An Empirical Analysis of Earnings and Employment Risk

Author(s): Luigi Guiso, Tullio Jappelli, Luigi Pistaferri

Source: Journal of Business & Economic Statistics, Vol. 20, No. 2 (Apr., 2002), pp. 241-253
Published by: American Statistical Association
Stable URL: http://www.jstor.org/stable/1392061
Accessed: 22/09/2008 09:15

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=astata.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal
of Business & Economic Statistics.

http://www.jstor.org
AnEmpirical Analysis of Earnings and
Employment Risk

LuigiGuiso
Universitdi Sassari and Ente Einaudi,Via Due Macelli73, 00187 Rome, Italy(guiso@tin.it)

Tullio JAPPELLI
CSEF,Departmentof Economics, Universitadi Salerno 84084 Fisciano (Salerno), Italy(tullioj@tin.it)

Luigi PISTAFERRI
Departmentof Economics, StanfordUniversity,Stanford,CA 94305-6072 (pista@stanford.edu)

The mean and highermomentsof the distributionof futureincome are crucialdeterminantsof individual
choices. These momentsare usually estimatedin panel data from past income realizations.In this article
we rely instead on subjective expectations available in the 1995 Survey of Household Income and
Wealth, a large random sample representativeof Italian households. The survey elicits informationon
the distributionof futureearningsand on the probabilityof unemployment.Analysis of this distribution
helps us understandhow individualuncertaintyevolves over the life cycle and if attitudestowardrisk
affect occupationalchoices and income riskiness.

KEY WORDS: Income risk; Subjectiveexpectations;Unemployment.

