SERIES

IZA DP No. 7083

PAPER

Income Inequality and Saving

Francisco Alvarez-Cuadrado
DISCUSSION

Mayssun El-Attar Vilalta

December 2012

Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
 Income Inequality and Saving

Francisco Alvarez-Cuadrado
McGill University
and CIREQ

Mayssun El-Attar Vilalta

McGill University
and IZA

Discussion Paper No. 7083

December 2012

IZA

P.O. Box 7240

53072 Bonn
Germany

Phone: +49-228-3894-0
Fax: +49-228-3894-180
E-mail: iza@iza.org

Any opinions expressed here are those of the author(s) and not those of IZA. Research published in
this series may include views on policy, but the institute itself takes no institutional policy positions.
The IZA research network is committed to the IZA Guiding Principles of Research Integrity.

The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
organization supported by Deutsche Post Foundation. The center is associated with the University of
Bonn and offers a stimulating research environment through its international network, workshops and
conferences, data service, project support, research visits and doctoral program. IZA engages in (i)
original and internationally competitive research in all fields of labor economics, (ii) development of
policy concepts, and (iii) dissemination of research results and concepts to the interested public.

IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.
IZA Discussion Paper No. 7083
December 2012

ABSTRACT

Income Inequality and Saving

Over the last three decades, average income for the bottom half of the US distribution
increased by 8% while their average saving rate decreased by eight percentage points. Over
the same period the US experienced a substantial increase in inequality and a continuous
decrease in the aggregate saving rate. We propose an explanation based on interpersonal
comparisons consistent with these trends. When households care about their consumption
relative to others, individual saving rates decrease with reference income while aggregate
saving decreases with income inequality. We provide evidence from the PSID and a panel of
OECD countries consistent with these predictions.

JEL Classification: D91, E21, C23

Keywords: relative consumption, income inequality, saving

Corresponding author:

Francisco Alvarez-Cuadrado
Department of Economics
McGill University
855 Sherbrooke St. West
Montreal, Québec H3A 2T7
Canada
E-mail: francisco.alvarez-cuadrado@mcgill.ca
1 Introduction
James Duesenberry, in his seminal work, Income, Saving and the Theory of Consumer Be-
havior (1949), introduces the relative income hypothesis in an attempt to rationalize the well
established differences between cross-sectional and time-series properties of saving rates. On
the one hand, a wealth of studies based on 1935-36 and 1941-42 cross-sectional budget surveys
present a saving ratio that increases with income. On the other hand, the time series data
on savings and income collected by Kuznets (1942) present a trend-less saving ratio. Duesen-
berry (1949) proposes an individual consumption function that depends not only on current
individual income but also on the level of income of the reference group. As a result “for any
given relative income distribution, the percentage of income saved by a family will tend to be a
unique, invariant, and increasing function of its percentile position in the income distribution.
The percentage saved will be independent of the absolute level of income. It follows that the
aggregate saving ratio will be independent of the absolute level of income” (Duesenberry 1949,
p. 3). Despite its empirical success, see for instance Mayer (1966, 1972), the relative income
hypothesis was quickly replaced by the well-known permanent income hypothesis (Modigliani
and Brumberg (1954), Friedman (1957)) as the economists’ workhorse to understand saving be-
havior.1 According to this view the cross-sectional correlation between saving rates and income
is driven by transitory deviations from permanent income, while in the aggregate, most tran-
sitory components cancel out, leading to the close relation between consumption and income
observed in time series data. In a recent paper, Dynan, Skinner and Zeldes (2004) resolve this
old empirical question. Applying new methods to a variety of saving measures derived from the
Panel Study of Income Dynamics (PSID), the Survey of Consumer Finances (SCF), and the
Consumer Expenditure Survey (CEX) they find a strong positive relationship between saving
rates and lifetime income.
The goal of this paper is twofold. First, we identify interpersonal comparisons as one of the
channels behind the documented gradient between saving rates and lifetime income. Second,
we explore the impact of income inequality on this gradient. For this purpose, we propose
a simple model of interpersonal comparisons. Our overlapping generations economy is popu-
lated by heterogeneous households who care about consumption relative to others and derive
a “warm-glow” from the bequest they leave to their descendents. Analyzing the saving choices
of one generation, we find several testable implications. First, individual saving rates increase
One should point out that Friedman’s view is more nuanced than the profession seems to believe “. . .
1

and finally, the evidence that we have cited seems to fit it (the permanent income hypothesis) somewhat
better. . . however, this evidence is by no means sufficient to justify a firm rejection of the relative income
hypothesis” (Friedman 1957, p. 169).

2
with lifetime income. Second, saving rates decrease with the level of income of the reference
group and they do so more for poor households. Third, an increase in income inequality leads
to an increase in inequality of saving rates. And fourth, once we allow for upward-looking
interpersonal comparisons, a mean preserving spread in the distribution of labor income leads
to a decrease in the aggregate saving rate.
Using several measures of saving derived from the PSID, we test the first three predictions.
We find, in line with Dynan et al. (2004), a strong positive gradient between saving rates and
income conditional on life-cycle and demographic characteristics. On average an increase in
income of 10% is associated with an increase in the saving rate that ranges between 0.4 and
0.5 percentage points depending on the measure of saving used. Approximating income of
the reference group by average income in the state of residence, we find that, after controlling
for household income and other individual characteristics, saving rates decrease in comparison
income and do so more for poor households. For instance, our results suggest that a 10%
increase in income of the reference group leads to a decrease in the total saving rate for a
household in the 10th income percentile of one fifth of a percentage point. This is a large
decrease, particularly if one considers that the average saving rate in the bottom decile is
close to zero. When we include a measure of income inequality, we find that those regions or
periods with higher income inequality are characterized by a more unequal distribution of saving
rates controlling for income. Furthermore, all these results remain robust when we instrument
for permanent income using average labor earnings, different measures of consumption, and
education. Most importantly, we conduct several exercises to rule out the possibility of reverse
causality and provide additional evidence that the impact of inequality on saving rates that we
identify is indeed due to interpersonal comparisons and not to other factors.
Finally, we test the last proposition constructing a panel that includes personal saving
rates for six developed countries compiled from OECD statistics and measures of inequality
from Alvarado, Atkinson, Piketty and Saez (2012). Our sample, although limited in terms
of country coverage, spans over 50 years and accounts for roughly 40% of world GDP. Our
results consistently suggest a negative correlation between saving rates and inequality. This
correlation persists when we include country fixed effects and other determinants of aggregate
saving such as age dependency ratios (Modigliani 1970), GDP growth (Carroll and Weil 1994),
or real interest rates (Boskin 1978).
Our empirical methodology is closely related to Dynan et al. (2004). While they focus on
documenting and exploring the robustness of the gradient between saving rates and permanent
income2 , our paper investigates one of the channels through which changes in permanent income
2
Specifally this gradient does not seem to be caused by life-cycle differences or variation in Social Security

3
lead to changes in saving rates, interpersonal comparisons, and studies the impact of income
inequality on this gradient. In a similar way, our work is related to Altonji and Villanueva
(2007) who find that the fraction of every extra dollar of lifetime resources that parents pass
to their children through inter vivos transfers or bequests increases with parental income. The
results of the empirical literature that tests the impact of the income distribution on aggregate
saving are scarce and inconclusive3 . On the one hand, Blinder (1975) using US time-series
data finds that higher income inequality raises aggregate consumption and therefore lowers
saving. Although the estimated effects are in general statistically insignificant. On the other
hand, Hong (1995) evaluates the effect of the income share of the top quintile on gross domestic
saving in a sample of developing and industrial countries. He finds that this share has a positive
effect on aggregate saving, suggesting that inequality increases saving. Finally, using data for
11 countries between 1921 and 2002, Leigh and Posso (2009) find no consistent relationship
between top income shares and national saving rates. A potential explanation for these mixed
results may lie in the use of different measures of saving, most of them derived from the current
account identity. In this sense, our measure of aggregate saving, the personal saving rate, draws
a clearer picture of the interaction between inequality and aggregate saving4 .
Our model is closely related to the extensive literature on relative consumption. Abel
(1990), Gali (1994), Corneo and Jeanne (1998), Carroll, Overland and Weil (2000), Liu and
Turnovsky (2005) and Wendner (2012) are just a few examples. In a context similar to ours,
Levine, Frank and Dijk (2010) present a model in which changes in one group’s spending shift
the frame of reference that defines consumption standards for others just below them in the
income scale, giving rise to expenditure cascades. Along these lines, a recent paper by Bertrand
and Morse (2012) finds that up to a quarter of the decline in the saving rate over the last three
decades could be attributed to these expenditure cascades. Our work also connects with the
recent literature on self-reported well-being. Early work by Easterlin (2010, 1995), and Oswald
(1997) found differences between the cross-section and time-series properties of happiness data
that are quite similar to those reported on savings data more than sixty years ago. Data
replacement rates (Huggett and Ventura 2000), or by a precautionary motive together with uncertain medical
expenses (De Nardi, French and Jones (2010), Kopecky and Koreshkova (2010)), or by the presence of dynastic
smoothing (Becker and Tomes, 1986) or by the inclusion of high-income entrepreneurs in their sample (Quadrini
(1999) and Hurst and Lusardi (2004)).
3
See Schmidt-Hebbel and Serven (2000) for a review of this literature.
4
Theoretical models on the relation between saving and inequality focus on household behavior suggesting
that the relevant aggregate variable in an empirical analysis should be close to the personal saving rate that we
use (see for instance, Deaton (1991), Alesina and Rodrik (1994), and Person and Tabellini (1994)). With the
exception of Blinder (1975) previous studies have used measures of saving that include public and corporate
saving.

4
on self-reported well-being shows that within a country at a given point in time those with
higher incomes are, on average, happier. However, average happiness in developed countries
has remained relatively constant over time despite sharp increases in per capita GDP. Clark,
Frijters and Shields (2008) highlight the importance of accounting for interpersonal comparisons
to account for the “Easterlin paradox”5 . Recent work has tried to estimate the direct impact
of interpersonal comparisons on self-reported well-being. Luttmer (2005), Ferrer-i Carbonell
(2005) and Dynan and Ravina (2007) find that people’s level of happiness depends positively on
how well they are doing relative to the average in their geographical area, even after controlling
for own income levels. Our model formalizes some of these insights and explores the implications
of relativity concerns for savings6 .
Finally, Dynan et al. (2004) find, not only a strong positive relationship between saving
rates and lifetime income but also evidence suggesting that the marginal propensity to save
out of lifetime income is higher for rich households than for poor ones. Figure 2(a) illustrates
these results. The rich not only save more as a fraction of their lifetime income but they also
save a higher fraction of each additional dollar of income. Other things equal, an increasing
marginal propensity to save has at least two direct implications. First, as households become
richer their saving rates should increase. So at the individual level, we should find an increase in
saving rates as permanent income increases. Second, if we redistribute income from households
with a low marginal propensity to save to those with a high marginal propensity to save the
aggregate saving rate should increase. So at the aggregate level, we should find that increases
in income inequality should be associated with increases in the average saving rate. Using PSID
data from 1994 to 2007, figure 2(b) depicts the evolution of average income and the average
saving rate for households below the median of the US (lifetime) income distribution. Over
this period, their average income increased by 8% while their average saving rate fell by almost
8 percentage points7 . Figure 2(c) combines the evolution of the share of income of the top
5
See Stevenson and Wolfers (2008) for a recent reassestment of this paradox.
6
Our work is also related to the recent debate on the different patterns of income and consumption inequality.
Over the thirty years that preceded the Great Recession, the US experienced a sustained increase in income
inequality (Katz and Autor (1999); Autor, Katz and Kearney (2008)). Along similar lines Cutler and Katz (1992)
document that changes in the distribution of consumption parallel changes in the distribution of income. Slesnick
(2001) questions these findings and Krueger and Perri (2006) using CEX data find that despite the surge in
income inequality consumption inequality has increased only moderately. Attanasio, Hurst and Pistaferri (2012)
focusing on expenditure categories that are measured without error, such as a vehicles owned, or exploiting
differences in demand income elasticities across categories to account directly for the non-classical measurement
error, that seems to plague most expenditure categories in this dataset, find that consumption inequality has
increased by nearly the same amount as income inequality. Although our focus is not on documenting the
patterns on saving inequality our findings are informative on the effect of inequality in saving.
7
A similar pattern has been reported by Bosworth and Anders (2008). These authors show that, once some
extreme values are trimmed, average saving by individuals in the lower third of the income distribution has

5
5% of the distribution from the update of Piketty and Saez (2003) with the personal saving
rate from the Federal Reserve Economic Data. In the early 80s, personal saving were roughly
10% of disposable income falling to barely 2% by 2007. Over the same period the share of
income of the top 5% increased by more than 50% from 21% to 33%. Our model provides a
simple explanation, combining all three trends. In the presence of interpersonal comparisons,
individual saving depends not only on individual lifetime income but also on the lifetime income
of the reference group. As a result and despite the 8% increase in lifetime income of households
below the median, their saving rates have fallen as a consequence of the much faster increases
in income for those in the upper tail of the distribution. For instance, average income for the
top decile increased by roughly 38% over the same period. At the aggregate level, and despite
of the fact that the model’s marginal propensity to save is independent of the level of income,
an increase in the share of income received by high-income households induces a reduction in
the levels of saving of low-income households that goes beyond this propensity. The reason lies
in the increase in consumption of high-income individuals that stimulates envious consumption
further depressing saving for low-income individuals. As a result, increases in income inequality
are associated with decreases in the aggregate saving rate.
The remaining of the paper is organized as follows. In Section 2, we set out the basic model
and derive its implications for the interaction among saving, income and inequality. In Section
3, we use data from the PSID to test these implications. In Section 4, we extend the model
to allow for a upward-looking comparisons and explore its implications for aggregate saving.
These implications are tested in Section 5. In Section 6 we offer some concluding remarks,
while the Appendices provide additional details on the derivations and results.

