfolio share for RESTATE since 1989. The increases
in ACCOUNT, STOCK and RETIRE in 1998 offset the decline in HOUSE,
VEHICLE and RESTATE. This suggests that households have substituted
financial assets for nonfinancial assets.

Table 2.3 presents housing expenditures of homeowners and renters in


1998. The first column shows the share of households in different income, age,
wealth and children (the number of children living in the household) groups.

20
The second column indicates the percentage of each of these groups that are
homeowners. Average housing expenditures for homeowners and renters are
presented, respectively, in the remaining two columns of the table. The per-
centage of households who are homeowners increases with income, wealth and
the age of the household head. It also increases with the number of children,
reaching a peak among households with two children. The average housing
expenditure is $7,042 for homeowners and $6,030 for renters. The housing ex-
penditures of renters and homeowners also increase with income, wealth and
the number of children in the household. For homeowners, the expenditure
on housing declines after the age of 65. For renters, it declines after age 50.
Among households with wealth below $250,000 and income below $50,000,
renters spend more on housing than owners. This is due to an increase in the
value of residences and also to the tax deduction for property taxes and mort-
gage interest payments that decrease the opportunity cost of homeownership.

Tables 2.4 and 2.5 show the household portfolio composition in 1998
by household permanent income, age of the household head, wealth, and the
number of children. The first row of Table 2.4 shows the portfolio shares
of assets that homeowners and renters hold. Since the primary residence is
the largest part of homeowners’ wealth, accounting for 57.9 percent, there
are marked differences in household portfolios of renters and owners. First,
for renters, VEHICLE is the most important asset held (41.5 percent of total
assets) followed by ACCOUNT (26.0 percent). For homeowners, however,
VEHICLE is the third largest asset (7.8 percent of total assets) following

21
RETIRE (10.2 percent). The portfolio shares for other assets such as STOCK,
RETIRE and RESTATE are almost equal for renters and owners.

For higher levels of income, as shown in Tables 2.3 and 2.4, the frac-
tion of households who are homeowners increases, while the housing share of
portfolio declines. For example, of the households with income below $15,000,
42.7 percent are homeowners holding 75.9 percent of their total assets in hous-
ing. Of the households with income above $100,000, in contrast, 86.7 percent
are homeowners, but they hold only 42.6 percent of total assets in housing.
Another noteworthy finding is that the portfolio shares for STOCK and RE-
TIRE for both homeowners and renters rise with income. Also, we observe
striking differences in the composition of portfolios by the level of wealth.
For homeowners, the share of the portfolio allocated to RESTATE and for all
households, the share of the portfolio allocated to STOCK rise at a rapid rate
with wealth. For example, among homeowners that have wealth exceeding $1
million, STOCK is the most important asset category with a share equal to
25.2 percent of total assets while housing accounted for only 22.6 percent.

Table 2.4 also presents the life cycle patterns in household portfolios.
Not surprisingly, portfolio composition of households with heads over the age
of 65 differs considerably from other age groups’ portfolios. Several findings
are worth noting. First, the portfolio share for ACCOUNT almost doubles
both for homeowners and renters over the age of 65 compared to 50-64 year
old group. Also, accumulation in STOCK relative to other assets increases
over age 65. This suggests that households with heads over age 65 substitute

22
liquid assets for nonfinancial assets. Second, the portfolio share for HOUSE
declines with age among the households headed by persons below age 65, but
it stays steady after age 65.

Finally, Table 2.5 shows the portfolio shares by the number of children
living in the household. The table indicates a strong relation between chil-
dren and the share of portfolio allocated to housing. The portfolio share for
owner-occupied housing increases with the number of children. For example,
housing accounts for 56.0 percent of the wealth for households with no chil-
dren, 60.9 percent for households with 2 children, and 65.3 percent for those
with three or more children. Homeowners invest a smaller share of their port-
folio in interest-bearing accounts and stocks with an increase in the number
of children. Also, the presence of children increases the share of the portfolio
allocated to vehicles.

Tables 2.4 and 2.5 reveal striking differences in portfolio structures


across income, wealth, and age groups. While portfolio composition differs
considerably between renters and homeowners, the relative changes in portfolio
shares of assets by income, age and wealth are similar. Children are likely to
affect the portfolio structures in two ways. The first is their effect on the choice
of tenure, and the second is their effect on asset shares of portfolios conditional
upon ownership. Table 2.3 investigates the effect of children on the tenure
choice, and table 2.5 looks at the link between children and shares of assets in
both renters’ and homeowners’ portfolios. The results indicate that the number
of children living in the household affects the portfolio shares for assets and

23
the probability that a household owns a home. The empirical model below
investigates the effect of children on both asset shares and homeownership
decision.

2.4 Estimation and Results

The resulting set of equations constitutes an endogenous switching


model in the form of a multivariate regression model. Portfolio shares of the
J + 1 assets and housing sum to one, and the disturbance covariance matrix is
singular. Thus, I drop one group of assets, OTHER, and include ACCOUNT,
STOCK, RETIRE, HOUSE, VEHICLE, and RESTATE in the estimation of
the model. Then I solve for the parameters of OTHER from the other equa-
tions. Of 13,898 households, 410 report zero wealth holding.7 I exclude those
households from the sample and correct for sample selection.

Dummy variables indicating the number and the age of children living
in the household are included in X. The other variables in X are chosen to be
consistent with previous empirical studies. Portfolio choice theory has shown
the importance of age, permanent income and wealth in determining the asset
shares in household portfolios. Age and age-squared of the household head are
included to capture a possible change in portfolio behavior related to the life
cycle. Previous research also indicates that a household’s marginal tax rate
(MRT) has an effect on its asset allocation decisions. Moreover, the marital

7
93 households in 1989, 111 in 1992, 100 in 1995, and 106 in 1998 had zero wealth holding.

24
status and the gender of the household head and willingness to undertake risky
investments (RISKY) may also affect the household’s asset allocation.

All variables that enter X are also included in Xc and Xh , with two
exceptions. First, the marginal tax rate affects the tenure choice and home-
owners’ expenditure on housing since homeowners can claim tax deductions
for mortgage interest payments and property taxes. However, the marginal
tax rate is not expected to affect the housing expenditure of renters. Thus,
marginal tax rate is not included in Xc . Second, willingness to undertake risky
investment does not enter Xc because it has an effect on the tenure choice re-
garding the investment motive but not on the expenditures on rental housing.
In addition, the vector Xh includes the race of the household head.

Table 2.6 presents the estimates of the probit model of equation (2.9).
The estimates of the homeownership equation are consistent with previous
studies. As a household’s permanent income rises, the probability of home-
ownership increases. Age of the household head increases the probability of
ownership until age 74. The coefficients for WHITE and MARRIED are signif-
icant and positive, indicating that at the sample mean, households with white
heads are 10.2 percent more likely to own than households with non-white
heads, and those that are married are 26.1 percent more likely to own than
those that are not. The coefficients on the variables showing the number of
children are positive and significant. Households with one child are 6.3 per-
cent, and those with two children are 10.8 percent, more likely to own relative
to households with no children. The probability of owning starts to decrease

25
after the second child, household with three or more children are only 9.6 per-
cent more likely to own relative to households with no child. The probability
of being a homeowner also increases with the household’s marginal tax rate,
suggesting that the tax-deductibility of property taxes and mortgage interest
is more valuable at a higher marginal tax rate.

Tables 2.7- 2.10 show the coefficients and the standard errors for each
of the seven asset equations and the housing expenditure equation for home-
owners. Permanent income has significant but small marginal effects on the
structure of homeowners’ portfolio. The share of the portfolio allocated to
RETIRE, HOUSE and VEHICLE increase with income, while the share allo-
cated to ACCOUNT, STOCK and RESTATE decreases with income. Higher
levels of wealth are associated with higher shares in ACCOUNT, STOCK, RE-
STATE, OTHER, and lower shares in HOUSE and VEHICLE. The marginal
effect of wealth on the share allocated to STOCKS, HOUSE and RESTATE is
large. A 10 percent increase in assets would increase the share of the average
portfolio allocated to STOCK by 0.62 percentage point. A similar increase
in assets would induce 1.25 percentage point decrease in HOUSE and 0.66
percentage point increase in RESTATE.

Age is an important determinant of portfolio shares in a homeowner’s


portfolio, and the results in Table 2.7 and 2.8 reveal a quadratic relationship in
terms of age. Portfolio shares for RETIRE, HOUSE and RESTATE increase
with age, reaching a peak at the age of 50, 63 and 50, respectively. Portfolio
shares for ACCOUNT and STOCK, however, decrease with age until the age of

26
50 and 43, respectively. This relation between age and portfolio shares suggests
that the structure of a household’s portfolio changes when the household head
reaches middle age. For example, households headed by persons above the age
of 45 start substituting liquid assets for nonfinancial assets such as HOUSE
and RESTATE.

The coefficients on the number and age of children suggest that the
presence of children plays a significant role on the portfolio structure of home-
owners’. Several results are of particular interest. First, relative to households
with no children, households with one child have a 5.6 percent higher portfolio
share of HOUSE, controlling for age and permanent income. Similarly, house-
holds with two and three or more children have 8.9 and 9.2 percent greater
portfolio shares in HOUSE. Second, the portfolio shares for ACCOUNT, RE-
TIRE, and VEHICLE decrease with an increase in the number of children.
Controlling for the number of children, households with all the children older
than age 13 hold a smaller portfolio share in HOUSE and a greater share in
VEHICLE and RESTATE.

Finally, homeowners that are willing to undertake risky investments


hold a greater share of risky financial assets, such as STOCKS and RETIRE,
and a smaller share of less risky assets, such as ACCOUNT and HOUSE.
All other things held constant, the portfolio shares allocated to ACCOUNT
and RESTATE have declined in 1998. Households have substituted STOCK,
RETIRE and VEHICLE for the other asset categories since 1995. An increase
in the marginal tax rates leads to an increase in the portfolio share allocated

27
to HOUSE and VEHICLE. It leads to a decrease in the share allocated to
ACCOUNT, RESTATE and OTHER.

Tables 2.9 and 2.10 present the estimates of the equations (2.11) for
renters. The effect of children is less pronounced for renters than for homeown-
ers. As renters have two or more children, the share for ACCOUNT decreases,
and the share for VEHICLE is significantly higher for households with three or
more children. More permanent income is associated with a higher share for
ACCOUNT, STOCK, RETIRE and RESTATE, and a lower share for VEHI-
CLE. An increase in total assets leads to an increase in the share for STOCK,
RESTATE and OTHER and a decrease in the share for ACCOUNT and VE-
HICLE. The quadratic relationship observed between the shares of assets in
homeowners’ portfolio and the age of the head holds true for the financial
assets in a renter’s portfolio. The portfolio share for RETIRE increases with
age until the age of 58, while the portfolio share for ACCOUNT and STOCK
decreases until the age 40 and 43, respectively. Since 1995, renters have shifted
toward RETIRE in their portfolio. Compared to 1989, for example, the 1998
portfolio share for RETIRE is 5.0 percent higher in renters’ portfolio.

Tables 2.7-2.10 report coefficients of the selectivity variables. Self-


selection occurred in households’ tenure choice. The coefficients on the se-
lection terms in equations for ACCOUNT, RETIRE and HOUSE for home-
owners are all statistically significant. For these assets, homeownership would
not have the same effect on renters, should they choose to buy homes. The
estimates of the Mills ratios for renters are significantly different from zero

28
for ACCOUNT, RETIRE, and RESTATE. This implies that other than in re-
gards to these three assets, there were no significant differences in the average
behavior of the two groups prior to home purchase.

The last two columns in Tables 2.8 and 2.10 present the estimates of the
housing expenditure equation. For homeowners, the expenditure on housing
increases with the number of children, but the number of children has no effect
on renters’ expenditure. Homeowners with one child have 11.9 percent higher
housing expenditure than homeowners with no child. The housing expenditure
of homeowners increases 8.3 percent with the second child. After the second
child, having more children increases the housing expenditures of homeowners
by only 3.2 percent. The age of the children in the household has no effect
on the housing expenditure of renters nor homeowners. For both renters and
owners, the significance and the same sign of the selection terms indicate
that self-selection occurred in a hierarchical sorting: the positive selectivity
bias indicates that those who own a house spend less compared to average
household had it chosen to own. On the other hand, the negative selectivity
bias for renters’ implies the reverse: renters spend less on housing compared
to average household of the sample had it chosen to rent.

I use the estimated coefficients and the variables of the model to calcu-
late the portfolio share for each asset by the number of children and the age
of the household head. Table 2.11 presents the estimates of shares for assets
that a typical homeowner holds. By a typical household, I mean a household
headed by a white married, all of the children in the household are younger

29
than age 13. The household head is willing to take risky investments and holds
mean wealth ($188,160) and permanent income ($46,690) and has a 15 per-
cent marginal tax rate. As mentioned above, children have two effects on the
portfolio structure of households. First, an increase in the number of children
increases the probability that a household owns a home. Second, conditional
on the tenure choice, children change the demand for each asset. The portfolio
shares of assets calculated in Table 2.11 include both of these effects. At all
ages, HOUSE is the most important asset, and its importance in the portfolio
increase with the number of children living in the household. VEHICLE is
the second most important asset in the portfolio when the household head is
30 years old. As the household head reaches middle age, more is invested in
RETIRE, and the share allocated to RETIRE becomes the second largest in
the portfolio. The number of children has a negative effect on the portfolio
share for RETIRE.

2.5 Conclusion

Using the 1989, 1992, 1995 and 1998 SCF, this chapter investigates
how the number and the age of children living in the household influence
the portfolio composition of households. One contribution of this chapter is
to study the effect of the portfolio constraint imposed by the consumption
demand for housing on the portfolio shares in housing and other assets. The
chapter examines the impact of children on the homeownership decision and
the constraint of consumption demand for owner-occupied housing. Using a

30
switching regression model that takes into account the consumption demand
for housing, the chapter compares the determinants of portfolio allocation of
homeowners to that of renters.

The results show that the number of children living in the household has
a significant effect on the tenure choice and on the housing demand of home-
owners. As homeowners have more children, the portfolio share for financial
assets such as interest-bearing accounts and retirement accounts decreases,
and the portfolio share for housing increases. However, the ratio of retirement
accounts to total assets in renters’ portfolios does not significantly decrease
with the number of children. This result suggests that, for households with
children, the consumption demand for housing is higher than the investment
demand. Since households cannot separate the level of consumption of housing
services from their investment in housing as an asset, the ratio of housing to
total assets increases as the number of children increases.

Considerable research has focused on whether U.S. households are sav-


ing enough for retirement. An important implication of the findings of this
chapter is that the constraint imposed by the consumption demand for housing
decreases the share of portfolio allocated to retirement wealth as the number
of children in a household increases. Therefore, the policies that change the
cost of housing and affect ownership decision influence not only the portfolio
share for owner-occupied housing but also the portfolio share for retirement
assets.

One direction for further research is to include the liabilities and bor-

31
rowing constraints of households into the model of portfolio choice. Most
households finance their home purchases with mortgage debt. The impact of
children on the portfolio share for housing may be an important determinant
of household mortgage debt.