1. INTRODUCTION and possibly higher moments of the individual distribution

Economistsroutinelyproposemodels in which currentdeci-
sions depend on expectationsof futurevariables.For instance,
theories of intertemporal choice with incomplete markets
posit that people react to expected income. When the strong
assumptionsthat lead to certaintyequivalenceare relaxed,the-
ory also predicts that people respond to higher moments of
the distributionof futureincome (Kimball 1990). The relevant
moments are those of the subjective income distributioncon-
ditional on informationavailableat the time the decisions are
made.
Only under the extreme hypothesis of complete marketsis
the measurementof individual income risk not an issue. But
when idiosyncratic shocks matter,measuring microeconomic
uncertaintybecomes a crucial issue in applied econometrics
and in calibrationof general and partial equilibriummodels.
In a recent survey, Browning, Hansen, and Heckman (1999)
argue that calibratingeconomic models with imperfect insur-
ance "requiresa measure of the magnitudeof microeconomic
uncertainty,and how that uncertaintyevolves over the busi-
ness cycle [...]. This introduces the possibility of additional
sources of heterogeneity because different economic agents
may confront fundamentallydifferentrisks."
These remarkshave implicationsfor many areasof research.
Measuring individual uncertainty is crucial when trying to
determinethe importanceof precautionarysaving. Individual
uncertaintyaffects the width of the inactionband in Ss models
of durable demand and housing investment. Income risk can
lead prudentindividualsto demand a higher risk premiumon
risky assets, and it can affect portfolio choice and the demand
for insuranceagainst insurablerisks. More generally, income
risk can impact labor supply,educationand occupationchoice,
job search, and many other economic decisions.
Two approacheshave emerged in the literatureto extract
moments of the distributionof futureincome from observable
variables. One relies on panel data and infers expectations
241
from past income realizations. To be valid, this method
requires assuming that individuals condition on the same set
of variablesto form expectations,that the individualsand the
econometricianhave the same information set, and that the
econometrician knows the stochastic process that generates
individual expectations. It is an unhappy feature of applied
economics that implausible assumptions and procedures get
accepted for lack of sound alternatives.
A second strand of literaturehas recently proposed rely-
ing on survey questions, not retrospectivedata, to elicit infor-
mation on the conditional distributionof future income. The
main advantageof survey questions over inference based on
realizations is that they do not require the econometricianto
know the variables that individuals consider in forming their
expectations.
Following this line of research,we rely on subjectiveexpec-
tations drawnfrom the 1995 Survey of HouseholdIncome and
Wealth (SHIW), a large representativesample of the Italian
population. As will be seen, to estimate the moments of the
income distributionfrom survey data we must rely on some
assumptionsand imputationsspecific to our dataset.
Our contributionto the literatureis on method and sub-
stance. On the methodological side, we take explicitly into
account that the distributionof future income results from
three distinct elements: the probability of job loss, the dis-
tribution of future wages and the distributionof unemploy-
ment compensation. Depending on the institutional features
of the labor market, each of these elements may be more
or less importantin determiningthe overall income distribu-
tion. For instance, if job search is costless and wages and
prices are fully flexible, future income depends only on wage
? 2002 American Statistical Association
Journal of Business & Economic Statistics
April2002, Vol. 20, No. 2
242
fluctuations,but if wages are sticky or fixed, income variability
depends mainly on fluctuationsin employment status. Previ-
ous studies focus mainly on wages or income from all sources.
The distributionof wages neglects the impact of the probabil-
ity of unemploymentand the distributionof benefits. Expec-
tations about income from all sources make it hard (if not
impossible) to assess the separateimpact of wage and unem-
ployment uncertaintyon overall income uncertainty.Subjec-
tive informationon future employmentprospects available in
the 1998 SHIW allows us to tackle the first issue directly.
Moreover,we can use external informationto impute unem-
ployment benefits as a function of various demographicand
labor status variables,thus accountingfor the second element.
On substance,we provideevidence thatis useful for various
branchesof researchthat, directly or indirectly,need to make
assumptions about various moments of the conditional dis-
tributionof individual incomes. First, we constructempirical
profiles of income uncertaintyand of the probabilityof unem-
ployment over the working career.The estimated age profiles
help us understandhow uncertaintyevolves over the life cycle.
The analysis is importantbecause simulation studies always
neglect age-relatedheterogeneityin income risk. Second, we
compare unemploymentrisk in Italy and in the United States
using comparable survey questions. This allows us to high-
light the role of labor marketflexibility and other institutional
factors. Finally, we relate unemploymentrisk, the coefficient
of variationof futureincome, and an index of the asymmetry
of the income distributionto a set of demographiccharacteris-
tics and to an index of risk aversion.The correlationbetween
income risk and risk aversion allows us to assess the sever-
ity of the self-selection problemthat potentiallyplagues many
empiricalstudies of precautionarysaving and portfolio choice.
The relation between our measures of earnings uncertainty
and workers'characteristicshelps identify variablesassociated
with job security.
The rest of the article is organized as follows. In Section
2 we survey the literaturebased on subjective income expec-
tations and describe the main characteristicsof the survey.
In Section 3 we estimate the overall individual labor income
distributioncombining the probabilityof unemploymentwith
informationon future earnings and imputationof unemploy-
ment benefits. In Section 4 we provide ample descriptionof
the cross-sectionalcharacteristicsof the income distributions.
In Section 5 we present the age profiles of the probability
of unemploymentand of the coefficient of variationof future
income. The cross-country comparison between unemploy-
ment risk in Italy and the United States is taken up in Section
6. In Section 7 we test if individuals that face lower income
risk also are more risk-averse.Section 8 summarizesour main
findings.
2. PREVIOUS EVIDENCEAND SAMPLE DESIGN
In order to derive empirical measuresof subjectiveincome
expectationsand income risk, one must design appropriatesur-
vey questions to characterizeeither the density or the cumu-
lative distribution function of future income. In the litera-
ture both approacheshave been taken. Guiso, Jappelli, and
Terlizzese (1992) is based on survey questions posed in the
Journalof Business & EconomicStatistics,April2002
1989 SHIW eliciting informationabout the density of future
earnings.Recently,Dominitz and Manski(1997a, b), Dominitz
(1998), and Das and Donkers (1999) have followed the alter-
native approachand so does the design of the 1995 SHIW.
The 1995 SHIWhas data on income, consumption,financial
wealth, real estate wealth, and several demographicvariables
for a representativesample of 8135 Italian households (see
Appendix A). A special section of the survey was designed to
characterizethe distributionof future income and the proba-
bility of unemployment.To our knowledge, the only two other
surveys containing informationon employmentprospects are
the Health and Retirement Survey (HRS), conducted at the
University of Michigan since 1992, and the Survey of Eco-
nomic Expectations (SEE), conducted at the University of
Wisconsin-Madison since 1994. The SEE is a national tele-
phone survey of the U.S. population. The survey is limited