2 Individual Saving and Income Inequality

Consider an OLG economy where the wages per unit of effective labor, wt , and the gross return
to saving, Rt , are exogenously determined.8
Individuals live for two periods, “youth” and “old-age”. At the end of their youth,
each individual gives birth to a single offspring and therefore the labor force is constant. Each
generation consists of a set of agents, indexed by i, distributed uniformly along the unit interval
with a total population mass of 1. Our agents are altruistic toward their children, deriving a
“warm-glow” from the bequest, bit+1 , that they leave to their descendants as in Adreoni (1989)9 .
fallen over the period 1984-2005.
8
See Alvarez-Cuadrado and Long (2012) for a similar model with endogenous factor price determination.
9
There are alternatives to this approach of modelling the motives for intergenerational transfers. Barro

6
 Individuals within a given generation differ in their endowment of productive labor, lti , and
in the bequest they inherit from their parent, bit . Specifically, we assume that their labor
productivity is the realization of a stationary i.i.d. random variable with mean ¯l = 1 and
standard deviation σ l . The resulting distribution of wages for period t has mean w̄t = wt and
standard deviation σtw = wt σ l .
Let’s focus on the i-th individual born in period t. When young, he inelastically supplies his
labor endowment, earning an income wti = lti wt that together with his inherited wealth, bit , is
divided between first-period consumption, cit , and saving, sit . His first-period budget constraint
is given by,

cit + sit = bit + wti ≡ yti (1)

where we denote by yti his lifetime resources.

In the second period of his life, the individual is retired. His only source of income is
the return on the savings he made when young, Rt+1 sit , which he allocates to second-period
consumption, dit+1 , and bequests, bit+1 ,

Rsit = dit+1 + bit+1 (2)

The preferences of an individual born in period t are given by the following life-cycle utility
function,
� � � � � � � � ��
U ĉit , dˆit+1 , bit+1 = ln cit − γḡti + β ln dit+1 − ηōit+1 + µ ln bit+1 (3)

where 0 < β < 1 is the subjective discount factor and µ > 0 governs the importance of the
bequest motive.
Our key behavioural assumption is that the satisfaction derived from consumption does
not depend on the absolute level of consumption itself but rather on how it compares to the
consumption of some reference group. Following Ljungqvist and Uhlig (200) we adopt an
�1
additive specification for relative consumption where, for the moment, ḡti = c̄t = 0 cjt dj and
�1
ōit+1 = dt+1 = 0 djt+1 dj are the reference consumption levels when young and old respectively
for any household born at time t. The importance of positional concerns when young and old is
(1974) considers pure altruism, while in Abel (1985) accidental bequests arise from market incompleteness, and
Bernheim, Schleifer and Summers (1985) propose a bequest-as-exchange model. The empirical evidence for the
reasons for inter-generational transfers is mixed. Nonetheless, the evidence reviewed by Arrondel, Parelman
and Pestieau (1994) and Davies and Shorrocks (1999) suggests that an important fraction of the observed
inheritances seems to reflect some kind of impure altruism close to the warm glow approach that we adopt.

7
given 0 < γ < 1 and 0 < η < 1 respectively10 . Concerning the relative magnitude of γ to η, one
may consider the work of development psychologists and sociologists (Coleman (1961), Simmons
and Blyth (1987), Corsaro and Eder (1990)) which suggests that interpersonal comparisons and
peer effects are more pronounced early in life. In the “first-period” of life, people work, find
partners, and raise children, being exposed to and therefore influenced by a wide variety of
social networks. More direct evidence comes from Charles, Hurst and Roussanov (2009) who
find that differences in visible consumption disappear when they restrict their sample to older
households, suggesting that the relative importance of interpersonal comparisons decreases with
age. In line with this evidence, we assume the following relationship between the degrees of
envy in both periods of life, η = ξγ, where 0 < ξ ≤ 1.
An important assumption concerns the non-positionality of bequests. Their limited observ-
ability and their concentration in the upper tail of the wealth distribution suggest that bequests
possess important non-positional features. Along these lines, Moav and Neeman (2010) present
a signalling game of status where agents derive utility from consumption, status, and a bequest,
which they model as non-observable. The non-observability of bequests in a signalling game is
equivalent to its non-positionality in a model of relative consumption. Heffetz (2011) conducts
a survey on the degree of visibility of 31 goods and services ranging from cars and watches to
medical insurance and education. Although bequests are not included in his list, the closest
expenditure item surveyed, life insurance, ranks penultimate in terms of visibility. The high
concentration of bequests on the upper tail of the wealth distribution is a well documented
empirical regularity. Mulligan (1997) estimates that in the U.S. the proportion of estates be-
queathing sufficient wealth to be subject to inheritance tax was between 2 and 4 percent in
the period 1960-1995. The fact that most households do not leave a substantial bequest is
consistent with the limited positionality of these transfers.
Finally, we place restrictions on the stationary distribution of productive endowments to
guarantee that everyone’s relative consumption is positive.

2.1 Saving and Inequality

The i-th individual of the generation born in period t takes prices11 and the choices of the
other members of his generation as given and chooses consumption in both periods of his life
10
As (Frank 1985, p. 111) points out, ”the sociological literature on reference group theory stresses that an
individual’s personal reference group tends to consist of others who are similar in terms of age”. Along these
lines our specification restricts interpersonal comparisons to individuals within the same generation. See Abel
(2005) and Alonso-Carrera, Caballe and Raurich (2008) for an alternative specification of the reference group.
11
Since we will focus on a single generation we drop time subscripts from prices.

8
to maximize (3) subject to (1) and (2). Let’s solve the problem in two stages. Given the level
of saving, sit , an old individual at time t + 1 chooses dit+1 and bit+1 to maximize

� � � �
V ≡ ln dit+1 − ξγ d¯t+1 + µ ln bit+1

subject to (2). The solution to this problem is

1 � i �
dit+1 = Rst + ξγµdt+1 (4)
1+µ

µ � i �
bit+1 = Rst − ξγdt+1 (5)
1+µ
and therefore
� � � �
� � 1 µ
V (Rsit ) = (1 + µ) ln Rsit − ξγ d¯t+1 + ln + µ ln
1+µ 1+µ
The young individual at time t chooses cit and sit to maximize

� �
ln cit − γc̄t + βV (Rsit ) (6)

subject to (1). The necessary conditions for this problem imply

Rsit − ξγdt+1
= βR (7)
(cit − γc̄t ) (1 + µ)
with the standard interpretation that along an optimal solution the i-th agent equates the
marginal rate of substitution between first and second period choices to its relative price, the
interest rate adjusted for the rate of time preference. Combining (7) and (1) we reach

ξγdt+1 ξγdt+1
cit [1 + β (1 + µ)] = wi + bit + β (1 + µ) γct − = yti + β (1 + µ) γct − (8)
R R
We begin by characterizing the optimal behavior of the average household, i.e. the household
�1
with average income, ȳt = 0 bjt dj + w̄.
Combining (4), (5), (7) and (8) we obtain the level of saving of the average individual,

s̄t = (1 − γ) β (1 + µ (1 − ξγ)) χȳt (9)

1
where χ ≡ .
(1 − ξγ) + (1 − γ) β (1 + µ (1 − ξγ))
Saving for the average household is a constant fraction of his lifetime income. Combining
(4)-(9) we obtain the remaining choices for the average household,

9
 (1 − ξγ) (1 − ξγ)
ct = dt+1 = bt+1 = (1 − ξγ) χȳt (10)
Rβ (1 − γ) Rβµ (1 − γ) (1 − ξγ)
We can use these results to characterize the behavior of the i-th individual of the same
generation. Substitute (10) into (8) to reach first-period consumption. Combine it with (7) to
reach his level of saving,

β (1 + µ) � i �
sit = yt − γφs ȳt , (11)
1 + β (1 + µ)
((1 + µ) (1 − ξγ) − ξ (1 − γ)) χ
where φs ≡ > 0.
(1 + µ)
It is straightforward to solve for the remaining optimal choices as functions of individual
and average lifetime income,

1 � i �
cit = yt + γφc ȳt (12)
1 + β (1 + µ)

Rβ � i �
dit+1 = yt + γφd ȳt (13)
1 + β (1 + µ)

Rβµ � i �
bit+1 = yt − γφb ȳt (14)
1 + β (1 + µ)
where

φc ≡ β ((1 + µ) (1 − ξγ) − ξ (1 − γ)) χ > φd ≡ (ξµ (1 − γ) β + ξ − 1) χ > 0,

and
φb ≡ ((1 − ξγ) + ξ (1 − γ) β) χ > 0.

Consumption of the i-th household, (12) and (13), is made out of two components. The first
increases in the household’s lifetime income. The second reflects the influence of interpersonal
comparisons and increases in the lifetime income of his reference group. As a result saving
and bequests depend on relative, rather than absolute, income. When individual satisfaction
depends on consumption comparisons across households the relevant variable driving saving
choices is the comparison between individual i’s lifetime income and the lifetime income of his
reference group. The agents populating our economy are not only “disposed, as a rule and on
the average, to be forward-looking animals” as those in (Modigliani and Brumberg 1954, p.
430) or Friedman (1957), but are also outward-looking animals as their choices are partially
driven by the choices of other members of the community they live in.

10
 The following propositions summarize the implications of the model for saving rates and
their interaction with income inequality.
Proposition 1. Under our assumptions, the (individual) saving rate increases with lifetime
income.

∂sirate β (1 + µ) γφs ȳ
= >0
∂y i 1 + β (1 + µ) (y i )2
Proof. Equation (11) implies the following saving rate out lifetime income,
� �
si β (1 + µ) ȳ
sirate ≡ i = 1 − γφs i (15)
y 1 + β (1 + µ) y
The result follows from differentiation of (15).
In the absence of interpersonal comparisons, γ = 0, the saving rate is proportional to lifetime
income and therefore, as in the standard version of the permanent income hypothesis, the saving
rate is independent of lifetime resources. Once we allow for consumption externalities, γ > 0,
the saving rate varies across the income distribution with poor households saving a smaller
fraction of their resources than rich households. The introduction of interpersonal comparisons
diverts resources from less positional uses –savings– to more positional ones –consumption when
young. Furthermore, this diversion varies across the income distribution being stronger for poor
households than for rich ones as the following proposition illustrates12 .
Proposition 2. Under our assumptions, individual saving rates decrease with the level of
income of the reference group and they do so more (less) for poor (rich) households.

∂sirate ∂ 2 sirate β (1 + µ) γφs 1

Proof. Differentiate (15) to reach = −y i = − < 0.
∂ ȳ ∂ ȳ∂y i 1 + β (1 + µ) y i
Finally, we can use this simple model to explore the impact of income inequality on individ-
ual saving. For simplicity, assume there is a generation, possibly the first, where the inherited
bequest is equal to zero. Under this simplifying assumption income inequality is determined
by the distribution of wages. A convenient measure of the dispersion of this distribution is
the coefficient of variation, i.e. the ratio between the standard deviation and the mean of the
12
There are two intuitive interpretations of these results. First, along the lines suggested by Alvarez-Cuadrado
and Long (2012), in the presence of interpersonal comparisons the income elasticities of demand for consumption
and saving vary across the income distribution. See their proposition 1 for details. Second, in line with the
work of Atkeson and Masao (1996, 1997), the introduction of interpersonal comparisons induces an intertemporal
elasticity of substitution (IES) that increases with income.
� To�illustrate this point assume that ξ = µ = 0. The
dit+1 c̄t
Euler equation for household i is given by i = β 1 − γ i R where the term in brackets is proportional to
ct ct
the IES.

11
 σw
wage distribution, CV (wi ) = . With average saving given by (9), we combine the standard
w̄
deviation of wages with (11) to derive the standard deviation of saving,

β (1 + µ)
σs = σw (16)
1 + β (1 + µ)
iσw
As a result the coefficient of variation of saving is given by CV (s ) = Θ with Θ ≡ (1+µ)/
w̄
[(1 + β(1+))(1 + µ(1 − ξγ))(1 − γ)χ] ≥ 1. So, an increase in income inequality leads to a
more than proportional increase in saving inequality. Nonetheless, we are interested in the
relationship between income inequality and saving rates. The following proposition illustrates
this relationship.
Proposition 3. An increase in income inequality, measured as an increase in the coefficient
of variation of wages, leads to an increase in inequality of saving rates.
Proof. Applying the Delta method to (15) we approximate the standard deviation for the
β (1 + µ) φs σ w
saving rate by σ srate ≈ and therefore the coefficient of variation of the saving
1 + β (1 + µ) w̄
rates is given by,

� � γ ((1 + µ) (1 − ξγ) − ξ (1 − γ)) σw

CV sirate ≈ (17)
(1 + β (1 + µ)) (1 − γ) (1 + µ (1 − ξγ)) w̄
As opposed to the standard version of the permanent income hypothesis, γ = 0, where saving
rates are independent of the level of income and therefore unaffected by the degree of income
inequality, in the presence of interpersonal comparisons, γ > 0, changes in the distribution of
income lead to changes in the distribution of saving rates. Specifically an increase in income
inequality is associated with an increase in the coefficient of variation of saving rates.
Remark to Proposition 3. The aggregate (average) saving rate is independent of the
income distribution.
Since preferences are quasi-homothetic and capital markets are perfect saving is an affine
function of income, a property that ensures that the distribution of income does not affect
aggregate outcomes13 . Along this dimension, our modelling approach is similar to that of
Chatterjee (1994) and Caselli and Ventura (2000).
In the next section we will use PSID data to test these propositions.
13
Notice that the timing of income is equivalent to perfect capital markets. See Alvarez-Cuadrado and Long
(2012) for a characterization of stationary distributions of wealth in this model.