32
Table 2.1: Descriptive Statistics by Year

1989 1992 1995 1998

Income and Assets


INCOME 47,968 46,054 49,319 50,658
ASSETS (Mean) 222,151 203,328 206,815 258,191
ASSETS (Median) 92,684 92,525 101,829 116,750
MRT 0.158 0.164 0.154 0.131
Eh 5,660 12,985 6,664 6,695

Demographics
AGE 47.8 48.5 48.3 48.9
MARRIED 0.59 0.58 0.59 0.59
FEMALE 0.28 0.27 0.28 0.28
NCHILD 0.83 0.80 0.75 0.75
CHAGE13 0.14 0.12 0.11 0.12
HOMEOWN 0.64 0.65 0.65 0.66
RISKY 0.51 0.50 0.55 0.61

Number of observations 2,900 3,509 3,773 3,807


% with positive wealth 0.97 0.97 0.97 0.97

Source: Survey of Consumer Finances, 1989-1998.


Notes: 1) Tabulations are weighted using sample weights. 2) All dollar values are reported
in 1998 dollars. The text defines total assets, permanent income and net worth. All variables
are defined in Appendix A.2.

33
Table 2.2: Mean Asset Shares by Year

1989 1992 1995 1998

Portfolio Shares
ACCOUNT 0.143 0.130 0.112 0.132
STOCK 0.043 0.043 0.047 0.059
RETIRE 0.057 0.072 0.094 0.105
HOUSE 0.415 0.432 0.410 0.394
VEHICLE 0.197 0.196 0.208 0.186
RESTATE 0.068 0.059 0.053 0.050
OTHER 0.076 0.067 0.076 0.072

Source: Survey of Consumer Finances, 1989-1998.


Notes: 1) Tabulations are weighted using sample weights. 2) The text defines the assets
called ACCOUNT, STOCK, RETIRE, HOUSE, VEHICLE, RESTATE, and OTHER.

34
Table 2.3: Expenditure on Housing, 1998
Eh
1998 dollars
%HH %HO HO RR
All households 100 66.78 7,042 6,030
Income
Below $15K 10.35 42.69 3,866 4,293
$15-30K 22.50 51.69 4,400 5,081
$30-50K 29.79 64.54 5,764 6,378
$50-100K 29.12 80.22 7,741 7,883
Above $100K 8.24 86.72 14,456 11,748

Age
Under 35 22.40 36.29 6,183 6,024
35-49 34.08 68.72 7,078 6,475
50-64 22.03 78.42 7,931 5,564
Above 65 21.49 78.89 6,496 5,489

Wealth
Below $50K 32.90 12.93 1,263 5,546
$50K-100K 12.38 80.72 3,065 7,555
$100K-250K 29.26 93.04 5,438 8,587
$250-1000K 21.29 95.43 9,843 9,486
Above 1000K 4.16 95.90 19,847 15,645

Children
CHILD0 61.17 64.22 6,677 5,973
CHILD1 15.46 64.46 6,976 6,081
CHILD2 14.28 72.77 7,803 6,002
CHILD3 9.09 67.55 8,195 6,391

Source: Survey of Consumer Finances, 1998.


Notes: 1) Tabulations are weighted using sample weights. 2) HH represents all households,
HO represents homeowners and RR represents renters.

35
Table 2.4: Mean Asset Shares, 1998

ACCOUNT STOCK RETIRE HOUSE VEHICLE RESTATE


HO RR HO RR HO RR HO RR HO RR HO RR
All households 0.071 0.260 0.064 0.049 0.102 0.112 0.579 0 0.078 0.415 0.054 0.043
Income
Below $15K 0.093 0.389 0.022 0.019 0.020 0.061 0.759 0 0.061 0.287 0.015 0.039
$15-30K 0.090 0.295 0.046 0.032 0.064 0.080 0.630 0 0.091 0.453 0.040 0.014
$30-50K 0.066 0.205 0.056 0.051 0.091 0.127 0.608 0 0.083 0.490 0.047 0.050
$50-100K 0.062 0.201 0.069 0.086 0.132 0.157 0.541 0 0.080 0.359 0.062 0.071
Above $100K 0.075 0.256 0.128 0.112 0.147 0.221 0.436 0 0.049 0.181 0.091 0.111
Age

36
Under 35 0.040 0.238 0.028 0.049 0.074 0.092 0.694 0 0.107 0.489 0.018 0.021
35-49 0.048 0.213 0.055 0.041 0.122 0.149 0.587 0 0.088 0.413 0.047 0.068
50-64 0.062 0.251 0.077 0.031 0.135 0.137 0.535 0 0.072 0.401 0.068 0.052
Above 65 0.129 0.454 0.083 0.090 0.055 0.059 0.559 0 0.056 0.190 0.065 0.046
Wealth
Below $50K 0.042 0.281 0.004 0.027 0.025 0.089 0.730 0 0.172 0.485 0.002 0.011
$50K-100K 0.054 0.135 0.007 0.122 0.047 0.212 0.750 0 0.101 0.143 0.017 0.183
$100K-250K 0.073 0.199 0.033 0.149 0.089 0.234 0.645 0 0.087 0.086 0.028 0.162
$250-1000K 0.082 0.166 0.111 0.165 0.151 0.192 0.445 0 0.047 0.022 0.097 0.293
Above 1000K 0.075 0.078 0.252 0.374 0.162 0.270 0.226 0 0.019 0.010 0.158 0.109

continued on the next page.


Table 2.5: Mean Asset Shares, 1998: Continued
ACCOUNT STOCK RETIRE HOUSE VEHICLE RESTATE
HO RR HO RR HO RR HO RR HO RR HO RR
Children
CHILD0 0.089 0.288 0.074 0.061 0.098 0.117 0.560 0 0.070 0.373 0.062 0.040
CHILD1 0.052 0.222 0.055 0.026 0.122 0.102 0.577 0 0.098 0.495 0.038 0.044
CHILD2 0.044 0.195 0.049 0.019 0.110 0.115 0.609 0 0.087 0.471 0.044 0.069
CHILD3 0.038 0.200 0.048 0.040 0.087 0.088 0.653 0 0.085 0.512 0.045 0.032

Source: Survey of Consumer Finances, 1998.


Notes: 1) Tabulations are weighted using sample weights. 2) All dollar values are reported in 1998 dollars. 3) HO repre-
sents homeowners and RR represents renters. The text defines the assets called ACCOUNT, STOCK, RETIRE, HOUSE,

37
VEHICLE, and RESTATE.
Table 2.6: Results from Probit Estimation
HOMEOWN
Coefficient Standard Errors Marginal Effects
CONSTANT -4.404 0.137 **
AGE 0.118 0.005 ** 0.014
AGE2 /100 -0.080 0.005 **
MARRIED 0.747 0.042 ** 0.261
FEMALE 0.137 0.044 ** 0.046
CHILD1 0.193 0.049 ** 0.063
CHILD2 0.341 0.047 ** 0.108
CHILD3 0.304 0.052 ** 0.096
CHAGE13 0.000 0.053
INCOME/10,000 0.030 0.004 ** 0.010
MTR 2.174 0.153 ** 0.007
RISKY 0.210 0.030 ** 0.072
WHITE 0.287 0.034 ** 0.102
YEAR92 -0.085 0.041 * -0.029
YEAR95 -0.052 0.040
YEAR98 -0.143 0.040 ** -0.050

Notes: 1) ** indicates significance at 1 percent level, and * indicates significance at 5 percent


level. 2) Variables are defined in Appendix A.2. The number of observations N=13,579.

38
Table 2.7: Results: Asset Shares and Housing Expenditure of Homeowners

ACCOUNT STOCK RETIRE HOUSE VEHICLE


Coef SE Coef SE Coef SE Coef SE Coef SE
CONSTANT 0.432 0.054 ** -0.455 0.069 ** -0.190 0.063 ** 1.515 0.087 ** 0.398 0.029 **
AGE -0.009 0.001 ** -0.006 0.001 ** 0.007 0.001 ** 0.005 0.002 ** 0.000 0.001
AGE2 /100 0.009 0.001 ** 0.007 0.001 ** -0.007 0.001 ** -0.004 0.001 ** -0.001 0.000
MARRIED -0.033 0.006 ** -0.026 0.007 ** 0.001 0.007 0.059 0.010 ** 0.019 0.003 **
FEMALE 0.008 0.005 -0.013 0.007 0.005 0.006 0.061 0.009 ** -0.022 0.003 **
CHILD1 -0.016 0.006 * -0.006 0.007 -0.013 0.006 * 0.056 0.009 ** -0.008 0.003 **
CHILD2 -0.026 0.006 ** -0.006 0.007 -0.024 0.006 ** 0.089 0.008 ** -0.010 0.003 **

39
CHILD3 -0.021 0.007 ** -0.013 0.007 -0.032 0.006 ** 0.092 0.009 ** -0.008 0.003 **
CHAGE13 -0.001 0.006 -0.010 0.006 0.003 0.005 -0.020 0.008 ** 0.013 0.002 **
L INCOME -0.007 0.002 ** -0.009 0.003 ** 0.012 0.003 ** 0.021 0.005 ** 0.005 0.002 **
L ASSET 0.004 0.001 * 0.062 0.002 ** -0.001 0.002 -0.125 0.003 ** -0.031 0.001 **
MTR -0.067 0.018 ** -0.016 0.021 0.005 0.020 0.307 0.020 ** 0.024 0.011 *
RISKY -0.010 0.003 ** 0.024 0.004 ** 0.023 0.004 ** -0.027 0.003 ** -0.003 0.002
YEAR92 -0.010 0.003 ** 0.006 0.005 0.007 0.004 -0.002 0.006 -0.003 0.002
YEAR95 -0.027 0.003 ** 0.008 0.004 0.023 0.004 ** -0.012 0.006 * 0.008 0.002 **
YEAR98 -0.027 0.004 ** 0.021 0.005 ** 0.037 0.004 ** 0.005 0.006 0.004 0.002 *
MR:home -0.115 0.018 ** -0.029 0.023 -0.049 0.020 * 0.198 0.021 ** 0.011 0.008
MR:+ wealth -0.077 0.031 * 0.112 0.054 * -0.063 0.049 0.056 0.039 -0.063 0.010 **

continued on the next page.


Table 2.8: Homeowners: Continued
RESTATE OTHER log Eh
Coef SE Coef SE Coef SE
CONSTANT -0.584 0.080 ** -0.116 0.049 * -0.246 0.220
AGE 0.004 0.002 ** -0.002 0.001 0.019 0.004 **
2
AGE /100 -0.004 0.001 ** 0.001 0.001 -0.015 0.003 **
MARRIED -0.010 0.009 -0.012 0.006 0.181 0.024 **
FEMALE -0.036 0.009 ** -0.004 0.006 0.177 0.025 **
CHILD1 -0.014 0.008 0.001 0.005 0.119 0.025 **
CHILD2 -0.021 0.008 ** -0.002 0.005 0.202 0.024 **
CHILD3 -0.010 0.008 -0.007 0.005 0.234 0.025 **
CHAGE13 0.017 0.007 * -0.002 0.004 -0.023 0.023

40
L INCOME -0.018 0.003 ** -0.004 0.002 0.068 0.011 **
L ASSET 0.066 0.002 ** 0.025 0.002 ** 0.568 0.007 **
MTR -0.158 0.025 ** -0.094 0.018 **
RISKY -0.004 0.005 -0.003 0.003
YEAR92 0.010 0.005 * -0.008 0.004 * 1.204 0.018 **
YEAR95 0.000 0.005 0.001 0.004 0.307 0.017 **
YEAR98 -0.032 0.005 ** -0.008 0.004 * 0.189 0.017 **
MR:home -0.001 0.025 -0.015 0.015 0.393 0.052 **
MR: + wealth 0.043 0.053 -0.008 0.036 -0.503 0.120 **

Notes: 1) ** indicates significance at 1 percent level, and * indicates significance at 5 percent level. 2) The text defines
the assets called ACCOUNT, STOCK, RETIRE, HOUSE, VEHICLE, RESTATE, and OTHER. All variables are defined in
Appendix A.2. MR represents Mills Ratio. The number of observations N=10,002.
Table 2.9: Results: Asset Shares and Housing Expenditure of Renters

ACCOUNT STOCK RETIRE VEHICLE


Coef SE Coef SE Coef SE Coef SE
CONSTANT 0.243 0.173 -0.285 0.097 ** -0.793 0.154 ** 2.059 0.243 **
AGE -0.014 0.003 ** -0.004 0.001 ** 0.015 0.002 ** -0.001 0.004
AGE2 /100 0.017 0.002 ** 0.005 0.001 ** -0.013 0.002 ** -0.004 0.003
MARRIED -0.086 0.021 ** -0.035 0.011 ** 0.021 0.017 0.130 0.026 **
FEMALE 0.026 0.013 * 0.010 0.007 0.027 0.010 ** -0.059 0.018 **
CHILD1 -0.032 0.017 0.005 0.010 0.002 0.012 0.025 0.020
CHILD2 -0.054 0.019 ** 0.000 0.011 0.019 0.014 0.034 0.023
CHILD3 -0.077 0.021 ** -0.005 0.015 0.010 0.015 0.061 0.024 *

41
CHAGE13 -0.003 0.022 -0.023 0.016 0.001 0.014 0.029 0.026
L INCOME 0.080 0.014 ** 0.022 0.007 ** 0.033 0.011 ** -0.109 0.020 **
L ASSET -0.072 0.004 ** 0.017 0.002 ** 0.014 0.003 ** -0.027 0.006 **
MTR 0.204 0.079 * 0.000 0.038 0.319 0.064 ** -0.120 0.103
RISKY 0.011 0.012 0.019 0.007 ** 0.028 0.009 ** -0.068 0.015 **
YEAR92 -0.015 0.014 0.001 0.008 0.018 0.014 -0.006 0.019
YEAR95 -0.033 0.015 * -0.001 0.008 0.051 0.012 ** 0.015 0.019
YEAR98 0.014 0.014 0.008 0.008 0.050 0.013 ** -0.030 0.019
MR:home -0.089 0.038 * -0.029 0.019 0.079 0.031 * 0.074 0.048
MR:+ wealth 0.070 0.039 0.039 0.037 0.052 0.035 -0.381 0.053 **

continued on the next page.


