Social Security and Households' Saving

Social Security and Households' Saving
Author(s): Orazio P. Attanasio and Agar Brugiavini

Source: The Quarterly Journal of Economics, Vol. 118, No. 3 (Aug., 2003), pp. 1075-1119
Published by: Oxford University Press
Stable URL: http://www.jstor.org/stable/25053931
Accessed: 05-03-2018 05:41 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access
to The Quarterly Journal of Economics
This content downloaded from 166.104.72.89 on Mon, 05 Mar 2018 05:41:08 UTC
All use subject to http://about.jstor.org/terms
SOCIAL SECURITY AND HOUSEHOLDS' SAVING*
Orazio P. Attanasio and Agar Brugiavini
This paper provides new evidence on the substitutability between private and
pension wealth by exploiting the Italian pension reform of 1992. We use a differ
ence-in-difference estimator that exploits the differential effects of the reform on
individuals belonging to several year-of-birth cohorts and different occupational
groups. We find convincing evidence that saving rates increase as a result of a
reduction in pension wealth. By allowing for the possibility that substitutability
changes with age, we find that substitutability is particularly high (and precisely
estimated) for workers between 35 and 45.
I. Introduction
The characterization of the determinants of households' sav
ing decisions is crucially important both for providing a frame
work capable of explaining the accumulation of wealth and for a
wide variety of policy issues. According to the life cycle theory,
individual savings depend, among other things, on the amount of
resources available (through public pensions) after retirement,
when earnings typically peter out. While the inverse relation
between future benefit entitlements and household saving seems
intuitive and natural, whether public pension wealth constitutes
a perfect substitute for financial savings is an empirical question
that has not been settled. The aim of this paper is to measure the
elasticity of household saving to changes in future pension enti
tlements. This elasticity is obviously related to the degree of
substitutability between pension and bequeathable wealth. Such
an exercise is of crucial importance to public policy, especially
when, as in most Western economies, major reforms of the pen
sion systems are being considered.
There are several reasons why pension wealth might not be
a good substitute for financial saving: future pension benefits are
* The authors wish to thank Margherita Borella, who skillfully organized the
data in the initial part of the project. Erich Battistin, Nicola Rossi, Ignazio Visco,
and Guglielmo Weber have discussed with us many aspects of the research; Alan
Auerbach, Luigi Guiso, and Reinhold Schnabel made many useful comments.
Useful comments were also received at the meeting "Ricerche Quantitative per la
Politica Econ?mica" organized jointly by the Bank of Italy and the Centro ?nter
universitario Discipline Economiche, S.A.DI.BA, Perugia, November 1995. We are
particularly grateful to Costas Meghir for some very useful suggestions. Agar
Brugiavini expresses gratitude to the Consiglio Nazionale di Ricerca (grant
96.01418.CT10) and to the European Union (Training and Mobility of Research
ers, contract FMRX CT90016) for generous financial support. The usual dis
claimer applies.
? 2003 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
The Quarterly Journal of Economics, August 2003
1075
1076 QUARTERLY JOURNAL OF ECONOMICS
not liquid and cannot be borrowed against; individuals might be

liquidity constrained at some points in time; the implicit rate of
return on pensions is not the same as that on financial savings.1
In other words, financial and insurance markets arrangements
(including tax treatment of pension and financial instruments)
and different welfare provisions may produce a variety of ob
served saving patterns and of substitutability results.
Given the importance of the topic, it is somewhat surprising
that little empirical evidence exists on the relationship between
the provision of public pensions and individual household sav
ings. Feldstein [1974] provided some time series evidence on such
a link for the United States. He estimated a simple consumption
function that included, in addition to current income and finan
cial wealth, an aggregate estimate of household social security
wealth and found a significant and positive effect of all three
variables. The coefficient on social security wealth was, in some
cases, larger than that on financial wealth.
King and Dicks-Mireaux [1982], instead, considered the re
lationship in a microeconomic data set. As they considered a
single Canadian cross section, they focused on the relationship
between the stock of saving and the present discounted value of
future benefits. In their study, they expressed both of these vari
ables as the ratio to an estimate of permanent income. Their
methodology, which consisted of estimating the degree of substi
tutability between financial and pension wealth, was subse
quently used and extended by Hubbard [1986] on U. S. data and
by Brugiavini [1987] and Jappelli [1995] on Italian data. Dia
mond and Hausman [1984] and Samwick [1997] have also used
micro data for the United States and the estimated offset between
the two stock measures is about 20 percentage points. However,
with the exception of the recent papers by Gustman and Stein
meier [1998] and Gale [1998], not many recent advances have
been made through direct household level data, and no consensus
has been reached on the order of magnitude of the substitutabil
ity parameter. It is particularly surprising that while a flurry of
pension reforms is taking place in Europe, the evidence on the
1. Contributions to the pension system, whether perceived as forced saving or

as taxes, can distort labor supply choices, particularly if the pension system is not
actuarially fair. To be more precise, the relationship between contributions on the
one hand and accrual rate of expected pension benefits on the other hand may
determine the timing of retirement decisions as well as the level of the labor
supply.
SOCIAL SECURITY AND SAVING 1077
relationship between bequeathable wealth and pension wealth

and on the likely impact of the reforms on asset accumulation is
scanty and inconclusive. Recent papers by Alessie, Kapteyn, and
Klijn [1997] and Euwals [2000] reach mixed conclusions on the
relationship between the two forms of wealth. As the authors
point out, there is a basic problem with the true source of vari
ability available in the estimated pension wealth variable.
This lack of evidence is partly due to the inherent difficulty in
measuring public pension (or social security2) wealth and partly
to the fact that observed cross-sectional differences in saving
behavior can be due to a variety of factors observed and unob
served. The presence of unobserved heterogeneity makes it diffi
cult to identify the effect that different endowments of social
security wealth have on different individuals' saving behavior. In
this paper we use changes in public pension wealth to identify the
effect of social security entitlements on household savings. For
such a purpose we use a time series of Italian cross sections, some
collected just before and others after a major reform of the social
security system.
As discussed below, the Italian social security system,
started in the early 1970s, became progressively more generous
and, by the end of the 1980s, clearly unsustainable. In September
1992 the Amato government introduced the first of a series of
important reforms. By changing retirement age, eligibility, early
retirement arrangements, and so on, the 1992 reform induced a
substantial reduction in the pension wealth of many Italian
households. What is more relevant for our exercise, however, is
that the reform changed the pension wealth of observable groups
of households in different ways. Typically, public sector, younger,
and educated employees were affected more than relatively older,
private sector, and uneducated employees. In what follows, we
exploit this variability to identify the elasticity of personal sav
ings to public pension wealth by comparing the changes of these
different groups.
Italian households have always been characterized by a re
markably high saving rate. In the last two decades, however, the
personal saving rate has declined by about 10 percentage points.
2. The term social security seems to have different uses in the United States
on the one hand, and in the United Kingdom and the rest Europe, on the other
hand. In this paper we will be using the U. S. meaning and consider social security
and public pension wealth ("pension wealth") as synonymous. In fact, in Italy
there are basically no occupational pension plans.
Several authors have tried to explain both the remarkable levels

of Italian households' saving rates and their subsequent large
fall. Rossi and Visco [1994, 1995] have explained part of the
decline that occurred during the 1980s, as directly caused by the
reforms of the social security system which took place during the
1970s. While these seemed minor changes at the time of their
introduction, they had a major impact on the system, by making
current and future benefits to retirees more generous than the
contributions paid to the system by the same individuals. The
immediate effect of these changes has been a substantial growth
of social security wealth for some groups of the population and
particularly for retirees and generations of active individuals
close to retirement (as opposed to future generations). If Italian
households perceived social security wealth as a close substitute
for private life-cycle-wealth, this increased generosity could have
resulted in a significant increase in the propensity to consume for
those who benefited most from the system.
Such a generous defined benefit system, based on a pay-as
you-go financing method and coupled with the recent demo
graphic trends, was bound to generate a growing deficit of the
Italian social security system. Hence, the 1992 reform has been
the first (much-needed) reaction to the structurally unstable sit
uation of the Italian old age public insurance system. It has
undoubtedly marked a turning point in the dynamic behavior of
the social security wealth: many individuals had the value of
their future pension benefits substantially curtailed. Further
more, the 1992 reform received much attention by the press and
was largely debated on television and radio programs: the impor
tance of the changes brought about by the reform could not go
unnoticed. In 1993, the year following the Amato Reform, the
saving rate of Italian households increased considerably. Obvi
ously, it is not possible to attribute the change entirely to the
reform of the social security system. Many things happened be
tween 1992 and 1993, including the fact that the economy entered
a substantial recession. With the same Act with which it changed
the social security system, the Government introduced a number
of measures aimed at re-equilibrating the Italian fiscal situation.3
For this reason, we rely on a difference-in-difference estimator
3. Miniaci and Weber [1999] provide a detailed analysis into the causes of the
1993 recession and found a relevant role played by the 1992 social security reform.
that exploits the differential effect of the reform on different

groups' pension wealth.
The rest of this paper is organized as follows. In Section II we
briefly summarize the existing literature on the relationship be
tween saving and pension wealth and present a very simple
model that will provide the framework to specify and interpret
our empirical results. In Section III we start by illustrating some
recent trends of Italian household saving. We then sketch the
main changes introduced by the 1992 reform and use some of the
features of the law to present some first evidence of the effect that
it had on household savings. In particular, we use the fact that
the law had very different effects on individuals with more than
or less than fifteen years of contribution tenure in 1992 to esti
mate the effect on saving rates by regression discontinuity de
sign. We then move on to illustrate the changes that the law had
on our estimates of individual pension wealth. In Section IV we
present our results on the effect of social security wealth on
individual saving rates. To identify these effects, we directly
exploit the differential effect that the law had on different groups
of Italian households.
II. On the Relationship between Pension Wealth and Saving: A

Simple Theoretical Framework and Empirical Specification
Our approach is different from those used in the papers in

the saving literature for two reasons. First, we consider the
relationship between public pension wealth, as measured by the
present discounted value of future entitlements, on the saving
rate of individual households. That is, unlike Feldstein [1974] and
subsequent papers, we use individual rather than aggregate data,
and unlike King and Dicks-Mireaux [1982], Hubbard [1986], Bru
giavini [1987], and Jappelli [1995], we use flows instead of stocks.
Second, we use the variation in pension wealth induced by a large
legislative change that occurred in Italy in 1992 to identify the
parameters of interest. This approach allows us to control for the
possible presence of fixed effects both at the individual and at the
group level. Our econometric approach and identification strategy
are similar to those used by Krueger and Pischke [1992] to ana
lyze the labor supply effects of the unexpected "benefit notch"
affecting U. S. workers born after 1916. Krueger and Pischke use
differences in the changes in social security across generations
and over time to study labor supply responses.
In this section we first sketch a very simple model, which is

useful to justify the empirical specification we use and interpret
the empirical results we present below. The model is nothing new.
However, it is useful to make the (somewhat obvious) point that
changes in future pension entitlements have different effects on
saving rates for people at different points in their lifetime. We
need to make this relationship precise as it will inform our em
pirical specification, which we present in subsection H.A. Finally,
we discuss our econometric strategy.
H.A. A Simple Model

Within the framework of a life cycle model, there is an obvi
ous relationship between publicly provided pension benefits and
the saving rates during the working life of a generic household. A
permanent and unexpected reduction in an individual's social
security wealth implies an increase in her savings and a reduc
tion in her current consumption. This behavior is explained by an
increased desire to provide resources for consumption after re
tirement. More generally, for individuals who are not at a corner,
that is, those who are not consuming the maximum they can,
changes in their pension wealth would induce a change in their
consumption and saving behavior.4 As we want to estimate such
a relationship, however, it is worth it to make it precise so as to
determine the variables we want to consider in the empirical
specification and how to interpret the coefficients we estimate.
As we are interested in obtaining closed-form solutions to
characterize mainly saving for retirement behavior, we consider a
version of the model without uncertainty. This choice, which
simplifies our discussion considerably, is not fully consistent with
the consideration of an "unexpected" change in pension wealth at
some point in the life cycle. We do not think, however, that a
complete treatment of uncertainty would add much to the ideas
we want to consider.5
Because we want to consider the differential effect that
changes in pension wealth have on individuals of different ages,
4. As we argue in footnote 1 above, in principle, labor supply can be affected.