in scope and small in size (1300 households interviewed in
1996). The drawbacksof the HRS data are that the surveyhas
no questions about income expectationsand that respondents'
age is deliberately restricted to the 51-61 range (in 1992),
that is, individuals approachingretirementfor whom unem-
ploymentrisk could be negligible or altogetherabsent.Explic-
itly considering unemploymentprobabilitiesat younger ages
is important,because employment risk is one of the major
determinantsof future income prospects.
Our survey questions focus on earnings ratherthan dispos-
able income and on individualsratherthan households. Focus
on earnings avoids mixing labor income and capital income
uncertainty.Focus on individualsavoids relying on one person
to evaluatethe income prospectsof otherhousehold members.
Unlike the SEE, the 1995 SHIW households report the dis-
tributionof after-taxincome, rather than gross income. One
advantage of using after-tax income is that most household
choices ultimately depend on disposable income, not income
before taxes. Furthermore,since in Italy income taxes and
social security contributionsare withheld at source, employ-
ees are betterinformed about their after-taxearnings.
Questions on income expectationswere asked to half of the
overall sample after excluding the currentlyretired and peo-
ple not in the labor force (a total of 4799 individuals, ran-
domly chosen among about 10,000 survey participants).Both
the employed, the unemployed,and the job seekers are asked
to state, on a scale from 0 to 100, their chances of having a
job in the 12 months following the interview. Each individ-
ual assigning a positive probabilityto being employed is then
asked to report the minimum (ym) and the maximum (YM)
incomes he or she expects to earn if employed, and the prob-
ability of earningless than the midpointof the supportof the
distribution,Prob(y < (ym+YM)/2) = Ir. The wordingof these
questions is reportedin Appendix B.
The employmentprobabilityquestion aims at obtainingthis
probability for both the currently employed and the unem-
ployed, taking into account job mobility, that is, that some
respondentsplan to quit or to change job. However, in prac-
tice the interpretationof the questioncould be differentfor the
currentlyemployed and for the unemployed. Moreover,it is
not clear that the respondent,if employed, reportsonly invol-
untaryjob losses ratherthan any change in employmentstatus
(includingjob mobility). Thus, the question could be subject
Guiso, Jappelli,and Pistaferri:An EmpiricalAnalysisof Earningsand EmploymentRisk
to measurement error and misrepresent true unemployment
risk. In Section 4 we will thereforecross-examinethe reliabil-
ity of the unemploymentquestion by comparing the average
subjective unemploymentprobability with actual unemploy-
ment rates drawnfrom labor force statistics;we also compare
unemploymentprobabilities with actual labor market transi-
tion probabilities.
In the next section we describe how we combine the infor-
mation derived from the survey questions to estimate the dis-
tributionof futureincome, both conditionaland unconditional
on working. From the original 4799 observations,we exclude
209 who expected to retire or to drop out from the labor
force within a year and 385 individualswith missing data on
the relevant questions. The nonresponserate is therefore 8%
(385/4590), and the final sample includes 4205 individuals.
3. THEINDIVIDUAL
DISTRIBUTIONS
The distributionof future income depends on the distri-
bution of future earnings if the individual is employed and
on the distributionof unemploymentbenefits if he or she is
unemployed.The two distributionshave to be weighted by the
probabilityof the two states, so that the distributionof future
income is the mixture of two distributions:
Yi (1- pi)
withprobability
=Ibi with probabilitypi,
where yi is earningsif employed, bi is unemploymentbenefits
if unemployed,pi is the probabilityof unemployment,and i
refers to the ith individual in the sample. We will denote by
f (y) the individual distributionof future earnings, or simply
the earnings distribution,g(bi) the individual distributionof
benefits, and h(xi) = (1-pi)f(yi) +pg(bi) the individualdis-
tributionof future income, or simply the income distribution.
As noticed, if the individualis currentlyemployed, pi is the
probabilityof losing the currentjob and not finding one in the
12 months following the interview; if the individual is cur-
rently unemployed, pi is the probabilityof not finding a job.
For the self-employed pi can be interpretedas the probabil-
ity of personal bankruptcy.Unemploymentbenefits bi should
include not only unemploymentcompensation, but also any
other resources that are formally or informally transferredin
case of unemployment.
Dropping individual subscripts, the expected value of [ym,(Ym+ yM)/2] and [(ym + YM)/2, yM]. If IT = .5 the dis-
income of individual i is then
E(x) = (1 - p)E(y) + pE(b), (2) the distributionis triangularover the same two intervals; if
where E(y) and E(b) denote the expected values of y and lar distributionover the [ym, YM] interval. The expressions to
b, respectively. The variance of the mixture income distribu- compute the mean and the variance of the triangulardistri-
tion is
= (1
Var(x) - p) f[y-E(x)]2f(y)dy
+ p f[b- E(x)]2g(b) db
= (1 - p) Var(y) + pVar(b)
+p(1 - p)[E(y) - E(b)]2. (3) lar distribution.We also checked the sensitivity of the results
243
To make Equations(2) and (3) operationalwe need to make
explicit assumptions about the distributionof benefits g(b)
and about the earnings distributionf(y). In fact, the SHIW
provides informationon p, some informationon f(y), and no
informationon g(b).
For benefits we assume that each individual forms point
expectations about b. Based on the rules governing Italian
welfare programs,we thus impute a value b to each individ-
ual in the sample. We use survey data and aggregateinforma-
tion to determine eligibility requirementsand welfare bene-
fits. The lattervary substantiallyacross populationgroups.For
instance, the self-employed and the long-termunemployedare
entitled to very few welfare programs.Since in Italy informa-
tion on eligibility requirementand welfare benefits is coded in
the legislation and widely known to the public, the assumption
of point expectations is reasonable.However, our imputation
does not take into accountother privatetransfers(monetaryor
in kind), which may representan importantincome source in
case of unemployment.The imputationprocedureis detailed
in Appendix C.
It is straightforwardto allow also for uncertainunemploy-
ment benefits. For instance, one could assume that not all
workersfile for the benefits they are, at least in principle,enti-
tled to. Alternatively,one could assume that help from rel-
atives or friends in case of layoff is uncertain. Finally, one
could assume that compensation received by young unem-
ployed depends on the numberof eligible unemployed in the
region of residence. But each of these scenarios is inherently
arbitrary,so we prefer to stick to the case in which benefits
are not a randomvariable.
As far as f(y) is concerned,recall that the survey provides
informationon the supportof the distribution[ym, YM] and on
the probabilitymass to the left of the midpointof the support,
Prob(y < (y, + YM)/2)= iT. Knowing the supportof the dis-
tribution,we can express the expected value and variance of
y as
E(y) = y f(y) dy (4)
Ym
Var(y) = yy2 f(y) dy - ( y f(y) dy . (5)
We consider two assumptionsconcerningf(y). The first is
that y is uniformly distributedover each of the two intervals
tributioncollapses to a single uniform distributiondefined in
the [ym, YM] interval. A second possibility is to assume that
IT = .5 the distributionagain collapses to a single triangu-
butions are reportedin Appendix D. Note that in both cases
E(y) and Var(y) depend only on the three known parameters,
Ym,YM,and T. The triangulardistribution(shown in Fig. 1) is
a more plausible descriptionof the probabilitydistributionof
earnings, because outcomes further away from the midpoint
receive less weight. For this reason in the remainderof the
article we reportstatistics computed accordingto the triangu-
244 Journalof Business & Economic Statistics,April2002