Table 2.10: Renters: Continued
RESTATE OTHER log Eh
Coef SE Coef SE Coef SE
CONSTANT -0.334 0.130 * 0.109 0.173 4.580 0.365 **
AGE 0.004 0.002 0.001 0.003 -0.004 0.004
2
AGE /100 -0.004 0.002 * -0.001 0.002 -0.002 0.003
MARRIED 0.000 0.015 -0.030 0.020 -0.064 0.036
FEMALE -0.017 0.010 0.014 0.014 0.110 0.033 **
CHILD1 0.011 0.011 -0.012 0.016 -0.061 0.035
CHILD2 -0.001 0.012 0.001 0.017 -0.067 0.037
CHILD3 -0.005 0.014 0.017 0.018 0.027 0.038
CHAGE13 -0.005 0.010 0.001 0.020 0.083 0.044

42
L INCOME 0.038 0.003 ** -0.022 0.014 0.315 0.030 **
L ASSET -0.038 0.054 0.030 0.005 ** 0.062 0.006 **
MTR -0.001 0.008 -0.365 0.073 **
RISKY 0.007 0.009 0.010 0.011
YEAR92 -0.021 0.009 * -0.005 0.014 0.069 0.030 *
YEAR95 -0.022 0.010 * -0.011 0.014 0.045 0.032
YEAR98 -0.014 0.024 -0.020 0.014 0.125 0.029 **
MR: home 0.102 0.033 ** -0.020 0.035 -0.285 0.042 **
MR: + wealth -0.004 0.012 0.118 0.038 ** -0.189 0.073 **

Notes: 1) ** indicates significance at 1 percent level and * indicates significance at 5 percent level. 2) The text defines
the assets called ACCOUNT, STOCK, RETIRE, HOUSE, VEHICLE, RESTATE, and OTHER. All variables are defined in
Appendix A.2. MR represents Mills Ratio. The number of observations N=3,577.
Table 2.11: Portfolio Shares for Assets by the Number of Children and Age

CHILD0 CHILD1 CHILD2 CHILD3


AGE=30
ACCOUNT 0.038 0.036 0.036 0.038
STOCK 0.064 0.061 0.064 0.056
RETIRE 0.093 0.086 0.079 0.070
HOUSE 0.594 0.626 0.642 0.650
VEHICLE 0.112 0.102 0.100 0.101
RESTATE 0.043 0.030 0.023 0.033

AGE=40
ACCOUNT 0.049 0.043 0.040 0.043
STOCK 0.056 0.053 0.055 0.047
RETIRE 0.128 0.120 0.111 0.103
HOUSE 0.552 0.590 0.611 0.617
VEHICLE 0.105 0.096 0.094 0.095
RESTATE 0.057 0.044 0.037 0.047

AGE=50
ACCOUNT 0.058 0.049 0.044 0.048
STOCK 0.058 0.054 0.055 0.047
RETIRE 0.141 0.132 0.122 0.114
HOUSE 0.534 0.577 0.602 0.607
VEHICLE 0.099 0.090 0.088 0.089
RESTATE 0.063 0.049 0.043 0.053

Notes: The text defines the assets called ACCOUNT, STOCK, RETIRE, HOUSE, VEHI-
CLE, and RESTATE.

43
Chapter 3

The Effect of Precautionary Motives on


Household Saving and Fertility

3.1 Introduction

Many recent studies have recognized the role of precautionary motives


on household saving behavior.1 Precautionary saving models predict that un-
certainty about future income may cause households to reduce their current
consumption in order to raise their stock of precautionary saving. As an exten-
sion to the traditional life-cycle model, these models are able to explain some
of the empirical consumption puzzles.2 For example, the standard life-cycle
model suggests that households smooth consumption and spread resources
across periods of high and low income. In many household-level data sets,
however, consumption profiles over age are hump-shaped, tracking the age-
earnings profile. Carroll [8] shows that this kind of consumption profile is
consistent with a precautionary saving model in which individuals face uncer-
tainty about their future earnings.

Yet, empirical work on the strength of precautionary saving has pro-


vided mixed evidence. Skinner [48], Dynan [18] and Starr-McCluer [50] find lit-

1
See Zeldes [58], Kimball [39], Hubbard et al. [32] and Carroll [9].
2
Deaton [15] and Browning and Lusardi [6] give a list of empirical puzzles.

44
tle or no evidence for precautionary motive, whereas Carroll and Samwick [11],
[12] and Lusardi [44] find more support for the precautionary motive. Brown-
ing and Lusardi [6] and Carroll et al. [10] suggest that the mixed results might
be due to the difficulties in empirically testing for precautionary saving.3 One
problem that has not been mentioned in the literature is that all of these
empirical models try to explain the effect of income uncertainty on house-
hold savings, ignoring the effect of uncertainty on household composition. By
examining the implications of uncertainty on the fertility decisions of house-
holds and incorporating fertility decisions into household saving decisions, this
chapter extends the empirical work on precautionary saving.

Most of a household’s saving motives can be grouped into one of three


categories: life-cycle motives, precautionary motives, and bequest motives. It
seems reasonable that these motives are affected by the presence of children.
For example, the life-cycle motive includes saving for children’s education,
the precautionary motive includes saving to protect the well-being of children
against income fluctuations, and, finally, the bequest motive includes saving to
leave assets to children. Yet the causal effect might go in the opposite direction;
that is, fertility might be affected by uncertainty or income fluctuations, given
precautionary and other motives. Furthermore, household income or the age
of the head might affect household saving and fertility simultaneously. This
chapter takes account of the fact that children are endogenous along with the

3
The problems include proxying certainty, finding an appropriate instrument, and incor-
porating the restrictions of the theoretical model. See Browning and Lusardi [6] and Carroll
et al. [10] for the details.

45
saving behavior when estimating the effect of children on savings.

Table 3.1 presents the proportion of households citing the following


motives -‘rainy’ days, retirement, buying a home and education of children-
as the most important reasons for saving in the 1983 SCF (data come from
the panel of 1983-89 SCF and is discussed at length in section 3.3). The most
frequently reported reason for saving was to increase resources for ‘rainy days’
such as unemployment and unexpected needs. More than 32 percent reported
that ‘rainy days’ were an important motivation for saving. The second most
frequent reason was saving for retirement, with 18 percent. The proportion of
households citing saving for children’s educational expenses and home purchase
were 5.7 and 4.1 percent, respectively. When disaggregated into age groups, all
of the four reasons reveal a hump shape: saving for ‘rainy days’ peaking in the
41-50 age group, saving for retirement peaking between age 51 and 60, saving
for a home purchase peaking below age 31, and saving for the education of
children peaking between age 31-40. This suggests that the relative importance
of saving for each motive depends highly on the composition and the life-cycle
stage of the household.

This chapter also addresses a neglected topic in the childbearing liter-


ature, namely, the effect of income uncertainty on fertility over the life-cycle.
Most life-cycle fertility models incorporate some types of uncertainty.4 For
example, Wolpin [57] estimates a dynamic stochastic model of fertility within

4
See Hotz et al. [30] for a survey of life-cycle fertility models.

46
an environment where infant survival is uncertain.5 Hotz and Miller [31] in-
tegrate the life cycle fertility and labor supply, and consider a number of
uncertainties such as the outcome of the contraceptive effort, the time path
of the husband’s income, and transitory shocks to the wife’s wage. None of
these studies, however, have specifically analyzed whether uncertainty about
earnings is a significant factor on the choice of whether or not to have a child.6

This chapter examines whether income uncertainty is associated with


lower fertility and higher savings. Using the data from the panel of 1983-89
SCF, I find that households with higher income uncertainty are less likely
to have a child. Yet the prediction of the precautionary view of savings is
not validated: income uncertainty actually reduces savings of households with
either high or low wealth holdings, and does not affect savings of the rest of
the population, even after controlling for the fact that saving is endogenous
to the fertility behavior. The finding is consistent with previous studies that
found little or no effect of precautionary motive on savings. However, there
is evidence that income uncertainty has a direct effect on fertility and family
size. This chapter also examines whether having a child has an effect on

5
Wolpin [57] presents a model in which income is stochastic but his model also assumes
that households have quadratic utility, and thus the variance of income does not appear in
the decision function.
6
Becker [1] suggests that children can be viewed as durable goods yielding psychic income
to the parents. In a study that addresses whether unemployment risk is an important factor
in the timing of the purchase decision of durable goods, Dunn [17] finds consumers respond
to increases in the unemployment risk by postponing purchases of a home or a vehicle. Thus,
this chapter can be viewed as a combination of those two prior works, treating children as a
durable good the demand for which is found to respond to increases in unemployment risk
(like other durable goods in Dunn.)

47
household savings. The results show that having a child appears to reduce
savings of households with young heads and to increase savings of those with
middle-aged heads.

The remainder of this chapter is organized as follows. Section 3.2 ex-


amines both the theoretical and the empirical model. Section 3.3 describes the
data set and the variables used in the empirical work. The empirical results
are reported in Section 3.4, and a summary of the findings with conclusions
are provided in Section 3.5.

3.2 The Relationship between Fertility and Saving

Households are assumed to maximize a lifetime utility function that is


additively separable over time. The utility of household i at age t depends on
the number of children, Mit , and a composite consumption good, Cit :
T
X
U= β t U (Mit , Cit ) (3.1)
t=0

where β is the discount factor and T is the time of death. The household
faces two decisions at each period: whether to have a child, and how much to
consume. If parents give birth to a child at age t, then ∆CHILDit = 1, and
= 0 otherwise. The number of children at age t, Mit , is the sum of all births
until age t.

The household is able to borrow and lend across time periods at a real
interest rate. Savings at age t, Sit , depend on the household income, the cost of
consumption good, and the cost of children. The household income is assumed

48
to be stochastic. Thus the household faces uncertainty about future income.
Depending on the utility function, income uncertainty can affect the fertility
and consumption decisions of the household.

This utility maximization problem, in general, is intractable and does


not deliver closed-form solutions without imposing structural assumptions con-
cerning the utility function. This makes deriving testable implications impos-
sible, even for a two-period model. The construction of the model, however,
shows how fertility and saving decisions can be determined simultaneously.

The lack of testable implications from the theoretical model allows me


to examine a general form of saving and fertility behavior. For the empirical
specification, I assume that the level of savings of a household i at time t, (Sit ),
is a linear function of the variability of the household’s income (Φit ), birth of
a child (∆CHILDit ), and a set of observable variables (Xits ) that measure
the life-cycle stage of the household. The matrix Xits includes the number of
children living in the household, permanent and transitory income and other
household demographics. Permanent income is defined as the expected income
for year t conditional on the demographics of the household, and transitory
income is defined as the difference between realized and expected income for
year t. Savings of a household i at time t can be thus represented as:

Sit = γ0 + Φit γ1 + ∆CHILDit γ2 + Xits γ3 + u1it (3.2)

where γ0 , γ1 , γ2 and γ3 are the parameters to be estimated, and u1it is an error


term representing unobservable variables.

49
The precautionary saving model predicts that saving is increased by a
combination of a positive third derivative of the utility function and uncer-
tainty about the future income. Therefore, a positive value for γ1 is implied
by a utility function with a positive third derivative (as with constant abso-
lute risk aversion (CARA) or constant relative risk aversion (CRRA) utility
functions). For a quadratic utility function (for which the third derivative is
zero), saving behavior does not respond to income variability, and in this case,
γ1 should be zero.

The life-cycle model suggests that a household that gives birth to a child
at time t saves less (due to an increase in necessary consumption). Households
with younger heads may save even less with an additional child because their
current (expected) income is less than the annuity value of their lifetime in-
come, and the difference between their income and expenditure is even greater.
Such a model suggests γ2 should be negative, and the coefficient of the inter-
action of ∆CHILDit with the age of the household head should be positive.

The childbearing decision of a fecund household is specified as a func-


tion of Φit and a set of household specific variables that affect the preferences
for a child, Xitc . A household is considered to be fecund if the wife is younger
than age 49 or if the head of the household is a female younger than age 49.
The decision to have an additional child is represented as

∆CHILDit∗ = η0 + Φit η1 + Xitc η2 + u2it (3.3)

50
where

∆CHILDit = 1 if ∆CHILDit∗ > 0

= 0 otherwise.

where η0 , η1 and η2 are parameters to be estimated, and u2it is an error term


representing unobservable variables. If consumers react to increases in uncer-
tainty by cutting down their consumption, as suggested by the precautionary
saving model, then they should also reduce their ‘consumption’ of children.
This implies that households with higher income variability are less likely to
have a child, and η1 should be negative.

Note that the model is identified even if u1it and u2it are not independent
and Xits includes all the variables in Xitc . The model is estimated using a two-
stage estimation procedure described in Maddala [45]. First, I restrict the
sample to the fecund population and get an estimate γ̂2 of γ2 by using the
probit ML method for the equation (3.3). Then I estimate the equation (3.2) by
OLS for all of the sample after substituting γ̂2 for γ2 for the fecund population
and 0 for the other households. Maddala [45] shows that the resulting estimates
of the coefficients are consistent.

3.3 Data

The data set used for estimation is the 1983 and 1989 panel of the SCF.
This data set contains detailed information on household assets, liabilities,
income and characteristics in 1983 and 1989. The 1983 SCF interviewed a

51
sample of 4, 103 households and 1, 497 of them were reinterviewed in 1989.7

Household saving is derived as a first difference in net worth between


1983-89, and this amount is divided by six to get the annual household saving.
Net worth (NWORTH) is the total value of household’s assets minus its total
liabilities. Total assets is the sum of financial assets and nonfinancial assets,
where the first includes liquid assets (checking assets, money market deposit
accounts, saving accounts, call accounts, certificates of deposit and saving
accounts), stocks, bonds, mutual funds, individual retirement accounts, Keogh
accounts, trusts, cash value of life insurance and the later includes residential
property, other real estate, business equity, vehicles and other real assets like
art and precious metals. Total liabilities include mortgage debt, home equity,
loans, other loans for property, credit card debt, automobile loans, balances
outstanding on lines of credit and loans on consumer durables.

The first saving measure, which will be called SAVE1, includes capital
gains, which makes it difficult to distinguish between active and passive sav-
ing. The 1989 SCF also asked households to report major changes in asset
holdings since 1983. This information could be used to exclude both realized
and unrealized capital gains. However, substantial inconsistencies are observed
between reported net investments in assets and measured changes in holdings.

7
The 1983 SCF consists of a dual sample. In addition to a standard multi-stage area
probability sample, a list sample was drawn from tax information provided by International
Revenue Service. An oversample of 438 high-income households came from this list in 1983,
and 361 of them were reinterviewed in 1989. See Kennickell and Starr-McCluer [37] for a
general description of the 1983-89 panel.

52
The inconsistency seems to be lower for home purchases (Kennickell and Starr-
McCluer [38]). To exclude the capital gains, I adjusted SAVE1 as follows to
obtain a measure called SAVE2: whenever a household did not buy or sell a
house that was the family’s primary residence, I kept the value of the primary
residence constant.

The income measure comes from the question, ‘In [the preceding cal-
endar year] how much was the total income you (and your family living here)
received from all sources, before taxes and other deductions were made?’ In-
come of the households for 1983, 1984, and 1985 are drawn from the 1986 wave
of SCF which was conducted with a large subset of 1983 respondents using a
shorter questionnaire.8 Income values for 1982, 1986, 1987 and 1988 are drawn
from the 1983-1989 panel. All values are converted to 1989 dollars using the
Consumer Price Index Research Series Using Current Methods(CPI-U-RS).

Using the panel dimension of income observations in the data, I de-


fine two measures of income uncertainty. The first measure assumes that
households have knowledge about their future income and expect their in-
come to change over time as household characteristics change. To remove the
predictable component of income growth, I regress log income on age, year
dummies, education, dummies representing asset holdings in 1983, household
9
demographic variables and age-interaction terms. Income uncertainty at-

8
Of the 4, 103 households in the 1983 SCF, 2, 822 were reinterviewed in 1986 using a
shorter questionnaire.
9
The income measure includes both capital and non-capital income. The precautionary
saving model predicts that income risk regarding capital income might have a different effect

53
tributed to each household is equal to the variance of residual log income
(VRLI) for the 1982-88 period.10 Household permanent income (PERINC)
is defined as the mean of predicted income over the seven year period, while
transitory income (TRANSINC) is the mean of residuals from the earnings
equation. The mean of the reported income over the 1982-88 period (MEAN-
INC) is also used as another measure of income.