In this paper we do not consider these effects. Neither do we explicitly consider the
possibility that intergenerational altruism undoes the changes in social security
wealth.
5. It is obviously possible that the precautionary motive induced by uncer
tainty could interact with the retirement motive. This is particularly true if the
higher retirement savings induced by a reduction in public pension wealth could
be used to smooth out business cycle fluctuations and therefore reduce the motive
for precautionary saving.
we have to consider at least three periods. Moreover, as we want

to consider individuals who have been affected by a (presumably
unanticipated) reform in the year we observe them, and individ
uals who were affected by such a reform in the past, we have to
consider at least four periods. We assume that people work dur
ing the first three periods of their lives and then retire. Labor
supply is exogenous and is compensated with a deterministic
level of earning. When people retire, they receive a certain level of
benefits b. For simplicity, we assume that the utility function is
isoelastic. Formally, a generic individual solves the following
maximization problem:
(Cl)1" a. D (C2)1" a. ?2 fe)1"' _, ?3 bJ1-"

(1) maxT3^ + ?rr^ + ?2T^ + ?3Tr^
subject to
3^2 y3 b
yi + 1 + r (1 + rf +
c2 c3 c4
^ + 7T?,-T + 7^?-v>. +
'* (1 + r) (1 + rf (1 + rf
Earnings in the three periods (y?, i - 1, 2, 3) should be
preted as net earnings, r is the constant interest rate,
discount factor, and I/7 the elasticity of intertemporal sub
tion. The solution of this problem is trivial. In particular, it is
to show that
y 1 ~ C\
(2) ---=1
yx 1 + d + d2 + ds
X
y2/yi yJyi blyx
1 + 1 + r + (1 + r)2 (1 + r
where d = ?1/7 (1 + r)a l)/y. For a consumer aged

plan for the second and third periods are given by
y2 ~~ ^2 d
(3) ~~JT~ =1~l + d + d2 + d3
x
(1 + r)yx y3/y2 b/
y2 (1 + r) (1 +
y 3 - C3
(4) l
1 + d + d2 + d3
X
(1 +
+
r)Vi
+1 + (1
6/y3 +
3^3 3^3 d + r)
Equations (3) and (4) can be used to derive the effect on second
and third-period saving (and consumption) of an anticipated
change in pension wealth. However, if we want to consider the
effect of an unanticipated change in pension wealth on consump
tion and wealth for a consumer aged two or three, at the current
age, we have to consider the reoptimization problem of this indi
vidual who has saved a certain amount in period 1 (and 2). The
reoptimization will yield saving rates as
y2-c2
(3') y2 1 + d + d2
b/y2
X 1 + (1 + r)A1 [ yjy2 t
3>2 1 + r (1 + r)?
3^3 - c3
(4') (1 + r)A2 + + b/y3
ys 1+d ys 1+r
where A? is accumulated private wealth in the ?th period.

For a consumer who does not reoptimize in period 2 or in
period 3, equations (3') and (4') will be equivalent to equations (3)
and (4), respectively, as it can be checked by substituting the
expressions for the current level of assets Ax and A2. For an
age-two consumer who is affected by a surprise change in the
second period of her life, the relevant expression is (3'), that takes
the decisions made in period 2 as given. Analogously, for an
age-three consumer, the relevant expression is (4').
Finally, if we consider an age-three consumer who, in the
second period of her life was affected by a surprise change in
pension wealth, one gets the expression
y s - es
(4") i + d + d2
x (1 + r)2A, | (1 + r)y2 | i
ys ys l + r
Once again, if we assume that there are
tion (4") is equivalent to (4') and (4). Howev
equation (4") is the relevant one for a consumer who has received
a surprise in period 2 and is observed in period 3.
The expressions above make clear, within the framework of a
very simple model, how to discount future benefits and how to
relate them to the current level of saving rates. The point we want
to stress is that the relationship depends not only on how far
away the consumer is from retirement, but also on when in her
life cycle she experiences the reform and when she is observed by
the econometrician relative to the surprise reform. Gale [1998]
has recently stressed the importance of the different planning
horizon for individuals of different ages. Our framework also
stresses (and our empirical work exploits) a related implication of
a standard life cycle model. The change in saving generated by a
change in pension wealth (i.e., by the reform) is greater, the closer
the individual is to retirement age. This reflects the fact that
younger individuals have a longer horizon over which to absorb
the "unexpected" shock to pension wealth.
The way in which the coefficient on the pension wealth
income ratio changes with age depends on the values of the
parameters that characterize the utility function. When we con
struct our "pension wealth" variable, we take them into explicit
consideration. In particular, in order to compute the implied
adjustments (appropriately modified to consider an arbitrary
length of life and a variable retirement age), we will be making
specific assumptions about the rate at which individuals discount
the future and about their preferences. It is easy to show that in
a more general model with N periods of work before retirement,
the coefficient of the ratio of the present discounted value of
benefits to current income in equations (3) and (4) is equal to a
factor k(a,d), defined as follows:
1 1-d
where ra is retirement age, assu

current age.6 Equations (3'), (4'), a
similar fashion.
In the simple model we have c
accrues in one period and is, for a
guishable from private financial w
6. We are assuming that ?1^ (1 + r)a

instance, this is obviously true, as in that
tioned above, this might not be the case. For one thing, public
pension wealth is substantially less liquid than financial wealth.
Furthermore, it accrues over the retirement period with a prede
termined pattern and, therefore, cannot be adjusted to particular
changes in needs.7 It is therefore plausible to assume that public
pension wealth is not a perfect substitute of private savings.
Indeed, many of the empirical studies of these issues, in the
tradition of the King and Dicks-Mireaux [1982] paper, measured
explicitly the degree of substitutability between pension and fi
nancial wealth. For these reasons, in the empirical application
below, we allow the coefficient on our definition of pension wealth
to be age dependent and check whether and how much it differs
from minus one. In other words, we interpret the coefficient on
our variable as a measure of the substitutability between pension
and financial wealth.
II.B. Empirical Specification

Our model is supposed to give a basic conceptual framework
that allows us to interpret the results of our empirical specifica
tion. In particular, we model saving rates and, following the logic
of the model, we regress them on the present discount value of
pension benefits divided by current income and adjusted for the
factors discussed above. However, as reality is more complex than
the simple model we propose, we also allow for additional con
trols. In particular, to capture macroeconomic shocks, we allow
for year dummies, while to control for permanent differences in
saving behavior across different groups, maybe induced by differ
ences in earning profiles or preferences, we allow for group dum
mies. Notice that as year-of-birth cohorts are part of the group
definition, allowing for group and year dummies is equivalent to
allowing for flexible age effects.8 Moreover, to capture more fully
the possibility of life cycle effects, we also control for various
demographic variables.
The other set of factors that are bound to be important for
saving rates, according to the simple model we sketched, is the
expected value of future earnings normalized by the level of
current income. Individuals who expect low future earnings
7. The model we have considered also neglects uncertainty, which might

affect the pattern of saving rates. Uncertainty about the sustainability of the
social security system complicates the relationship further.
8. We experiment with the introduction of a term in age, however, to capture
the effect of age differences within each cohort.
should save more than individuals who expect high future earn
ings. We capture these effects in two different ways. As we use
projection of future earnings to compute expected pension wealth,
we have as a by-product of our procedure some estimates of
expected future earnings. As they are likely to be error-ridden, we
instrument for them using the same interactions of group and
time dummies that we use for pension wealth. Alternatively, we
simply assume that variation in future earnings is captured by
the flexible age and cohort effects that we allow in our estimation.
It turns out that the two procedures do not yield very different
results.
To be more precise, we estimate the following equation:
(6) SRiJt = Xt? + Haht)FEi}t + 0 (ai>t)PWiit + xt + fc + eM,

where SRit is the saving rate for household i at time t, defined as
income minus consumption divided by income, FE is the ratio of
future to current earnings, and PWit is the ratio of pension
wealth to current earnings adjusted following the procedure de
scribed above. 0(a??) and <\>(a??) are age and time-dependent
parameters, where time dependence is relative to the date of the
reform: in order to capture their variability with age, we model
them both as polynomials in age and as a step function. The term
xt represents time effects, and fc are group effects. X is a vector of
controls that includes demographic variables such as family com
position dummies.
U.C. Econometric Issues
The estimation of equation (6) on individual data poses a
number of problems. As pension wealth, that is, the present
discounted value of future benefits, is not typically observed, one
has to estimate it. However precise the method used in estimat
ing such a quantity, the resulting variable is likely to be affected
by measurement error, which is likely to introduce a bias in the
estimated coefficients. Moreover, an additional source of bias is
introduced by the possibility that individual (and unobserved)
heterogeneity in saving behavior is related to the individual vari
ables used to estimate pension wealth. To take this type of prob
lem into account, one can use an Instrumental Variable tech
nique, using as instruments group dummies interacted with year
dummies. Groups are defined so to capture systematic differences
in social security wealth and, in particular, differences in the time
evolution of such a variable. Our proposed strategy, therefore, is
an application of a difference-in-difference estimator.9 Identifica

tion requires that PWit (the pension wealth variable) varies in a
way that is not fully explained by group and time dummies. This
hypothesis can be easily tested by checking the significance of
interaction terms in a regression of PWit on group, time dummies,
and their interaction. The 1992 reform, which occurs in the mid
dle of our sample period, gives us the possibility of implementing
such a technique as it induced changes in pension wealth that
were systematically different across the groups we considered.
This variation across groups is obviously crucial for the identifi
cation of 0, and therefore, we form groups in such a way as to
maximize it.
As we discuss below, we have a time series of four cross
sections. The first two are before the reform, while the last two
refer to the first year after the reform and to the third year after
the reform. The effect of pension wealth depends both on the age
of an individual at the time of the reform and on when the
individual is observed after the reform. This is apparent if we
compare the expressions for age three saving rates given in equa
tions (4') and (4"). The former refers to a consumer who experi
ences a surprise change in period 3 of her life, while the latter
refers to a consumer aged three who experiences the surprise
change in period 2. Therefore, in our sample, among the "postre
form" households, we have to distinguish households of a given
age that are observed one (in 1993) or three years (in 1995) after
the reform. If we exclude the last year of the data (i.e., we drop the
year 1995), we can avoid this last adjustment, as we would have
only one year "after the reform." In what follows, we present
results for both cases. Significant differences between the two
sets of results could be due to delays in the adjustment of saving
rates to the changes induced by the reform.10
Estimating social security wealth is obviously quite hard.
Our procedure requires knowledge of a number of legislative
details, as well as some assumptions on expected earning profiles,
retirement age, and so on. Furthermore, as we use social security
wealth as a determinant of individual saving behavior, we would
9. See Heckman and Robb [1985] for a discussion of the method and Eissa
[1995] and Blundell, Duncan, and Meghir [1998] for recent applications of the
method to female labor supply behavior. Meyer [1995a, 1995b] provides a general
discussion and applications on the effects of unemployment insurance.
10. Note that our sample is, by and large, made of different individual
households observed over time. However, the sample also contains a small panel
component.
ideally require an estimate of social security wealth as perceived