12
3 Empirical Analysis: Individual Saving and Inequality
3.1 Empirical Methodology
There are several issues that we need to address before testing the empirical implications of the
model presented in the previous section.
The first issue concerns the appropriate definition of saving. We use three different measures
that capture different angles of savings behaviour. First, we use a broad measure of saving,
the change in wealth between two consecutive waves of the PSID. Among other things this
measure includes realized and unrealized capital gains on housing, other real estate, financial
assets, and owned businesses. Second, we add estimates of saving through Social Security and
private pensions to the change in wealth to obtain our most comprehensive measure of saving.
Finally, we compute a measure of active saving as the difference between income exclusive of
capital gains and consumption.
Second, we would like to recover the effect of permanent income on savings, but we do
not observe permanent income. This is the key problem tackled by Dynan et al. (2004) and
we will follow their approach. The PSID reports current income which is a noisy measure
of permanent income since it includes transitory components of income and possibly some
measurement error. We deal with this problem using consumption, food consumption, years of
education of the head of the household, and average labor income as instruments for permanent
income. As pointed out by Dynan et al. (2004), different instruments may suffer from different
problems. For instance, to be able to use average labor income we need to assume that the
transitory component in income is uncorrelated over time. One could also think that education
could have a direct effect on saving decisions (Lusardi 1999), and this is why we include years
of education of the head of the household as an explanatory variable in our regressions and only
exclude it when we use education as an instrument. Furthermore instrumental variables provide
an estimate for a specific group, those individuals whose behaviour is affected by the instrument
(Imbens and Angrist 1995). Since not every individual is affected by the instruments in the
same way, the estimated coefficients will vary according to the instruments used. Because of the
different potential problems associated with each instrument and the heterogenous treatment
effects, we perform our empirical analysis using one instrument at a time. Our goal is to show
that the predictions of the model are robust to changes in specification.
Since current income, rather than permanent income, is also used in the construction of the
saving rate similar, although less acute, problems arise. We reduce the impact of transitory
income on the measured saving rate by computing savings over multi-year intervals and using
averages of current income over these same periods. As a result, the remaining transitory

13
income component or the presence of classical measurement error in our dependent variable is
unlikely to cause a bias in our estimates14 .
Two other key variables for testing the propositions derived from our model are reference
income and inequality. We measure income of the reference group as average income in the
state of residence and we use the coefficient of variation of disposable income in the state of
residence as our measure of inequality. Both variables vary across states and over time.
In contrast to Dynan et al. (2004), whose main focus is on the identification of the causal
effect of income on saving rates, our focus is on the impact of reference income and inequality
on saving rates and on the saving-income gradient. A causal interpretation of our results
requires ruling out the possibility of reverse causality. One concern about reverse causality
arises from the fact that saving rates vary systematically with income. As a consequence,
it may be the case that a portion of the changes in income inequality are the result of past
variation in saving rates. To deal with this concern, we show that results are similar when we
use a measure of inequality calculated using labor income, which excludes capital income and
therefore is independent of past saving choices15 . The existence of factors that simultaneously
affect both the dependent and independent variables could also affect the causal interpretation
of our results. To deal with this, we include state dummies in the regression to capture potential
unobserved drivers of differences in saving across states (e.g. cultural differences, say between
Minnesota and California). To rule out a potential relationship between saving and inequality
driven by business fluctuations we also include time dummies16 .
Once the causal nature of our relationships of interest is established, we will provide addi-
tional evidence that they are indeed due to interpersonal comparisons and not to other factors.
Even if the relationship between inequality and saving is causal, it could still be driven by fac-
tors other than interpersonal comparisons. For instance, differences in inequality across states
may be associated with differences in redistributive policies that also affect saving. In section 4,
we address this concern by including political variables in the regression. Moreover, we verify
that our results are not driven by interstate variation in housing prices that may be caused by
differences in inequality across states (Matlack and Vigdor 2008). In addition to this, results
14
Notice that the effect of the transitory component of income in the two wealth-based measures of the saving
rate will tend to induce a negative correlation between saving rates and income biasing the coefficient of income
downward. The concerns discussed so far were also faced by Dynan et al. (2004). We refer the interested reader
to their paper for a more detailed discussion.
15
Education is a particular type of saving and labor income reflects the returns to educational investment.
Since education is not included in our measure of savings, reverse causality through this channel does not seem
to be an issue. Nonetheless, in all our regressions we control for years of education.
16
Furthermore, Runkle (1991), using PSID data, finds that aggregate shocks are not very important for
individual consumption (and therefore saving) choices.

14
on reference income and saving also provide evidence on the role of interpersonal comparisons.
Finally, all our regressions include as controls several factors that may affect saving (or
consumption) decisions. We include age and age squared of the head of the household to
control for life-cycle variation in saving rates (Huggett and Ventura 2000), years of education
(Lusardi 1999), and a measure of household size (Attanasio and Browning 1995).
For most of our results we estimate a median regression with the savings rate as our depen-
dent variable, and a measure of income, reference group income or regional inequality, and the
other controls as explanatory variables. When we implement the IV estimation to recover the
effect of permanent income we use a two-stage procedure. Standard errors are bootstrapped
using 500 replications.

3.2 Data
We obtain all our information from the PSID, which is unique in providing a measure of assets
for a panel of households. We use 8 wealth supplements covering the years 1984, 1989, 1994,
1999, 2001, 2003, 2005, and 2007. Using the panel dimension of the data, we estimate each
household’s change in wealth and active saving. As noted by Bosworth and Anders (2008), very
wealthy households are underrepresented in the PSID, but for the bottom 95 percent of the
wealth distribution, the PSID wealth data yields very similar results to data from the Survey
of Consumer Finances (SCF) where this sample bias is not an issue.
We select our sample using the same criteria as Dynan et al. (2004). We restrict the
sample to households with a head between 30 and 60 years of age, measured at the middle of
the preceding 5-year period. As a result we avoid dealing with issues relevant to very young
households, such as liquidity constraints, and to very old ones, such as retirement or acute
health problems. We select households with an absolute value of active saving not larger than
$1,050,000 and with disposable income of at least $1,400.17 From the remaining sample, we
drop households in the extreme 0.5% of the distribution of saving. We also drop observations
for households that had a change in head. All the variables are adjusted for inflation using the
price deflator for personal consumption expenditures and are reported using 2007 dollars. We
end up with a sample of 24,036 observations about 8,063 households, with a minimum of 2,815
households in any given year.
For ease of comparison, we construct saving rates following the procedure outlined by Dy-
nan et al. (2004). For our first measure of saving, ∆ Wealth, we use information on the change
17
These figures, expressed in 2007 dollars, are equivalent to those in Dynan et al. (2004), expressed in 1994
dollars.

15
of eight asset categories net of their associated liabilities. These assets are home equity, other
real estate, private business/farm, vehicles, transaction accounts, corporate equities, annu-
ities/IRAs, and other savings. We also include information about non-collateralized debt. The
resulting measure of saving is very broad and even includes realized and unrealized capital gains
on housing.
Our second measure extends the previous one to include contributions to Social Security and
company pension schemes, ∆ Wealth + pensions. We include this as a separate measure since
not all of these contributions result from individual choices. To calculate these contributions
we add PSID data on employee’s contributions to private pension plans and expected Social
Security benefits.18 We compute these benefits subtracting net marginal Social Security tax
rates reported by Feldstein and Samwick (1992) and Cushing (2005) from gross Social Security
tax rates. While the gross Social Security tax rate is easy to compute –11.2% (12.5% after 1995)
of labor income up to $50,000– the net marginal tax rate (and the expected benefits) varies
with income, age and marital status of the tax payer. For instance, in 1994 a household in
Tennessee with four members whose head is 35 years old and earns the median income, around
$46,000, faced a net marginal Social Security tax rate of 5.59%. As a result, his Social Security
savings (or expected benefits) were 5.61% of his labor income.
The third measure, Active Saving, corresponds to the difference between income excluding
capital gains and consumption. This measure includes net investments in private pensions and
annuities, home equity, businesses, vehicles, stocks, cash assets and other financial assets plus
redemption of other debt. It is also adjusted for transfers of assets in and out of the household19 .
For the denominator of the saving rate we use either five-year or two-year averages of
household income, depending on the periodicity of the PSID wealth supplements. Household
income is total income after taxes and transfers including income from labor earnings, asset
flows, private transfers, private pensions, public transfers, and Social Security pensions minus
total taxes. When the numerator of the saving rate includes unrealized capital gains or pensions,
its denominator, the measure of income, is adjusted accordingly. For each household, the saving
rate is calculated as average saving over a multi-year period divided by average income over the
same period. Finally, we refer to the three saving rates associated to these measures of saving
as total saving rate, total saving rate plus pensions, and active saving rate, respectively.
Table 1 presents summary statistics for each of these variables. To reduce the influence of
18
Unfortunately the PSID does not include employer’s contributions to private pension plans. Since these
contributions are mostly concentrated in high-income households, this omission will tend to reduce the gradient
between saving rates and income.
19
See Appendix I for a more detailed description of the construction of our saving measures.

16
observations with very low income, the reported mean is computed dividing average annualized
saving by average income over the period. Since the income measures used to compute the
different saving rates differ, median income differs across measures. The reported median and
mean saving rates suggest that the measures of saving differ in breadth and that the distribution
of saving rates is skewed to the right. Comparing rows 1 and 2 one can see that the inclusion of
pensions increases the saving rate for the median saver by 4 percentage points. The summary
statistics for the distribution of saving rates are similar to those reported by Dynan et al. (2004)
and Bosworth and Anders (2008).
Finally, this table also presents summary statistics for households’s disposable income and
the other variables included in our regressions. We report different measures of income that
will be used at different stages of our analysis. Average labor income is computed using all
observations on labor income for the time period that a household remains in the sample
(1968-2007). On average this measure includes fourteen observations of yearly income.

3.3 Saving Rates, Current Income and Inequality

All the regressions include age, age squared and years of education of the head of the household,
size of the household, and year and state dummies as controls. Although not reported, their
associated coefficients are statistically significant throughout all our specifications20 . In Table 2,
we test the relationship between income and the saving rate using our three different measures
of saving. The first two columns show that the total saving rate increases with income. When
we introduce a quadratic income term the relationship is convex. This is consistent with the
findings of Dynan et al. (2004), who capture this non-linearity by allowing for quantile specific
slopes. Column 1 indicates that an increase in income of 10% translates into an increase in the
saving rate of half of a percentage point. Hence, increasing income of a household roughly at
the median of the income distribution ($46,000) by $20,000 increases his saving rate by 2.1
percentage points from about 10% to about 12%.21 Using the estimates of Dynan et al. (2004),
this same household would increase his saving rate by 2.2 percentage points22 . When we extend
our saving measure to include Social Security benefits and private pensions, saving rates become
slightly less responsive to income, but the relation remains positive and significant. Combined
with a larger constant (not reported), this indicates that the inclusion of Social Security raises
20
Furthermore, the coefficients on the socio-demographic controls have the expected signs. The effect of age
is hump-shaped. The effect of education is positive, while that of household size is negative.
21
The saving rate is computed for a 45-year old head of a household with 4 members who has 12 years of
education and lives in Oregon.
22
When using the PSID, Dynan et al. (2004) find that increasing a household’s income by 10,000 1982 dollars
(about 20,000 2007 dollars) implies a change in the saving rate of 2.2 percentage points.

17
the saving rate of the poor, weakening the saving-income relationship23 . The relationship
between saving and income remains positive and significant, although marginally weaker, when
we consider active savings, which excludes capital gains.
Proposition 2 states that individual saving rates decrease with the level of income of the ref-
erence group, and more so for poor households. To test this proposition, we regress the saving
rate on own income, our measure of income of the reference group –average state income–, and
an interaction between the two variables. This interaction aims at capturing the differential
impact of reference income across the income distribution. Our theory predicts that the coeffi-
cient on income of the reference group should be negative and the coefficient on the interaction
term positive. Results from this regression are reported in Table 3. Since the log of each income
measure ranges roughly from 8 to 13, the marginal effect of own income is positive for all agents
above the bottom percentile. Considering an individual with income and regional income equal
to the median of our sample, we find that the estimated marginal effects of income are similar
to those reported in Table 2. The sign of the marginal effect of average state income varies
across the income distribution. It is negative for poor households and positive for richer ones.
This positive effect on the saving rates of richer households, particularly those with incomes
above the mean, may be driven by the fact that average state income is an inadequate measure
of their reference income. In the next section, once we instrument for permanent income, we
will consider different reference groups for different individuals, all involving upward-looking
comparisons, and show that Proposition 2 holds for a much larger fraction of the income dis-
tribution. Finally, the signs, size, and significance of the coefficients are similar for different
measures of the saving rate and when we include a quadratic income term.
Proposition 3 states that an increase in income inequality leads to an increase in inequality
of saving rates. In a way, our results in Table 2, that simply reproduce those of Dynan et al.
(2004), already suggest this through a compositional effect. If the rich save more, increasing
income inequality will increase saving inequality. In our empirical exercise, we can go one
step further and test whether income inequality has a direct effect, beyond its impact through
income, on saving rates and the saving-income relationship. To test this, we compute the
coefficient of variation of disposable income for each year and state. We then regress the
household saving rate on income, the coefficient of variation, and the interaction between the
two. Our theory predicts that the coefficient on this interaction term should be positive, i.e.
23
Dynan et al. (2004) find that adding estimates of saving through Social Security and private pensions to
the total saving rate steepens the saving-income gradient. This difference might arise from the fact that their
estimates of private pensions are more accurate than ours since they include employer’s contribution to private
plans.

18
as inequality increases the gradient between saving rates and income becomes steeper. As
always, we also include our set of controls which includes state dummies. We conduct this
exercise for each one of the measures of saving. Table 4 shows the results of these regressions.
Our coefficient of interest is positive and significant as predicted by the theory. The negative
coefficient associated with inequality together with the positive interaction term implies that
income inequality has a negative effect on saving rates for households with low levels of income.
This negative effect becomes weaker, and after a certain point positive, as income increases.
Since our results control for the level of income they suggest that income inequality has more
than just a compositional effect on saving inequality24 .
Results are similar when we use other measures of inequality. Table 5 reproduces this
analysis using the Gini coefficient, the 90/10 percentile ratio and the standard deviation of
the log of labor income. The basic picture that emerges from these extensions is the same:
Inequality steepens the gradient between saving rates and income.

3.4 Saving Rates, Permanent Income and Inequality

In this section we reproduce the previous exercises using data on consumption, food consump-
tion, average labor income, and education as instruments for permanent income. We follow
a two-stage estimation procedure. In the first stage, we regress current income on one of the
instruments, the set of control variables, and measures of reference income or inequality. The
fitted values of this regression are the measures of permanent income that we use in the second
stage, where we estimate a median regression with the saving rate as our dependent variable
and predicted permanent income, reference income or inequality, and the other controls as
explanatory variables.
Tables 6, 8 and 10 report these results. While point estimates vary with the choice of
instrument, the magnitude and direction of the effects remain similar across specifications.
The high value of the F-test of the excluded instrument implies that all the instruments are
powerful. In the tables, we report results for the total saving rate and include a linear income
term as a regressor. We obtain similar results for other measures of saving or when we include
a quadratic income term. We report bootstrapped standard errors using 500 repetitions.
The first two columns of Table 6 present results using either consumption or food consump-
24
In our model a mean-preserving spread in the distribution of income increases inequality in saving rates
only through a compositional effect (since rich households become richer and poor households poorer). This is
simply a consequence of the fact that this spread does not affect the level of reference income, which is simply
average income. In the data changes in inequality most likely affect reference income. This change in reference
income is behind the direct effect of inequality on saving.