The second measure of uncertainty is the variance of log income for the
1982-89 period (VLI).11 This measure assumes households have no information
to forecast future income aside from their current income. However, households
probably expect their income to change over time and know when some of
these changes will happen. Not excluding such expected changes biases this
VLI measure of uncertainty upward. VRLI may suffer from the same deficit if
income change is due to a factor that the household has information about but
is not controlled for in the income regression. As pointed out in Lusardi [44]
and Browning and Lusardi [6], variability measures like VLI and VRLI might
be poor proxies for uncertainty. Most studies use instrumental variables for the
uncertainty proxy using information on occupation, education and industry.
However, using instrumental variable estimators is not useful when the first
stage instruments are poor. In addition, finding an appropriate instrument

on household saving behavior than that of earnings. Dummies representing the amount of
assets that households hold by 1983 are included in the regression to control for this effect.
10
Female labor supply decisions are correlated with household fertility decisions. There-
fore, I control for the employment status of the spouses and female heads in the earnings
regression. Unfortunately, this information is only available for 1983, 1986 and 1989.
11
Another income variability measure, which is not reported in this chapter, is the coef-
ficient of variation of log income. The empirical results hold true for this measure too.

54
to exclude for identification is problematic. Therefore, I use VRL and VRLI
without an instrumental variable estimator.

The sample selection criteria for the sample are as follows. A household
is only included in the analysis if it remained intact between 1983 and 1989 (i.e.,
households that experienced a change in composition such as marriage, divorce,
separation or the death of either head or spouse are excluded.) This eliminates
the income variability or net worth change caused by family separation or
family creation. I also exclude those households with net worth greater than
$10 million in 1983 or 1989 or for whom the absolute value of the change
in net worth per year is more than $600, 000.12 To calculate an accurate
measure of income uncertainty, those households with more than three missing
or non-positive income values are dropped. In the panel 1983-89 SCF, 299
households out of 1,479 experienced a major change in family composition
and were dropped from the sample. Of the remaining 1,180 households, 84
were dropped because of outlying net worth or saving values and 66 were
dropped because of missing income values. The final sample consists of 1, 035
households with the heads between the age of 22 and 88 in 1983.13

Table 3.2 illustrates the composition of the sample in detail. All vari-
ables are described in detail in Appendix B.1. The variable ∆CHILD indi-

12
The cut-off net worth of $10 million and saving of $600, 000 is somewhat arbitrary.
However, this exclusion or a similar one is necessary when working with means which are
affected by outliers.
13
The sample design in 1983 specifically excluded households with the heads under the
age of 22. See Kennickell and Starr-McCluer [37] for details.

55
cates the fertility of the household between 1983 and 1989, i.e., ∆CHILD = 1
if the household experienced at least one birth of a child.14 I refer to the
households that had a child as households with an additional child. Columns
(1) and (2) provide the variable means and standard deviations of all of the
households in the sample and the fecund households, respectively. Most of the
differences in net worth and saving between the two groups can be attributed
to the fact that these two groups are at different stages of their life cycles.
Fecund households are headed by younger persons, had less net worth in 1983,
but had higher income between 1983 and 1989, and saved more compared to
the other households in the sample. Also, fecund households are faced with
higher income uncertainty than the rest of the sample.

Columns (3) and (4) of Table 3.2 provides the variable means by house-
hold fertility of the fecund households. According to the SCF data, 18.6 per-
cent of the fecund households had a child between 1983 and 1989. Among
fecund households, comparisons across the two groups of households - with
and without an additional child - are plausible: households with an additional
child are younger, mostly married (89.1 percent), have higher expected income
and have a higher number of young (0-6 years old) children in 1983. The av-
erage net worth for households with an additional child is $82,611 in 1983,
which is $37,179 less than the mean net worth of the rest of the fecund house-

14
Only 3.3 percent of the families experienced more than one birth during that time
period-2.3 percent had two children and 1.0 percent had three. Therefore, I use a probit
model and a dummy variable to indicate the fertility choice instead of using a count data
model.

56
holds. On the other hand, households with an additional child saved more
than the rest of the fecund sample. According to SAVE1 and SAVE2, they
saved $11, 170 and $8, 690, respectively, while fecund households without an
additional child saved $7, 648 and $6, 462. The remarkable difference in the
housing tenure choice of the two groups shows the link between the decisions of
having a child and purchasing a house. The homeownership proportion among
households with an additional child was 53.7 percent in 1983 and rose to 77.8
percent in 1989. The homeownership proportion for the rest of the fecund sam-
ple is higher in 1983 but compared to the households with an additional child,
the increase is insignificant, from 67.0 percent in 1983 to 72.5 in 1989. Also,
when we compare the income uncertainty of the two groups, we observe that
households who had a child are faced with lower income uncertainty (0.132
versus 0.191).

The measures of income uncertainty by household characteristics are


given in Table 3.3. Uncertainty estimates are greater for households with mean
income below $10, 000 and above $60, 000. This suggests that households
at the tails of the income distribution face higher uncertainty. The same
argument is true for net worth and SAVE2: households in the bottom 25 and
top 10 percent of the net worth distribution in 1983 have the highest income
variability. Households in the bottom 25 and top 10 percent of the SAVE2
distribution face higher income variability than the rest of sample. When
households are grouped by SAVE1, the bottom 25 percent of the distribution
faces a lower income variability than the households in the 25-50 percent of

57
the SAVE1 distribution. For other SAVE1 groups, income uncertainty is lower
except the top 10 percent of the distribution, for whom it reaches its highest
value. When grouped according to the number of children, the estimates of
income uncertainty decrease as the number of children living in household in
1983 increases.

Table 3.4 represents household saving, income and income uncertainty


by childbearing decisions and the age of the household head. Households who
had a child between 1983-89 are different from other households in terms of
their saving, income and income variability. Almost 36 percent of the house-
holds with heads below age 31 in 1983 had a child during the following six year
period, versus 16 percent of the households with heads between age 31-40, and
1 percent among the age 40 and above group. Households that had a child save
more, regardless of how savings were measured, and their permanent income is
higher. Difference between the savings of households with and without an ad-
ditional child increase as the age of the household head increases. Considering
the income uncertainty, households with young and middle age heads that had
a child face lower income variability, regardless of the uncertainty measure.

3.4 Estimation and Results

Table 3.5 shows the results of the probit analysis of the fertility de-
cision of the fecund sample. The dependent variable is ∆CHILD = 1 if
the household had a child between 1983-89, 0 otherwise. The right-hand
variables include factors that are expected to affect the demand for a child;

58
namely, age (AGE), marital status (MARRIED), race of the household head
(WHITE), an income risk measure (VRLI and VLI), a household income mea-
sure (MEANINC, PERMINC and TRANSINC), number of adults living in the
household (NADULT), a dummy indicating whether spouse works full time at
paid employment in 1983 (SPFULLT), homeownership in 1983 (HOWN83),
number of young (YOUNGCH), middle (MIDDCH) and older (HIGHSCH)
children in 1983 and the interaction terms for age (AGE83×HOWN83 and
AGE×YOUNGCH). The analysis in column (1) uses VRLI as the income un-
certainty measure and permanent and transitory income as the income mea-
sures; column (2) uses VLI and mean income and finally, column (3) uses
VRLI and mean income.

The results in table 3.5 show that other things being equal, the prob-
ability of having another child declines with income variability (regardless of
the measure) and the number of children in each age group living in the house-
hold. Evaluated at the sample mean values, a 0.1 increase in VRLI and VRL
decreases the probability of having a child by 0.5 and 0.6 percent, respec-
tively. The probability of having another child is lower for a household that
has a full-time working spouse or that is headed by a white person. The co-
efficient of age is highly significant and negative, indicating that being one
year older reduces the probability of having another child by 1 percent. The
signs of the age interaction terms imply that older homeowners are less likely
to have a child, whereas older households with small children are more likely
to experience another birth. A married household is 8 to 10 percent more

59
likely to have another child, compared to households headed by an unmarried
person. Finally, the probability of having a child seems to increase with in-
come; both permanent income and mean income in columns (1) and (2) are
significant. However, transitory income in column (1) has a negative effect
and mean income in column (3) is insignificant.

Estimates of the saving equations are presented in tables 3.6 and 3.7 for
SAVE1 and SAVE2. Columns (1)-(3) in tables 3.6 and 3.7 use the same in-
come and uncertainty measures as columns (1)-(3) in table 3.5. The predicted
\
probability of having a child, ∆CHILD, is included as a right-hand variable
with other factors that might affect the saving behavior; namely, a measure
of income uncertainty (VRLI and VRL), income (PERMINC, TRANSINC
and MEANINC), age of the head (AGE), the number of adults and chil-
dren living in the household (NADULT and NCHILD), a self-described ex-
pectation to leave a bequest (BEQUEST83), the change in the number of
adults between 1983-1989 (∆NADULT), a dummy indicating having 1983
net worth in the top 10 percent and bottom 25 percent (NWORTH90 and
\
NWORTH25), age interaction terms (AGE×NCHILD and age×∆CHILD)
and income uncertainty interaction terms (VRLI (VRL)×NWORTH25 and
VRLI (VRL)×NWORTH90). The top 10 percent and bottom 25 percent net
worth holdings in 1983 are included to address the saving behavior of the
wealthy and the not wealthy. Similarly, the income uncertainty interaction
terms show whether or not , in terms of the effect of uncertainty, the behavior
of the wealthy and the not wealthy are different than the rest of the population.

60
The results for two measures of savings are quite similar (SAVE1 in
Table 3.6 and SAVE 2 in Table 3.7). Income uncertainty reduces savings of
the households in the top 10 percent and bottom 25 percent of the wealth
distribution and does not affect the rest of the population. For example,
evaluated at the sample average of VRLI, which is 0.162, households in the
top 10 percent of the wealth distribution save almost $11,500 less than the rest
of the population whereas households in the bottom 25 percent of the wealth
distribution save $3,500 less than the rest of the population as a result of an
increase in income uncertainty. Having 1983 net worth in the top 10 percent is
associated with higher levels of SAVE1 and SAVE2 in all specifications. The
results in Table 3.6 show that households in the top 10 percent of the wealth
distribution and with VRLI of 0.162 save about $15,000 more than the rest of
the sample.

Saving also increases with income, regardless of the measure. However,


the results in Table 3.6 and Table 3.7 do not the support the idea that house-
holds save a higher fraction of transitory income. The estimated coefficient
of the propensity to save out of transitory income is 0.24 in column (1) of
Table 3.6, and it is significantly lower than the estimated propensity to save
out of permanent income, which is 0.34.

Before we examine the effect of children on savings, let us look at the


effect of the number of adults living in the household. Both SAVE1 and
SAVE2 reduce with the number of adults living in the household, around
$7,132-$8,447, respectively. Changes in the number of adults between 1983

61
and 1989, however, do not affect SAVE1 but appear to reduce SAVE2.

The overall effect of having an additional child on household savings


depends on the age of the head: households with heads younger than age
29 save less compared to households with heads age 29 and older. Having
an additional child reduces savings; however being one year older and having
an additional child increases savings. The same is true for the number of
children living in the household. Households with children save less when the
household head is below age 35 and save more above that age. Controlling
for the number of children already living in the household, age does not affect
the savings behavior of those without children. This result highlights the
importance of the interaction between household composition and the age of
the household head. When we control for permanent income as in column
(1) of Tables 3.6 and 3.7, we observe that the effects of the children and age
interaction terms decrease but do not disappear. Also, households expecting
to leave a bequest save significantly more (around $12,555-$12,796 more) than
household that do not expect to leave a bequest. That is another impact of
children on household savings.

This chapter also estimates average savings of households who did not
have a child between 1983-89 and compares it with what they would have
saved if they had chosen to have a child. For this, the sample is restricted
to only fecund households, and the fertility decision is modeled as an endoge-

62
nous switching model.15 The results are given in Table 3.8. Overall, average
SAVE1 of the households that did not have a child is around $12,695-13,133
according to the results of the three regressions in Tables 3.6. The results
show that households would have saved around $2,297-4,066 less if they had
chosen to have a child. This finding suggests that the overall effect of children
on household saving is negative.

3.5 Conclusion

This chapter estimates the effect of the precautionary motive on house-


hold fertility and savings by relating income uncertainty to the changes in
the number of children and household net worth. In estimating this effect, I
take into account the fact that fertility decisions are endogenous to household
saving decisions.

The empirical results suggest that income uncertainty directly affects


the probability of having a child, even after controlling for several demographic
characteristics. Income uncertainty actually decreases savings of the house-
holds with high or low wealth holdings and does not affect the saving behavior
of the rest of the population. Finally, changes in the number of children and
children already living in the household reveal a significant effect on house-
hold savings. The direction of the response, however, depends on the age of
the household head, implying that younger households save less whereas older

15
See Maddala [45] for the models with self-selectivity.

63
households save more with an increase in the number of children.

The main finding of this chapter is consistent with the life-cycle theory
of saving and consumption. Household composition is an important factor of
life-cycle savings. After controlling for the number of children living in the
household and the expectation of leaving a bequest, the age effect on savings
disappears. At the same time, the findings are not consistent with the pre-
dictions of the precautionary saving model that agents faced with uncertainty
about future income increase their savings.

64
Table 3.1: Saving Motives by Age Groups, 1983

Rainy Days Retirement Home Children


All 0.323 0.176 0.041 0.057

By Age
Below 31 0.326 0.011 0.117 0.065
31-40 0.362 0.122 0.052 0.111
41-50 0.383 0.192 0.016 0.047
51-60 0.289 0.345 0.004 0.047
61-70 0.273 0.230 0.017 0.010
70 and over 0.206 0.250 0.000 0.000

Source: Survey of Consumer Finances, 1983-1989.


Notes: The table reports the proportion of households citing the selected motives as the
most important reason for saving as ‘rainy days’, retirement, buying home and education of
children respectively. Observations are weighted to reflect the U.S. population as a whole.
Number of observations: 1035.

65
Table 3.2: Descriptive Statistics by Household Fertility Decision

All HH Fecund HH ∆CHILD = 0 ∆CHILD = 1


SAVE1 7699 8282 7648 11170
SAVE2 6080 6868 6462 8690
MEANINC 37668 43109 44122 38852
PERMINC 36339 40171 39921 41210
NWORTH 140628 112861 119790 82611

AGE 45.8 36.6 38.1 30.4


EDUC 12.4 13.1 12.9 13.7
WHITE 0.795 0.780 0.788 0.747
MARRIED 0.652 0.791 0.767 0.891
NCHILD 0.845 1.334 1.424 0.938
YOUNGCH 0.246 0.407 0.372 0.556
MIDDCH 0.297 0.490 0.537 0.289
HIGHSCH 0.302 0.437 0.516 0.092
∆CHILD 0.112 0.186
NADULT 2.001 2.120 2.135 2.055
BEQUEST 0.460 0.487 0.456 0.624
HOWN83 0.672 0.645 0.670 0.537
HOWN89 0.725 0.734 0.725 0.778
VRLI 0.162 0.180 0.191 0.132
VLI 0.169 0.189 0.197 0.151

N 1035 509 422 87

Note: ∆CHILD=1 if the household had a child between 1983 and 1989 (0 otherwise).
The table reports means of the variables. All variables are described in Appendix B.1.
Observations are weighted using the sample weights. All dollar values are in 1989 dollars.