f(y) o Benefits= 80%of earnings A Benefits= 50%of earnings

40-

E
o
8(1-t )0(y -y)/(yM--ym)2 o 30-
..........................

8n(y-ym)/(y m)2 ........................... 0

.o
S20-
0
t-I.

Ym (ym+YM)/2 YM Y
0
S10-
Figure1. The Triangular
Distribution
of Earnings.

0-
on the assumptionof a uniform distribution.All results were
0 20 40 60 80 100
very similar and are not reportedfor brevity. Probabilityof unemployment
If people have point expectationsaboutbenefits,Var(b) = 0
and we can rewriteEquations(2) and (3) as Figure2. TheEffectof the Probabilityof Unemploymenton the Coef-
ficientof Variation
of Income. Thefigurereportsthe relationbetween the
E(x) = (1 - p)E(y) + pb (6) probabilityof unemploymentand the coefficientof variationof income
CV(x). In the examples average earnings E(y) and the variance of
Var(x) = (1 - p)Var(y) +p(l - p)[E(y) -b]2. (7) earnings Var(y) are set at 100 and 16, respectively.In the upper line
unemploymentbenefitsare set at 80%of E(y), in the lowerline at 50%.
The coefficient of variation provides a convenient mea-
sure of income risk that is particularlyuseful for compar-
ison between different individuals, groups or samples. The a maximum at 10% for an individual with p = .5, and then
coefficient of variation of earnings, CV(y) = Sd(y)/E(y), is declines for individuals with higher values of p. The higher
immediately obtainedby the ratio between the square root of curve refers to individualswho are entitled to b = .50 x E(y);
Equation(5) and Equation(4). It is of course affected by dis- as explained in Appendix C, this situationis typical of small
tributionalassumptions.The coefficient of variationof income firm employees. The much larger value of [E(y) - b] raises
is the weight of p in determiningthe riskiness of futureincome.
For an individualwith p = .2, CV(x) is now 22%. The impact
CV(x) Sd(x) of the probabilityof unemploymentis maximumfor an indi-
E(x) vidual with p = 70%, CV(x) = 35%.
Figure 2 shows that CV(x) is very sensitive to the level of
(1 - p)Var(y) + p(1 - p)[E(y) - b]2 ( unemploymentbenefits at all levels of p. The imputationof
= (8) b can be
(1 - p)E(y) + pb questionable,particularlybecause it neglects infor-
mal transfers.Thus in the next section we reportinformation
Equation (8) highlights that the relation between the prob- on both CV(y), which is independentfrom b and p, and on
ability of unemployment and overall income risk, as mea- CV(x).
sured by CV(x), is non linear. The standard deviation of Since we have an estimate of the distributionsat the indi-
income Sd(x) equals the standarddeviation of earnings if vidual level, we can easily check if the distributionsare sym-
p = 0 and zero if p = 1, is concave in p, and is maximized metric.We thus constructand analyze an index of skewness of
when p = .5 x {1 - Var(y)/[E(y) - b]2} = p* < .5. Thus an the distributions,AS(y) = [M(y) - E(y)]/Sd(y) for earnings
increase in p above p* reduces expected income and raises and AS(x) = [M(x)- E(x)]/Sd(x) for income, where M(.) is
Sd(x). The probabilityp affects also the expectationof income the median of the distribution.The formula for the median is
[the denominator of Eq. (8)], so that the relation between reportedin Appendix D. In Section 7 we will consider more
p and CV(x) too is a non linear function of Var(y), E(y), in detail how this index varies with individualcharacteristics.
and b.
This is illustratedin Figure 2, where we set E(y)= 100
4. CROSS-SECTIONALDISTRIBUTIONS
and Var(y) = 16, so that CV(y) = 4%; as will be seen in
Section 4, this value is close to the median coefficient of vari- The foregoing definitionsand assumptionsallow us to com-
ation of earnings in our sample. The figure reports CV(x) of pute the mean and varianceof both futureearningsand future
individuals with the same earnings distributionbut different income for each individual in the sample, and therefore to
unemploymentprobabilities.The lower curve plots CV(x) as obtain a cross-sectional distributionof individual means and
a function of p if b = .8 x E(y), approximatelythe level of variances.These cross-sectionaldistributionsare conveniently
benefits large firm employees are entitled to. For an individual summarizedin Table 1 by the cross-sectional distributionof
with p = .2 (about the sample average of the probabilityin the probabilityof unemploymentand of the individualcoeffi-
our sample), CV(x) is 8%. The coefficient of variationreaches cient of variation.
Guiso, Jappelli,and Pistaferri:An EmpiricalAnalysisof Earningsand EmploymentRisk 245