19
tion as an instrument. To measure consumption, we follow Dynan et al. (2004) and use a
weighted average of food at home, food away from home, rental payments and imputed hous-
ing flows. The estimated relationship between income and saving rates becomes steeper and
the negative impact of reference income in the lower tail of the income distribution becomes
stronger. Next, in column 3 we use average labor income as an instrument for permanent
income. In order to avoid problems of reverse causality, this instrument excludes any source of
capital income that could be correlated with past saving decisions. Furthermore, this measure
also excludes labor income from the current period. The signs and statistical significance of our
coefficients of interest remain unchanged. Finally, the last column of this table reports results
using years of schooling of the household head as an instrument25 . Again the same picture
emerges, saving rates increase with income, decrease with reference income, and more so for
low-income households.
Because of the interaction, it is hard to read marginal effects of Table 6 directly. Therefore,
the marginal effects of average state income at different percentiles of the income distribution are
reported in Table 7. A marginal effect of -0.39 (row 1, column 1) indicates that for households
in the 1st percentile of the income distribution, an increase in reference income of 10% reduces
their saving rate by 0.04 percentage points. Although small, this effect is far from negligible,
particularly if one considers that the median saving rate for households around that percentile
is zero. This negative effect disappears between the 35th and 45th percentiles depending on the
choice of instrument. As we pointed out in the previous section, this may be a consequence of
the fact that average state income is not an appropriate reference level for richer households. In
Table 8 we repeat the previous exercise restricting the sample to those households with income
above the mean and recalculating reference income as the average income of this restricted
sample. The sign, size, and significance of the coefficients remains unchanged. Nonetheless,
the marginal effect of reference income now remains negative for households up to the 90th
percentile of the (whole) income distribution, suggesting that increases in reference income
reduce saving rates for most households. Table 9 summarizes these marginal effects for the
restricted sample.
Table 10 explores the effects of income inequality on saving rates. The results are similar
to those reported in the previous section. After controlling for permanent income, inequality
has a negative impact on saving rates. This impact decreases with permanent income. Figure
2 illustrates the relationship between income and the saving rate implied by the coefficients
reported in column 3 of Table 10. The solid line represents the income-saving relationship in
25
In this case we exclude years of schooling from the set of controls.

20
a state with low inequality (Oregon) and the broken line that in a state with high inequality
(Texas). The gradient between saving rates and income is positive and becomes steeper in
states with high inequality. The effect of inequality on saving rates is substantial. For instance,
if we take a Texan at the median of the US income distribution, his total saving rate is 5
percentage points or 40% lower than that of a household with the same level of income living
in Oregon. Notice that inequality reduces saving rates for most of the income distribution.

3.5 Further Analysis

3.5.1 Sample composition and reverse causality

In this section we explore the robustness of our results to differences in the sample composition.
We recalculate our measures of reference income and inequality for each sample. In the analysis
that follows we will use average labor income as an instrument for permanent income and we
will focus on total saving –change in wealth– as our measure of saving. Nonetheless, results are
similar for the other measures of saving or when we use alternative instruments for permanent
income.
Excluding some observations at the top and bottom of the distribution of saving rates we
obtain some indication of the sensitivity of our results to extreme values. Results for different
trimmed samples are reported in Table 11. For ease of comparison, Column 1 reproduces
the results from our baseline sample. In the remaining columns, we use a sample that drops
households with saving rates at the top and bottom 1% and 5% of the distribution. Not
surprisingly, since we are using median regression, results are unaffected by this trimming.
Table 12 presents results after restricting the sample to households with different levels of
permanent income. On the one hand, excluding households with permanent income below a
certain threshold, $10,000 per year, we ensure that our results are not driven by the behavior
of the very poor, whose saving rates may simply reflect the presence of a subsistence level of
consumption. On the other hand, several authors have pointed out difficulties in rationalizing
the saving behavior of very rich households in terms of a standard life-cycle model, see for
instance Carroll (2000). Following a path that goes back, at least, to Weber (1958), Carroll
(2000) argues that the pursuit of wealth for its own sake may be behind the saving choices of
the very rich. Along these lines, excluding high-income households from our sample we ensure
that our results are not driven by what Carroll (2000) refers to as the “spirit of capitalism”. In
both cases there is always a fraction of the income distribution that saves less as average income
or inequality increase. These results suggest that our previous estimates were not driven by
the behavior of very poor or very rich households.

21
 Trimming our sample by income is also useful to infer whether the presence of asset-based
means tested government programs affects our results. These programs distort the saving in-
centives for marginally eligible households. In our context, this would be particularly important
if the eligibility criteria or coverage change over time and this change is somehow correlated
with inequality at the state level. By restricting our sample to households with permanent
income over $35,000, well above the eligibility criteria for most of these programs, we work
with a set of households whose saving decisions should not be affected by the presence of these
plans26 . Columns 3 and 6 in Table 12 report the results of this exercise. The estimates remain
similar to those reported using the full sample. In view of this we conclude that the presence
of asset-based means tested government programs is unlikely to be driving our results.
Results in Table 13 are obtained using a sample that excludes those states with the highest
and lowest levels of inequality. The three states with the highest level of inequality are Texas,
New York and California. Those with the lowest level are Oregon, Washington and Wisconsin.
Once more, the basic results remain unchanged.
Finally, one might think that differences in inequality across states may result from differ-
ences in the distribution of capital income that arise from different saving patterns across the
income distribution. If this was the case, past differences in saving would be at the root of
current differences in inequality. To show that our results are robust to this specific type of
reverse causality, we repeat our basic regression constructing our measures of inequality using
only labour income. Results, reported in Table 14, are consistent with our baseline estimates.

3.5.2 Does redistribution matter?

There are two, somewhat opposite, concerns related to the degree of redistribution at the state
level. First, along the lines suggested by Alesina and Rodrik (1994), states with more inequality
may engage in more redistribution of goods and services for which households typically save,
for instance health care or higher education. This may reduce saving in that state, particularly
for poor households since health costs represent a higher fraction of their income. In this
case inequality has an effect on saving rates but it does so through a channel different from
the one illustrated in our theoretical model. Inequality affects saving through redistribution
rather than through interpersonal comparisons. A second, and probably less relevant, concern
arises if more equal states are so as a result of redistribution, this redistribution mainly takes
26
Although the eligibility criteria of these programs is often stated in terms of income and wealth, we restrict
our sample only in terms of permanent income. By doing so, we assume that permanent income is highly
correlated with wealth. Finally, it is worth noticing that these programs tend to have a stronger impact on the
saving decisions of older households which are already excluded from our baseline sample.

22
the form of consumption goods and services, and these goods are necessities, i.e. they have
a low income elasticity of demand. One can think of food stamps or public transportation as
goods that satisfy these requirements. If this is the case, redistribution will reduce more than
proportionally the consumption needs of poor households providing them with additional room
for saving. As a result, low-income households may save more in more equal states.
This second concern arises if differences in inequality do simply reflect differences in redis-
tribution at the state level, i.e. more equal states are so because they redistribute more. Since
we have data on pre-tax labor income and income after taxes and transfers we can test the
extent to which redistribution drives the patterns of inequality across states27 . We regress the
estimated coefficient of variation of after-tax income on the coefficient of variation of pre-tax
income, a time trend and regional dummies. The adjusted-R squared of this regression is very
high, 0.95, and the coefficient on the coefficient of variation of pre-tax income is 0.83. This
suggests that pre-tax levels of inequality and state fixed characteristics account for almost all
the variation in inequality in the data.
The first concern requires that more unequal states engage in more redistribution. To
explore this hypothesis, we construct an index variable that indicates the political party of the
state governor during the period in which the savings decision took place. This variable goes
from 0, the state governor was a Democrat during the entire period, to 1, the state governor
was a Republican during the entire period. As a first approximation, this variable should be
a good proxy for the degree of generosity in the public provision of goods and services at the
state level.
The correlation between regional inequality and our governor index is positive (10%), indi-
cating that on average, states governed by Republicans have higher inequality. The scenario
described in the first concern thus appears questionable. In addition, Table 15 shows that when
the governor index is included in our regressions, our results remain unchanged. Moreover, this
index is only weakly significant, both on its own and when interacted with individual income.
To the extent that differences in state governor capture differences in the generosity of the wel-
fare state, our estimates of the effect of inequality on the saving-income relation do not seem
to be driven by differences in redistribution at the state level.
27
Notice that the denominator of any of our saving rates is income after taxes and transfers. This ensures
that we do not overstate the saving rates of low-income households in more redistributive states inducing an
spurious correlation between saving rates and inequality.

23
3.5.3 Housing prices

It is possible that income inequality leads to increases in the prices of certain goods, particularly
those that are locally in fixed supply. One can think of housing as a prime example. For
instance, focusing on tight housing markets, Matlack and Vigdor (2008) find evidence consistent
with this hypothesis. To the extent that this happens, higher housing prices would make
residents of more unequal states poorer in real terms. As a consequence, they will tend to save
less in line with the results of Dynan et al. (2004).
In Tables 16 and 17 we use a measure of real estate prices at the state level to test if the
effect of inequality on savings results from differences in price levels. We include state level
housing prices in our two standard specifications. Table 16 shows results using our standard
measure of saving, while Table 17 shows results for a measure that excludes home equity.
Results in Table 16 show that higher house prices are associated with higher saving. This
is only natural, as increases in real estate prices directly feed into higher saving rates through
capital gains on housing. The effects of reference group income and inequality become slightly
smaller but remain significant. When home equity is excluded from the measure of saving,
house prices have a negative effect on saving, particularly for the poor. The effect of reference
income weakens slightly although remains significant, while the effect of inequality remains
qualitatively similar but loses significance. Note that while this measure of saving is useful to
abstract from mechanical effects of changes in housing prices on saving through capital gains,
it also misses most asset holdings for the majority of households and therefore is unlikely to
yield strong results.28

3.5.4 Envy through life

In line with our theoretical assumption, there is evidence that suggests that the importance of
interpersonal comparisons may decrease over the life-cycle. The influence of reference groups
seems to be more important for people who are starting their professional careers or are still
trying to establish a certain status in society. As a result, this influence should become less
important as people age as we assume in our theoretical model. In this section, we explore
whether the data supports this assumption. For this purpose, we include in our benchmark ex-
ercise an interaction term between the key variable, income of the reference group or inequality,
and age of the head of the household. The rest of the specification remains unchanged and, in
28
Similarly, Bertrand and Morse (2012) argue that it is unlikely that inequality affects consumption through
home prices. They find that households consume a larger share of their income in more unequal places inde-
pendently of whether they are homeowners (who could finance consumption by loans against home equity) or
renters.

24
particular, it still includes age and age squared as controls29 .
Results are reported in Table 18. The effect of the interactions involving age is always
positive and significantly different from zero when using the income of the reference group.
This indicates that the negative effect of reference income on saving rates decreases as the head
of the household becomes older.

3.5.5 The marginal propensity to save under upward-looking comparisons

In this section we follow the procedure outlined by Dynan et al. (2004) to estimate the marginal
propensity to save (MPS). In line with our simple model of interpersonal comparisons we have
used average state income as reference income for all the household in each state. Nonetheless,
Veblen (1899), Duesenberry (1949), and Frank (2007) eloquently argue that the behavior of
successful individuals or groups sets the standard for the rest of the community. Ferrer-i
Carbonell (2005) provides convincing microeconometric evidence on the importance of upward-
looking comparisons as a determinant of subjective well-being. In view of this evidence, a more
realistic setup would allow for individuals to compare themselves with those who are ranked
above them in the income distribution30 . Under upward-looking comparisons, state average
income is only a sensible reference point for those households in the lower part of the income
distribution. Along these lines in this section we consider a different measure of reference
income. We classify households in a given state and year into permanent income quintiles. We
then estimate the reference income of each household as the average income of those quintiles
above him. We do that for all the individuals except those in the top quintile, for whom we
measure reference income as the average income of their own quintile31 .
We exploit the panel dimension of our dataset and regress changes in total saving over time
for a given household on changes in its income and changes in reference income. The coefficients
associated with the income quintile dummies are our measures of the quintile specific MPS.
Table 19 summarizes the results of this exercise. In both specifications the coefficients are
statistically different from zero and, except for the bottom two quintiles, different from each
other. In the second column we include the change in income of the reference group as a
regressor. Its coefficient is negative and statistically significant. To illustrate the implications
29
To obtain more variation in age, we extend our sample slightly by including households whose heads are
between 60 and 66 years old, still below the official retirement age of 67.
30
We will explore the theoretical implications of such an environment in the next section.
31
Results excluding the top quintile are similar. Another possibility is to use average income of individuals with
higher income as the reference for individuals in the highest income group. Although the results are unaffected,
this specification suffers from important problems of multicollinearity between variation in households’ level of
income and variation in income of the reference group.

25
of our result, consider again Figure 2(b). Between 1984 and 2007, for households in the lower
half of the income distribution, average income increased by $2,018 and average saving decreased
by $1,844. As discussed in the introduction, a MPS that increases with income is difficult to
reconcile with these patterns. Using the estimates in column 1 that abstract from the impact
of changes in reference income, one would predict that savings of the average household with
income below the median should have increased by $129 over this period. Results in column
2 allow us to control for the evolution of reference income. Since this measure increased by
$13,284 for this household, our estimates predict a drop in savings of $688.32

4 Aggregate Saving and Income Inequality

In the previous model we have characterized the savings behavior of an heterogeneous-agent
economy under the restrictive assumption that the reference group was common and equal to
the average household in the economy. In this section we relax this assumption and revisit the
relationship between aggregate savings and inequality. For the sake of illustration, we set ξ = 1
and assume there are only two homogeneous income groups, denoted by the superscripts H
(rich) and L (poor), with their wages given by wH and wL respectively, with wH > wL . The
population is evenly distributed between these two groups33 . In line with the evidence discussed
in the previous section, we assume that the reference group of the rich households is made up
only of rich households while the reference group of poor households is composed of a weighted
average of poor and rich households, with ρ being the weight placed on poor households. As a
result, reference consumption levels for the two groups are given respectively by

ḡtH = cH and ōH ¯H

t t+1 = dt+1 (18)

ḡtL = ρcLt + (1 − ρ)cH and ōLt+1 = ρd¯Lt+1 + (1 − ρ)d¯H

t+1 (19)

The levels of saving for a rich and a poor household are given by34 ,

sH H
t = β (1 + µ (1 − γ)) πyt (20)
� �
sLt = ψβ (1 + µ(1 − γρ)) ytL − µγ(1 − ρ)πytH (21)
32
This presupposes that the household with average income in the bottom 50% of the income distribution is
in the second quintile of the distribution. For this household, reference income therefore is the average income
of the top three quintiles.
33
For simplicity let’s assume that in our generation of interest the bequest levels are common across types
and equal to zero.
34
See Appendix I for a detailed derivation.