66
Table 3.3: Mean Income Uncertainty by Household Demographics

% HH VRLI VLI
All 100 0.162 (0.426) 0.169 (0.433)
MEANINC
Below $10,000 14.0 0.1715 (0.4232) 0.1912 (0.4472)
$10,000-29,999 37.7 0.1479 (0.2642) 0.1559 (0.2740)
$30,000-59,999 36.1 0.1194 (0.2657) 0.1184 (0.2210)
Above $60,000 12.2 0.3234 (0.9087) 0.3362 (0.9431)
NWORTH
Below $10,265 25.0 0.2713 (0.6504) 0.2920 (0.6786)
$10,265-52,232 25.0 0.0988 (0.3334) 0.1027 (0.3333)
$52,233-12,7517 25.0 0.1284 (0.3230) 0.1330 (0.2825)
$127,518-314,446 15.0 0.1088 (0.2086) 0.1056 (0.1996)
Above $314,446 10.0 0.2117 (0.3279) 0.2134 (0.3524)
SAVE1
Below (-$1,078) 25.0 0.1462 (0.2274) 0.1468 (0.2269)
(-$1,078)-1,725 25.0 0.1855 (0.5537) 0.2013 (0.5836)
$ 1,726- 10,056 25.0 0.1716 (0.4780) 0.1803 (0.4881)
$ 10,057-30,818 15.0 0.0902 (0.1794) 0.0998 (0.1808)
Above $30,818 10.0 0.2374 (0.5512) 0.2290 (0.4940)
SAVE2
Below (-$739) 25.0 0.2281 (0.6129) 0.2299 (0.6357)
(-$739)-1,489 25.0 0.1312 (0.2371) 0.1436 (0.2449)
$1,490-7,125 25.0 0.1363 (0.2579) 0.1451 (0.2738)
$7,126-24,492 15.0 0.0977 (0.3980) 0.1095 (0.4092)
Above $24,492 10.0 0.2391 (0.5530) 0.2330 (0.4963)
NCHILD
No Children 55.1 0.1832 (0.5066) 0.1889 (0.5245)
1 Child 17.2 0.1527 (0.2171) 0.1707 (0.2458)
2 and more 27.7 0.1269 (0.3376) 0.1296 (0.2973)

Note: Standard deviations are given in parentheses. Observations are weighted using the
sample weights (N=1305). VRLI is the variance of residual log income, and VLI is the
variance of log income. All variables are defined in Appendix B.1.

67
Table 3.4: Savings, Income and Income Uncertainty by Age and Fertility

AGE ∆CHILD %HH SAVE1 SAVE2 PERMINC VLRI VLI


Below 31 0 12.1 4,528 4,331 23,628 0.2920 0.3064
1 6.8 10,272 9,220 34,422 0.1507 0.1839
31-40 0 20.5 4,619 3,477 40,769 0.1727 0.1774
1 3.9 11,049 5,390 52,899 0.0841 0.0812
41 and above 0 56.3 8,812 6,884 36,487 0.1369 0.1408
1 0.5 24,636 27,257 43,912 0.2376 0.2261

Note: ∆CHILD=1 if household had a child between 1983-1989 (0 otherwise). VRLI is the variance of residual log income,
and VLI is the variance of log income. All variables are defined in Appendix B.1.

68
Table 3.5: Probit: Fertility Decision of Fecund Households

(1) (2) (3)


Coef StdE Coef StdE Coef StdE
CONSTANT 1.768 0.749** 1.389 0.720* 1.357 0.717*
AGE -0.077 0.021** [-.0099] -0.063 0.020** [-.0095] -0.063 0.020** [-.0094]
VLRI -0.448 0.225** [-.0577] -0.397 0.186** [-.0594]
VLI -0.418 0.203** [-.0625]
PERMINC/1000 0.006 0.001** [0.0008]
TRANSINC/1000 -0.001 0.001* [-.0001]
MEANINC/1000 0.001 0.000* [ 0.0001] 0.001 0.000
MARRIED 1.228 0.392** [ 0.0829] 1.366 0.384** [ 0.1037] 1.368 0.385** [ 0.1041]
WHITE -0.714 0.225** [-0.1297] -0.545 0.208** [-0.1040] -0.543 0.208** [-0.1038]

69
NADULTS 0.073 0.134 0.047 0.131 0.047 0.130
SPFULLT -0.564 0.178** [-0.0675] -0.566 0.175** [-0.0786] -0.565 0.175** [-0.0787]
HOWN83 2.232 0.953** [-0.0920] 1.980 0.886** [-0.0844] 1.998 0.886** [-0.0845]
HAGE -0.068 0.027** -0.060 0.024** -0.060 0.024**
YOUNGCH -1.844 0.623** [-0.2375] -2.236 0.627** [-0.3340] -2.229 0.627** [-0.3338]
MIDDCH -0.331 0.132** [-0.0426] -0.369 0.130** [-0.0550] -0.369 0.130** [-0.0553]
HIGHSCH -0.420 0.166** [-0.0541] -0.409 0.154** [-0.0612] -0.410 0.154** [-0.0613]
YAGE 0.049 0.019** 0.062 0.019** 0.060 0.024**
Likelihood -152.20 -160.00 -160.14
Pseudo R2 0.35 0.31 0.31

Note: Coef reports coefficients and StdE reports standard errors. Marginal effects are given in the brackets. YAGE is
YOUNGCH×AGE, and HAGE is HOWN83×AGE. ** indicates significance at 5 percent level, and * indicates significance
at 10 percent level. Number of observations N=509.
Table 3.6: Regressions of SAVE1 on Income Uncertainty with Endogenous Fertility Decision

(1) (2) (3)


Coef StdE Coef StdE Coef StdE
CONSTANT 19212 19254 15228 19373 15240 19296
VRLI 2250 2252 2258 2365
VRL 2180 2441
NWORTH25 343 3520 -500 3222 -464 3230
NWORTH90 26829 12463 ** 36903 11508 ** 37290 11447 **
VRLI× NWORTH25 -21824 6015 ** -22195 5914 **
VRLI× NWORTH90 -71146 29321 ** -71517 28439 **
VLI× NWORHT25 -21218 5702 **
VLI× NWORHT90 -63955 26900 **

70
AGE -208 294 -131 306 -131 305
\
∆CHILD -134139 93086 -185235 83042 ** -186203 83131 **
\
AGE ×∆CHILD 4256 3304 6412 2952 ** 6474 2950 **
PERMINC 343 102 **
TRANSINC/1000 240 59 **
MEANINC/1000 255 53 ** 257 53 **
NADULTS -7554 3838 ** -7132 3916 ** -7244 3908 **
NCHILD -22542 11934 * -24479 11818 ** -24818 11791 **
AGE× NCHILD 646 320 ** 698 319 ** 707 319 **
BEQUEST83 12654 5751 ** 12916 5780 ** 12952 5786 **
∆ NADULT -7664 5177 -7860 5120 * -7845 5112
R2 .22 .22 .22

Note: Coef reports coefficients and StdE reports standard errors. ** indicates significance at 5 percent level, and * indicates
significance at 10 percent level. Number of observations=1,035.
Table 3.7: Regressions of SAVE2 on Income Uncertainty with Endogenous Fertility Decision

(1) (2) (3)


Coef StdE Coef StdE Coef StdE
CONSTANT 23200 19365 17753 19528 17718 19448
VRLI 1501 2042 1550 2091
VRL 1583 2137
NWORTH25 712 3511 142 3237 182 3244
NWORTH90 24093 12070 ** 32041 11432 ** 32484 11382 **
VRLI× NWORTH25 -20171 6222 ** -20204 6023 **
VRLI× NWORTH90 -61036 29489 ** -60981 28519 **
VLI× NWORTH25 -19347 5780 **
VLI× NWORTH90 -53719 26977 **

71
AGE -257 293 -161 306 -160 305
\
∆CHILD -126939 91894 -175653 82756 ** -177076 82940 **
\
AGE×∆CHILD 3945 3267 6075 2966 ** 6149 2969 **
PERMINC/1000 292 100 **
TRANSINC/1000 210 63 **
MEANINC/1000 221 56 ** 223 57 **
NADULT -8479 3843 ** -8168 3935 ** -8264 3929 **
NCHILD -21818 11811 * -23842 11737 ** -24089 11699
AGE× NCHILD 624 320 * 679 320 ** 686 319 **
BEQUEST83 11545 5760 ** 11712 5766 ** 11736 5772 **
∆NADULT -9888 5374 -10090 5329 * -10074 5321 *
R2 .18 .18 .18

Note: Coef reports coefficients and StdE reports standard errors. ** indicates significance at 5 percent level, and * indicates
significance at 10 percent level. Number of observations=1,035.
Table 3.8: The Effect of a Change in the Fertility Decision on SAVE1

Fecund HH (1) (2) (3)


E(SAVE1|∆CHILD=0) 13,133 12,695 12,672
E(SAVE1|∆CHILD=1) 8,527 10,154 10,375

N 422 422 422

Notes: E(SAVE1|∆CHILD=0) denotes average SAVE1 of the households that did not
have a child between 1983-1989 and E(SAVE1|∆CHILD=1) denotes average SAVE1 of the
households had they chosen to have a child.

72
Chapter 4

Saving for Children’s College Education

4.1 Introduction

The purpose of this chapter is to analyze an important life-cycle saving


motive: saving for children’s college education. Understanding the effect of fi-
nancing children’s college education on household saving behavior is important
at least for three reasons.

First, parents contribute a significant amount to their children’s col-


lege costs. According to the 1996 National Postsecondary Student Aid Survey
(NPSAS), 90 percent of dependent undergraduate’s parents contributed to
their children’s college costs. While the percentage was lower for those in the
lower income group (income below $35,000), for those in the higher income
group (income above $70,000) it was 98 percent (Presley and Clery [47]). Ac-
cording to the 1987 NPSAS, 65 percent of the parents contributed a positive
amount to their children’s college costs as a gift, and the average amount of
their support was about $3,900 (Choy and Henke [14]). Of those contributing
to their children’s college costs, about 65 percent reported using some previ-
ous savings, and 80 percent reported using some current income. Using the
1983-86 SCF, Gale and Scholz [23] estimate that the annual flow of parental

73
contributions totaled about $35 billion. Gale and Scholz [23] convert the flow
of college support to a stock of wealth using steady-state assumptions. Accord-
ing to their estimation, contributions to children’s education yield a wealth of
$1,441.5 billion, which is 12 percent of the aggregate net worth in 1983.

Second, families who save for college reduce their eligibility for financial
aid. The college financial aid system imposes an implicit tax on the savings
of households that are potentially eligible for financial assistance. Edlin [19],
Feldstein [22], Dick and Edlin [16] and Long [43] have recently examined the
adverse effect of the means-tested student aid process on household asset ac-
cumulation. According to Edlin [19] and Feldstein [22], the financial aid tax
rate on capital income can be as high as 50 percent. However, as shown in
Long [43], the results in Edlin [19] and Feldstein [22] depend on a variety of
assumptions such as the number of children enrolled in college, anticipated
college costs and the amount of aid that is received and so on. Dick and
Edlin [16] use data on financial aid awards to calculate a marginal tax rate
and find that families with children attending average-priced colleges face a
financial aid tax ranging from 2 percent to 16 percent. Using alternate but
also plausible assumptions, Long [43] finds that the effect of the financial aid
tax on asset holdings is smaller than the effect in the prior literature. To
date, the focus has been on calculating the financial aid tax and measuring its
negative impact on household asset accumulation. Using the data on actual
expenditures on children’s college education, this chapter examines the effect
of anticipated educational expenses on household savings.

74
Third, the quality-quantity model of fertility behavior assumes that
parents have preferences both for the expenditure per child and the number
of children. The analyses in Willis [55] and Becker and Lewis [2] show that
parents with few children have substituted quality for quantity. In the empiri-
cal investigation of this model, different forms of parental expenditure such as
children’s schooling, child care and bequests have been used as the qualitative
measure. The estimates in Tomes [54] show that family size and children’s
schooling are jointly determined. Behrman et al. [3] develop a model relating
children’s schooling to family size, with and without equal access to financing
for education, and test predictions of their model using the veterans sample
of white male twins and the sample of their adult offspring. Without unequal
access to schooling, they find an inverse relationship between family size and
children’s schooling. Tomes [53] empirically tests whether parental bequests of
wealth and human capital investments represent substitute forms of parental
transfer. The results of his model confirm that investments in children’s hu-
man capital, which are measured by children’s income and years of schooling,
are negatively related to subsequent levels of inheritance. The estimates in
Tomes [53] confirm the prediction of the quantity-quality model that bequests
and children’s income are negatively related to family size. Using the National
Longitudinal Survey of the High School Class of 1972, Steelman and Pow-
ell [51] investigate the relationship between the structure of the sibling group
and parental financial support for children’s college education. Specifically,
they analyze the influence of size and ordinal position of siblings on the like-

75
lihood and amount of parental support. Their results show that the number
of siblings significantly decreases both the likelihood and amount of parental
contribution to children’s college education. Moreover, ordinal position alters
parental support in favor of later-born children. This chapter also uses the
amount of parental expenditure on children’s college education as a measure
of child quality.1 Given the rapidly rising cost of college tuition, an analysis
of financing college education and family size highlights an important aspect
of the quality-quantity model.

A number of studies have analyzed motives for saving such as saving


for retirement, saving for ‘rainy days’ and saving for bequests and inter vivos
transfers.2 In addition, using Japanese household data, Horioka and Watan-
abe [29] analyze the amount of gross saving and dissaving for each of twelve
motives including saving for retirement, bequests, emergencies, education and
so on. The results of their analysis show that retirement and precautionary
motives account for 25.7 and 28.1 percent of gross saving, respectively. Saving
for children’s education is the third most important saving motive after saving
for retirement and ‘rainy days’ and accounts for 9.2 percent of gross saving.
Their findings also show that the importance of each saving motive depends
on the age and the life-cycle stage of the household.

1
The data set used in the chapter does not provide information on the ordinal position
of the child attending college. Steelman and Powell [51] argue that later-born children are
more favored relative to earlier-born ones due to the family life cycle. Parents have more
resources when later-born children reach college age. It would be of interest to investigate
this effect on the level of parental support using the information on household savings.
2
See Browning and Lusardi [6] for a survey of the literature.

76
Although saving for retirement, income fluctuations and bequests have
motivated substantial research, the motive of saving for children’s education
has not been much investigated. One exception is Souleles [49]. Using the
Consumer Expenditure Survey, Souleles [49] examines consumption of house-
holds as they pay for the college expenses of their children. His findings are
consistent with the life-cycle theory of consumption and saving. His results
show that households smooth their consumption into the academic year and do
not cut their consumption in the 6-9 months before the academic year starts.

The SCF contains a question that asks the household’s most impor-
tant reason for saving. Table 4.1 shows the percentage of households in the
1983 survey citing retirement, ‘rainy days’ (emergencies and unemployment),
education of children, buying a home and other reasons as the most impor-
tant reason for saving. Other reasons for saving include saving for ordinary
living expenses, medical and dental expenses, and buying durable household
goods, taking vacations and so on. The last column of Table 4.1 shows the
percentage of households reporting that they cannot or do not save. The table
provides the responses of the sample used in this chapter. The sample includes
households with nonretired heads and spouses (The SCF and restrictions on
the sample are discussed in Section 4.3). Among the households saving for
retirement, ‘rainy days,’ home purchase and children’s education, saving for
‘rainy days,’ is the most cited reason. While 35.5 percent of households list
‘rainy days’ as the most important reason for saving, 15.3 percent list retire-
ment and 5.3 percent list education as the most important reason for saving.

77
The percentage of households saving for ‘rainy days,’ retirement and other rea-
sons show a systematic trend relative to the total number of children. As the
number of children increases, the percentage of households citing ‘rainy days’
and other reasons as the most important reason decreases, and the percentage
of households citing retirement as the most important reason increases. An
almost equal number of households with one or two children and with three
or more children report saving for children’s education as the most important
reason.