by the individual. If the social security system before (or after, for
that matter) the Amato reform was widely perceived to be unsus
tainable, our estimates of social security wealth might deviate
substantially from perceived wealth. This argument constitutes
probably the most convincing reason to use average changes in
estimated social security wealth to identify the relationship of
such a variable with the personal saving rate. In 1991 the Italian
social security system was one of the most generous in the world
and clearly unsustainable in the long run. As these issues were
widely debated in the time up to the reform of 1992, it is not
unreasonable to think that a change in the legislation was widely
expected. However, it should be stressed that the type of reform
that would eventually be implemented was not obvious. In par
ticular, it was not obvious who would be affected more by the
reform and by how much. The limitations of the reform itself and
the difficulties of subsequent governments in implementing fur
ther and more radical changes in 1994 and 1995 show that the
political economy of the pension reform was extremely complex
and characterized by outcomes that were ex ante far from obvious
and anticipated. If changes in pension wealth were not fully
anticipated, they could identify the relationship between public
pension provision and individual saving. In any case, if house
holds perceived the system existing before 1992 as unsustainable,
the reform would not have changed their perception of social
security wealth. This would make our test less powerful in that it
would make it harder to identify a relationship between changes
in pension wealth before and after the reform and changes in
savings.
III. The Amato Reform and its Differential Effects

on Social Security Wealth
III.A. Household Savings in Italy

The saving behavior of Italian households has been charac
terized for a long time by very high saving rates. Different studies
have proposed several explanations, with different degrees of
plausibility. Some authors have stressed the limited development
of financial markets, and in particular of a market for housing
mortgages as the most likely explanation for Italy's high saving
rates. Other authors have also appealed to the presence of strong
bequest motives.11 While the fact that for Italian households it

has been until very recently very hard to borrow, either to finance
consumption or to purchase a house, is well established, it is
difficult to make a convincing connection between these facts and
observed saving rates.
From the beginning of the 1980s, however, the saving rate of
Italian households has decreased considerably. The national sav
ing rate (adjusted for inflation and for durable purchases) has
declined from a peak of 24 percent at the beginning of the 1960s
to just below 10 percent in the early 1990s. The private saving
rate exhibits a similar trend, though it is characterized by stron
ger fluctuations. The Italian social security system was at the
time (and still is) a patchwork of different public funds: each
characterized by its own rules in terms of eligibility and benefit
payments. Almost invariably changes in the social security leg
islation taking place after the 1950s went in the direction of
increasing the generosity of the system: policy-makers engaged in
a perverse struggle in order to provide advantages to one fund or
the other in turn. Almost at the same time the system underwent
a complete transition from being funded to becoming a Pay-As
You-Go system. Not surprisingly, Rossi and Visco [1995] propose
an explanation for the decline in the saving rate driven by the
increase in public pension wealth that took place starting in the
early 1970s. Their analysis is based on time series regressions
and on the identification of long-run relationships between sav
ing, growth, and private wealth and pension wealth. Such an
approach involves computing an estimate of aggregate financial
pension wealth and necessarily neglects aggregation issues. The
main idea of this test is to assess the degree of substitutability of
private and pension wealth by comparing the coefficients on the
two variables in a long-run consumption function. The hypothesis
of full substitutability is that the two coefficients are equal. Such
a hypothesis implies that large changes in social security wealth
are reflected in changes in consumption.
A few studies have attempted to address these issues using
Italian micro data [Brugiavini 1987; Jappelli 1994]. Both these
authors attempt to estimate the degree of substitutability of
financial wealth and pension wealth using just one wave of the
same survey we use in this paper and obtain estimates of the
11. See, for example, Guiso, Jappelli, and Terlizzese [1994] for the former
argument and Guiso and Jappelli [1994] for the latter.
crucial parameter much lower than those obtained by Rossi and

Visco. The problems with the estimation of this type of parame
ters are well-known: Auerbach and Kotlikoff [1983], for instance,
discuss the difficulties with the identification of these parameters
both in time-series and cross-sectional studies. The structural
instability of the parameters is due to the changes in population
structure and to the fluctuations in interest rates as well as in
contribution rates during the transition after the introduction of
a new pension system and its subsequent changes.12
In this paper we propose tests that take a different direction.
Rather than rely on estimates of the degree of substitutability
between private and pension wealth, our tests look directly at
how changes in saving rates are related to changes in pension
wealth. As we discuss above, the Amato reform gives us the
instrument to identify this relationship.
III.B. Institutional Details
The 1992 reform of the Italian social security system13
mainly concerned the basic social security system and set out
some of the principles for the partial introduction of fully funded
occupational pensions. The Italian social security system is made
up of three major funds: private sector employees (contributing to
the INPS fund14), public sector employees, and the self-employed.
The changes introduced by the reform focused on the criteria of
eligibility for old age and early retirement pensions and the size
of the benefit outlays. In particular, the reform envisioned, once
phased in, the following:
an increase in the normal retirement age in the private
sector and for the self-employed, from 60 to 65 for men and
from 55 to 60 for women, for the public sector the normal
retirement age was already 65.
an increase in the minimum number of years of social
security tax payments from fifteen to twenty.
12. Samwick [2000] also looks, in a panel of countries, at structural reforms

which go in the direction of reducing the PAYG component of the old age insur
ance program in favor of a funded component. However, his data set presents
similar aggregation problems of the type discussed by Auerbach and Kotlikoff,
and the author himself questions the robustness of his results.
13. The text of the Law can be found in Istituto Poligrafico e Zecca dello Stato
(1992). The system was changed again in 1995 by the Dini government. We will
not discuss the changes introduced by the Dini reform here.
14. INPS stands for Istituto Nazionale per la Previdenza Sociale (National
Institute for Social Security). Private sector employees are in fact a subset of this
fund, known as private sector employees fund: INPS-FPLD.
the limit of 35 years of contribution for eligibility to an

early retirement option was extended to all funds, hence
restricting access to early retirement particularly for work
ers in the public sector.15
a reduction in benefits for all three funds, achieved by
replacing a pension calculation, which was previously of a
final salary type (average of last five years' gross earnings),
with a career average earnings calculation (benefits to be
computed as a fraction of the average earnings over the
entire working life of an individual).16
Table I summarizes the "steady state" effects of the Amato
reform for different groups of workers. However, it should be
stressed that the reform was supposed to be fully phased in only
after 2032 and a transitional period was envisioned. The rules in
place for the transitional period affected the normal retirement
age, the benefit calculation, and the access to early retirement on
the basis of seniority, i.e., on the basis of accrued rights in 1992.
For senior workers (those who had accrued fifteen years of con
tributions in 1993) the increase in normal retirement age was
introduced only gradually, the benefit calculation rules were al
most untouched and, most importantly, restrictions on eligibility
to early retirement were implemented very gradually.
Hence the transitional period left almost unaffected social
security rights for workers who were on the verge of retirement
while greatly affecting younger workers. Younger workers were
potentially losing a substantial share of their pension wealth
from the reform, particularly if their age-earnings profile was
sufficiently steep. It is relevant to note that the seniority criterion
outlined above does not affect only the very young, i.e., those
entering the labor market in 1993, but to a larger extent those
who had contributed to the system for a substantial number of
years in 1992.
The transition period, therefore, implied that different
groups of workers were affected, in terms of pension wealth, in
radically different ways, over and above the differential effect
15. This is a peculiar feature of the Italian Social Security system: the early
retirement option allowed workers to retire at any age if they had completed a
given number of years of tax payments. This eligibility requirement varied be
tween funds: 35 years for men and women in the private sector, twenty years for
men in the public sector and fifteen years for women in the public sector.
16. Years with particularly low income were excluded from the computations.
Past earnings were converted to current prices using the rate of inflation of the
CPI.
TABLE I
Some Key Features of the Social Security System before and after (S
Prereform
Private sector FPLD Public sector Private
Normal retirement age 60 (men) 65 (all)a 6

55 (women) 60
Pensionable earnings Average of final five Last month's real Career ave
years' real earnings earnings13 (conver
(converted to real values t
values through price index +
index) year of
Pension benefit Fraction of pensionable Fraction of pensionable Fraction o
earnings given by a earnings given by a earning
factor of 2% for each factor of 2.33% for each factor o
year in the system (at year in the system (at year in
most 40 years) most 40 years) most 40
Indexation of pension Cost of living plus real Cost of living plus real Cost of li
earnings growth earnings growth
Years of contributions for 15 15
eligibility
Early retirement Any age if 35 years' SS Any age if 20 years' SS Any age
provision taxes taxes (15 years for taxes
women)
a. To be more specific, this is 65 for Central Government Employees and 60 for Local Government Employees. Ho
eligibility requirement due to the early retirement option.
b. The pensionable earnings calculation is actually made on a monthly basis both in the private sector (where we t
and in the public sector.
c. For further details on the Amato reform, see Brugiavini [1999] and Brugiavini and Fornero [2001].
implied by the way in which the reform changed the existing

legislation in the steady state. In subsection III.D we describe in
detail the changes experienced in 1992 by different groups.
III.C. Some Preliminary Evidence