26
 1 1
where 0 < ψ ≡ <π≡ < 1.
1 + µβ (1 − γρ) + β 1 + µβ (1 − γ) + β
Although this model has similar implications for individual saving than the previous one,
its aggregate implications differ. The following proposition summarizes the impact of income
inequality on aggregate savings in the presence of upward-looking comparisons.
Proposition 4. An increase in income inequality, a mean preserving spread in the distri-
bution of wages, leads to a decrease in the aggregate (average) saving rate.
yH
Proof. Defining the share of total income received by the rich households by ysH ≡ H ,
y + yL
we combine (20) and (21) to derive the aggregate (average) saving rate as,

� � L � �
sagg H H H
rate = ys srate + 1 − ys srate = ψβ (1 + µ(1 − γρ)) − 2µγ(1 − ρ)πysH (22)

which is decreasing in the share of income received by rich households, ysH , and therefore in
inequality.
In this framework, where upward-looking interpersonal comparisons matter, a mean pre-
serving spread in the distribution of wages reduces the aggregate level of saving and therefore
the aggregate saving rate. Although an increase in the wage of rich households leads to a pro-
portional increase in their level of saving, leaving their saving rate unchanged, it is clear from
(21) that this increase leads to a decrease not only in the level of saving of the poor but also
in their saving rate. This decrease in saving of poor households is larger than the increase in
saving of the rich.
Remark to Proposition 4. A distributionally-neutral increase in income, i.e. an in-
crease in income that leaves unchanged the shares of aggregate income earned by rich and poor
households, leaves the aggregate saving rate unchanged.
It is evident from (22) that the aggregate saving rate remains constant as long as the share
of income of rich households remains constant. This implies this last remark.

5 Empirical Analysis: Inequality and Aggregate Saving

This last proposition states that an increase in inequality should reduce the aggregate saving
rate. Our previous results suggest that this relationship holds in US micro data. For instance,
Figure 4 indicates that individuals below the 85th percentile of the permanent income distri-
bution save less in more unequal states. In this section, we test again Proposition 4 using panel
data on six OECD countries. We estimate a simple saving specification that includes a measure
of income inequality together with other standard determinants of saving. Our theory predicts
that the coefficient on inequality should be negative.

27
 In contrast to previous cross-country studies on the determinants of aggregate saving that
focus on measures of saving derived from the current account and therefore include elements
of public or corporate saving (Edwards (1996), Schmidt-Hebbel and Serven (2000),Li and Zou
(2004), Leigh and Posso (2009)), we will use a measure of the personal saving rate, from
several OECD yearbooks, as our dependent variable. This measure is closer to our theory,
which focuses on the determinants of household saving and abstracts from other components of
national saving. Reasonably long series for the personal saving rate are not available for many
countries, so we are forced to restrict our sample to Canada, France, Germany, Japan, the UK,
and the US. Although the number of countries in our sample is not large, the time dimension
is, covering the years 1955 to 2007. Over this period our sample accounts for roughly 40% of
world GDP.
Our measure of inequality is the fraction of total income accruing to the top five per cent of
the income distribution.35 We obtain this measure from Alvarado et al. (2012). This variable
is available at a yearly frequency, except for Germany, where it is measured on a tri-annual
basis. We follow Schmidt-Hebbel and Serven (2000) and include the age dependency ratio, the
rate of growth of real GDP per capita, and the real interest rate as additional regressors.36
Results are reported in Table 20. Column 1 shows a negative association between aggre-
gate saving rates and income inequality, controlling for other determinants of saving. This
relationship remains robust to the inclusion of country-fixed effects, as shown in column 2. In
both specifications, the coefficient on inequality is negative and significant. The results from
column 2 suggest that a percentage point increase in the share of income accruing to the top
five percent of the income distribution is associated with a reduction in the personal saving
rate of 0.26 percentage points. The evolution of the share of income of the top five percent,
reported in Figure 2(c), together with this estimate suggests that around half of the decrease in
the US personal saving rate (3.38 percentage points) can be associated to the process of income
concentration that began more than three decades ago.37
In the next two columns, we make sure that our results are not due to unique events in
35
We have estimated the same specification using other conventional indicators of inequality like the Gini
coefficient or the ratio of the income shares of the richest 5 percent and the bottom 95 percent, with similar
results.
36
Notice that these authors include the level of per capita GDP and its square to capture the non-linear
relationship between income and savings. This is particularly important when the variation in income through
time or across countries is substantial. This is not the case in our sample. Nonetheless, the sign and significance
of our coefficient of interest remains unchanged when we replace the rate of growth of per capita GDP by its
level.
37
The remaining coefficients also have the expected signs. The dependency ratio has a negative effect on the
aggregate saving rate while in most specifications the real interest rate and the GDP growth rate have a positive
effect.

28
certain periods or business cycles. We do so by including decade dummies (column 3) and by
using 5-year averages of our yearly data (column 4). In both cases, the negative relationship
between saving and inequality remains significant.38
At the aggregate level, the growth rate of per capita GDP and the interest rate are jointly
determined with the saving rate. As a result, endogeneity becomes a concern. To deal with
this issue, we estimate the model using Blundell and Bond’s (1998) system GMM estimator
with one lag. To further reduce endogeneity concerns, we use 5-year averaged data to extend
the length of the lag. Columns 5 and 6 show that the coefficient on inequality remains negative
and significant.
Finally, it is worth noticing that the timing of income in our theoretical model is equivalent
to the assumption that credit markets are perfect. In this sense, one should expect the negative
effect of inequality on aggregate saving through interpersonal comparisons to be reduced by
the presence of credit market imperfections. These imperfections would prevent low income
households from borrowing against their future income to emulate the consumption of richer
households. To allow for this relationship, we control for a measure of financial development and
its interaction with inequality in column 7. Among the several popular measures of financial
development in the literature, we choose stock market capitalization because this measure,
unlike ratios of credit or deposits to GDP, is less directly related to the saving rate. Results are
consistent with the predictions of the model. The relationship between inequality and saving
remains negative, and more so for higher levels of financial development.39
Despite limited availability of data – which limits the set of econometric techniques that
can be used – we can thus establish a negative relationship between inequality and saving not
only in individual-level data, but also in aggregate data. We see these results as promising;
with more data on private saving, we could probe the robustness of these results with more
detail and take the analysis to the level of the micro data analysis.

6 Conclusions
We have presented two models, four propositions, and some statistical evidence on the im-
portance of interpersonal comparisons and its connection with the positive gradient between
saving rates and income documented by Dynan et al. (2004). The intuition behind our theo-
38
Furthermore, we estimate the model dropping one country at a time to verify that our results are not driven
by the behaviour of any individual country. The basic results remain unchanged.
39
Unfortunately, we are limited to using panel fixed effects with yearly data in this regression due to limited
data availability. The World Bank’s measure of stock market capitalization only covers the years 1989 to 2007.
Taking five-year averages would leave us with only two observations per country.

29
retical model is simple. When households care about relative consumption, the comparison of
their consumption with that of their peers, the level of individual saving is determined by the
interaction of two opposing forces. First, saving increases in the household’s lifetime income, as
in the permanent income hypothesis. Second, as a result of interpersonal comparisons, saving
decreases in the lifetime income of the reference group, as in the relative income hypothesis.
In addition, our model suggests that the negative effect of reference income on saving rates
is stronger for low-income households, that increases in income inequality translate into in-
creases in inequality of saving rates, and that, in the presence of upward-looking interpersonal
comparisons, changes in income inequality affect aggregate saving.
We use more than two decades of PSID data and a long, although thin, panel of OECD
countries to test these implications. The data supports the basic predictions of our theory. Like
Dynan et al. (2004), we find that saving rates increase in permanent income. Approximating
income for the reference group by average state income, we find that, after controlling for own
income, saving rates decrease in comparison income and they do so more for poor households.
When we include a measure of income inequality, we find that those regions or periods with
higher inequality are characterized not only by a more unequal distribution of saving rates but
also by lower saving rates for most of the income distribution. These results do not seem to
be driven by measurement error, reverse causality, differences in redistribution across time or
states, or variation in real estate prices. Finally, although preliminary, our cross-country results
suggest that the distribution of income may have important aggregate implications.
In their seminal paper, Dynan et al. (2004) conclude that “the rich do, indeed, save more”
or, equivalently, that non-rich households save less. Our analysis suggests one of the reasons
for these lower rates of saving, the importance of interpersonal comparisons.

30
7 Tables and Figures

Figure 1: Motivation Figures.

.35
.3
.25
.2
.15
.1
.05
Saving
5
0
50000
150000
100000
permanent
0000 rate income
(change in wealth)
.35
.3
Saving rate (change in wealth)

.25
.2
.15
.1
.05

0 50000 100000

permanent income

(a) Savings across the income distribution. (b) Evolution of income and savings for households
with low income.
Saving
Aggregate
Share
35
30
25
1
5
0
20
15
10
1960
1970
1980
1990
2000
2010
0 of
Rate
income
SavingofRate
the top 5%
20

35

Share of income of the top 5%

15

30
Saving Rate
10

25
5

20
0

1960 1970 1980 1990 2000 2010

Aggregate Saving Rate Share of income of the top 5%

(c) Evolution of the household saving rate and inequal-

ity in the US.

Note: (a) Saving is the change in wealth over a period. Permanent income is the average labor income of the
household over the period in which the household is in the sample (in 2007 dollars). (b) Saving rate is the
average change in wealth over a period for households below the median of permanent income divided by their
average disposable income. Disposable income is the average after tax household income in 2007 dollars for
households below the median of permanent income. The figure reports 5-year moving averages except 1999,
which is an average of 1994, 1999 and 2003. (c) Inequality is measured as the share of income of the top 5%
of the distribution from the update of Piketty and Saez (2003). The personal saving rate is obtained from the
Federal Reserve Economic Data.

31
 Table 1: Summary statistics.
saving median mean median income
total saving rate 0.07 0.21 47,115
total saving rate + pension 0.11 0.24 49,323
active saving rate 0.04 0.12 46,199
income median mean std. dev.
disposable income (after tax) 46,526 56,497 54,137
disposable income (pre tax) 56,332 72,093 83,346
labour income (pre tax) 53,510 69,304 82,824
avg. labour income 52,288 61,570 56,512
demographics median mean std. dev.
head’s years education 12 13.09 2.44
head’s age 45 45.53 8.05
household size 3 3.02 1.52

Note: All saving and disposable income measures are expressed in 2007 dollars. Measures of income used to
compute the saving rates: ∆Wealth: disposable income + capital gains from housing, other real estate, business
and stocks, ∆Wealth + pensions: disposable income + capital gains + employer contributions to Social Security
and pensions. Median saving rate equals median of the ratio of saving to income. Average saving rate equals
average saving divided by average income.

Table 2: Saving rates and income.

∆Wealth ∆Wealth + pensions Active
ln(income) 0.049*** -0.418*** 0.038*** -0.369*** 0.039*** -0.299***
(0.002) (0.024) (0.002) (0.027) (0.002) (0.021)
ln(income)2 0.023*** 0.020*** 0.017***
(0.001) (0.001) (0.001)
Pseudo R2 0.010 0.012 0.010 0.010 0.001 0.001
N 24,036 24,036 24,036 24,036 24,036 24,036

Note: The column heading indicates the saving measure used. ∆Wealth refers to the change in wealth or total
saving. Active saving is defined as total saving minus capital gains. Income is current after tax household income
in 2007 dollars. All regressions also include controls for age, age squared, household size, head’s education and
year and state dummies which are not reported. Standard errors are reported in parentheses. *** significant
at less than 1%; ** significant at 5%; * significant at 10%

32
 Table 3: Saving rates, income and reference group.
∆Wealth ∆Wealth + pensions Active
ln(income) -0.974*** -1.022*** -0.899*** -1.002*** -0.742*** -0.761***
(0.089) (0.086) (0.093) (0.095) (0.087) (0.078)
ln(income)2 0.021*** 0.017*** 0.016***
(0.001) (0.001) (0.001)
ln(state income) -0.969*** -0.619*** -0.902*** -0.678*** -0.712*** -0.427***
(0.089) (0.090) (0.093) (0.099) (0.087) (0.081)
ln(state income)× 0.095*** 0.061*** 0.087*** 0.065*** 0.073*** 0.044***
ln(income) (0.008) (0.008) (0.009) (0.009) (0.008) (0.008)
Pseudo R2 0.011 0.012 0.010 0.010 0.001 0.001
N 24,036 24,036 24,036 24,036 24,036 24,036

Note: The column heading indicates the saving measure used. ∆Wealth refers to the change in wealth or total
saving. Active saving is defined as total saving minus capital gains. Income is current after tax household income
in 2007 dollars. State income is the state level average of this measure. All regressions also include controls
for age, age squared, household size, head’s education and year and state dummies which are not reported.
Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at 5%; * significant
at 10%

Table 4: Saving rates, income and state inequality.

∆Wealth ∆Wealth + pensions Active
ln(income) 0.037*** -0.417*** 0.028*** -0.360*** 0.030*** -0.302***
(0.005) (0.024) (0.005) (0.024) (0.004) (0.025)
ln(income)2 0.023*** 0.020*** 0.017***
(0.001) (0.001) (0.001)
inequality -0.124*** -0.053** -0.117*** -0.039** -0.104*** -0.047*
(0.043) (0.027) (0.043) (0.023) (0.038) (0.027)
inequality × 0.011*** 0.004** 0.010** 0.003* 0.010*** 0.004
ln(income) (0.004) (0.002) (0.004) (0.002) (0.004) (0.003)
Pseudo R2 0.011 0.012 0.010 0.010 0.001 0.001
N 24,036 24,036 24,036 24,036 24,036 24,036

Note: The column heading indicates the saving measure used. ∆Wealth refers to the change in wealth or total
saving. Active saving is defined as total saving minus capital gains. Income is current after tax household
income in 2007 dollars. Inequality is measures as the coefficient of variation of disposable income per region an
year. All regressions also include controls for age, age squared, household size, head’s education and year and
state dummies which are not reported. Standard errors are reported in parentheses. *** significant at less than
1%; ** significant at 5%; * significant at 10%

33
 Table 5: Saving rates and current income – Other measures of inequality.
cv(income) gini p90/p10 s.d(ln(income))
ln(income) 0.037*** 0.012 0.049*** 0.033***
(0.005) (0.017) (0.007) (0.010)
inequality -0.124*** -0.982** -0.001 -0.183*
(0.043) (0.434) (0.008) (0.110)
ln(income)× 0.011*** 0.087** 0.0002 0.017*
inequality (0.004) (0.040) (0.001) (0.010)
Pseudo-R2 0.011 0.011 0.011 0.011
N 24,036 24,036 24,036 24,036

Note: The column heading indicates the measure of inequality used. As a measure of saving we use change in
wealth. All regressions also include controls for age, age squared, household size, head’s education and year and
state dummies which are not reported. Standard errors are reported in parentheses. *** significant at less than
1%; ** significant at 5%; * significant at 10%

Table 6: Saving rates, permanent income and reference group.