Table 4.1 also shows the percentage of households citing each saving mo-
tive by the number of children and net worth in 1983. Among the households
in the bottom 25 percentile of the wealth distribution, a higher percentage
of the households with one or two children report saving for children’s edu-
cation than those with three or more children (8.3 percent vs. 5.8 percent).
Among the households in the higher wealth groups, the effect of the number of
children on the percentage of households reporting saving for children’s edu-
cation disappears. The percentage of households in the 25-75 percentile of the
wealth distribution citing ‘rainy days’ as the most important saving motive
is higher than the percentage of households in the lower and higher wealth
groups. Controlling for the number of children, the percentage of households
saving for retirement increases with wealth. For example, among households
with 1 or 2 children, while only 2.7 percent of those in bottom 25 percentile
of wealth distribution report saving for retirement, 24.4 percent in the top 25
percentile report saving for retirement. This table shows that the number of

78
children has a significant effect on saving motives. We continue to observe this
effect even after controlling for the household wealth.

In this chapter, I introduce life-cycle savings into the quality and quan-
tity model of fertility and derive predictions concerning the effect of expected
educational expenditures on household savings. I also obtain predictions con-
cerning the simultaneous determination of family size and college expenditures
per child. The data from the 1983-86 SCF is used to estimate two equations
in which the dependent variables are household savings and educational ex-
penses. Using the actual college expenses reported in the SCF, the effect of
anticipated educational expenses on household savings are estimated.

The results show that an increase in the number of children decreases


the per child college expenditures paid by households by approximately by
$317 in 1986 dollars. Further, households save for their children’s college ex-
penditures, and the amount of savings increases with the age of the household
head. Other things constant, a household with a 43 year old head expecting
to have $2,000 in children’s college expenses saves $8,000 more than it would
had it not expected to have any college expenses. The results are consistent
with the predictions of the Life-Cycle Theory of saving and consumption that
households save in advance for children’s college expenditures. Also, the empir-
ical findings provide an answer to why saving is concentrated among wealthier
households. Households with higher income and wealth expect to have higher
educational expenses, and they save in advance for these expenses.

The remainder of this chapter is organized as follows. Section 4.2 an-

79
alyzes a model of the quality-quantity interaction of fertility with household
savings. Section 4.3 describes the 1983-86 SCF. Section 4.4 provides a frame-
work for the empirical analysis of the interaction between savings and college
expenditures. Section 4.5 estimates the determinants of college expenditures
and uses these estimates to investigate the effect of expected college expendi-
tures on household savings. Finally, a summary and conclusions are presented
in Section 4.6.

4.2 A Model of Saving for College

This section considers a world in which individuals (parents) live for


two periods. In the first period, a couple earns y1 , chooses to have n children,
and the family consumes together c1 and saves A to earn interest at the rate
of r. In the second period, the return on accumulated assets (1 + r)A and
second period wage income y2 are divided between consumption c2 and paying
for children’s college education e. For simplicity, per capita college investment
is assumed to be equal for all n children.

Parents choose first-period consumption, second-period consumption,


investment to each child’s education, and the number of children to maximize

U = U (c1 , c2 , e, n) (4.1)

subject to

c1 = y1 − A (4.2)

c2 = y2 + (1 + r)A − πen (4.3)

80
where π is the price of education. Substituting (4.2) and (4.3) into (4.1) yields
the following unconstrained maximization problem:

U = U (y1 − A, y2 + (1 + r)A − πen, e, n) (4.4)

where the three choice variables are accumulated assets (A), educational ex-
penses (e) and the number of children (n). The first-order conditions are

−U1 + (1 + r)U2 = 0 (4.5)

−πnU2 + Ue = 0 (4.6)

−πeU2 + Un = 0 (4.7)

where U1 and U2 are the marginal utility of consumption in the first and second
periods, respectively; Ue is the marginal utility of children’s education, and Un
is the marginal utility of family size.

If the utility function is CES with equal elasticity of substitution be-


tween all arguments, equation (4.5) can be written as follows:

y1 − A 1
= (1 + r) γ−1 . (4.8)
y2 + (1 + r)A − πen
where γ is the elasticity of substitution. This expression implies that an in-
crease in educational expenses decreases the second-period consumption rel-
ative to first-period consumption. Since the right-hand side is a constant, a
decrease in second-period consumption is likely to decrease the first-period
consumption, which results in an increase in accumulated assets.

81
To extend the analysis to account for uncertainty, let us assume that the
second period wage income y2 is stochastic. Solving the consumer’s problem
yields the following equation

−U1 + (1 + r)E1 [U2 ] = 0

where E1 represents the expected marginal utility of consumption in the second


period conditional on all information available in first period. This condition
shows that greater uncertainty is linked to greater saving in the first period
when the third derivative of the utility function is positive.3 The combination
of a positive third derivative of the utility function and uncertainty about fu-
ture income reduces consumption in the first period. When uncertainty about
future income is assumed, household saving can be associated with two differ-
ent saving motives: saving for uncertainty about future income (precautionary
saving) and saving for children’s education. The empirical specification of the
model described below controls for precautionary saving while it estimates the
effect of educational expenses on household savings.

The interaction of the quality and quantity dimensions of choice is


reflected in the fact that the marginal costs of education and family size depend
on the level of each other in equations (4.6) and (4.7). This interdependence
implies an inverse relationship between the number of children and educational
expenses.

3
If the third derivative is positive, then U2 is a convex function. In this case, E1 [U2 ]
exceeds U2 [E1 ].

82
4.3 Data

The empirical analysis uses data from the 1983-1986 SCF. The 1983
survey contains interviews from a random sample of 3,824 households and a
high-income supplement of 438 households. In 1986, 2,822 of these households
were reinterviewed. The SCF contains detailed information on household as-
sets, liabilities, income and demographic characteristics. The variables used in
the empirical analysis are classified into four groups: fertility, college expendi-
ture, savings and other control variables.

The fertility variable (CHILD) is the number of children of either the


respondent or spouse, including those not living in the household. This variable
includes children of previous marriages living with former spouses.4

The college expenditure variable (COLLEXP) captures the quality di-


mension associated with the expenditure per child. In 1986, respondents were
asked if they had any children attending college from 1983-85 and if they had
any college expenses on the behalf of their children. The respondents were
also asked how many years of college their children completed from 1983-85.
I use the number of children attending college and the number of years they
attended to normalize college expenditures. The college expenditure variable
is the outlay of college education per child.

Household savings are measured in two ways. The first measure, SAVE1,

4
Unfortunately, the data does not differentiate between children away in college or living
on their own and with former spouses.

83
is the change in net worth between 1983 and 1986 divided by the number of
years. Net worth (NWORTH) is the total value of household’s assets minus
its total liabilities.5 SAVE1 includes the realized and unrealized capital gains.
In order to exclude unrealized capital gains on the primary residence, a second
measure of savings (SAVE2) is used. Whenever a household did not buy or
sell a house that was the family’s primary residence, the value of the primary
residence in 1983 is kept constant.

Other controls include variables that affect savings, educational expen-


ditures and fertility decisions. These are age, race, marital status, gender and
the educational level of the household head, and the educational level of the
spouse, and the demographic characteristics associated with tastes (urban res-
idence, reasons for borrowing and saving, and whether or not the household
head is willing to undertake risky investments).

Using the reported household income for 1982, 1983, 1984 and 1985, I
estimate household permanent and transitory income. Household permanent
income (PERINC) is defined as the predicted income in 1985 obtained from
regressing the log of total income on age, education, race, gender of the house-
hold head and other household characteristics. Transitory income (TRINC)

5
Total assets is the sum of financial assets and nonfinancial assets, where the first in-
cludes liquid assets (checking assets, money market deposit accounts, saving accounts, call
accounts, certificates of deposit and saving accounts), stocks, bonds, mutual funds, individ-
ual retirement accounts, Keogh accounts, trusts, cash value of life insurance and the later
includes residential property, other real estate, business equity, vehicles and other real assets
like art and precious metals. Total liabilities include mortgage debt, home equity, loans,
other loans for property, credit card debt, automobile loans, balances outstanding on lines
of credit and loans on consumer durables.

84
is the difference between reported income in 1985 and estimated permanent
income. Appendix C.1 gives a detailed definition of the variables used in the
estimation of the model.

The sample is restricted to families that did not change composition


from 1983-86. Estimating the relationship between savings and educational
expenditures is complicated for families who experienced a major change in
composition such as marriage and divorce. The sample is also constrained to
include only the households with nonretired household heads and their spouses
if the head is married. Households with retired household heads are assumed
to be in the life-cycle stage of dissaving. Also, households with family income
above $100,000 are excluded to avoid the difficulty of modeling the relationship
between educational expenditures and savings. These restrictions leave us with
a sample containing 1,690 households.

Table 4.2 presents summary statistics of the variables used. The aver-
age household in the sample is headed by a forty two year old married high
school graduate and includes two children. The average household net worth
in 1983 is $81,575 in 1986 dollars. Permanent and transitory incomes in 1985
are $25,931 and $3,296, respectively. According to the first measure of sav-
ings, the typical household saves $5,806 and according to second measure, it
saves $4,811. The head of the median household reports that it is all right to
borrow money for educational expenses. For households with nonzero college
expenditures, the average expenditure is $2,005.

Table 4.3 presents average household savings and college expenses by

85
the number of children attending college. As the number of children in college
increases, college expenditures per child decreases. However, household savings
increase with the number of children in college. Table 4.3 also breaks down
savings and college expenses by the number of children in college and net
worth in 1983. The data show that college expenses increase with net worth.
Households in all three wealth groups (bottom 25 percentile, 25-75 percentile
and top 25 percentile) spend less per child as the number of children attending
college increases. Savings of households with children in college increase with
net worth. Except the households in the bottom 25 percentile of the net worth
distribution with two or more children attending college, household savings
increase with the number of children in college. Households in the top 25
percentile of the wealth distribution with two or more children in college save
on average $18,077, more than twice as much the households in the same wealth
group with one child in college. The data show that households continue to
save while children are in college. However, the number of children in college
is inversely related to the expenditure per child as predicted by the quantity-
quality model.

4.4 Empirical Specification

The simultaneous relation between educational expenses (e∗ ) and house-


hold savings (a) is specified as follows:

e∗i = δ1 ni + x01i κ1 + u1i (4.9)

ai = η2 e∗i + x02i κ2 + u2i (4.10)

86
where x1i and x2i are the vectors of exogenous variables, and κ1 and κ2 are the
vectors of parameters to be estimated. The model predicts that an increase
in the number of children decreases the anticipated and actual educational
expenditure, i.e. δ1 < 0. Another prediction of the model is that educational
expenses increases household savings, i.e. η2 > 0. The structural disturbances
ui = (u1i , u2i ) are assumed to be randomly drawn from a 2-variable distribution
P
with E(ui ) = 0 and E(ui u0i ) = .

The theoretical model derives predictions concerning the effect of the


completed lifetime fertility on the educational expenses. However, the data
on household fertility gives the number of children ever born to a household
headed by a person of a certain age. Therefore, the completed fertility, n, is
obtained to estimate the expected educational expenses. The expected com-
pleted family size is given by

E[n] = exp(x03i κ3 + φi ) (4.11)

where x3i is a vector of demographic characteristics, κ3 is a vector of param-


eters, and φi is an age specific factor. I use estimates of the parameters of a
Poisson regression model to construct the completed fertility profile when the
household head is 55 years old.6

Information on educational expenses is available only if the household


has a child attending college and if the household spends a positive amount

6
The model is also estimated using the predicted number of children obtained from the
Poisson model. The empirical results hold true for this measure of fertility too.

87
on financing her education. Let gi = 1 indicate that the household has a child
attending college, gi = 0 indicate that none of the children are attending col-
lege. Then e∗i is observed to be ei if e∗i > 0 and gi = 1. I obtain expected
educational expenses as follows. First, I estimate a Tobit model for the edu-
cational expenditures of the households with children attending college. Then
I use the estimates of those parameters to construct the profile of anticipated
educational expenses.

4.5 Estimation and Results

Table 4.4 contains estimates of the Poisson regression model of the


fertility equation. The right-hand variables include household demographics
expected to affect the number of children: namely, age (AGE), martial status
(MARRIED), gender (FEMALE), race (BLACK) and the education of the
household head (HIGHSCH and COLLEG), permanent income (PERINC),
a dummy indicating whether the spouse works for a full time job in 1983
(FSPOUSE), and education of the spouse (HIGHSCHSP and COLLEGSP),
and a dummy indicating whether the household does not live in a SMSA area
(NSMSA).

Estimates of the coefficients in Table 4.4 are consistent with previous


studies. An increase in the permanent income increases the number of chil-
dren. Controlling for permanent income, households headed by high school
and college graduates have fewer children than those headed by persons with-
out a high school degree. Married households have more children. However,

88
controlling for marital status, households with spouses working full-time and
with high school and college degrees have fewer children. I use the estimates of
the regression to predict the completed household fertility when the household
\
head is 55 years old (CHILD). \ is 3.03.
The average CHILD

Of the 1,690 households in the sample, 338 had a child attending col-
lege between 1983-86, and 252 reported contributing a positive amount to
their children’s college expenses. Households with children attending college
between 1983-86 are included in the estimation of the Tobit regression. Ta-
ble 4.5 reports estimates of the equation (4.9). Columns 1 and 2 contain the
\ and columns 3 and 4 contain the results with CHILD.
results with CHILD
The right-hand variables also include other factors that might affect the college
expenditures; namely, age (AGE) and education (COLLEG) of the household
head, and permanent and transitory income (PERINC and TRINC), a dummy
indicating whether or not the household head believes it is all right to borrow
money for college expenses of children (BEDUCAT) and dummies indicating
the most important reason for saving (SRETIRE, SEMERG, SCHEDUC and
SHOME).

As predicted by the quantity-quality model, the amount of college ex-


penditure decreases with the number of children. The instrumental estimate
of the coefficient on the number of children is almost three times as large as the
OLS estimate (-$459 vs. -$187). The partial derivative of the expected college
expenditure with respect to the number of children is calculated at the mean
\ and other explanatory
values of the estimated number of children (CHILD)

89
variables as follows:
δ̂1 ∗ Φ((δ̂1 ni + x01i κ̂1 )/σ̂), (4.12)

where σ̂ is the estimate of the standard error, Φ is the standard normal cumu-
lative distribution, and δ̂1 and κ̂1 are the estimates of δ1 and κ1 , respectively.
Using the approximation, an additional child results in a drop of $317 in ex-
pected college expenditures at the mean of values.

Estimates in Table 4.5 show that increases in permanent and transitory


income increase the level of expenditures for educational expenses. Households
with heads who believe that it is all right to borrow for educational expenses
have higher expenditures. Households citing saving for children’s education
as the most important reason for saving spend more than other households.
Households citing saving for retirement and buying a home as the most im-
portant reason for saving spend less on children’s education. I use estimates
\ to calculate
of the Tobit model and the expected completed fertility, CHILD,
\ ). The average COLLEXP
the expected college expenditures (COLLEXP \ is
$1,789 per child.