In what follows, we estimate for the households in our data
set the value of pension wealth, taking into account all the insti
tutional details discussed above and relate these estimates to
their saving behavior. Before embarking on the discussion of our
estimates of household pension wealth, it is worth exploiting
some of the aspects of the reform which naturally divide our
households into groups on which the effect of the reform was
substantially different. In particular, individuals with more than
or less than fifteen years of contributions experienced a differen
tial treatment in the overall generosity and eligibility require
ments. While, in what follows, we exploit these differences to
maximize the differences in pension wealth changes across our
groups, in this subsection we use regression discontinuity design
to estimate the effect of the reform on saving rates. The advan
tage of this procedure is that it does not require an estimate of
pension wealth. The possible disadvantage relative to the more
structural evidence we present below is that this procedure might
yield less precise estimates and does not quantify the degree of
substitutability between pension wealth and saving.
We split the sample between the households headed by an
individual with more than or less than fifteen years of contribu
tions in 1992. We then run a nonparametric regression of saving
rates (obtained as income minus consumption over income) on
tenure (defined as the number of years of contributions to the
social security system for the head of the household) for each of
these two groups in 1993 and 1991 and evaluate the conditional
expectation at sixteen and fourteen years of tenure, respec
tively.17 If we define with sry(i,t) the saving rate of a household
belonging to group y (less than fifteen years of tenure in 1992)
with i years of tenure at time t and with sr0(i,t), the correspond
ing quantity for a household belonging to group o (more than
fifteen years of tenure in 1992) the effect of interest is obtained as
17. We experimented with different values of the bandwidth parameter, and

we chose the standard value obtained for the normal Kernel whereby the band
width depends on the sample size and the standard deviation of the independent
variable h = 1.06o-(n)~ , as this did not seem to impose excessive smoothing.
[*ry(16,93) - *r0(16,93)] - [sry(14,91) - sr0(14,91)],
where sry and sr0 are the fitted values of the nonparametric
regressions for the individuals with less (y) and more (o) than
fifteen years of contributions in 1992. We compute the standard
errors of these effects by bootstrapping our sample, also taking
into account individual level clusters for the panel component of
the sample.
When applied to the whole sample of private sector and
public sector employees, we get an effect of 0.049 with a standard
error of 0.098. As the effect of the reform was much higher on
public sector employees, we redo the exercise for this group and
for the private sector employees separately. For the former we
obtain an effect of 0.17 with a standard error of 0.091. For the
latter group we obtain an effect of 0.090 with a standard error of
0.080.
This constitutes a first piece of evidence, albeit imprecise,
that the Amato reform had a differential effect on individuals
whose pension wealth changed differently as a consequence of the
reform. To better quantify this effect, we use explicit estimates of
pension wealth, whose estimation we now proceed to discuss.
III.D. Estimates of Public Pension Wealth and the

Effect of the Amato Reform
As our purpose is to relate individual public pension wealth
and individual saving rates, we need a microeconomic data set.
We use the 1989, 1991, 1993, and 1995 Survey on Household
Income and Wealth conducted by the Bank of Italy (SHIW). These
surveys, described in detail in the Appendix, contain information
on saving (defined as income minus consumption) as well as on a
variety of other variables (such as occupation, age, other demo
graphics, and labor market behavior). The information contained
in the survey allows us to compute, for each income recipient, an
estimate of his or her public pension wealth before and after the
reform.
Obviously, in estimating household pension wealth, we have
to make a number of strong assumptions. While not unreason
able, they can be criticized for a variety of reasons. It should be
stressed, however, that for the estimation strategy pursued in
this paper, as we use a difference-in-difference estimator, what is
relevant is not that we get the level of pension wealth exactly
right, but rather that we estimate without systematic biases the

differential changes in pension wealth induced by the 1992 pen
sion reform.
Finally, some would argue that Italian households were
"overannuitized," and the reform, by reducing annuitized re
sources, would have taken households closer to their optimal
provision for retirement with a reduced impact on current con
sumption and saving. It is hard to assess the validity of this
argument, but two considerations are in order: while it may be
true that some workers would rather move to a more funded
system, once contributions have been paid into the system (the
payroll tax was 22 percent of gross earnings at the time of the
reform), a reduction in future benefits was essentially a wealth
tax. Even worse, younger workers, who had not yet accrued
substantial rights, were expected to pay higher payroll taxes and
receive less generous pensions as a result of the reform. If any
thing, the attitude seemed to take advantage of the generosity of
the system by claiming early retirement as soon as possible (see
Brugiavini [1999]).
Each component of the Amato reform has a different effect on
each of the groups we consider. In many cases it is not obvious to
compute the net combined effect of the many changes induced by
the reform on pension wealth. Some broad patterns, however, are
quite clear and have been discussed in the literature.
Peracchi and Rossi [1995] talk of a "wealth tax" imposed by
the reform on the active population, referring to the change in the
benefit computation method, which mainly affects younger gen
erations. The redistribution brought about by the reform across
sectors is a bit more complex: it is clear that the change in the
minimum number of years of contributions to obtain early retire
ment benefits hits the public sector employees. Furthermore, it
should be stressed that the changes in the early retirement bene
fit for the public sector were, at the time following the reform, one
of the most debated issues. It can safely be assumed that the
households headed by a public sector employee were acutely
aware of the implications of the new rules. More generally, it can
be argued that public-sector employees were the group that most
enjoyed the generosity of the Italian social security system in the
prereform regime and were presumably most affected by the
changes. It is this type of variability that allows us to perform our
tests.
It should be added that an abrupt recession episode took

place in 1993 in Italy, marked by a fall in consumer expenditure
and in household disposable income. The implicit assumption we
are making is that the recession episode did not affect the groups
we are considering in different ways, at least for those variables
that are important determinants of savings. Miniaci and Weber
[1999] analyze in detail this episode and write explicitly about the
pension reform and identify it as one of the main forces behind
the change in observed saving rates.
The research strategy we use is based on dividing the sample
into a number of groups chosen to maximize the variation, across
groups, in pension wealth changes induced by the reform.18 The
institutional details discussed above were therefore crucial in
forming the definition of the groups. We divide the sample on the
basis of year of birth (cohort)19 and sector of activity of the
household head. We divide the active population into three large
groups: private sector employees (who participate in what is
known as FPLD?the Italian acronym for private employees
fund?see note 4 above), public sector employees, and other,
including mainly the self-employed.20 In addition, we have the
pensioners' group, which is affected by the reform only through
the indexation of future benefits. As far as the definition of co
horts is concerned, we consider four large groups defined on the
basis of the year of birth of the household head: those born in
1903-1934, those born in 1935-1944, those born in 1945-1957,
and those born after 1957. These intervals were chosen so to
match the crucial seniority levels used in the new regime dis
cussed above. Further details on the definition of cohorts are
provided in the Appendix. Mean values for the relevant variables
are contained in Table VIII also in the Appendix.
As we have information on education levels and occupation of
18. The hypothesis that group membership is constant over time is particu
larly problematic for young and old households, as household headship might
change endogenously with age at the beginning and at the end of the life cycle, as
household formation and dissolution are endogenous to saving decisions. This
consideration should make us cautious in interpreting the results we present
below.
19. While in estimating earnings growth, we use up to seven cohorts, we use
only four cohorts in forming groups to obtain instruments for the changes in
pension wealth. These were chosen according to the "seniority" rules described
above.
20. In this group we also have included employees who do not belong to the
major social security administration Private Sector Employees Fund (INPS), such
as managers, journalists, actors, and a few other employees.
TABLE II
Mean Net Household Pension Wealth (Panel A) and Mean Saving Rates
(Panel B) for Different Groups (Pension Wealth in Millions
of 1994 Italian lira, Saving Rates in Percentage Points)
Panel A: Mean net pension wealth

1989 1991 1993 1995
Occupation and cohort
Private sector employees
Born after 1957 1737 1715 1240 1193
1945-1957 999 985 817 743
1935-1944 796 732 626 500
1903-1934 664 646 525 280
Public sector employees
Born after 1957 1322 1496 996 899
1945-1957 1030 1162 846 729
1935-1944 858 892 737 591
1903-1934 842 731 599 429
"Other" active head of household
Born after 1957 1518 1642 1032 1040
1945-1957 1179 1263 803 702
1935-1944 915 1006 712 637
1903-1934 784 793 600 500
Retired
Born after 1957
1945-1957 696 725 614 472
1935-1944 657 707 577 497
1903-1934 415 414 341 287
both spouses, in principle it would be possible to define finer and

more homogeneous groups. In particular, households could be
divided, for instance, also on the basis of the spouse occupation,
age, and education and not only on the characteristics of the
household head. The problem with such a strategy, of course, is
that of having extremely small, if not empty, cells. A much larger
sample would allow us a much more efficient identification of
groups.
In Table II we report averages of net household pension
wealth and saving rates for the years before and the years after
the reform for different cohorts and occupational groups; while in
Table III we compute percentage changes in mean values for the
same variables (net pension wealth and saving rates) between
two years of the survey for each group. The main message from
these tables is that there is a substantial decline in pension
wealth between 1991 and 1993. Moreover, there are marked
TABLE II
(continued)
Panel B: Mean saving over mean income X 100

1989 1991 1993 1995
Born after 1957 13.2 10.9 4.7 10.0
1945-1957 17.2 15.7 14.7 15.3
1935-1944 22.0 19.8 20.6 8.0
1903-1934 25.5 20.8 16.6 14.2
Born after 1957 14.8 10.7 16.0 12.5
1945-1957 18.6 15.9 17.6 16.5
1935-1944 21.3 22.1 23.0 22.6
1903-1934 24.9 22.0 29.9 25.4
Born after 1957 19.2 11.2 -9.0 -4.5
1945-1957 19.3 17.3 2.5 1.9
1935-1944 11.7 19.4 4.6 14.4
1903-1934 24.0 25.7 21.5 29.1
Retired
Born after 1957
1945-1957 2.9 13.5 9.6 -2.8
1935-1944 17.9 19.4 15.9 17.3
1903-1934 21.5 21.9 16.3 17.2
Mean saving rates are computed as mean saving over mean income. Net social security wealth i
present discounted value of future benefits minus contributions due. For the youngest cohort there are
heads of household collecting benefits, but these are typically not old age benefits. This is why the row for t
youngest cohort in the group "retired" appears empty both in Panel A and in Panel B. The group of re
households is not used in the econometric analysis.
differences in how the reform affects different cohorts and occ

pational groups. Notice, for instance, the large negative effec
the pension reform on younger cohorts and on public sector
ployees (but young self-employed workers are also substantia
affected). In Panel B of Table III we notice that, in 1993, for s
groups there is an increase in saving rates, while for others ther
is a decline. Public sector employees, for instance, are the gr
that shows the largest positive changes in saving rates.
The numbers in Table III can be further explored to provid
before moving to our more formal analysis in the following
tion, some preliminary evidence about the relationship betw
pension wealth and saving. In Figure I we plot the change in
median saving rate for each of the cohort/occupation groups
consider against the change in average pension wealth for t
same groups. The analysis in Section IV identifies in a m
TABLE III
Changes in Mean Net Household Pension Wealth (Panel A) and
Changes in Mean Saving Rates (Panel B) for Different Groups
(Percentage Change for Two-Year Intervals)
Panel A: Changes in mean net household pension wealth

1991 1993 1995
Occupation and cohort
Born after 1957 -1.2 -27.6 -3.8
1945-1957 -1.3 -17.0 -9.1
1935-1944 -8.1 -14.3 -20.1
1903-1934 -2.7 -18.6 -16.6
Born after 1957 11.0 -32.1 -9.7
1945-1957 12.8 -27.1 -13.8
1935-1944 3.9 -17.4 -19.7
1903-1934 -11.2 -18.1 -28.3
Born after 1957 8.2 -30.1 0.7
1945-1957 7.1 -26.3 -12.5
1935-1944 9.9 -14.3 -10.4
1903-1934 1.1 -14.4 -16.7
Retired
Born after 1957 ? ? ?
1945-1957 4.2 -15.3 -23.0
1935-1944 7.6 -18.3 -13.9
1903-1934 -0.3 -17.4 -15.9
Panel B: changes in mean saving over mean income