Consumption Food Avg Labor Education
Consumption Income
ln(perm. income) -1.623*** -1.575*** -1.552*** -1.424***
(0.250) (0.271) (0.174) (0.193)
ln(state income) -1.745*** -1.676*** -1.565*** -1.508***
(0.240) (0.262) (0.171) (0.188)
ln(perm.income)× 0.164*** 0.157*** 0.150*** 0.143***
ln(state income) (0.023) (0.025) (0.016) (0.018)
Pseudo-R2 0.008 0.008 0.010 0.008
F -excluded instruments 2,753 760 10,492 5,088
N 20,578 20,195 21,360 24,036

Note: The column heading indicates the instrument used in the regression. As a measure of saving we use change
in wealth. Perm. income is a measure of permanent income estimated using an OLS regression, including the
instrument and our set of controls. State income is the state level average of this measure. All regressions also
include controls for age, age squared, household size, head’s education and year and state dummies which are
not reported. Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at 5%;
* significant at 10%

34
Table 7: Marginal effect of state income on saving rates across the income distribution. Using
permanent income.
perm. income distribution marginal effect of state income
percentile household consumption food average education
income consumption labor income
1 $ 3,664 -0.396 -0.385 -0.329 -0.336
5 $ 9,191 -0.245 -0.241 -0.190 -0.205
10 $ 14,597 -0.169 -0.168 -0.121 -0.139
15 $ 19,640 -0.120 -0.121 -0.076 -0.097
20 $ 23,882 -0.088 -0.091 -0.047 -0.069
30 $ 31,181 -0.044 -0.049 -0.006 -0.031
40 $ 38,533 -0.010 -0.015 0.025 0.000
50 $ 46,526 0.021 0.014 0.054 0.026

Note: Marginal effects computed using results from table 7. All saving and disposable income measures are
given in 2007 dollars.

Table 8: Saving rates, permanent income and reference group– changing the sample and the
reference group
Consumption Food Avg Labor Education
Consumption Income
ln(perm. income) -2.175*** -4.060*** -2.466*** -3.613***
(0.664) (0.783) (0.537) (0.518)
ln(state income) -2.956*** -2.762*** -4.734*** -4.215***
(0.687) (0.560) (0.818) (0.543)
ln(perm. income)× 0.243*** 0.405*** 0.242*** 0.362***
ln(state income) (0.061) (0.072) (0.049) (0.048)
Pseudo-R2 0.020 0.015 0.017 0.015
F -excluded instruments 1,391 283 4,459 869
N 9,632 9,600 10,244 11,291

Note: The column heading indicates the instrument used in the regression. As a measure of saving we use change
in wealth. Perm. income is a measure of permanent income estimated using an OLS regression, including the
instrument and our set of controls. State income is the state level average of this measure. All regressions also
include controls for age, age squared, household size, head’s education and year and state dummies which are
not reported. Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at 5%;
* significant at 10%

35
Table 9: Marginal effect of state income on saving rates across the income distribution. Using
permanent income and restricting the sample above average.
perm. income distribution marginal effect of state income
percentile household consumption food average education
income consumption labor income
55 $ 51,038 -0.324 -0.343 -0.136 -0.291
60 $ 55,778 -0.302 -0.307 -0.114 -0.259
65 $ 60,343 -0.283 -0.275 -0.095 -0.230
70 $ 65,582 -0.263 -0.241 -0.075 -0.200
75 $ 71,578 -0.242 -0.206 -0.054 -0.168
80 $ 78,636 -0.219 -0.168 -0.031 -0.134
85 $ 87,967 -0.192 -0.122 -0.004 -0.094
90 $ 101,830 -0.156 -0.063 0.032 -0.041
95 $ 127,992 -0.101 0.030 0.087 0.042

Note: Marginal effects computed using results from table 8. All saving and disposable income measures are
given in 2007 dollars.

Table 10: Saving rates, permanent income and inequality.

Consumption Food Avg Labor Education
Consumption Income
ln(perm. income) 0.132*** 0.103*** 0.038*** 0.090***
(0.018) (0.022) (0.013) (0.010)
inequality -0.149 -0.227* -0.339*** -0.254***
(0.127) (0.127) (0.134) (0.104)
ln(perm. income) × 0.013 0.020 * 0.030*** 0.023***
inequality (0.012) (0.012) (0.012) (0.009)
Pseudo-R2 0.009 0.008 0.010 0.008
F -excluded instruments 2,574 781 10,522 5,722
N 20,578 20,195 21,360 24,036

Note: The column heading indicates the instrument used in the regression. As a measure of saving we use change
in wealth. Perm. income is a measure of permanent income estimated using an OLS regression, including the
instrument and our set of controls. Inequality is measured as the coefficient of variation of disposable income per
state and year. All regressions also include controls for age, age squared, household size, head’s education and
year and state dummies which are not reported. Standard errors are reported in parentheses. *** significant
at less than 1%; ** significant at 5%; * significant at 10%

36
 Figure 2: The saving rate by income percentile in states with different levels of inequality.

0.25

0.2

0.15
(change in wealth)
Saving Rate

0.1

0.05

0
10 20 30 40 50 60 70 80 90

-0.05

-0.1
Income Distribution

Low Inequality (cv=0.51) High Inequality (cv=3.71)

Note: Percentiles of the permanent income distribution on the x-axis. Saving rates are predicted values from
the regression in column 4 in Table 10, for 2005, for a 45-year old head of a household with 4 members and 12
years of education. Percentiles of the permanent income distribution used in the prediction refer to the entire
sample. Therefore, the level of income at each point is the same for both groups of states. The values of the
coefficient of variation correspond to the minimum and maximum in the sample across regions. (note: low being
Oregon and high being Texas.)

37
Table 11: Saving rates and permanent income – altering the sample composition: eliminating
extreme values of saving rates.
reference group inequality
baseline excluding excluding baseline excluding excluding
1% 5% 1% 5%
ln(perm. income) -1.552*** -1.629*** -1.538*** 0.038*** 0.044*** 0.039***
(0.174) (0.180) (0.160) (0.013) (0.013) (0.011)
ln(state income) -1.565*** -1.643*** -1.538***
(0.171) (0.177) (0.157)
ln(perm. income)× 0.150*** 0.158*** 0.149***
(state income) (0.016) (0.017) (0.015)
inequality -0.339*** -0.291*** -0.305***
(0.134) (0.131) (0.119)
ln(perm. income)× 0.030*** 0.026*** 0.027***
inequality (0.012) (0.012) (0.011)
Pseudo R2 0.010 0.026 0.038 0.010 0.025 0.036
N 21,360 21,162 19,472 21,360 21,162 19,472

Note:The column heading indicates the way the sample has been altered. The percentage reported has been
eliminated from both sides of the distribution. As a measure of saving we use change in wealth. Perm. income
is a measure of permanent income estimated using an OLS regression, including the instrument and our set of
controls. Inequality is measured as the coefficient of variation of disposable income per state and year. State
income is the state level average of this measure. All regressions also include controls for age, age squared,
household size, head’s education and year and state dummies which are not reported. Standard errors are
reported in parentheses. *** significant at less than 1%; ** significant at 5%; * significant at 10%

38
Table 12: Saving rates and permanent income – altering the sample composition: restricting
the sample by income.
Dependent variable: ∆Wealth
reference group inequality
Income> Income< Income> Income> Income< Income>
10,000 100,000 35,000 10,000 100,000 35,000
ln(perm.income) -2.514*** -1.763*** -3.950*** 0.053*** 0.027** 0.136***
(0.304) (0.285) (0.657) (0.017) (0.013) (0.033)
ln(state income) -2.544*** -1.685*** -4.177***
(0.300) (0.280) (0.659)
ln(perm. income)× 0.240*** 0.173*** 0.372***
ln(state income) (0.028) (0.027) (0.059)
inequality -0.432*** -0.362*** -0.199*
(0.062) (0.047) (0.105)
ln(perm.income)× 0.039** 0.033*** 0.016*
inequality (0.016) (0.004) (0.010)
Pseudo R2 0.010 0.007 0.011 0.010 0.006 0.010
N 20,287 18,959 14,180 20,287 18,959 14,180

Note: The column heading indicates the way the sample has been altered. As a measure of saving we use change
in wealth. Perm. income is a measure of permanent income estimated using an OLS regression, including the
instrument and our set of controls. State income is the state level average of this measure. Inequality is measured
as the coefficient of variation of disposable income per state and year. State income is the state level average
of this measure. All regressions also include controls for age, age squared, household size, head’s education and
year and state dummies which are not reported. Standard errors are reported in parentheses. *** significant
at less than 1%; ** significant at 5%; * significant at 10%

39
Table 13: Saving rates and permanent income – altering the sample composition: eliminating
most and least unequal regions.
baseline excluding excluding excluding
low inequality high inequality high&low
ln(perm. income) 0.038*** 0.042*** 0.025 0.038***
(0.013) (0.011) (0.017) (0.016)
inequality -0.339*** -0.285*** -0.408** -0.249*
(0.134) (0.114) (0.195) (0.140)
ln(perm. income) × 0.030*** 0.025*** 0.038** 0.022
inequality (0.012) (0.010) (0.018) (0.017)
Pseudo-R2 0.010 0.010 0.009 0.008
N 21,360 20,235 17,218 16,093

Note: The column heading indicates the regions that have been excluding from the sample: high inequality
regions are Texas, New York and California. Low inequality regions are Oregon, Washington and Wisconsin. As
a measure of saving we use change in wealth. Perm. income is a measure of permanent income estimated using
an OLS regression, including the instrument and our set of controls. Inequality is measured as the coefficient
of variation of disposable income per state and year. State income is the state level average of this measure.
All regressions also include controls for age, age squared, household size, head’s education and year and state
dummies which are not reported. Standard errors are reported in parentheses. *** significant at less than 1%;
** significant at 5%; * significant at 10%

40
 Table 14: Saving rates and permanent income – using only labour income.
reference group inequality
baseline labour income baseline labour income
ln(perm. income) -1.553*** -1.166*** 0.038*** 0.023***
(0.174) (0.182) (0.013) (0.013)
ln(state income) -1.565*** -1.187***
(0.171) (0.175)
ln(perm. income)× 0.151*** 0.114***
ln(state income) (0.016) (0.017)
inequality -0.339*** -0.457***
(0.134) (0.132)
ln(perm. income)× 0.030*** 0.042***
inequality (0.012) (0.012)
Pseudo-R2 0.010 0.009 0.010 0.010
N 21,360 21,360 21,360 21,360

Note: The column heading indicates if the income of the reference group and the coefficient of variation of
income (inequality measure) have been computed using total disposable income or only labour income per state
and year. As a measure of saving we use change in wealth. Perm. income is a measure of permanent income
estimated using an OLS regression, including the instrument and our set of controls. Inequality is measured as
the coefficient of variation of disposable income per state and year. State income is the state level average of
this measure. All regressions also include controls for age, age squared, household size, head’s education and
year and state dummies which are not reported. Standard errors are reported in parentheses. *** significant
at less than 1%; ** significant at 5%; * significant at 10%

41
 Table 15: Saving rates and permanent income – controlling for redistribution.
baseline reference group inequality
ln(perm. income) 0.067*** 0.062*** -1.503*** 0.036***
(0.006) (0.005) (0.161) (0.012)
governor index 0.008* -0.111 0.006 0.007*
(0.005) (0.083) (0.004) (0.004)
ln(perm. income)× 0.011*
governor index (0.007)
ln(state income) -1.521***
(0.156)
ln(perm. income)× 0.146***
ln(state income) (0.015)
inequality -0.359***
(0.126)
ln(perm. income)× 0.032***
inequality (0.012)
Pseudo-R2 0.009 0.009 0.010 0.010
N 21,076 21,076 21,076 21,076

Note: As a measure of saving we use change in wealth. Perm. income is a measure of permanent income
estimated using an OLS regression, including the instrument and our set of controls. The governor index goes
from 0, the state governor was a Democrat during the period in which the saving decision took place, to 1, the
state governor was a Republican during the entire period. Inequality is measured as the coefficient of variation
of disposable income per state and year. State income is the state level average of this measure. All regressions
also include controls for age, age squared, household size, head’s education and year and state dummies which
are not reported. Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at
5%; * significant at 10%

42
 Table 16: Saving rates and permanent income – controlling for regional housing prices.
reference group inequality
baseline local prices baseline local prices
ln(perm. income) -1.553*** -0.947*** 0.038*** -0.066***
(0.174) (0.262) (0.013) (0.029)
ln(state income) -1.565*** -0.920***
(0.171) (0.244)
ln(perm. income)× 0.151*** 0.085***
ln(state income) (0.016) (0.022)
inequality -0.339*** -0.159*
(0.134) (0.089)
ln(perm. income) 0.030*** 0.014*
inequality (0.012) (0.008)
housing prices -0.006*** -0.008***
(0.002) (0.002)
ln(perm income)× 0.001*** 0.001***
housing prices (0.0002) (0.0003)
Pseudo-R2 0.010 0.010 0.010 0.010
N 21,360 19,542 21,360 19,542

Note: As a measure of saving we use change in wealth. Perm. income is a measure of permanent income
estimated using an OLS regression, including the instrument and our set of controls. Housing prices is a measure
of real estate prices at the state level. Inequality is measured as the coefficient of variation of disposable income
per state and year. State income is the state level average of this measure. All regressions also include controls
for age, age squared, household size, head’s education and year and state dummies which are not reported.
Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at 5%; * significant
at 10%

43
Table 17: Saving rates (excluding home equity) and permanent income – controlling for regional
housing prices.
reference group inequality
baseline local prices baseline local prices
ln(perm income) -0.491*** -0.339*** 0.022*** -0.005
(0.119) (0.160) (0.007) (0.010)
ln(state income) -0.497*** -0.343***
(0.118) (0.162)
ln(perm income)× 0.048*** 0.032***
ln(state income) (0.012) (0.013)
inequality -0.061 -0.032
(0.048) (0.052)
ln(perm income)× 0.005 0.003
inequality (0.005) (0.006)
housing prices -0.001*** -0.002***
(0.0003) (0.0003)
ln(perm income)× 0.0001*** 0.0002***
housing prices (0.00003) (0.00005)
Pseudo-R2 0.003 0.003 0.003 0.003
N 21,360 19,542 21,360 19,542