Table 4.6 presents actual and estimated college expenses by household


net worth in 1983. The amount of contribution to children’s college education
increases with wealth. While the average contribution of households in the
bottom 25 percentile of wealth distribution is $1,436, the average contribution
of the top 25 percentile is $3,093 per child. Estimated contributions of the
households with children in college are very close to the actual expenses. The
last two columns of Table 4.6 show SAVE1 and SAVE2 for households with

90
and without children in college. Households in the bottom 25 percentile of
wealth distribution save significantly less than those without children in col-
lege. However, households with greater wealth save more if they have a child
attending college. Interestingly, households in top 25 percentile of the wealth
distribution save almost five times more if they have a child in college. Similar
to Table 4.3, the data in Table 4.6 show that wealthier families contribute
more to their children’s education and continue to save while their children
are in college.

Table 4.7 presents the effect of expected college expenditures on house-


hold savings. Columns 1 and 2 contain the estimates for SAVE1, and columns
3 and 4 contain the estimates for SAVE2. Explanatory variables include age
(AGE), gender (FEMALE) of the household head, permanent and transitory
incomes (PERINC and TRINC), nonurban residence (NSMSA) and dummies
indicating household net worth in 1983 (NWORTH25 and NWORTH75), a
dummy indicating whether or not the household had a windfall greater than
$3,000 between 1893-86, two other reasons for saving, which are retirement
and emergencies (SRETIRE and SEMERG), a dummy indicating whether the
household head is willing to take risky investments (RISKY), and the number
of children attending college between 1983-86 (NCHCOLL).

The estimates of SAVE1 and SAVE2 are very similar. In estimates of


\ )
both equations, the coefficient of expected college expenditure (COLLEXP
\ )
is negative and the coefficient of age interaction term (AGE×COLLEXP
is positive, indicating that an increase in expected college expenditure raises

91
savings after age 28. Permanent income increases both SAVE1 and SAVE2.
The effect of transitory income on both measures of savings is positive and
significant, showing that households save approximately 39 percent of their
transitory income. Households citing saving for retirement as the most impor-
tant reason save more. However, saving for emergencies does not significantly
affect savings. Also, households in the bottom 25 percentile of the wealth dis-
tribution save $2,493 more and households in the top 25 percentile save $11,398
less than those in the middle of the wealth distribution . Finally, households
with heads who are willing to undertake risky investments save $7,833 more
than other households. Finally, the number of children attending college does
not significantly decrease household savings. This result does not necessarily
mean that households are not saving for children’s college education. The sav-
ing behavior of a household with a child in the first year of college in 1983 can
be quite different from a household with a child finishing up college in 1983.
Unfortunately, the data does not have detailed information on the years that
children were attending college between 1983 and 1986.

Figure 4.1 shows the effect of the age of the household head on SAVE1.
Using the estimates in Table 4.7, household savings are calculated in five year
intervals. For each age group, the figure first calculates savings of a typical
household expecting to contribute $2,000 to college expenses and compares
it to what it would have saved, had it not expected to contribute a positive
amount. By typical, I mean a household in the 25-75 percentile of the wealth
distribution, citing a motive other than retirement or emergencies as the most

92
important saving motive, headed by a male, who is not willing to undertake
risky investments and did not receive a windfall greater than $3,000 between
1983-86. The household is assumed to have average permanent and transitory
incomes for their age group. The results show that savings of the household
with an anticipated $2,000 college expenses increase with age. The effect
of expecting to contribute $2000 on household savings is $8,000 at the age
of 43. If the household does not expect to contribute to children’s college
expenses, savings decline to zero at the age of 43. This striking result is
due to the assumption that this household is assumed not to cite saving for
retirement as the most important reason. The results in Table 4.7 show that
the effect of saving for retirement on household saving is positive and raises
household savings by $4,894. Saving motives change with age and household
composition. If we assume, for example, that this household starts saving for
retirement when the household head is 43 years old, this will increase its saving
by $4,894. This figure only shows that controlling for other factors, the effect
of anticipated college expenses on savings is positive and significant, and it
increases with the age of the household head.

4.6 Conclusion

This chapter examines the effect of saving for children’s college edu-
cation on household savings. I introduce life-cycle savings into the quality
and quantity model of fertility and derive predictions concerning the effect
of educational expenditures on household savings. I also obtain predictions

93
concerning the simultaneous determination of family size and college expendi-
ture per child. Using the actual college expenditures reported in the 1983-86
Survey of Consumer Finances, I estimate expected expenditures on children’s
college education. The model uses the expected expenditures and other con-
trol variables that affect savings to estimate an equation of savings. I analyze
the effect of educational expenditures on two different measures of savings,
which are the change in net worth between 1983 and 1986, and the change in
net worth excluding the capital gains on primary residence.

The main finding of this chapter is that households save in advance for
children’s college expenditures. The amount of savings for college expenses
increases with the age of the household head. Other things constant, the
difference between savings of households with and without college expenses
can be as high as $8,000 at the age of 43.

The results are consistent with the predictions of the life-cycle theory
of saving and consumption that households save in advance for expected ex-
penses. The results are also consistent with the findings in Souleles [49], which
show that despite large college expenses, households smooth consumption into
the academic year and do not cut consumption in the 6-9 months before the
academic year starts. By focusing on household savings, this present chapter
examines the effect of college expenditures over the life-cycle and finds that
most of the saving done by wealthier households can be attributed to saving
to finance their children’s college expenses.

94
Table 4.1: Saving Motives By the Number of Children

SRETIRE SEMERG SCHEDU SHOME SOTHER NOSAVE


HH 0.153 0.355 0.053 0.035 0.380 0.023

CHILD
0 0.074 0.404 0.002 0.049 0.455 0.016
1-2 0.120 0.377 0.066 0.042 0.381 0.013
3 or more 0.223 0.310 0.064 0.021 0.345 0.037

NWORTH
0-25p 0 0.049 0.337 0.000 0.043 0.543 0.029
1-2 0.027 0.383 0.083 0.079 0.401 0.027

95
3 or more 0.053 0.278 0.058 0.060 0.469 0.082

25-75p 0 0.087 0.471 0.004 0.057 0.382 0.000


1-2 0.102 0.397 0.057 0.042 0.391 0.010
3 or more 0.212 0.351 0.063 0.009 0.341 0.023

75 to 100p 0 0.115 0.369 0.000 0.033 0.440 0.043


1-2 0.244 0.330 0.070 0.009 0.339 0.007
3 or more 0.341 0.264 0.068 0.016 0.279 0.032

Source: Survey of Consumer Finances, 1983.


Notes: This table reports the proportion of households citing the selected motives as the most important reason for saving.
SRETIRE: saving for retirement, SEMERGE: saving for ‘rainy days,’ SCHEDU: saving for the education of children, SHOME:
saving to buy a home, SOTHER: saving for other reasons and NOSAVE: cannot/does not save. Tabulations are weighted
using the sample weights. The number of observations N=1690.
Table 4.2: Descriptive Summary of Variables

Variables Mean Std. Deviation


CHILD 2.47 2.14
COLLEXP> 0 2005.32 2817.36
SAVE1 5806.13 35397.47
SAVE2 4811.41 33402.64
NWORTH 81575.08 161860.86
PERINC 25931.11 12491.75
TRINC 3296.32 15127.68
AGE 42.43 14.10
FEMALE 0.24 0.43
COLLEG 0.24 0.43
HIGHSCH 0.53 0.50
BLACK 0.13 0.33
MARRIED 0.65 0.48
BEDUCAT 0.86 0.35
WINDF 0.08 0.28
RISKY 0.17 0.37
NSMSA 0.28 0.45

Source: Survey of Consumer Finances, 1983-86.


Notes: Tabulations are weighted using sample weights. All dollar values are reported in 1986
dollars. The number of observations N=1690. All variables are described in Appendix C.1.

96
Table 4.3: Savings and College Expenses by the Number of Children in College

SAVE1 SAVE2 COLLEXP


NCHCOLL
0 5041 4206 0
1 6661 5277 2236
2 or more 12357 10317 1657

NWORTH
0-25p 0 3762 3705 0
1 4961 5005 905
2 or more 1829 1695 797

25-75p 0 5804 4348 0


1 6032 5305 1882
2 or more 7707 5695 951

75-100p 0 4870 4551 0


1 7745 5312 2937
2 or more 18077 15577 2355

Source: Survey of Consumer Finances, 1983-86.


Notes: NCHCOLL shows the number of children attending college between 1983-86. Tab-
ulations are weighted using sample weights. All dollar values are reported in 1986 dollars.
The number of observations N=1690.

97
Table 4.4: Poisson Regression: Number of Children

Coefficient Std. Error


CONSTANT -1.408 0.134 **
AGE 0.022 0.145 **
FEMALE 0.930 0.126 **
FSPOUSE -0.106 0.039 **
HIGHSCH -0.435 0.057 **
COLLEG -0.964 0.097 **
BLACK 0.461 0.062 **
MARRIED 1.078 0.128 **
NONSMSA 0.032 0.038
PERINC/1000 0.038 0.004 **
HIGHSCHSP -0.717 0.090 **
COLLEGSP -0.389 0.062 **

N OBS 1690
Mean of dependent variable 2.43
Log L -2971.48
R2 0.293

Source: Survey of Consumer Finances, 1983-86.


Notes: ** indicates significance at 5 percent level, and * indicates significance at 10 percent
level. Variables are described in Appendix C.1.

98
Table 4.5: Tobit Estimates of College Expenditure Equation

Coefficient Std. Error Coefficient Std. Error


CONSTANT -601.4 1589.5 -1335.6 1452.8
AGE -7.4 23.0 0.5 22.9
\
CHILD -458.8 232.9 **
CHILD -187.4 92.2 **
PERINC/1000 88.2 19.9 ** 67.5 18.1 **
BEDUCAT 1063.5 551.7 * 997.1 550.6 *
SRETIRE -1124.8 473.8 ** -1103.5 472.6 **
SEMERG -407.0 409.8 -372.6 408.7
SCHEDUC 1467.1 611.7 ** 1391.3 609.1 **
SHOME -4936.3 2444.9 ** -4903.2 2524.9 *
COLLEG -45.4 555.9 557.8 468.1
TRINC/1000 72.4 10.0 ** 72.0 9.9 **
SIGMA 2912.3 134.0 ** 2905.6 133.7 **
N OBS 338
Proportion of
+ observations .746
Log L -2429.81 -2429.70

Source: Survey of Consumer Finances, 1983-86.


Notes: ** indicates significance at 5 percent level, and * indicates significance at 10 percent
level. All variables are described in Appendix C.1.

99
Table 4.6: College Expenditures and Savings by the Number of Children in College

NWORTH CHCOLL %HH COLLEXP \


COLLEXP SAVE1 SAVE2
0-25p 0 23.65 0 1278 3800 3732
1 1.29 1436 1445 2802 2989
25-75p 0 44.52 0 1614 5570 4256
1 5.59 2334 1960 8864 6712
75-100p 0 17.02 0 2219 3731 3599
1 7.93 3093 3064 15904 12709

Source: Survey of Consumer Finances, 1983-86.


Notes: Tabulations are weighted using sample weights. All dollar values are reported in 1986 dollars. CHCOLL=1 if the
household has a child attending college between 1983-86 (0 otherwise). The number of observations N=1690.

100
Table 4.7: Effect of Anticipated College Expenses on Savings

SAVE1 SAVE2
Coefficient Std. Error Coefficient Std. Error
CONSTANT 22618.9 11080.5 ** 22327.6 10944.2 **
AGE -1310.3 416.3 ** -1284.6 407.3 **
AGE2 10.2 3.6 ** 10.2 3.4 **
PERINC/1000 388.8 163.8 ** 336.8 160.3 **
TRINC/1000 382.8 140.2 ** 309.9 139.2 **
a
\
COLLEGEXP -10.1 4.5 ** -9.8 4.7 **
a
\
AGE×COLLEGEXP 0.3 0.1 ** 0.3 0.1 **
SRETIRE 4894.4 2538.0 * 4679.4 2434.0 *

101
SEMERG -790.2 1820.5 382.9 1728.3
NWORTH25 2493.5 1325.0 * 3028.0 1271.6 **
NWORTH75 -11398.3 2954.0 ** -9329.1 2796.2 **
WINDF 5607.2 3422.0 5041.8 3349.8
RISKY 7833.8 3199.8 ** 6820.1 3139.0 **
FEMALE 2783.9 2474.6 2460.5 2409.8
NSMSA -315.6 2026.3 761.3 2016.8
NCHCOLL -774.0 1642.5 -615.6 1516.8
R2 .107 .084

Source: Survey of Consumer Finances, 1983-86.


Notes: ** indicates significance at 5 percent level, and * indicates significance at 10 percent level. All variables are described
in Appendix C.1.
a
Predicted value of the variable from Tobit regression of educational expenditures.
Figure 4.1: The Importance of Educational Expenses on Savings

SAVE16

10, 045− ?
?
?
?


3, 736− • ?

?


?| | | | | | |
23 28 33 38 43 48 53 Age
−1, 068− •

−5, 791− •

? savings of a household with $2000 college expenses.


• savings of a household with no college expenses.

102
Appendices

103
Appendix A

Appendix for Chapter 2

A.1 Estimating Marginal Tax Rates

The marginal tax rate of each household is computed using the tax
Form 1040 and the information on sources of income. The SCF collects in-
formation on many components of total income, including wage and salaries,
taxable interest, tax-exempt interest, dividends, alimony received, rents, roy-
alties, business income and farm income. Components of income such as other
gains and IRA distributions that are not reported in the SCF are set to zero.
The sum of household income from all sources gives the adjusted gross income
(AGI).

In determining filing status and personal exemptions, I use the infor-


mation on marital status, number of dependents, and age of the household
head and the spouse. All married couples are assumed to file a joint return.
The SCF does not contain information on some possible deductions such as
medical expenses, state and local income taxes, job expenses and moving ex-
penses. Thus, households are assumed to claim standard deductions instead
of itemizing deductions.

Subtracting the standard deduction and exemptions from the AGI

104
yields the taxable income. I then apply the appropriate tax rate schedule
to calculate the household’s tax liability. The marginal tax rate is computed
by running this method twice - once with AGI and then with AGI minus 100.
The difference in total tax liabilities divided by 100 gives the marginal tax
rate.

A.2 Definition of Variables

Name Description
INCOME Estimated earnings of the household head and spouse at
the age of 45. See Appendix A.3.
ASSET Total assets of the household.
MTR Marginal tax rate of the household. See Appendix A.1.
Eh Consumption demand for housing. For homeowners, it is
the opportunity cost of owning a house.
AGE Age of the household head in years.
MARRIED =1 if the household head is married.
FEMALE =1 if the household head is a single female.
NCHILD Number of children younger than age 22 who live in
the household.
CHILD0 =1 if no children are living in the household.
CHILD1 =1 if only one child is living in the household.
CHILD2 =1 if two children are living in the household.
CHILD3 =1 if three or more children are living in the household.

105
CHAGE13 =1 if the youngest child is older than age 13.
HOMEOWN =1 the household is a homeowner.
WHITE =1 if the household head is white.
RISKY =1 if the household head reports that he is willing to take
risky investments.
YEAR92 =1 if the household is included in the 1992 survey.
YEAR95 =1 if the household is included in the 1995 survey.
YEAR98 =1 if the household is included in the 1998 survey.

A.3 Estimating Permanent Income

The measure of permanent income is constructed using the method


outlined in King and Dicks-Mireaux [40]. This measure is defined as predicted
earnings at the age of 45 plus an individual-specific effect.