Table1. Deciles of the Cross-SectionalDistribution
of the Probabilityof Unemploymentand of the
Coefficientof Variation
Deciles p CV(y)
0 .00
I
II 0 .99
III 0 1.67
IV 0 2.24
Median 10 3.14
VI 10 3.87
VII 30 4.91
VIII 50 6.26
IX 80 8.84
Mean 22.13 4.13
NOTE: Thetable reportspercentagevalues of the deciles of the cross-sectionaldistributions
of the probabilityof unemployment,the coefficientof variationof earnings,and the coefficient
of variationof income.Thecoefficientof variationof earningsis definedas CV(y)= Sd(y)/E(y),
whereSd(y) is the standarddeviationand E(y) the mean of earnings.The coefficientof vari-
ationof income is definedas CV(x)= Sd(x)/E(x) where Sd(x) is the standarddeviationand
E(x) is the mean of income;mean and standarddeviationin this case take into accountthe
probability of unemploymentand expected benefits.The numberof observationsis 4205. All
statisticsare computedusing sample weights.
Column (1) of Table 1 displays the cross-sectional distri-
bution of p. For over 40% of the sample p = 0, a signal of
substantialrigidity in the labor market;the incidence of p = 0
among the employed is even higher. On the other hand, only
3% of the sample is certainto be unemployedin the year fol-
lowing the interview (p = 1). A sizable fraction of the sam-
ple reportssubstantialunemploymentuncertainty:for 20% the
probabilityexceeds 50%. As we arguein Section 2, due to the
wording of the questions, for some individualsa high p does
not necessarily reflect worsening employment prospects. For
instance, women who anticipatehaving a child or young men
expecting compulsory military service may correctly report
temporaryexit from the labor force in the year following the
survey, ratherthanjob dismissal or inability to find a job.
Column (2) displays the deciles of the coefficient of vari-
ation of earnings, CV(y) = Sd(y)/E(y). The cross-sectional
distribution of the coefficient of variation CV(y) is right
skewed, as shown by the positive difference between the
cross-sectional mean, equal to 4.13%, and the cross-sectional
median, 3.14%.
The cross-sectional distributionof CV(y) can be compared
with previous evidence for Italy. Guiso et al. (1992) use sub-
jective expectations from the 1989 SHIW. Since no question
on employment prospects was asked in 1989, a proper com-
parison between the 1995 and the 1989 SHIW must focus on
CV(y). In thatsurveyrespondentswere asked a ratherdifferent
set of questions about earnings prospects. They had to assign
probabilityweights, summing to 100, to a set of intervals of
income changes over the 12 months following the interview.
It is not obvious whetherasking questions aboutthe density
function of future income is more effective (in terms of min-
imizing the probabilityof nonresponseand in eliciting mean-
ingful data) than questions about the cumulative distribution
function. We can provide some evidence on this important
issue comparing nonresponses to the questions on expecta-
tions in the 1989 and 1995 SHIW. In 1989 5954 of those
interviewed did not answer the questions on the subjective
income density function (a nonresponserate of 43%). In 1995
the fraction of nonrespondentsto the questions on the cumu-
lative distributionwas only 8%. In contrast,in 1989 a much
higher fractionof respondentsreportedno income risk (34%),
CV(x) while the same fraction in 1995 was only 13%. The much
higher response rate suggests that the 1995 questions concern-
.00
1.24 ing the cumulativedistributionfunction are easier to grasp and
2.14 thus provide more reliable information.Furtherevidence (not
3.71 reportedfor brevity) indicates that the probabilityof nonre-
5.83
10.21 sponse in 1989 is statisticallysignificantlylower for the more
18.00 educated,the residentin the North, and the young. In contrast,
39.64 the same demographicvariablesdo not significantlyaffect the
75.17
24.29 probabilityof nonresponsein 1995.
Regardless of the assumptions on the shape of the dis-
tribution, the cross-sectional average of CV(y) is higher in
1995 than in 1989 (about2%), reflectingdifferencesin sample
design and risk across sample periods. In fact, the 1995 inter-
views were completed between May and October of 1996 (a
recession year), whereas the 1989 interviews were completed
in the spring of 1990, at the end of an upswing. Nonethe-
less, CV(y) in both years is fundamentallycharacterizedby
the small magnitudeof income risk.
Column (3) of Table 1 reports the cross-sectional distri-
bution of the coefficient of variation of income, CV(x) =
Sd(x)/E(x), which combines the variance of earnings with
informationon the probabilityof unemploymentand the impu-
tation of benefits [see Eq. (8)]. For the bottom part of the
distribution, where p is near or exactly zero, there is not
much difference between CV(y) and CV(x). But already at
the median the impact of p is substantial:the cross-sectional
median of CV(x) is 5.83%. In the top two deciles the impact
of p and of imputed benefits is dramatic,because the coef-
ficient of variation exceeds 40%, so that the overall cross-
sectional mean is 24.29%. The high values in the top two
deciles often refer to self-employedor high-incomepeople for
which benefits are low relativeto earnings.Given the substan-
tial asymmetry of the cross-sectional distributionsof CV(y)
and particularlyof CV(x), in the remainingof the article we
use the median as location parameterand rely on regressions
estimatedby least absolutedeviations,which are robustto out-
lying observations.
It is not easy to compare CV(y) or CV(x) with measuresof
income risk obtainedby regressionanalysis. MaCurdy(1982)
estimatesa univariateincome process and reportsthatthe stan-
dard deviation of the growth rate of income is 23.5% in the
PSID. Although this numberis close to the average CV(x) in
our sample, partof the income variabilityin panel data is cer-
tainly due to measurementerror and unobservedheterogene-
ity. Furthermore,the time-series error of the income process
estimatedwith panel data does not necessarilyreflectthe inno-
vation faced by individuals,who might consider a much larger
set of variables than the econometrician and therefore have
superiorinformationabouttheir income prospects.Finally,the
standarderrorsestimatedby univariateincome functions may
reflectinequalityratherthantrueex-anteuncertainty.Moments
of the income distributionestimated with survey questions
may thereforebe more reliable than panel data estimates.
Intertemporalchoice models (for instance, models with pre-
cautionarysaving or portfolio allocations) emphasize the role
of lifetime income uncertainty rather than uncertainty one
246 Journalof Business & EconomicStatistics,April2002

period ahead, which is the focus of the article. One can show
that under a set of reasonableassumptions(finite horizon and
constant conditional variances), the conditional variance of
lifetime income is proportionalto the conditionalvarianceone
period ahead. If the income process is an arithmeticrandom
walk, the proportionalityfactor is a deterministicfunction of
age and of the discount rate. If instead the income process is
the sum of a permanentrandomwalk componentand a transi-
tory white noise component (a popularcharacterizationof the
income process), then lifetime income uncertaintyis still pro-
portionalto uncertaintyone period ahead,providedone knows
how much of the total income variance is due to transitory
or permanentshocks. With serial correlationin the transitory
shock, however,the proportionalityno longer holds.
We now examine how p and the moments of the income
distributionvary across demographicgroups. Averages of the
probabilityof unemploymentare presentedin Table 2 for the
employed, the unemployed, and the total sample. If work-
ers are identical in all characteristicsand unemploymentwere
purely voluntary,one should observe no large differences in
unemploymentprobabilitiesbetween the two groups. But this
is clearly not the case: p is much higher for the currently
unemployed(64 against 15%), suggesting strong state depen-
dence or heterogeneity in employment status. Furthermore,
the large differences within demographicgroups suggest that
layoffs are not generated by random draws but are strongly
related to marketand individual characteristics.Interestingly,
these figures are close to labor markettransitionprobabilities
obtainedfrom panel data of the LaborForce Surveyrunby the
National StatisticalOffice. In 1999 for those currentlyunem-
ployed, the probabilityof remainingunemployedwas 63.6%,
quite close to that reported in Table 3. For those currently
employed the probabilityof becoming unemployed over the
year was 5.2%, substantiallylower than that reportedin the
table. One explanationis that some of the currentlyemployed
interpretthe unemploymentquestion as referringto job mobil-
ity, not unemployment.Alternatively,another explanation is
that the employed are more pessimistic about job prospects
thanjustified on the basis of official statistics.
With the exception of gender, health, and job search sta-
tus, the patternof p by demographicgroup is similar,regard-
less of employment status. The probabilityis higher for the
Table3. Cross-SectionalMediansby Selected
DemographicCharacteristics