1991 1993 1995
Born after 1957 -17.5 -55.9 58.2
1945-1957 -8.2 -6.5 3.9
1935-1944 -10.3 4.3 -61.0
1903-1934 -18.2 -20.3 -14.3
Born after 1957 -27.7 50.0 -21.4
1945-1957 -14.4 10.7 -6.4
1935-1944 3.5 25.1 -1.9
1903-1934 -11.5 25.9 -15.1
Born after 1957 -41.4 -101.0 -50.6
1945-1957 -10.3 -85.8 -23.9
1935-1944 25.3 -76.2 112.5
1903-1934 7.2 -16.2 35.0
Retired
Born after 1957 ? ? ?
1945-1957 30.5 -29.0 -129.7
1935-1944 8.3 -17.9 8.3
1903-1934 1.9 -25.5 5.4
changes in saving rates
-.2 0~ .2
changes in pension wealth
Figure I
Changes in Median Saving Rate against Changes in Median Pension Wealth
Figure I shows changes in median saving rates against changes in median
pension wealth on the sample of active households (nonpensioners) for the four
years (four data points in each group are lost in taking differences). Pension
wealth has been rescaled before taking the differences and is in thousands of 1994
million lira. The numbers in the scatter diagram represent the groups: for exam
ple, 13 stands for cohort 1 (youngest cohort) in occupation 3 (occupation other).
The regression is based on 33 observations, and the estimated slope is -0.153 (s.e.
= 0.099).
rigorous fashion and quantifies the relationship between savings

and pension out of the relative variation in pension wealth. In
Figure I, however, we show that such a relationship is already
visible in this simple picture: there is a negative relationship
between the changes in saving rates and the changes in pension
wealth. Although a clear pattern does not emerge for all groups,
a cluster of cohorts whose members are employees in the public
sector (groups 12, 22, 32, and 42, respectively) tend to have
marked negative changes in median pension wealth associated
with positive changes in saving rates.
IV. The Effects of Pension Wealth on Saving Rates
Having measured pension wealth for each household in the

sample before and after the Amato reform and having described
how average changes in this variable are related to average

changes in saving for different groups, we are now in a position to
move to a more structured analysis. As we discussed above, we
use a simple regression framework. In particular, for each house
hold in the sample, we relate saving rates to pension wealth and
future earnings, both of which are corrected by the factor speci
fied in equation (5). These two variables are interacted, in most
specifications, with age dummies. The interaction with age of the
pension wealth variable allows the degree of substitutability be
tween pension and financial wealth to be age dependent. In
addition, we add year dummies to capture aggregate effects,
group dummies to control for fixed group effects and a term in age
to capture differences in age within each cohort.21
In order to account for the effects of demographic variables on
saving rates, we experimented with two sets of variables. In one
specification we make use of the number of children in the house
hold (as a percentage of total household members) and number of
income recipients (also in percentage terms). This is to capture
the possibility that within-household redistribution takes place,
particularly if more income recipients are coresident. We also use
dummies for education attainment (no education or elementary
education, secondary education, university degree), these should
be relevant both for shaping preferences for saving, e.g., vis-?-vis
precautionary saving, and to proxy for a general "level of welfare"
measure. We cannot argue, as in Gale [1998], that specifically
financial literacy would play a role, as we are not looking in any
way at private pension plans, nor do we consider portfolio com
position. Hence education plays a role more as a proxy for welfare
conditions, which is why we experiment with different sets of
demographic variables and, taken as a group, these could also
help to capture the existence of liquidity constraints.
We estimate our specifications on the subsample of the active
population including all workers and report our estimation re
sults in Tables IV to VII. Each table contains four columns cor
responding to the cases with and without the year 1995 and with
and without future earnings. When we exclude future earnings,
we assume that their effect is captured by group and time dum
mies. We have a slight preference for the results that exclude
21. As groups are defined, among other things, by year-of-birth cohorts,

group, and time effects span, for the most part age effects. Age only captures
differences in age within a cohort, as this is defined by an interval of years.
1995, as for the households observed in that year, we should take

into account the fact that they experienced the reform three years
in advance, while those in 1993, experienced the reform the
previous year.
In all specifications the left-hand side of the equation is the
saving rate defined as income minus consumption divided by
income. Pension wealth and future human wealth (when in
cluded) are both adjusted with the factor reported in equation (5),
with 7 = 1 and ? = 0.98. In the tables we only report the
coefficients on pension wealth and future earnings, the coeffi
cients on the year dummies, and the other relevant explanatory
variables. To facilitate the interpretation of the results, we also
provide the total sample average of the coefficient on pension
wealth.
We estimate the coefficients of interest by Instrumental Vari
ables, using as instruments the interaction of time and group
dummies. To check the rank condition on this estimator, we
regress our estimates of pension wealth on time and group dum
mies and their interactions and test for the significance of the
interaction terms. We reject the null at any sensible level of
significance (F(33,18546) = 12.28). In Table IV we regress saving
rates on pension wealth and on demographics by using the full set
of instrumental variables, which include the interaction of year
dummies with cohort-occupation dummies. With the exception of
column 1 (all years, no future earnings), the coefficient of (ad
justed) pension wealth is not significantly different from zero and
its point estimate is small (0.03).
In Table V we adopt the same procedure as in Table IV, but
let the coefficient on pension wealth be dependent on age, to allow
for the possibility that the degree of substitutability between
financial and pension wealth changes over the life cycle. The
results we obtain indicate a substantial amount of substitutabil
ity for some age groups and little for others, as we strongly reject
the null that the coefficient on pension wealth is the same for
different age groups. We plot the coefficients on pension wealth
of columns 1 and 3 in the top panels of Figures II (a similar
picture can be obtained for the coefficients in columns 2 and 4).
We notice a somewhat puzzling U-shaped age pattern, in that the
youngest and oldest households seem to have the lowest degree of
substitutability between pension and financial wealth. While one
could imagine that the low degree of substitutability for young
TABLE IV
Estimating the Effect of Pension Wealth on Saving
Dependent Variable: Saving Rate
Baseline Specification with No Interaction Terms
Excluding expected Including expected

future earnings future earnings
(1) All years (2) 89-91-93 (3) All years (4) 89-91-93
(N = 18598) (N = 14522) (N = 18598) (N = 14522)
Pension wealtht -0.328 -0.037 -0.037 -0.037
(0.023) (0.045) (0.032) (0.042)
Expected value of 0.089 0.004
future earnings (0.244) (0.443)
share of earners 0.310 0.278 0.324 0.278
(0.037) (0.040) (0.043) (0.075)
Share of children 0.099 0.054 0.086 0.053
(0.044) (0.049) (0.062) (0.077)
Secondary school 0.016 0.060 0.016 0.060
(0.007) (0.013) (0.007) (0.017)
College or more 0.073 0.118 0.076 0.118
(0.014) (0.017) (0.019) (0.018)
Year 1991 -0.014 -0.014 -0.015 -0.014
(0.003) (0.002) (0.003) (0.002)
Year 1993 -0.080 -0.080 -0.073 -0.079
(0.031) (0.034) (0.033) (0.050)
Year 1995 -0.083 -0.069
(0.032) (0.039)
Group dummies Y Y Y Y
Number of IV 51 39 51 39
Saving rates are defined as (Income - Consumption)/Income. Pension wealt

wealth over current income by 1000. Expected future earnings are the present v
current earnings by 1000. Reference categories for the dummy variables are "no
and "year 1989." Standard errors in parentheses are robust standard errors, allow
on cohort-occupation groups.
households could be explained by liquidity c

harder to tell a story to explain the pattern o
households. It should be stressed that, at least
cations, the degree of substitutability for the m
holds is very high, and close to -1. These nu
higher than what was found by previous resea
possible exception of Gale [1998] and, on aggre
data, Feldstein [1974].
The variability that identifies the paramete
Table V is the interaction between time and gr
groups are defined by the interaction of coho
groups. The identifying assumption is that the
TABLE V
Specification with Full Interaction Terms in the Instruments Set

(N = 18598) (N = 14522) (N = 18598) (N = 14522)
(Age 20-35)* PWit -0.052 -0.043 -0.104 -0.195

(0.073) (0.074) (0.082) (0.102)
(Age 36-45)* PWit -0.092 -0.268 -0.754 -1.149
(0.128) (0.174) (0.286) (0.498)
(Age 46-55)* PWit -0.041 -0.153 -0.237 -0.830
(0.083) (0.098) (0.217) (0.152)
(Age 56-60)* PWit -0.045 -0.131 -0.055 -0.390
(0.043) (0.067) (0.131) (0.125)
(Age 61-65)* PWit -0.030 -0.088 -0.070 -0.152
(0.010) (0.040) (0.035) (0.035)
(Age 20-35)* FEit -0.119 -0.804
(0.542) (0.449)
(Age 36-45)* FEit 1.410 2.195
(0.610) (1.500)
(Age 46-55)* FEit 0.199 0.839
(0.241) (0.349)
(Age 56-60)* FEit -0.232 0.390
(0.483) (0.353)
(Age 61-65)* FEit 0.323 0.271
(0.265) (0.225)
Share of earners 0.319 0.254 0.372 0.275
(0.039) (0.060) (0.087) (0.134)
Share of children 0.126 0.097 0.100 -0.008
(0.068) (0.096) (0.111) (0.158)
(0.008) (0.016) (0.013) (0.023)
College or more 0.073 0.123 0.090 0.123
(0.014) (0.017) (0.046) (0.067)
Year 1991 -0.015 -0.003 -0.024 -0.037
(0.006) (0.010) (0.012) (0.012)
Year 1993 -0.094 -0.107 -0.178 -0.313
(0.055) (0.056) (0.106) (0.102)
Year 1995 -0.104 -0.224
(0.073) (0.160)
Group dummies Y Y Y Y
Average coefficient on
pension wealth -0.060 -0.160 -0.352 -0.717
Saving rates are defined as (Income - Consumption)/Income. Pension wealth (PW) is defined as net
pension wealth over current income by 1000. Expected future earnings (FE) are the present value of future
earnings over current earnings by 1000. Reference categories for the dummy variables are "no education or
low education" and "year 1989." Standard errors in parentheses are robust standard errors, allowing for
cluster effects based on cohort-occupation groups.
time-specific effects on savings. In Table VI we somewhat relax

this assumption and include as controls in the regression dum
mies for the interaction between years and cohorts only. There
estimated coefficient estimated coefficient
Figure II
Life Cycle Pattern of the Estimated Substitution Coefficient
Figure II contains the estimated coefficient against age for the full sample of
four years and for the specifications containing interaction terms of pension
wealth with age dummies. Top-left panel: specification with no future earnings
(Table V, column 1); top right panel: specification with future earnings (Table V,
column 3); (3) bottom-left panel: specification without future earnings with year *
cohort used as control (Table VI, column 1); and (4) bottom right panel: specifi
cation with future earnings and year * cohort used as controls (Table VI, column
3).
fore, the parameters of interest are identified through the year

and occupation interaction.22
This specification, which would be the right one in the pres
ence of cohort-specific effects of, say, the 1993 recession, yields
larger substitutability for young groups and a limited effect for
older groups. The interactions of time and cohort dummies are
statistically significant, perhaps signaling that different cohorts
were indeed affected by the recession in different ways. We plot
22. The rank condition is again satisfied. The F-statistic for the hypothesis
that time * occupation effects in pension wealth are zero after controlling for time,
group, and time * cohort interaction is F(6, 18573) = 2.29 which has ap-value
of 0.0328.
TABLE VI
Specification with Interaction Terms for Years and Cohorts
Included as Controls