The permanent income Y for individual i is defined as

Ln Yi = Zi βp + εpi − c(AGEi ), (A.1)

where Zi is a vector of observable characteristics, βp is the parameter vector,


εpi is an unobservable variable measuring characteristics such as ability (εpi
has zero mean and variance of σs2 ), and c(AGEi ) is a cohort effect.

Observed earnings are assumed to differ from permanent income in two


ways. The first is due to the movements along the age-earnings profile over
the life cycle, and the second is transitory changes in earnings. Thus, earnings

106
in year t are
Ln Eit = Ln Yi + e(AGEit − 45) + uit , (A.2)

where e(.) measures the log of the age-earnings profile; AGEit is the age of the
respondent and uit is the log of the transitory component (uit has zero mean
and variance of σu2 , and is assumed to be uncorrelated with εpi ).

To construct an estimate of permanent income, I need the estimates of


βp , εpi and c(AGEi ). I combine (A.1) and (A.2) and estimate the resulting
earnings equation using each wave of SCF separately, and this provides the
estimate βˆp . Since age-earnings profile e(AGEit − 45) and c(AGEi ) cannot be
identified for this estimation, King and Dicks-Mireaux [40] use outside data
to impose a cohort effect. Instead, I assume that the cohort effect, ĉ(AGEi ),
is zero. Finally, to get an estimate of εpi , I calculate the minimum variance
estimator of εpi using
ε̂pi = α(εpi + uit ), (A.3)

where α = σs2 /(σs2 + σu2 ). Following King and Dicks-Mireaux [40], I assume
that α = 0.5.

Earnings equations are estimated separately for household heads and


spouses. The selectivity-adjusted earnings functions are estimated for the
sample consisting of individuals with nonzero earnings. For heads with zero
earnings, permanent income is calculated from Zi β̂p . The same procedure is
used for spouses, with one exception. Their permanent income is adjusted for

107
non-participation at different stages of the life cycle as follows:

Yiw = Ŷi P rob(Ei > 0),

where Ŷi is the permanent income estimate, and the probability of nonzero
earnings is computed for each spouse from the probit estimates. Household
permanent income is the sum of the estimates of permanent income for the
head and spouse.

108
Appendix B

Appendix for Chapter 3

B.1 Definition of Variables

Name Description
SAVE1 First difference in net worth between 1983-89 divided
by 6.
SAVE2 First difference in net worth between 1983-89 controlling
for capital gains in home prices divided by 6.
VRLI Variance of residual of log income from the earnings
equation.
VRI Variance of log income.
MEANINC Mean of reported income between 1982-88.
PERMINC Mean of the predicted income from the earnings regression.
TRANS Mean of the residual income from the earnings regression.
NWORTH Household net worth in 1983.
NWORTH25 =1 if the household in the bottom 25 percent of the net
worth distribution in 1983.
NWORTH90 =1 if the household in the 10 percent of the net worth
distribution in 1983.
AGE Age of the household head in 1983.

109
EDUC Years of education of the household head in 1983.
WHITE =1 if the household head is white.
MARRIED =1 if the household head is married.
NCHILD Number of children living in the household in 1983.
YOUNGCH Number of children between age 0-6 in 1983.
MIDDCH Number of children between age 7-12 in 1983.
HIGHSCH Number of children between age 13-18 in 1983.
∆CHILD =1 if the household had an additional child
between 1983-89.
NADULT Number of adults living in the household in 1983.
∆NADULT Change in the number of adults between 1983-89.
BEQUEST =1 if the household head is planning to leave a bequest.
HOWN83 =1 if the household owns a home in 1983.
HOWN89 =1 if the household owns a home in 1989.
SPFULLT =1 if the spouse is working fulltime in 1983.

110
Appendix C

Appendix for Chapter 4

C.1 Definition of Variables

Name Description
CHILD Number of children ever born to the household head.
COLLEXP Amount of expenditure on the college education of a child
in 1986 dollars.
NWORTH Net worth in 1986.
NWORTH25 =1 if the household is in the bottom 25 percentile of the
wealth distribution.
NWORTH75 =1 if the household is in the top 25 percentile of the wealth
distribution.
PERINC Predicted 1985 household income.
TRINC Difference between total income in 1985 and permanent income.
SAVE1 Difference between net worth in 1986 and 1983 divided by 3.
SAVE2 Difference between net worth in 1986 excluding the capital
gains on primary residence and net worth in 1983 divided by 3.
AGE Age of the household head in 1983.
FEMALE =1 if the household head is female.
HIGHSCH =1 if the household head has a high school degree.

111
COLLEG =1 if the household head has a college degree.
HIGHSCHSP =1 if the spouse has a high school degree.
COLLEGSP =1 if the spouse has a college degree.
FSPOUSE =1 if the spouse is working at a full-time job.
BLACK =1 if the household head is African-American.
MARRIED =1 if the household head is married.
RISKY =1 if the household head is willing to undertake risky
investments.
WINDF =1 if the household received a windfall greater than $3,000
between 1983-86.
BEDUCAT =1 if the household head thinks it is all right to borrow for
education.
SRETIRE =1 if retirement is the most important reason for saving.
SEMERGE =1 if emergencies are the most important reason for saving.
SCHEDU =1 if children’s education is the most important reason for
saving.
SHOME =1 if saving to buy a home is the most important reason for
saving.
SOTHER =1 if the household cited another reason as the most
important reason to save.
NSMSA =1 if the place of residence is not in a SMSA.
NCHCOLL Number of children attending college between 1983-86.

112
Bibliography

[1] Gary S. Becker. A Treatise On the Family. Cambridge: Harvard Uni-


versity Press, 1991.

[2] Gary S. Becker and Gregg H. Lewis. On the interaction between quantity
and quality of children. Journal of Political Economy, 81(2):S279–S288,
1973.

[3] Jere R. Behrman, Robert A. Pollak, and Paul Taubman. Family re-
sources, family size, and access to financing for college education. Journal
of Political Economy, 97(2):398–419, 1989.

[4] James Berkovec and Don Fullerton. A general equilibrium model of


housing, taxes, and portfolio choice. Journal of Political Economy,
100(2):390–429, 1992.

[5] Carol C. Bertaut. Stockholding behavior of U.S. households: Evidence


from the 1983-1989 Survey of Consumer Finances. Review of Economics
and Statistics, 80(2):263–275, 1998.

[6] Martin Browning and Annamaria Lusardi. Household saving: Micro


theories and micro facts. Journal of Economic Literature, 34(4):1796–
1855, 1996.

113
[7] Jan K. Brueckner. Consumption and investment motives and portfolio
choice of homeowners. Journal of Real Estate Finance and Economics,
15(2):159–180, 1997.

[8] Christopher D. Carroll. How does future income affect current consump-
tion? Quarterly Journal of Economics, 109(1):111–148, 1994.

[9] Christopher D. Carroll. Buffer stock saving and the life-cycle/permanent


income hypothesis. Quarterly Journal of Economics, 112(1):1–56, 1997.

[10] Christopher D. Carroll, Karen D. Dynan, and Spencer D. Krane. Un-


employment risk and precautionary wealth: Evidence from households’
balance sheet. Technical Report 15, Board of Governors of the Federal
Reserve System, Finance and Economics Discussion, 1999.

[11] Christopher D. Carroll and Andrew A. Samwick. The nature of precau-


tionary wealth. Journal of Monetary Economics, 41(1):41–71, 1997.

[12] Christopher D. Carroll and Andrew A. Samwick. How important is


precautionary saving? Review of Economics and Statistics, 80(3):410–19,
1998.

[13] Ngina S. Chiteji and Frank P. Stafford. Portfolio choices of parents and
their children as young adults: Asset accumulation by African-American
families. American Economic Review, 89(2):377–380, 1999.

114
[14] Susan P. Choy and Henke R. Robin. Parental financial support for un-
dergraduate education. Technical report, National Center for Education
Statistics, U.S. Department of Education, Washington, 1992.

[15] Angus S. Deaton. Understanding Consumption. New York: Oxford


University Press, 1992.

[16] Andrew W. Dick and Aaron S. Edlin. The implicit taxes from college
financial aid. Journal of Public Economics, 65(3):295–322, 1997.

[17] Wendy E. Dunn. Unemployment Risk, Precautionary Saving, and Durable


Goods Purchase Decisions. Technical Report 48, Board of Governors of
the Federal Reserve System, Finance and Economics Discussion, 1998.

[18] Karen Dynan. How prudent are consumers? Journal of Political Econ-
omy, 101(6):1104–13, 1993.

[19] Aaron Edlin. Is College Fiancial Aid Equitable and Efficient? Journal
of Economic Perspectives, 7(2):143–158, 1993.

[20] Marjorie Falvin and Takshi Yamashita. Owner-occupied housing and


the composition of the household portfolio over the life cycle. Technical
Report 1998, University of California, San Diego.

[21] Eugene. F. Fama and Merton H. Miller. The Theory of Finance. Dryden
Press, Hinsdale, IL, 1972.

115
[22] Martin Feldstein. College scholarship rules and private saving. American
Economic Review, 85(3):552–66, 1995.

[23] William G. Gale and John Karl Scholz. Intergenerational transfers and
the accumulation of wealth. Journal of Economic Perspectives, 8(4):145–
160, 1994.

[24] Allen C. Goodman. Demographics of individual housing demand. Re-


gional Science and Urban Economics, 20(1):83–102, 1990.

[25] Allen C. Goodman and Masahiro Kawai. Permanent income, hedonic


prices, and demand for housing: New evidence. Journal of Urban Eco-
nomics, 12(2):214–237, 1982.

[26] Donald R. Haurin, Patric H. Hendershott, and Dongwook Kim. Housing


decisions of American youth. Journal of Urban Economics, 35:28–45,
1994.

[27] J. Vernon Henderson and Yannis M. Ioannides. A model of housing


tenure choice. American Economic Review, 73(1):98–113, 1983.

[28] J. Vernon Henderson and Yannis M. Ioannides. Owner occupancy: In-


vestment vs consumption demand. Journal of Urban Economics, 21(2):228–
241, 1987.

[29] Charles Yuji Horioka and Wako Watanabe. Why do people save? A
micro-analysis of motives for household saving in Japan. Economic Jour-
nal, 107(442):537–552, 1997.

116
[30] V. Joseph Hotz, Jacob Alex Klerman, and Robert J.Willis. Handbook of
Population and Family Economics, chapter The Economics of Fertility in
Developed Countries, pages 275–347. Amsterdam: Elsevier, 1997.

[31] V. Joseph Hotz and Robert A. Miller. An empirical analysis of life cycle
fertility and female labor supply. Econometrica, 56(1):91–118, 1988.

[32] R. Glenn Hubbard, Jonathan Skinner, and Stephen P. Zeldes. Pre-


cautionary saving and social insurance. Journal of Political Economy,
103(21):360–399, 1995.

[33] Keith Ray Ihlanfeldt. An empirical investigation of alternative approches


to estimating the equilibrium demand for housing. Journal of Urban
Economics, 9(1):97–105, 1981.

[34] Yannis M. Ioannides. Dynamics of the composition of household asset


portfolios and the life cycle. Applied Finance, 2(3):145–159, 1992.

[35] Nancy A. Jianakoplas and Alexandra Bernasek. Are women more risk
averse? Economic Inquiry, 36(4):620–630, 1998.

[36] Richard Deitz John Robst and KimMarie McGoldrick. Income variabil-
ity, uncertainty and housing tenure choice. Regional Science and Urban
Economics, 29(2):219–229, 1999.

[37] Arthur B. Kennickell and Martha Starr-McCluer. Household saving and


portfolio change: Evidence from the 1983-89 SCF panel. Review of In-
come and Wealth, 43(4):381–399, 1997.

117
[38] Arthur B. Kennickell and Martha Starr-McCluer. Retrospective report-
ing of household wealth: Evidence from the 1983-89 Survey of Consumer
Finances. Journal of Business and Economic Statistics, 15(3):452–63,
1997.

[39] Miles S. Kimball. Precautionary saving in the small and in the large.
Econometrica, 58(1):53–73, 1990.

[40] Mervyn A. King and L-D. L. Dicks-Mireaux. Asset holdings and the
life-cycle. Economic Journal, 92(366):247–267, 1982.

[41] Mervyn A. King and Jonathan I. Leape. Wealth and portfolio composi-
tion: Theory and evidence. Journal of Public Economics, 69(2):155–193,
1998.

[42] Lung-Fei Lee and Robert P. Trost. Estimation of some limited dependent
variable models with the application to housing demand. Journal of
Econometrics, 8:357–382, 1978.

[43] Mark Long. The impact of asset-tested college financial aid on household
savings. 2001.

[44] Annamaria Lusardi. On the importance of the precautionary saving


motive. American Economic Review, 88(2):449–453, 1998.

[45] G.S. Maddala. Limited Dependent and Qualitative Variables in Econo-


metrics. Cambridge University Press, 1983.

118
[46] James M. Poterba and Andrew A. Samwick. Household portfolio alloca-
tion and over the life cycle. Technical Report 6185, National Bureau of
Economic Research, 1997.

[47] Jennifer B. Presley and Suzanne B. Clery. Middle income undergradu-


ates: Where they enroll and how they pay for their education. Technical
report, National Center for Education Statistics, U.S. Department of Ed-
ucation, Washington, 2001.

[48] Jonathan Skinner. Risky income, life cycle consumption and precaution-
ary savings. Journal of Monetary Economics, 22(2):237–55, 1988.

[49] Nicholas S. Souleles. College tuition and household savings and consump-
tion. Journal of Public Economics, 77:185–207, 2000.

[50] Martha Starr-McCluer. Health insurance and precautionary savings.


American Economic Review, 86(1):285–295, 1996.

[51] Lala Carr Steelman and Brian Powell. Acquiring capital for college:
The constraints of family configuration. American Sociological Review,
54:844–855, 1989.

[52] Annika E. Sundén and Brian J. Surette. Gender differences in the allo-
cation of assets in retirement savings plan. American Economic Review,
88(2):207–211, 1998.

[53] Nigel Tomes. The family, inheritance and the intergenerational transmis-
sion of inequality. Journal of Political Economy, 89(5):928–958, 1981.

119
[54] Nigel Tomes. A model of fertility and children’s schooling. Economic
Inquiry, 19(2):209–234, 1981.

[55] Robert J. Willis. A new approach to the economic theory of fertility


behavior. Journal of Political Economy, 81(2):S14–S64, 1973.

[56] Edward N. Wolff. Recent trends in the size distribution of household


wealth. Journal of Economic Perspectives, 12(3):131–150, 1998.

[57] Kenneth I. Wolpin. An estimable dynamic stochastic model of fertility


and child mortality. Journal of Political Economy, 92(5):852–874, 1984.

[58] Stephen P. Zeldes. Optimal consumption with stochastic income: De-


viations from certanity equivalence. Quarterly Journal of Economics,
104(2):275–98, 1989.

120
Vita

Tansel Yilmazer was born in Izmir, Turkey on June 2, 1970, the daugh-
ter of Önder Yilmazer and Necla Yilmazer. She received her Bachelor of Arts
degree in Business Administration from Boĝaziçi University in January 1994.
She began her graduate studies at Boĝaziçi University, where she received a
Master of Arts degree in Economics in June 1997. She later continued her
education at the University of Texas at Austin. Effective August 2002, she
accepted an assistant professor position at Purdue University.

Permanent address: 834 Main Street Apt. B


Lafayette, IN 47901

This dissertation was typed by the author.

121

Vous aimerez peut-être aussi