Earnings
Table2. Probabilityof Unemploymentby Selected in 1995 Benefits E(x) CV(x)
DemographicCharacteristics
Age
<35 8264 1757 8781 7.95
Unemployed Employed Whole 35-44 11415 6611 11226 4.54
in 1995 in 1995 sample >45 11363 2479 11234 4.52
Gender
Age Male 10330 2063 10744 6.32
<35 61.82 17.61 27.86 Female 8264 2975 8729 5.41
35-44 59.46 13.01 15.99 Education
>45 76.50 13.62 18.81 Compulsory 8678 1375 8626 8.16
Gender HighSchool 10330 5165 10847 4.59
Male 65.50 14.50 21.63 College 12913 7748 13760 3.64
Female 61.10 15.96 22.90 Region
Education North 10330 2292 10589 5.87
Compulsory 67.19 18.31 26.13 Center 9710 2892 10261 4.94
HighSchool 61.67 12.81 19.43 South 7231 2292 7954 6.14
College 49.26 8.84 13.50 Healthstatus
Region Poor 8574 1986 7231 5.83
North 50.10 11.75 14.28 Fair 9814 1910 9477 5.81
Center 60.10 12.73 19.66 Good 9815 2475 10048 5.83
South 69.75 22.57 34.80 Sector
Healthstatus Public 12190 7748 12138 2.15
Poor 62.45 22.66 32.67 Private 9297 1223 9297 8.87
Fair 78.83 15.19 23.26 Occupation
Good 63.43 14.89 21.77 Employee 10847 7438 10778 4.44
Sector Self-employed 8662 68 10230 8.16
Public - 9.29 9.29 Unemployed 0 5165 5940 9.22
Private - 17.34 17.34 Firmsize
Firmsize <20 8729 1253 7748 28.20
< 20 - 21.30 21.30 20-99 9814 1757 9039 10.71
20-99 - 19.94 19.94 >99 12913 7748 12397 4.04
> 99 - 10.69 10.69 Job search
Job search Searching 1291 2066 6284 12.83
Searching 62.57 29.38 48.44 Not searching 10743 2445 10847 4.54
Not searching 68.32 13.44 15.36 Riskaversion
Riskaversion High 11911 2407 11622 4.41
High 66.91 13.62 17.47 Low 11467 1788 11622 6.26
Low 72.80 13.63 17.85
Totalsample 9814 2353 9900 5.83
Sample average 63.67 15.06 22.13
Numberof observations 659 3546 4205 NOTE: The table reports the cross-sectional medians by selected demographic characteris-
tics. E(x) denotes the mean of income; CV(x) is the coefficient of variation of income (percent-
NOTE: The table reports percentage values of the average probability of unemployment in age values), defined as CV(x) = Sd(x)/E(x), where Sd(x) is the standard deviation of income.
the year following the interview by selected demographic groups. All means are computed In columns 1-3 values are expressed in Euro. Cross-sectional medians are computed using
using sample weights. sample weights. The number of observations is 4205.
Guiso, Jappelli,and Pistaferri:An EmpiricalAnalysisof Earningsand EmploymentRisk 247

young, the less well educated,residentsin the South, and those
employed in small private firms. Stratifyingby employment
status, we find that active job seekers reporta lower p (63%)
than non searchers (68%); among the employed, the pattern
is reversed (29 against 13%). One possible interpretationis
that the unemployed who search are those with a reasonable
chance of finding a job, while those who do not are discour-
aged by a high perceivedprobabilityof remainingunemployed
anyway, consistent with the presence of fixed costs of search.
Alternatively,the unemployed who do not search may report
lower unemploymentprobabilitiessimply because job search
improves employment prospects. As for the employed, it is
likely that some of them search when still employed because
they anticipatelosing their job, and this is correctly reflected
in the higher reportedprobabilityof unemployment.
These findings have implications for simulation models of
consumer behavior.For instance, Carroll(1997) assumes that
the probability of unemploymentis constant in each period
and for each individualand simulates a buffer-stockmodel of
consumption.The evidence in Table 2 is strongly at variance
with this assumption.
The cross-sectionalmediansby demographicgroupsof real-
ized 1995 income, b, E(x), and CV(x) are reportedin Table 3.
Since interviews were completed between May and October
1996, they reflect expectations stretching well into 1997,
whereas income realizationsrefer to the calendar year 1995.
Furthermore,1995 earnings do not include benefits, while
E(x) does. The comparison is especially misleading for the
unemployed,which had no earningsin 1995 but are assumed
to receive unemploymentbenefits in the future. Given these
caveats, on average expected and realized incomes are quite
close (9814 and 9900 Euro, respectively).
Several groups expect an income decline, particularlythe
elderly, people self-reportingpoor health status, and employ-
ees of small firms. The pattern of unemployment benefits
across groups reflects the Italian welfare legislation program:
public sector employees and employees of large firmsare more
likely to receive substantial income support in the case of
unemployment.Very few welfare programssupport the self-
employed in case of drops in earnings.
The cross-sectional pattern of the coefficient of variation
confirms that employees of the private sector (particularlyof
small firms) and the self-employed perceive high risk (the
median of the coefficient of variationin the group is 8.16%),
while public sector employees perceive little risk (2.15%).
Active job seekers expect more volatile incomes than those
who are not currentlysearching. Finally, the young and resi-
dents in the South face comparativelymore risk.
Knowledge of the evolution over time of earnings and
employmentrisk mattersin a variety of contexts. It is impor-
tant in simulationstudies of lifetime wealth accumulationthat
feature precautionarysaving (for instance, Carroll 1997). It
helps in understandingthe age patternof the composition of
the household portfolio and of the willingness to hold risky
assets. Insofaras the availabilityof credit dependson the risk-
iness of income prospects, it can help in understandingthe
age profile of households facing liquidity constraints.
We reportseparateage profiles for the probabilityof unem-
ployment p, the coefficient of variation of earnings CV(y),
and the coefficient of variationof income CV(x). To compute
the age profiles, we run kernel regressions for each of these
variables on age using a Gaussianweight function. Since we
want to focus on the evolution of income risk over the work-
ing career,the sample is restrictedto the currentlyemployed
aged 20 to 50. The profiles are estimated for two education
groups, up to compulsory schooling and more than compul-
sory schooling. They representthe effect of age on p, CV(y),
and CV(x), without controlling for other age-relatedindivid-
ual characteristics.Since we use a purecross-section,we make
no attemptat disentanglingage effects from cohort effects.
It is clear that uncertaintyvaries considerablyover the life
cycle. The age profile of p in the top left of Figure 3 declines
for both education groups. One way to interpretthe profile is
that, due to asymmetric informationon workers' ability and
on-the-joblearning,workersperceive that employerspreferto
lay off young workers. That is, when employers choose to
lay off a worker,they pick from a group on which they have
little information,which often is the group with short tenure.
The decline of p with age is particularlystrongfor individuals
with higher education(in this group p declines from 25% for
the young to 5% for the 50 years old), possibly because firm's
knowledge of white-collarworkersis more stronglycorrelated
with tenure and because these jobs offer a wider spectrumof
careerpossibilities.
The age profile of CV(y) in the top right of Figure 2 is con-
cave for both educationgroups, and the shape is quite similar:
an increasing profile in the late 20s, followed by a decline in
the 30s, and a flat profile after 40. The age profile of CV(x)
in the bottom left of Figure 2 is dominatedby the age pattern
of p. Again, it signals that income riskiness tends to decline
with age, an effect that is particularlystrong for individuals
with higher education:in this group CV(x) declines from 40
to 10%.
6. COMPARISONOF UNEMPLOYMENTRISK IN
ITALYAND THE UNITEDSTATES