Note: As a measure of saving we use change in wealth excluding home equity. Perm. income is a measure of
permanent income estimated using an OLS regression, including the instrument and our set of controls. Housing
prices is a measure of real estate prices at the state level. Inequality is measured as the coefficient of variation
of disposable income per state and year. State income is the state level average of this measure. All regressions
also include controls for age, age squared, household size, head’s education and year and state dummies which
are not reported. Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at
5%; * significant at 10%

44
 Table 18: Saving rates and permanent income – Envy through life.
reference group inequality
baseline age baseline age
ln(perm income) -1.553*** -1.466*** ln(perm income) 0.038*** 0.036***
(0.174) (0.174) (0.013) (0.012)
ln(state income) -1.565*** -1.659*** inequality -0.339*** -0.369***
(0.171) (0.183) (0.134) (0.132)
ln(perm income) × 0.150*** 0.143*** ln(perm income) × 0.030*** 0.032***
ln(state income) (0.016) (0.016) ln(inequality) (0.012) (0.012)
Age× 0.004*** Age× 0.0003
ln(state income) (0.001) inequality (0.0006)
Pseudo-R2 0.010 0.010 0.010 0.010
N 21,360 21,360 21,360 21,360

Note: Columns 2 includes an interaction of average state income and age and column 4 includes an interaction
of age and inequality. As a measure of saving we use change in wealth. Perm. income is a measure of permanent
income estimated using an OLS regression, including the instrument and our set of controls. Housing prices
is a measure of real estate prices at the state level. Inequality is measured as the coefficient of variation of
disposable income per state and year. State income is the state level average of this measure. All regressions
also include controls for age, age squared, household size, head’s education and year and state dummies which
are not reported. Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at
5%; * significant at 10%

45
 Table 19: Marginal propensity to save.
change in saving
∆(∆Wealth)
Quantile 1 0.040 0.049
(0.024) (0.027)
Quantile 2 0.064 0.067
(0.016) (0.018)
Quantile 3 0.069* 0.078*
(0.009) (0.011)
Quantile 4 0.139* 0.111*
(0.008) (0.010)
Quantile 5 0.327* 0.328*
(0.002) (0.003)
∆income -0.062***
reference group (0.008)
Pseudo-R2 0.007 0.007
N 16,798 14,236

Note: This table shows estimates of marginal propensity to save for each quintile obtained from changes in
saving over time. The dependent variable is the change in household saving between periods divided by the
number of years in the period. Reported in the table are marginal effects of changes in household income
between periods by income quintile. They are computed from the coefficient on the interaction between the
change in income and the income quintile indicators. Column 2 shows the results of the specification in which
we included the change in income of the reference group, estimated as the average income of the quintiles above
the quintile to which the individual belongs for a given space and a given year. Note that for quintile 5 we use
the average income of the quintile as income of the reference group. Also included in the regression (coefficients
not reported in the table) are the a set of indicator variables for permanent income, age, a quadratic term in
age, the number of persons in the household, education and year and regional dummies. Standard errors are
reported in parentheses. * The coefficient is significantly greater than that for the previous quintile, on the
basis of a one-sided 5 percent test.

46
 Table 20: Aggregate saving rates and income inequality.

Pooled fixed effects GMM fixed effects

OLS time 5-years 5-years financial
baseline baseline dummies average baseline average development
inequality -0.297*** -0.262*** -0.162** -0.282* -0.032** -0.119** -0.596**
(0.066) (0.074) (0.080) (0.158) (0.015) (0.055) (0.293)
age dependency -0.311*** -0.410*** -0.237 -0.157 -0.054 -0.065** -0.705***
ratio (0.109) (0.120) (0.144) (0.294) (0.030) (0.130) (0.225)
GDP growth 0.226** 0.215** 0.192* 0.800** -0.166*** 0.333* 0.189
(0.095) (0.093) (0.098) (0.315) (0.046) (0.197) (0.119)
real interest 0.513*** 0.530*** 0.566*** 0.566*** 0.003 0.077 0.192
rates (0.072) (0.072) (0.109) (0.149) (0.115) (0.034) (0.161)
financial 0.088***
development (0.026)
inequality× -0.003**
fin.development (0.001)
Wald test
(p-value) 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Sargan test
(p-value) 0.792 0.230
AR(2)
(p-value) 0.187 0.651
Observations 216 216 216 41 214 39 86

Note: The dependent variable is the country-level household saving rate. Source: OECD yearbooks (1975,
1980, 1983, 1988, 1991 and 2001). The first independent variable captures inequality and is the share of income
received by the top 5% households. Source: Alvarado et al. (2012). The age dependency ratio is the population
aged 65 and over as share of total population. Source: Groningen Growth and Development Centre, Total
Economy Database, http://www.rug.nl/research/ggdc/data/total-economy-database-. GDP growth is
real per capita GDP growth from the World Bank National Accounts. The real interest rate is interest on bonds
minus inflation from the Federal Reserve Bank of St. Louis (Federal Reserve Economic Data). The measure
of financial development is stock market capitalization over GDP. Source: Beck and Demirguc-Kunt (2009).
The countries included in the sample are Canada, France, Germany, Japan, the UK and the US. The sample
consists of annual observations covering 1954-2007, except for column 3 in which we use 5-years averages for all
variables. Standard errors are reported in parentheses. *** significant at less than 1%; ** significant at 5%; *
significant at 10%

47
References
Abel, A. (1985), ‘Precautionary Saving and Accidental Bequests’, American Economic Review
75, 777–95.
Abel, A. (1990), ‘Asset Prices Under Habit Formation and Catching Up With the Joneses’,
American Economic Review 80, 38–42.
Abel, A. (2005), ‘Optimal Taxation When Consumers Have Endogenous Benchmark Levels of
Consumption’, Review of Economic Studies 72(1), 21–42.
Adreoni, J. (1989), ‘Giving with Impure Altruism: Applications to Charity and Ricardian
Equivalence’, Journal of Political Economy 97, 1447–1458.
Alesina, A. and Rodrik, D. (1994), ‘Distributive Politics and Economic Growth’, The Quarerly
Journal of Economics 109(2), 465–490.
Alonso-Carrera, J., Caballe, J. and Raurich, X. (2008), ‘Estate taxes, consumption externalities,
and altruism’, Journal of Public Economics 92, 1751–1764.
Altonji, J. G. and Villanueva, E. (2007), ‘The Marginal Propensity to Spend on Adult Children’,
The B.E. Journal of Economic Analysis Policy 7(1), 14.
Alvarado, Atkinson, Piketty and Saez (2012), ‘Income inequality in the United States, 1913-
1998’, The Quarterly Journal of Economics 118(1), 1–39.
Alvarez-Cuadrado, F. and Long, N. V. (2012), ‘Envy and Inequality’, Scandinavian Journal of
Economics 114(3), 949–973.
Arrondel, L., Parelman, S. and Pestieau, P. (1994), The effect of bequest motives on the
composition and distribution of assets in France, in T. Tachibanaki, ed., ‘Savings and
Bequests’, University of Michigan Press, Ann Arbor, pp. 229–244.
Atkeson, A. and Masao, O. (1996), ‘Wealth-varying intertemporal elasticities of substitution:
Evidence from panel and aggregate data’, Journal of Monetary Economics 38(3), 507–534.
Atkeson, A. and Masao, O. (1997), ‘Rate Of Time Preference, Intertemporal Elasticity Of
Substitution, And Level Of Wealth’, The Review of Economics and Statistics 79(4), 564–
572.
Attanasio, O. and Browning, M. (1995), ‘Consumption over the Life Cycle and over the Business
Cycle’, American Economic Review 85(5), 1118–1137.
Attanasio, O., Hurst, E. and Pistaferri, L. (2012), ‘The Evolution of Income, Consumption,
and Leisure Inequality in The US, 1980-2010’, NBER Working Papers, National Bureau
of Economic Research 17982.
Autor, D. H., Katz, L. F. and Kearney, M. S. (2008), ‘Trends in US wage inequality: revising
the revisionists’, Review of Economics and Statistics 90(2), 300–323.
Barro, R. J. (1974), ‘Are Government Bonds Net Wealth?’, Journal of Political Economy
82(6), 1095–1117.

48
Beck, T. and Demirguc-Kunt, A. (2009), ‘Financial Institutions and Markets Across Countries
and over Time: Data and Analysis’, World Bank Policy Research Working Paper No.4943
.
Bernheim, B., Schleifer, A. and Summers, L. (1985), ‘The Strategic Bequest Motive’, Journal
of Political Economy 93, 1045–75.
Bernheim, D. B., Skinner, J. and Weinberg, S. (2001), ‘What accounts for the variation in
retirement wealth among U.S. households?’, The American Economic Review 91(4), 832–
857.
Bertrand, M. and Morse, A. (2012), ‘Trickle-Down Consumption’, Chicago Booth Working
Paper .
Blinder, A. (1975), ‘Distribution Effects and the Aggregate Consumption Function’, Journal of
Political Economy 83(3), 447–75.
Blundell, R. W. and Bond, S. (1998), ‘Initial Conditions and Moment Restrictions in Dynamic
Panel Data Models’, Journal of Econometrics 87, 115–143.
Boskin, M. J. (1978), ‘Taxation, saving and the rate of interest’, Journal of Political Economy
76(2), 3–27.
Bosworth, B. and Anders, S. (2008), ‘Savings and wealth accumulation in the PSID, 1984-2005’,
CRR Working Paper 2008-2.
Carroll, C. D. (2000), ‘Portfolios of the Rich’, NBER Working Papers 7826, National Bureau
of Economic Research .
Carroll, C. D., Overland, J. and Weil, D. N. (2000), ‘Saving and Growth with Habit Formation’,
American Economic Review 90(3), 341–355.
Carroll, C. D. and Weil, D. N. (1994), ‘Saving and growth: a reinterpretation’, Carnegie-
Rochester Conference Series on Public Policy 40(1), 133–19.
Caselli, F. and Ventura, J. (2000), ‘A Representative Consumer Theory of Distribution’, Amer-
ican Economic Review 90(4), 909–926.
Charles, K. K., Hurst, E. and Roussanov, N. (2009), ‘Conspicuous Consumption and Race’,
The Quarterly Journal of Economics 124(2), 425–467.
Chatterjee, S. (1994), ‘Transitional Dynamics and the Distribution of Wealth in a Neoclassical
Growth Model’, Journal of Public Economics 54, 97–119.
Clark, A., Frijters, P. and Shields, M. (2008), ‘Relative Income, Happiness, and Utility: An
Explanation for the Easterlin Paradox and Othe Puzzles’, Journal of Economic Literature
46, 95–144.
Coleman, J. (1961), The Adolescent Society, Free Press, New York.
Corneo, G. and Jeanne, O. (1998), ‘Social organization, status, and savings behavior’, Journal
of Public Economics 70(1), 37–51.
Corsaro, W. and Eder, D. (1990), ‘Children’s Peer Cultures’, Annual Review of Sociology
16, 197–220.

49
Cushing, M. J. (2005), ‘Net Marginal Social Security Tax Rates over the Life-Cycle’, National
Tax Journal 58(2), 227–246.
Cutler, D. M. and Katz, L. F. (1992), ‘Rising Inequality? Changes in the Distribution of Income
and Consumption in the 1980’s’, American Economic Review 82(2), 546–51.
Davies, J. and Shorrocks, A. (1999), The distribution of wealth, in A. Atkinson and F. Bour-
guignon, eds, ‘Handbook of Income Distribution’, Elsevier, Amsterdam, pp. 605–675.
De Nardi, M., French, E. and Jones, J. B. (2010), ‘Why Do the Elderly Save? The Role of
Medical Expenses’, Journal of Political Economy 118(1), 39–7.
Deaton, A. (1991), ‘Saving and liquidity constraints’, Econometrica 59(5), 1221–1248.
Duesenberry, J. (1949), Income, Saving and the Theory of Consumer Behavior, Harvard Uni-
versity Press, Cambridge, MA.
Dynan, K. and Ravina, E. (2007), ‘Increasing Income Inequality, External Habits, and Self-
Reported Happiness’, American Economic Review 97(2), 226–231.
Dynan, K., Skinner, J. and Zeldes, S. (2004), ‘Do the rich save more?’, Journal of Political
Economy 112(2), 397–444.
Easterlin, R. (1995), ‘Will Raising the Incomes of All Increase the Happiness of All? ’, Journal
of Economic Behavior and Organization 27, 35–47.
Easterlin, R. (2010), Does Economic Growth Improve the Human Lot? Some Empirical Ev-
idence, in P. A. David and M. W. Reder, eds, ‘Nations and Households in Economic
Growth: Essays in Honour of Moses Abramovitz’, Academic Press, New York, pp. 89–125.
Edwards, S. (1996), ‘Why are Latin Americas savings rates so low? An International compar-
ative analysis’, Journal of Development Economics 51, 5–44.
Feldstein, M. and Samwick, A. (1992), ‘Social Security Rules and Marginal Tax Rates’, NBER
Working Papers 3962, National Bureau of Economic Research .
Ferrer-i Carbonell, A. (2005), ‘Income and well-being: an empirical analysis of the comparison
income effect’, Journal of Public Economics 89(5-6), 997–1019.
Frank, R. (1985), Choosing the Right Pond: Human Behavior and the Quest for Status, Oxford
University Press, New York and Oxford.
Frank, R. (2007), Falling Behind: How Income Inequality Harms the Middle Class, University
of California Press, Berkley.
Friedman, M. (1957), A Theory of the Consumption Function, Princeton University Press,
Princeton.
Gali, J. (1994), ‘Keeping Up with the Joneses: Consumption Externalities, Portfolio Choice,
and Asset Prices’, Journal of Money, Credit, and Banking 26, 1–8.
Heffetz, O. (2011), ‘A Test of Conspicuous Consumption: Visibility and Income Elasticities’,
Review of Economics and Statistics 93(4), 1101–1117.
Hong, K. (1995), Income distribution and aggregate saving, unpublished manuscript, Harvard
University Press, Cambridge, MA.