(N - 18598) (N = 14522) (N = 18598) (N = 14522)
(Age 20-35)* PWit -0.435 -0.418 -0.300 -0.421

(0.117) (0.150) (0.064) (0.077)
(Age 36-45)* PWit -0.428 -0.340 -1.964 -2.690
(0.204) (0.204) (0.266) (0.519)
(Age 46-55)* PWit -0.088 -0.270 -0.517 -0.486
(0.344) (0.191) (0.202) (0.109)
(Age 56-60)* PWit -0.179 -0.194 -0.248 -0.213
(0.141) (0.129) (0.224) (0.143)
(Age 61-65)* PWit -0.092 -0.103 -0.125 -0.106
(0.051) (0.052) (0.065) (0.044)
(Age 20-35)* FEit -2.731 -3.158
(0.708) (0.994)
(Age 36-45)* FEit 3.125 4.883
(0.454) (1.052)
(Age 46-55)* FEit 0.286 1.500
(0.259) (0.383)
(Age 56-60)* FEit 0.229 0.907
(0.636) (0.553)
(Age 61-65)* FEit 0.259 0.267
(0.238) (0.151)
Share of earners 0.389 0.282 0.444 0.706
(0.139) (0.091) (0.176) (0.280)
(0.176) (0.106) (0.125) (0.226)
(0.007) (0.015) (0.027) (0.079)
College or more 0.081 0.129 0.057 0.117
(0.025) (0.021) (0.061) (0.106)
Year 1991 -0.099 -0.096 0.033 0.035
(0.008) (0.127) (0.056) (0.076)
Year 1993 -0.489 -0.471 -0.522 -0.664
(0.097) (0.126) (0.063) (0.078)
Year 1995 -0.543 -0.981
(0.133) (0.104)
Group dummies Y Y
Interaction dummies
(Years * Cohorts) Y Y Y Y
pension wealth -0.289 -0.316 -0.898 1.157
low education" and "year 1989." Standard errors in parentheses are robust standard errors according to
TABLE VII
Specification with Interaction Terms for Years and
Occupation Included as Controls
Excluding expected future Including expected future

earnings earnings
(1) All years (2)89-91-93 (3) All years (4)89-91-93

(N = 18598) (N = 14522) (N = 18598) (N = 14522)
(Age 20-35)* PWit 0.048 0.046 0.041 0.012

(0.031) (0.030) (0.044) (0.023)
(Age 36-45)* PWlt 0.078 -0.046 0.097 0.077
(0.068) (0.068) (0.202) (0.110)
(Age 46-55)* PWit 0.056 -0.042 0.177 -0.223
(0.057) (0.035) (0.127) (0.040)
(Age 56-60)* PWit -0.006 -0.086 0.110 -0.107
(0.033) (0.035) (0.092) (0.040)
(Age 61-65)* PWit -0.014 -0.067 0.018 -0.059
(0.014) (0.017) (0.017) (0.011)
(Age 20-35)* FEit -0.454 -0.646
(0.134) (0.134)
(Age 36-45)* FEit -0.413 -0.742
(0.377) (0.300)
(Age 46-55)* FEit -0.700 -0.008
(0.166) (0.078)
(Age 56-60)* FEit -0.837 -0.368
(0.297) (0.129)
(Age 61-65)* FEit -0.330 -0.344
(0.238) (0.065)
Share of earners 0.311 0.257 0.236 0.162
(0.029) (0.035) (0.040) (0.039)
(0.039) (0.042) (0.046) (0.057)
(0.009) (0.011) (0.009) (0.010)
College or more 0.077 0.117 0.068 0.088
(0.013) (0.014) (0.024) (0.021)
Year 1991 -0.009 -0.001 -0.022 0.021
(0.008) (0.007) (0.011) (0.011)
Year 1993 -0.011 -0.014 -0.057 -0.096
(0.026) (0.017) (0.053) (0.024)
Year 1995 -0.004 -0.071
(0.037) (0.069)
Group dummies Y Y
Interaction dummies
(Years * Occupation) Y Y Y Y
pension wealth 0.053 -0.026 0.107 -0.052
low education" and "year 1989." Standard errors in parentheses are robust standard errors, allowing for
the coefficients on pension wealth in columns 1 and 3, which are

now identified only by the variation induced by the reform across
education groups, in the bottom panel of Figure II. The age
pattern of the substitutability between savings and pension
wealth is now monotone in the specification without future earn
ings, in that the only groups that seem to have low substitutabil
ity of pension and financial saving are, surprisingly, the older
consumers.
Finally, in Table VII we introduce the interaction of occu
tion and years as controls, while excluding the interaction of ye
and cohorts. Identification is now obtained through the inter
tion of years and cohorts. We find that future pension wealt
completely insignificant. On the other hand, the occupation * ye
effects are not statistically significant in our regression. Th
results might not be surprising as the identification of age-spec
coefficients on pension wealth is problematic, as we are usi
cohort-specific variation only at one point in time.
To summarize, when we allow the degree of substitutabili
between financial and pension wealth to be a function of age
both samples and for most specifications, we find a significant a
sizable offset of the pension wealth variable on the saving rate o
the expected sign. However, the point estimates are usually ab
minus one, indicating less than perfect substitutability betw
pension and financial wealth. Moreover, there is consider
variation in the size of this coefficient across ages and specif
tions. As is evident in Figure II, where we plot 6(a? t), the coe
cient on pension wealth is much higher in absolute value over th
40-50 age interval, as it may be expected if liquidity constra
are relevant for the youngest individuals. The consistently
substitutability between pension and financial wealth we es
mate for the older consumers (above 50) is somewhat puzzling
this group should not be affected by liquidity constraints.
possibility is that we are somehow misestimating the adjustm
factor (see equation (5)) and that this bias affects more the
horts who are closer to retirement. The degree of substitutab
for the youngest individuals depends on whether we add cont
for cohort * year effects. When we do, we find much high
substitutability.
As for the differences across specifications, the coefficient o
pension wealth is much closer to -1 in the specifications th
includes the estimates of future earnings, than in those where t
latter is proxied by age effects.23 In particular, when we exclude

future earnings, the estimated coefficient is never smaller than
-0.3 or -0.4, while in the second specification for the middle
aged groups, is not significantly different from -1.
It should be stressed that our results, which are directly
derived from the simple life cycle model sketched in Section II,
are not directly comparable to previous results obtained in the
literature, as we make use of the saving rate as a dependent
variable, rather than regressing the stock of private wealth
against the stock of pension wealth. To ease the interpretation,
we have attempted a miniature policy experiment. Take the av
erage individual, who is characterized by the mean value of
continuous variables and the reference category for the dummy
variables (e.g., year 1989, no or low education, etc.). Suppose that
the ratio of pension wealth over current income increases by 10
percent (for given current income this corresponds, for such rep
resentative individual, to an approximate increase of 150 million
lira in pension wealth). If we feed this change through our esti
mates for the average individual, everything else being equal, we
obtain a saving rate which goes from 0.0865 to 0.044 (a reduction
of approximately 50 percentage points). This is quite dramatic,
and it shows that the impact of the change is most felt. Obviously,
it cannot be generalized to the entire sample or to the population
at large. Given the nonlinearity of the relationships involved, it is
also hard to provide a round number for the effects in terms of
changes in the aggregate capital stock.
V. Conclusions
In this paper we have estimated the effect that public pen
sion wealth has on personal household saving. For this purpose
we have studied the 1992 Italian pension reform. This episode is
particularly useful for several reasons. First of all, we have two
large and consistent household surveys that immediately precede
and follow the reform. Second, the reform did not change the
nature of the pension system, in that the Italian system remained
an unfunded defined-benefits system. The 1992 reform, however,
23. It should be noted that the identification of the coefficients on future

earnings is problematic, as it is not clear that there exists exogenous variation of
the kind we use to identify the coefficient on pension wealth.
substantially changed the present discounted wealth of a large

majority of Italian households. Furthermore, and more impor
tantly for our purposes, the reduction in public pension wealth
was far from uniform across households. It is the variability in
the changes in pension wealth across well-defined groups of Ital
ian households that we exploit to identify the effect that pension
wealth has on saving rates.
The first step of our exercise consists of the estimation of the
level of pension wealth for each household included in our sam
ples, before and after the reform. This is done at the individual
level by using the information on the relevant pension law as well
as extrapolation of the levels of earning and some assumptions
about retirement ages.
The results indicate that pension wealth is a substitute for
private financial wealth (more specifically for private saving),
especially for individuals in the middle of their life cycle. Our
assessment of how good a substitute, however, depends cru
cially on the specification we use. When we include an explicit
estimate of future earnings along with our estimate of future
pension wealth, we obtain that for some age groups, pension
wealth is a perfect substitute for financial wealth. However,
when we proxy such a variable with age effects, we find much
lower estimates of the degree of substitutability. For the same
age group a lira of pension wealth is worth at most 0.4 lira of
saving.
Our estimates of the degree of substitutability depend also on
which of the two samples we use and on the particular parame
terization. For the larger sample it is on average -0.30 and -0.4
in the smaller sample when pension wealth is interacted with an
age polynomial, while when interacting pension wealth with age
dummies, for the larger sample we obtain an estimate of an
average effect of -0.35 (and as large as ?0.71 in the smaller
sample).
Our results constitute one of the first pieces of evidence in
the European literature derived from micro data on the rela
tionship between the provision of social security and household
saving. In this sense, they complement the time-series evi
dence provided by Feldstein [1974] and many other authors.
They also provide information on the importance of life cycle
saving, and more generally, on the validity of the life cycle
model of consumption.
Appendix
A. The Survey of Household Income and Wealth