comparison of overall income uncertaintyis

Intemrnational
5. CROSS-SECTIONALAGE PROFILES considered in Dominitz and Manski (1997b) and Das and
Donkers (1999). They show that perceivedincome uncertainty
Income and employment uncertaintychange over the life is much higher in the United States than in Europe (repre-
cycle. At the beginning of their career, people face a wide sented by Italy and the Netherlands).The most naturalexpla-
range of possible opportunitiesinvolving differentpatternsof nationfor the differencebetween perceivedrisk in Europeand
lifetime earnings,but also differentpatternsof lifetime income the United States is that it reflects tighter labor marketregu-
uncertainty.Lateron, by choice or chance some of the original lations and more generous welfare programsin Europe.
opportunitiesare no longer available and individuals eventu- In this section we complement their evidence by focusing
ally settle in jobs with well-defined characteristics. instead on perceivedunemploymentrisk. If indeed differences
248 Journal of Business & Economic Statistics, April 2002

O Higheducation ALoweducation 0 Higheducation A Loweducation

25- 5-

E
o0>

E 0

0 >
20 o0 40 C 2 30 4
a
40-0 0

20 30 40 50 20 30 40 50
t---
Age Age

40

o0

Age

Figure 3. Age Profiles of the Probability of Unemployment, Earnings and Income. The age profiles are estimated by a kernel regression using
a Gaussian weight function. Each regression is estimated separately for employed workers who completed junior high school (low education)
and
and those who completed
those who completed high school or
high school college (high
or college education).
(high education).

in income uncertaintybetween the United States and Europe
mainly stem from differencesin labor marketregulations,than
they should become manifest when comparingunemployment
probabilities.This comparisoncould also help one understand
better the source of the difference. In Figure 4 we reportthe
cumulative distributionfunction of the probability of unem-
ployment in Italy and in the United States. Manski and Straub
(2000) providedata for the United States. For comparisonwith
the U.S. study, we focus on those currentlyemployed. Apart
from the large difference at p = 0, the two cumulative dis-
tributionsare surprisinglysimilar: in both countries 70% of
individualsperceive p < 10 percent, and 10% face p > 60%.
The main difference in the two distributionsis at low levels
of p: the fraction of those facing no risk of job loss is much
higher in Italy than in the United States (60 against 30%),
a reflection of the differentinstitutionalcharacteristicsof the
labor market,with a tougherjob protectionlegislation in Italy
than in the United States, and of the larger size and stability
of public sector'sjobs in Italy.
The comparison between the two countries could be
affected by the high Italian unemploymentrates, particularly
in the Southern regions. To insulate the comparison from
cyclical and structuraldifferencesin the labor market,we drop
individuals living in provinces where the unemploymentrate
exceeds 7%, leaving us with a sample in which the overall
unemploymentrate is close to the 1993 U.S. nationalaverage
(we retainroughly 1000 observations).At low levels of p the
large difference between the two countries is quite substan-
tial; at higher levels of p the shape of the distributionfunction
tends to become similar.
Since the difference between Italy and the United States is
concentratedin the bottompartof the p distribution,in Table4
we focus on the demographiccharacteristicsof the subsam-
ples of those reportingp = 0 and p > 0. Regardlessof group,
the fraction of individualsreportingp = 0 is much higher in
Italy.The qualitativepatternacross groupsis similarin the two
countries:the group with p = 0 is largerfor those with longer
tenure (as proxied by age), those with college education, and
the self-employed. But note the very high premium for job
stability of Italian college graduates:over 72% report p = 0
against 57% in the group with low education. The compar-
ison confirms the different nature of the two labor markets:
Guiso, Jappelli,and Pistaferri:An EmpiricalAnalysisof Earningsand EmploymentRisk 249