50
Huggett, M. and Ventura, G. (2000), ‘Understanding why high income households save more
than low income households’, Journal of Monetary Economics 45(2), 361–397.
Hurst, E. and Lusardi, A. (2004), ‘Liquidity Constraints, Household Wealth, and Entrepreneur-
ship’, Journal of Political Economy 112(2), 319–347.
Imbens, G. and Angrist, J. D. (1995), ‘Identification and Estimation of Local Average Treat-
ment Effects’, Econometrica 62(2), 467–475.
Katz, L. F. and Autor, D. H. (1999), Changes in the wage structure and earnings inequality,
in O. Ashenfelter and D. Card, eds, ‘Handbook of Labor Economics’, Vol. 3, Elsevier,
chapter 26, pp. 1463–1555.
Kopecky, K. and Koreshkova, T. (2010), ‘The impact of medical and nursing home expenses
and social insurance’, Federal Reserve Bank of Atlanta Working Paper 2010-19.
Krueger, D. and Perri, F. (2006), ‘Does income inequality lead to consumption inequality?
evidence and theory’, Review of Economic Studies 73(1), 163–193.
Kuznets, S. (1942), Uses of National Income in Peace and War, National Bureau of Economic
Research, New York.
Leigh, A. and Posso, A. (2009), ‘Top Incomes And National Savings’, Review of Income and
Wealth 55(1), 57–74.
Levine, A., Frank, R. H. and Dijk, O. (2010), ‘Expenditure Cascades’. http://ssrn.com/
abstract=1690612orhttp://dx.doi.org/10.2139/ssrn.1690612.
Li, H. and Zou, H.-f. (2004), ‘Savings and Income Distribution’, Annals of Economics and
Finance, Society for AEF 5(2), 245–270.
Liu, W.-F. and Turnovsky, S. (2005), ‘Consumption externalities, production externalities, and
long-run macroeconomic efficiency’, Journal of Public Economics 89(5-6), 1097–1129.
Ljungqvist, L. and Uhlig, H. (200), ‘Tax policy and aggregate demand management under
catching up with the Joneses’, American Economic Review 90, 356–366.
Lusardi, A. (1999), Information, Expectations, and Savings for Retirement, in H. Aaron, ed.,
‘Behavioral Dimensions of Retirement Economics’, Brookings Institution Press and Russell
Sage Foundation, Washington, D.C.
Luttmer, E. F. (2005), ‘Neighbors as Negatives: Relative Earnings and Well-Being’, Journal of
Economics 120(3), 963–1002.
Matlack, J. L. and Vigdor, J. L. (2008), ‘Do rising tides lift all prices? Income inequality and
housing affordability’, Journal of Housing Economics 17(3), 212–224.
Mayer, T. (1966), ‘The Propensity to Consume Permanent Income’, American Economic Review
56, 1158–1177.
Mayer, T. (1972), Permenent Income, Wealth, and Consumption: A Critique of the Permanent
Income Theory, the Lif-Cycle Hypthesis, and Related Theories., Berkley University Press,
California.

51
Moav, O. and Neeman, Z. (2010), ‘Status and Poverty’, Journal of the European Economic
Association 8(2-3), 413–420.
Modigliani, F. (1970), he life-cycle hypothesis and intercountry differences in the saving ratio,
in M. F. S. W. A. Eltis and J. N. Wolfe, eds, ‘Induction, growth, and trade: essays in
honour of Sir Roy Harrod’, Oxford University Press, Oxford, pp. 197–225.
Modigliani, F. and Brumberg, R. H. (1954), Utility analysis and the consumption function: an
interpretation of cross-section data, in K. K. Kurihara, ed., ‘Post-Keynesian Economics’,
N.J. Rutgers University Press, New Brunswick, pp. 388–436.
Mulligan, C. (1997), Parental Priorities and Economic Inequality, The University of Chicago
Press, Chicago.
Oswald, A. J. (1997), ‘Happiness and Economic Performance’, The Economic Journal
107(445), 1815–31.
Person, T. and Tabellini, G. (1994), ‘Is Inequality Harmful for Growth?’, American Economic
Review 84(3), 600–621.
Piketty, T. and Saez, E. (2003), ‘Income inequality in the United States, 1913-1998’, The
Quarterly Journal of Economics 118(1), 1–39.
Quadrini, V. (1999), ‘Growth, learning and redistributive policies’, Journal of Public Economics
74(2), 263–297.
Runkle, D. E. (1991), ‘Liquidity constraints and the permanent-income hypothesis: Evidence
from panel data’, Journal of Monetary Economics 27(1), 73–98.
Schmidt-Hebbel, K. and Serven, L. (2000), ‘Does income inequality raise aggregate saving?’,
Journal of Development Economics 61, 417–446.
Simmons, R. and Blyth, D. (1987), Moving Into Adolescence: The Impact of Pubertal Change
and School Context, Aldine Press, New York.
Slesnick, D. (2001), Consumption and Social Welfare, Cambridge University Press, Cambridge.
Stevenson, B. and Wolfers, J. (2008), ‘Economic Growth and Subjective Well-Being: Reassess-
ing the Easterlin Paradox’, Brookings Papers on Economic Activity, Economic Studies
Program, The Brookings Institution 39(1-(Spring)), 1–102.
Veblen, T. (1899), The Theory of the Leisure Class: An Economic Study of Institutions, Modern
Library, New York.
Weber, M. (1958), The Protestant Ethic and the Spirit of Capitalism, Routledge Classics,
London, UK.
Wendner, R. (2012), ‘Ramsey, Pigou, heterogenous agents, and non-atmospheric consumption
externalities’, Graz Economics Papers GEP 01-2012.

52
8 Data Appendix (for online publication)

We use household and individual data and the wealth supplements from the PSID for the period
1984-2007. Unless otherwise noted, the same methodology is used to construct variables for

the different periods between 1984-2007. All the variables are adjusted for inflation using the
price deflator for personal consumption expenditures and are reported using 2007 dollars.

We use three different saving measures; active saving, change in wealth, and change in
wealth plus pension and social security. All of the wealth and active saving components are
from the wealth and active saving supplement files of the PSID, which give detailed information

on wealth levels and components of active saving for 1984, 1989, 1994, 1999, 2001, 2003, 2005
and 2007. The saving variables are constructed in the following way:

a. Change in Wealth: Computed as the difference between real net worth in a given period
and real net worth in the preceding period. Real net worth is constructed as the sum of
values of seven asset types: Home equity, other real estate, private business and farms,

vehicles, transaction accounts, corporate equities, annuities and IRAs and other savings,
net of debt. Net worth does not include either defined benefit or defined contribution
pension savings. If respondents failed to provide an estimate of a wealth component or if

they could only provide a range of values, the PSID staff uses imputation procedures.

b. Active Saving: this measure is designed to be similar to the traditional income minus
consumption measure of saving. It equals the change in wealth over a given period minus
capital gains, for example on housing or stocks. Since capital gains questions were not

directly asked, the measure is built in part from questions about saving flows.

First, consider active savings in the main residence: if a household has not changed

home during the period, this consists in the change in the mortgage liability. If the
home changed during the period, the change in home equity also enters. Secondly, active

53
 savings in assets subject to capital gains consist in purchases and sales of other real estate,

businesses and corporate equities. Finally, active savings in fixed income and other assets
consists in changes in the holdings of “other savings” and transaction accounts, minus
changes in non-collateralized debt.

c. Change in wealth plus pensions and social security: This variable adjusts the change in

wealth measure above for saving through pension and social security contributions. Its
construction follows Dynan et al. (2004). Since our sample is longer, we use imputed social
security taxes from Feldstein and Samwick (1992) for the years up to 1995, and estimates

from Cushing (2005) after that. These papers provide estimates of net marginal tax rates
through social security contributions for different demographic groups. The difference
between the social security contribution rate (a 11.2% payroll tax on labor income up to

$50,000 until 1995, 12.5% afterwards) and the implied tax rate constitutes saving. This
is straightforward for singles. Spouses may contribute separately or not; therefore we use
labor income of the higher-earning spouse in our calculation if the lower-earning spouses

contributed less than one third of household labor income.

We also include employee contributions towards defined contribution pension plans, cal-

culated as the reported contribution rate on the current job times labor income. The
PSID does not contain information on employer contributions.

Our measure of disposable income is the household real income after taxes. In the PSID,
this variable is constructed as “the sum of total family income from labor earnings, asset flows,

private transfers, private pensions, public transfers, and social security pensions minus total
household taxes. Labor earnings include wages and salary from all employment including self-
employment (farming, business, market gardening, and roomers and boarders), professional

practice or trade, and bonuses, overtime and commissions. Asset flows include income from
interest, dividends, and rent. Private transfers include child support, alimony, and income

54
from non-household members. Private pensions include retirement income from private pen-

sion plans, Veterans Administration pensions, and annuities. Public transfers include AFDC
payments, supplemental security income (SSI), unemployment compensation, workers compen-
sation, and the face value of food stamps. Social security pensions include social security

payments received by the head, partner, and other family members. Total household taxes
includes income taxes of the head, partner and other family members, as well as payroll taxes
of the head and partner.”

We calculate the saving rate by dividing active saving and the change in wealth by five or
two times average disposable income, depending on the period under consideration. For each

measure of saving we use the corresponding average disposable income measure. For the change
in wealth, we add capital gains in housing, stocks and businesses to the measure of disposable
income discussed above. For the change in wealth plus pensions and social capital, we also add

the worker’s contributions to social security and pension plans to disposable income.

As instruments we use:

a. Two measures of consumption: food consumption, and weighted consumption. To-

tal real annual food consumption is the sum of total real annual food consumed at home
(including the amount saved from food stamps) and annual real food consumed out of the

home. To construct the measure of weighted consumption we use the weights from Bern-
heim, Skinner and Weinberg (2001) and we calculate consumption as C = 1.930(food at
home) + 2.928(food away from home) + 1.828(rental payments if renter) + 0.1374(value

of house if home owner). Consumption measures are only available beginning in 1994,
since food questions were temporarily suspended in 1988.

b. Education: This is measured as the number of years of education completed by the head
of the household at the end of the saving period.

c. Average labor income: This is computed using all observations on labor income for the

55
 time period that a household remains in the sample. To compute this variable, we extend

our sample to all the years available (1968-2007), and not only those years for which we
have a wealth supplement. On average this measure includes fourteen observations of
yearly income.

The sample selection is as follows: we exclude households whose head was less than 30
years old or older than 60 years old at the middle of the saving period. We select households
with an absolute value of active saving not larger than $1,050,000 and with disposable income

of at least $1,400.17 in a given year. From the remaining sample, we drop households in the
extreme 0.5% of the distribution of saving. We also drop observations for households that had
a change in head. We end up with a sample of 24,036 observations about 8,063 households,

with a minimum of 2,815 households in any given year.

To test if our results could be driven by differences in redistribution across time or

states we construct an index variable that indicates the political party of the state governor
during the period in which the savings decision took place. This variable goes from 0, the state
governor was a Democrat during the entire period, to 1, the state governor was a Republican

during the entire period. To test if our results could be driven by differences in variation in
real estate prices we include state level housing prices in our standard specifications. For this
we use the monthly S&P/Case-Shiller Home Price index available on www.spindices.com. It

tracks repeat sales of houses and are a modified version of the original method by Case, Shiller
and Weiss.

In section 5, we use aggregate data for six different countries: Canada, France, Germany,

Japan, the UK and the US. The sample consists of annual observations covering 1954-2007,
except for Germany, for which the saving variable is measures at a tri-annual basis. The
variables used are the following:

56
a. Household saving rates. Obtained from different OECD yearbooks (1975, 1980, 1983,

1988, 1991 and 2001).

b. Inequality. Measured as the share of income received by the top 5% households. Source:
Alvarado et al. (2012).

c. Age dependency ratio: measured as the population aged 65 and over as share of total pop-
ulation. Source: Groningen Growth and Development Centre Total Economy Database,

http://www.rug.nl/research/ggdc/data/total-economy-database-.

d. GDP growth. Source: World Bank National Accounts.

e. The real interest rate is measured as the interest on bonds minus inflation. Source: Federal
Reserve Bank of St. Louis (Federal Reserve Economic Data), series INTGSxxSM193N,
where xx = {U S, DE, JP, CA, GB, F R}, and xxxCPIALLMINMEI, where xxx={U SA, DEU,

JP N, CAN, GBR, F RA}, available at http://research.stlouisfed.org/fred2.

f. Financial development: Stock market capitalization over GDP. Source: Beck and Demirguc-
Kunt (2009)

57
Appendix I (for online publication)

Finding the counterparts of (4), (5), (7), and (8) for the i-th household born in period t, where
ḡti and ōit+1 are the levels of consumption of his reference group when young and old respectively,

we obtain

� i � γōit+1
sit i
= β (1 + µ) ct − γḡt + (23)
R
γōit+1
cit (1 + (1 + µ) β) = yti + β (1 + µ) γḡti − (24)
R
1 � i �
dit+1 = Rst + µγōit+1 (25)
1+µ
µ � i �
bit+1 = Rst − γōit+1 (26)
1+µ
We proceed sequentially. First we solve (23)-(26) together with (18), noting that cH H
t = ct

and d¯H H
t+1 = dt , to obtain the optimal choices of the rich households,

cH H
t = πyt (27)

sH H
t = β (1 + (1 − γ) µ) πyt (28)

dH H
t+1 = Rβπyt (29)

bH H
t+1 = (1 − γ)µRβπyt (30)

We now solve the corresponding system for poor individuals, noting that cLt = cLt and
d¯Lt+1 = dLt . From (25) we reach

(1 + µ − µγρ)dLt+1 = RsLt + µγ(1 − ρ)dH

t+1 (31)

Using (23)

γρ L γ(1 − ρ) H
sLt = β(1 + µ)(1 − γρ)cLt − β(1 + µ)γ(1 − ρ)cH
t + dt+1 + dt+1 (32)
R R

58
Substitute (31) into (32)

(1 + µ) (1 − γρ) L γ(1 − ρ) (1 + µ) H
st − β(1 + µ)(1 − γρ)cLt = −γβ(1 + µ)(1 − ρ)cH
t + d (33)
1 + µ − µγρ R (1 + µ − µγρ) t+1

From (24) and (31),

γρ γ(1 − ρ) (1 + µ) H
sLt +(1+β(1+µ) (1 − γρ))cLt = ytL +β(1+µ)γ(1−ρ)cH
t − d (34)
1 + µ − µγρ R (1 + µ − µγρ) t+1

Write (33) and (34) using (27) and (29) as

Ax = b

where  
(1 + µ)(1 − γρ)
 −β(1 + µ)(1 − γρ) 
A ≡  1 + µγρ − µγρ 
1 + β(1 + µ) (1 − γρ)
1 + µ − µγρ
� L�
s
x ≡ Lt
ct
 � � 
µ(1 − γρ) H
 −βγ(1 + µ)(1 − ρ) 1 + µ − µγρ πyt 
b≡ L
� � 

µ(1 − γρ) H
yt + β(1 + µ)γ(1 − ρ) πyt
1 + µ − µγρ
that applying Cramer’s rule can be solved simultaneously for

� �
sLt = ψβ (1 + µ(1 − γρ)) ytL − µγ(1 − ρ)πytH
� �
cLt = ψ ytL + γµβ (1 − ρ) πytH

and then dLt+1 and bLt+1 are given by (31) and (26), respectively.

59