The Survey of Household Income and Wealth (SHIW) is
conducted, since 1987, every other year and has been widely used
by several researchers to analyze the saving behavior of Italian
households. For a detailed analysis of the features of the sample,
see Brandolini and Cannari [1994]. In this Appendix we briefly
describe the definitions and the selection criteria we use.
The survey contains detailed information on household in
come, consumption, and wealth, as well as a number of demo
graphic and economic variables. While the information on con
sumption is less detailed than that available in other surveys
(such as the U. K. Family Expenditure Survey and the U. S.
Consumer Expenditure Survey or even the Survey on Household
Consumption in Italy), it is, however, sufficient to construct sav
ings flows as disposable income minus consumption expenditure.
The data are generally of excellent quality and have been used for
a variety of different studies.24 Although the saving rate in the
survey is higher than the national saving rate, the difference
between the two estimates is broadly consistent over time and
should not affect our econometric exercise (see Brugiavini and
Padula [2001]). Since 1989, the survey also contains a (relatively
small) longitudinal component.
The main advantage of the Survey for our study is the fact
that it was conducted in 1989, 1991, 1993, and 1995, that is,
before and after the Amato reform of the pension system. We do
not make direct use of the longitudinal component of the survey
(even though we correct the standard errors for the fact that some
households appear in both time periods we use). In fact, the
advantage of having a longitudinal component is not a large one,
as we focus on differences across groups, and therefore it will be
sufficient to compare means across groups and not necessarily
follow the same individuals over time. In our empirical exercise in
addition to the demographic variables, we use the information
available on consumption and disposable income. Income is ob
servable in two components: labor earnings and other income.
24. Of particular interest, because of their originality, are the modules on

households' expectations and on cash balances. In what follows, we use the data
on expected retirement age.
Unfortunately, it is not possible to measure capital gains (or

losses) on assets owned by the households.25
Consumption is measured as an expenditure on durables and
nondurables and services over the last year. While this division is
not very fine, it allows us, at least in principle, to isolate the
saving component of the expenditure on durables. Ideally, one
would like to add to consumption the service flow from durables.
As there is almost no information on the stock of durables, this is
not feasible. We therefore use two alternative definitions of con
sumption and therefore saving. We first include durables in con
sumption and then in saving. However, we only report results for
the former definition as there is no appreciable change.26
In addition to consumption and income and the variables
that are used to define the groups (such as head of the household's
age, occupation, and level of education), we also use other vari
ables of particular interest for our analysis, such as the informa
tion on planned retirement age. Unfortunately, some of the vari
ables observed in 1991, which could have been useful for our
analysis, were not collected in all the other years. In particular, in
1993 there is no information on income uncertainty and on the
expected replacement rates at retirement.27 Information on this
last variable, in particular, would have allowed us, on the one
hand, to identify the households whose expected (or perceived)
pension wealth had declined the most as a consequence of the
1992 reform and, on the other hand, to relate these changes to
their saving behavior. The information on uncertainty would
have allowed us to check the hypothesis that the increase in
saving in 1993 was due to an increase in the precautionary motive
(although this question was not asked of the entire sample to
start with). The sample size effectively used (after estimation of
pension wealth) is 14,522 households in 1989, 1991, and 1993,
and 18,598 households in total for the four years. We excluded
households where some of the crucial information was missing,
25. We do not correct for inflation on assets denominated in nominal terms.

As inflation was relatively low over the period considered, we do not think this is
too important.
26. As we consider group means, it should be remembered that durable
expenditure contains, at the individual level, many zeros. A possible interpreta
tion of the procedure that includes durable expenditure into consumption is that
the average expenditure approximates durable services for that group.
27. Furthermore, only part of the sample answers the questions on income
uncertainty, and this would make the sample size too small for our application.
particularly the information necessary to identify groups, the

information on consumption, the information necessary to gross
up income (see below), and to compute social security wealth.
B. Group Definitions
There are two levels to the analysis. At a first level we
distinguish groups of workers in order to model earnings growth;
at a second level we define groups in order to estimate the effect
of changes in pension wealth on saving rates. The two sets of
definitions are consistent; however, we use a finer grouping at the
former level than at the latter. In particular, at the first level we
distinguish seven cohorts, gender, three occupational categories,
and three education levels. At the second level we focus our
attention on the household rather than on the individual, and
given the reduction in sample size, we form groups based only on
the age of the head of the household, and we reduce the number
of cohorts to four. The definition of cohort in the second stage was
chosen in order to match the seniority levels in 1993 that trig
gered changes in the treatment of past contribution according to
the Amato reform.
C. -Year-of-Birth-Cohorts
In the estimation of the age-earnings profile, we use seven
year-of-birth cohorts: the focus here was to define cohorts in order
to capture productivity changes. The choice of the size of each
cohort is partially constrained by the coding of age contained in
the SHIW before the 1984 Survey, where age was provided in
fixed bands of approximately ten years each. Ideally, in order to
measure earnings' growth, one would like to follow each cohort
back in time as far as possible. In the SHIW this is not possible in
a direct way because individuals within a given age-band could
belong to more than one cohort. In order to overcome this prob
lem, we adopted a simple weighted average measure of mean
(median) earnings, which provides unbiased estimates of the
actual mean (median) earnings of each cohort in each year. The
cohorts we use are the following: workers born before 1918, those
born between 1918 and 1924, those born between 1925 and 1933,
between 1934 and 1940, between 1941 and 1949, between 1950
and 1955, and after 1955.
A slightly different approach is taken when relating saving
rates to pension wealth (and other controls). As explained in the
text, for the instrumenting strategy we pursue, what matters

there is that we match the seniority levels in 1992 in order to
maximize the differences in the effect of the reform on pension
wealth. Hence we consider four groups: those born in 1903-1934,
those born in 1935-1944, those born in 1945-1957 and those born
after 1957.
D. -Occupational Category (Groups)

We group workers according to the social security fund they
contribute to. The SHIW provides occupation and industry for
each worker; however, these do not allow us to identify the work
er's social security fund in all cases. While it is relatively straight
forward to identify public sector employees and, to a lesser extent,
private sector employees, it is much harder to classify the self
employed. After several attempts we have decided to consider the
self-employed group as the residual group. This means that while
a large fraction of workers in this group are self-employed proper,
a small fraction belong to other funds. The details of this choice
are given below.
E. Public Sector Employees

These correspond to employees whose contributions are col
lected by the Treasury Fund and the Government Employees
Fund (known as INPDAP). Occupations are, for example, civil
servants, employees in the army, employees in the national
health service, etc. Unfortunately, the data do not allow us to
distinguish a group of private sector employees who provide
"house help" and are improperly included in the public sector.
F. Private Sector Employees

(INPS-FPLD-National Institute for Social Security-Private
Sector Employees Fund). This group mainly consists of blue
collar and white-collar employees. Between 1991 and 1993 a
change in the occupational categories coding occurs, which pro
vides a finer disaggregation. However, we could not exploit this
new feature of the survey, and for coherence, we had to stick to
the 1991 coding frame. We exclude from this private sector em
ployees group top-level managers: these pay contributions to a
special fund (known as INPDAI) which we include in the self
employed group.
G. Retired
We use both an earnings criterion and a self-reported char
acteristic. We consider retired all individuals who are not active
and are self-reported retired and do not receive earnings. These
are not necessarily all old-age pensioners, as there could be DI
pensioners or young people drawing a survivor's benefit.
H. Other
These are mainly self-employed individuals. A few other peo
ple receiving earnings from employment fall within this group:
top-level managers and other special funds (e.g., actors). Hence,
this is essentially a residual group. The reason for this choice is
twofold: on the one hand, we need groups that are exhaustive of
the population; on the other hand, it is crucial to have a satisfac
tory sample size. Nothing would prevent us from distinguishing a
smaller group of actual self-employed individuals, plus define a
fifth group of others. However, the information provided in SHIW
would not allow for the identification of some of the activities in
self-employment, particularly in the arts and crafts categories.
Hence by following this approach, we would suffer from two
drawbacks: on the one hand, the self-employed group would be
rather small; and, on the other hand, we would leave out some
important groups of self-employed workers.
/. Educational Groups
These enter both at the first level of the analysis (earnings
projections) and as instruments in some of the specifications we
estimate. The educational groups we consider are roughly com
parable to the following U. S. categories: (1) No education?high
school dropouts, (2) a secondary level education which we could
call "some college," and (3) college graduates and above. The age
earnings profiles for the highly educated group are much steeper:
this has direct implications for the earnings growth parameter
entering the computation of social security wealth.
K. Net Pension Wealth

Household's pension wealth is defined as the present ex
pected value of future social security benefits to which both
spouses are entitled, evaluated at the current age of the head of
the household. This is the sum of old age pension wealth of
husband and wife plus pension wealth to the surviving spouse.

We use both a gross pension wealth definition and a net pension
wealth one: the latter is obtained by subtracting from gross pen
sion wealth the expected value of future social security taxes that
a worker pays from his current age until retirement.28 We adopt
exactly the same definition as in Feldstein [1974]; however, for
the reasons given below, we do not require that the earnings
growth rate and the discount factor take the same value.
In order to estimate pension wealth for workers, we project
forward gross earnings to retirement age and then apply the rules
in place in each year to compute the first yearly benefit. This
requires a number of steps. First, the SHIW contains net earn
ings (these are after income tax and after social security tax),
while for benefit computation we need gross earnings. Hence, in
order to gross up earnings, we impute for each worker in the
survey the income tax and the social security tax paid by the
worker (excluding the tax paid by the employer). This procedure
is rather complex because the income tax schedule is highly
nonlinear and contains many different tax rebates (details can be
obtained from the authors upon request). Once gross earnings in
the survey year are obtained, we need to make assumptions on
earnings growth for each individual. We set up an econometric
specification delivering age-earnings profiles for different gender
occupation-education groups based on time series of cross sec
tions of the SHIW going back to 1978. This shows a marked
difference in the yearly growth rate of real earnings for different
cohorts: younger cohorts start higher and have steeper age-earn
ings profiles.
Hence we use cohort-specific growth rates: overall, these
range between around 1 percent to 3 percent for older cohorts to
about 6-7 percent for younger cohorts of individuals with higher
education. There is very little evidence on age-earnings profiles
for Italian workers, and very few studies account for cohort spe
cific growth rates. Our results are broadly consistent with what
was found by Lucifora and Rappelli [1995] in a study based on
28. The social security tax rate for employees ranges between 7.45 percent of
gross earnings in 1991 to 8.34 percent in 1993. Hence the present value of future
taxes, up to retirement age, is a rather small fraction of gross pension wealth. This
is because, as explained in the text, we gross up net earnings taking into consid
eration only the tax rate of the worker, while we disregard the tax rate of the
employer. Hence we also exclude future social security taxes of the employer,
which amount to approximately 17 percent of gross earnings.
panel data drawn from administrative records of private sector

employees. Their results are comparable to our estimates as they
also distinguish cohorts, gender, and occupational groups.
Finally, we fix the retirement age of each worker at the
expected retirement age directly elicited from respondents in the
survey questionnaire.
This procedure gave us the first yearly benefit for each indi
vidual by implementing the benefit calculation rules prevailing
under each legislation (we also took account of the existence of a
minimum-level-benefit, in those cases where the computed bene
fit was very low, and benefit capping, if applicable).
In order to compute pension wealth, we then applied the
appropriate growth rate to social security benefits?according to
the prevailing rule?up to a maximum age of 100. Italian gov
ernment actuaries have often used a real earnings growth rate of
1.5 percent in their macro-simulations of future social security
expenditures: we take this assumption for the earnings growth
rate entering the pension indexation rule for all retirees (only in
1991). We use a real discount rate of 3 percent, and apply survival
probabilities by age and gender drawn from the Italian Life Tables.
For pensioners the procedure is much simpler, as we only
need to project forward the actual benefit and apply the relevant
survival probability.
Social security taxes are the present (at retirement) value of
a constant fraction of earnings that workers are expected to pay
between the current age and retirement, conditional upon sur
vival. There is a discrete jump in the social security tax levied on
workers between the years 1991 and 1993, which makes the use
of net social security wealth interesting for this paper. Social
security taxes are certainly expected to grow in real terms over
the years, and by keeping the tax rate constant after 1993, we are
probably underestimating the value of social security taxes, par
ticularly for young workers.
Our estimates of the level of household social security wealth
(in 1994 lira) are higher than obtained in previous studies (see,
for example, Jappelli [1995]). However, this is easily justified: we
use gross earnings rather than net earnings; we use cohort
gender-occupation specific growth rates in earnings, rather than
a flat growth rate; and finally, we explicitly account for benefits to
the surviving spouse, which were previously neglected. Further
more, it is important to keep in mind that what matters for our
estimation strategy is the difference between groups and over
TABLE VIII
Mean Values of the Relevant Variables for the Entire Sample and the
Regression Sample (Pension Wealth in Millions of 1994 Italian lira,
Saving Rates in Percentage Points)
Regression
Entire sample sample
Number of observations 32212 18598
Age 53.6 43.5
Year of birth 1938 1948
Number of earners in the household 1.7
Number of children in the household 0.60
Private sector employees 0.24 0
Public sector employees 0.17 0.30
"Other" active head of household 0.17 0.28
Retired 0.41 ?
Net pension wealth 710.5 937.3
Expected future earnings 215.6 368.6
No education or elementary school 0.39 0.30
Secondary school 0.52 0.61
College degree 0.09 0.09
Born after 1957 0.12 0.20
Born 1945-1957 0.28 0.45
Born 1935-1944 0.21 0.27
Born 1903-1934 0.39 0.06
Saving rate 0.16 0.15
Expected retirement age 60.22 60.2
time in pension wealth, rather than the a

wealth. Our construction of pension wealth
tently in both years?even though allowing
about by the reform.
UNr^ERSiTY College London, Institute for Fiscal S

Universit? ca' Foscari di Venezia and Institute for
References
Alessie, Rob, Arie Kapteyn, and F. Klijn, "Mandatory Pensions and Personal
Savings in the Netherlands," De Economist, CXLV (1997), 291-324.
Auerbach, Alan J., and Laurence J. Kotlikoff, "An Examination of Empirical Tests
of Social Security and Saving," in Social Policy Evaluation: An Economic
Perspective, E. Helpman, A. Razin, and E. Sadka, eds. (New York: Academic
Press, 1983).
Blundell, Richard, Alan Duncan, and Costas Meghir, "Estimating Labor Supply
Responses Using Tax Reforms," Econometrica, LXVI (July 1998), 827-862.
Brandolini, Andrea, and Luigi Cannari, "Methodological Appendix: The Bank of
Italy's Survey of Household Income and Wealth," in A. Ando, L. Guiso, and I.
Visco, eds., Saving and the Accumulation of Wealth. Essay on Italian House
hold and Government Saving Behavior (Cambridge, UK: Cambridge Univer
sity Press, 1994).
Brugiavini, Agar, "Empirical Evidence on Wealth Accumulation and the Effects of
Pension Wealth: An Application to Italian Cross-Section Data," Financial
market group, Discussion Paper, No. 20, London School of Economics, 1987.
-, "Social Security and Retirement in Italy," in J. Gruber, and D. A. Wise, eds.,
Social Security and Retirement around the World (Chicago: The University of
Chicago Press, 1999).
Brugiavini, Agar, and Elsa Fornero, "Pension Provision in Italy," in R. Disney and
P. Johnson, eds., Pension Systems and Retirement Incomes across OECD
Countries (London: Edward Elgar, 2001), pp. 197-235.
Brugiavini, Agar, and Mario Padula, "Too Much Saving for Retirement? The Case
of Italy," Research in Economics, LV (2001), 39-60.
Diamond, Peter A., and Jerry A. Hausman, "Individual Retirement and Savings
Decisions," Journal of Public Economics, XXIII (1984), 81-114.
Eissa, Nada, "Taxation and the Labor Supply of Married Women: The Tax Reform
Act of 1986 as a Natural Experiment," NBER Working Paper, No. 5023,1995.
Euwals, Robert, "Do Mandatory Pensions Decrease Household Savings: Evidence
for the Netherlands," De Economist, CXLVIII (2001), 643-670.
Feldstein, Martin, "Social Security, Induced Retirement and Aggregate Capital
Accumulation," Journal of Political Economy, LXXXII (1974), 357-374.
Gale, William G., "The Effect of Pensions on Household Wealth: A R??valuation of
Theory and Evidence," Journal of Political Economy, CVI (1998), 706-723.
Guiso, Luigi, Tullio Jappelli, and Daniele Terlizzese, Wliy Is Italy's Saving Rate
so High? in A. Ando, L. Guiso, and I. Visco, eds., Saving and the Accumulation
of Wealth. Essay on Italian Household and Governement Saving Behavior
(Cambridge, UK: Cambridge University Press, 1994).
Guiso, Luigi, and Tullio Jappelli, "Intergenerational Transfers and Capital Mar
ket Imperfections. Evidence from a Cross-section of Italian Households," in A.
Ando, L. Guiso, and I. Visco, eds., Saving and the Accumulation of Wealth.
Essay on Italian Household and Governement Saving Behavior (Cambridge,
UK: Cambridge University Press, 1994).
Gustman, Alan L., and Thomas L. Steinmeier, "Effects of Pensions on Savings:
Analysis with Data from the Health and Retirement Study," NBER Working
Paper No. 6681, Cambridge, MA, 1998.
Heckman, James J., and Richard Robb, "Alternative Methods for Evaluating the
Impact of Interventions," in J. Heckman and B. Singer, eds., Longitudinal
Analysis of Labour Market Data. Econometric Society Monograph, 10 (Cam
bridge, UK: Cambridge University Press, 1985).
Hubbard, Robert Glenn, "Pension Wealth and Individual Saving," Journal of
Money, Credit and Banking, XVIII (1986), 167-178.
Istituto Poligrafico e Zecca dello Stato,?Gazzetta Uf?iciale della Repubblica Itali
ana?Legge 23 ottobre 1992, n. 421, Delega al Governo per la Razionalizza
zione e la Revisione delle Discipline in Materia di Previdenza (Rome: 1992).
Jappelli, Tullio, "Does Social Security Reduce Wealth Accumulation? Evidence
from Italian Survey Data," Ricerche Economiche, XLIX (1995), 1-32.
King, Mervyn A., and Louis Dicks-Mireaux, "Asset Holdings and the Life Cycle,"
Economic Journal, XCII (1982), 247-267.
Krueger, Alan B. and J?rn-Steffen, Pischke, "The Effect of Social Security on
Labor Supply: A Cohort Analysis of the Notch Generation," Journal of Labor
Economics, X (1992), 412-437.
Lucifora, Claudio, and Francesco Rappelli, "Evoluzione delle Retribuzioni nel
Ciclo di Vita: un'Analisi su Dati Longitudinali," Lavoro e Relazioni Indus
triali, Rivista di Econom?a Applicata, III (1995), 77-116.
Meyer, Bruce, "Lessons from the U. S. Unemployment Insurance Experiment,"
Journal of Economic Literature, XXXIII (1995a), 91-130.
-, "Natural and Quasi-Experiments in Economics," Journal of Business and
Economic Statistics, XIII (1995b), 151-161.
Miniaci, Raffaele, and Guglielmo Weber, "The Italian Recession of 1993: Aggre
gate Implications of Microeconomic Evidence," Review of Economics and
Statistics, LXXXI (1999), 237-249.
Ministero del Tesoro-Ragioneria Generale dello Stato. Ufficio Statistico, "Ten

denze Demografiche e Spesa Pensionistica: alcuni Possibili Scenario Quad
erno Monogr?fico n. 9 dei Conti Pubblici e Congiuntura Econ?mica (Rome:
1996).
Peracchi Franco, and Nicola Rossi, "Nonostante tutto ? una riforma," in Frontiere
della Pol?tica Econ?mica, F. Giavazzi, A. Penati, and G. Tabellini, eds. (Milan:
Il Sole 24 Ore, 1995).
Rossi, Nicola, and Ignazio Visco, "Private Saving and Government Deficit," in A.
Ando, L. Guiso, and I. Visco, eds., Saving and the Accumulation of Wealth.
Essay on Italian Household and Government Saving Behaviour (Cambridge,
UK: Cambridge University Press, 1994).
Rossi, Nicola, and Ignazio Visco, "National Saving and Social Security in Italy,"
Ricerche Economiche, XLLX (1995), 329-356.
Samwick, Andrew, "The Impact of Social Security Reform on Saving." National
Tax Association Proceedings?1997, pp. 178-186.
-, "Is Pension Reform Conducive to Higher Saving?" Review of Economics and
Statistics, LXXXII (2000), 264-272.

Social Security and Households&#39; Saving

Uploaded by

Copyright:

Available Formats

You might also like

Social Security and Households&#39; Saving

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Social Security and Households&#39; Saving

Uploaded by

Copyright:

Available Formats

Social Security and Households' Saving

Author(s): Orazio P. Attanasio and Agar Brugiavini

not liquid and cannot be borrowed against; individuals might be

1. Contributions to the pension system, whether perceived as forced saving or

relationship between bequeathable wealth and pension wealth

Several authors have tried to explain both the remarkable levels

that exploits the differential effect of the reform on different

II. On the Relationship between Pension Wealth and Saving: A

Our approach is different from those used in the papers in

In this section we first sketch a very simple model, which is

H.A. A Simple Model

4. As we argue in footnote 1 above, in principle, labor supply can be affected.

we have to consider at least three periods. Moreover, as we want

(Cl)1" a. D (C2)1" a. ?2 fe)1"' _, ?3 bJ1-"

where d = ?1/7 (1 + r)a l)/y. For a consumer aged

where A? is accumulated private wealth in the ?th period.

where ra is retirement age, assu

6. We are assuming that ?1^ (1 + r)a

II.B. Empirical Specification

7. The model we have considered also neglects uncertainty, which might

(6) SRiJt = Xt? + Haht)FEi}t + 0 (ai>t)PWiit + xt + fc + eM,

an application of a difference-in-difference estimator.9 Identifica

ideally require an estimate of social security wealth as perceived

III. The Amato Reform and its Differential Effects

III.A. Household Savings in Italy

bequest motives.11 While the fact that for Italian households it

crucial parameter much lower than those obtained by Rossi and

12. Samwick [2000] also looks, in a panel of countries, at structural reforms

the limit of 35 years of contribution for eligibility to an

Normal retirement age 60 (men) 65 (all)a 6

implied by the way in which the reform changed the existing

III.C. Some Preliminary Evidence

17. We experimented with different values of the bandwidth parameter, and

[*ry(16,93) - *r0(16,93)] - [sry(14,91) - sr0(14,91)],

III.D. Estimates of Public Pension Wealth and the

right, but rather that we estimate without systematic biases the

It should be added that an abrupt recession episode took

Panel A: Mean net pension wealth

both spouses, in principle it would be possible to define finer and

Panel B: Mean saving over mean income X 100

differences in how the reform affects different cohorts and occ

Panel A: Changes in mean net household pension wealth

Panel B: changes in mean saving over mean income

changes in saving rates

rigorous fashion and quantifies the relationship between savings

IV. The Effects of Pension Wealth on Saving Rates

Having measured pension wealth for each household in the

how average changes in this variable are related to average

21. As groups are defined, among other things, by year-of-birth cohorts,

1995, as for the households observed in that year, we should take

Excluding expected Including expected

Saving rates are defined as (Income - Consumption)/Income. Pension wealt

households could be explained by liquidity c

Excluding expected Including expected

(Age 20-35)* PWit -0.052 -0.043 -0.104 -0.195

time-specific effects on savings. In Table VI we somewhat relax

fore, the parameters of interest are identified through the year

Excluding expected Including expected

Social Security and Households' Saving

Social Security and Households' Saving

Social Security and Households' Saving

[ry(16,93) - r0(16,93)] - [sry(14,91) - sr0(14,91)],