o Italy AItaly,unemployment
rate<=7%
countrieswith more rigid labor marketinstitutionsand wages
o USA
(as Italy). According to Bertola and Ichino, the probability
1- of unemploymentdepends on employment status: the unem-
.9- ployed are more likely to find a job and the employed are
more likely to be laid off in countrieswith more flexible labor
.8-
markets.Using simulations,Flinn (2001) comparesthe impli-
.7- cations of lifetime welfare inequality of labor marketinstitu-
6.6 tions in Italy and the United States. He finds that the Ameri-
can flexible system is characterizedby higher cross-sectional
dispersion in earnings (and thereforehigher income risk) but
lower inequality in lifetime welfare, compared to the Italian
.-
inflexible system.
.14- Overall, the internationalcomparisonsupportssome of the
hypothesis advancedby Bertola and Ichino. Providedthat the
.1 - differences in p and overall income uncertaintydo not reflect
0- sample design and other measurementproblems,thereis com-
pelling evidence that in countries with greater labor market
0 10 20 30 40 50 60 70 80 90 100
flexibility on-the-jobwage uncertaintyis higher than in coun-
Probabilityof unemployment tries with more rigid labor markets. At the same time, our
analysis in this section shows that for the employedp is higher
Figure4. The Probabilityof Unemploymentin Italyand the United
States. Thefigurereportsthe subjectiveprobabilityof unemploymentin in environments with more flexible labor markets; unfortu-
Italyand the UnitedStates. Dataforthe UnitedStates are drawnfromthe nately, for the unemployed we cannot make the comparison
1994-1998 Surveyof Economic Expectations;see Manskiand Straub because we lack data for the United States.
(2000). The higherline for Italyexcludes individualslivingin provinces
where the average unemploymentrate (drawnfrom aggregate labor
force statistics)exceeds 7%in 1995. 7. INCOMERISKAND RISKAVERSION
One objectionto the use of income risk indicatorsin empir-
tighterregulationof Italian labor marketsreduce substantially ical tests of household behavior is that coefficient estimates
the employee's perceived risk of job dismissal relative to the can be biased by self-selection. This can happen if unob-
United States. served preferences correlate with income risk. For instance,
There is a growing literaturethat compares the effect of risk-aversepeople may choose low risk occupationsand avoid
labor market institutions on the amount of risk that workers risky jobs, such as self-employmentor employmentby small
face, earnings inequality, and the consequent welfare effects. firms. Considerthen an applied economist who wants to esti-
Bertola and Ichino (1995) argue that labor marketinstitutions mate the importanceof precautionarysaving and runs a regres-
are the main determinantsof the degree of risk perceived at sion of saving on income risk omitting risk aversion. The
the individual level. They make a strong case that workers coefficient of income risk will be biased downward by the
in countries where labor markets are highly flexible (as the endogeneity of the risk indicator:even an insignificantcoef-
United States) perceive higher income risk than workers in ficient can be consistent with precautionarysaving because
income risk is negatively correlatedwith an omitted variable
(risk aversion). See Dynan (1993) for an empirical example
Table4. Fractionof Respondents withZeroand PositiveProbabilityof and a discussion.
Unemployment: A ComparisonBetween Italyand the UnitedStates A question in the 1995 SHIW provides an opportunityto
measure individual attitudes toward risk and to gauge the
Italy UnitedStates
severity of the self-selection problem. Questions on risk atti-
p=0 p>0 p=0 p>0 tudes are also asked in the HRS and have been used by Barsky,
Age Juster,Kimball, and Shapiro (1997). Each household head in
<35 51 49 27 73 the SHIW was asked: "Suppose you have the opportunityto
35-44 64 36 32 68 invest in a risky asset. There is an equal chance that you will
>45 65 35 42 58
Education gain 5000 Euros or lose everything.At most, how much would
No college 57 43 30 70 you be willing to pay to purchase this asset?" The expected
College 72 28 32 68 value of such a lottery is 5000 Euro, and most households
Occupation
Employee 58 42 29 71 (95%) reporta price thatis strictlyless thanthe expected value
Self-employed 60 40 48 52 of the lottery;thatis, they are risk-averse.We then define three
Totalsample 59 41 31 69 indicators correspondingto low (willingness to pay strictly
less than 1250 Euro), medium (between 1250 and 2500), and
NOTE: The table reportsthe fractionof respondents(inpercentagevalues)withzero (p = 0)
and positive(p > 0) probabilityof unemploymentin Italyand in the UnitedStates in the year high (2500 or more) risk aversion.
followingthe interviewby selected demographicgroups.Dataforthe UnitedStates are drawn In Table 5 we regress p, CV(y), CV(x), and AS(x) on a set
fromthe Surveyof EconomicExpectations.We thankCharlesManskyand John Straubfor
makingavailableto us the distributionsforthe UnitedStates. Individualsolderthan 50 years of demographic characteristicsand the risk-aversionindica-
of age and the unemployedare excluded. tors. As the risk-aversionquestion is asked only to household
250 Journalof Business & EconomicStatistics,April2002

Table5. TheDeterminantsof the Probabilityof Unemployment,the Coefficientof Variation,

and the Asymmetryof ExpectedIncome

p CV(y) CV(x) AS(x)

Age -.0605 -.0176 -.0390 -.1402

(.0687) (.0097) (.0209) (.0982)
Male -2.8611 .5914 1.4471 -1.4646
(1.8140) (.2543) (.5507) (2.5929)
Education(in years) -1.2043 -.0682 -.3199 -.3008
(.1488) (.0210) (.0450) (.2127)
Residentof the North -.6061 -.2858 .2927 -4.7788
(1.6174) (.2275) (.4903) (2.3119)
Residentof the South 9.0034 .0205 .5486 .5706
(1.6712) (.2352) (.5067) (2.3889)
Poor health 1.0687 -.2120 -2.3892 2.5685
(4.3194) (.6022) (1.3007) (6.1742)
Good health -.9839 .0057 -1.1623 2.3533
(1.7631) (.2484) (.5365) (2.5201)
High riskaversion .0414 -.2712 -1.2599 8.3309
(1.7276) (.2422) (.5247) (2.4694)
Moderateriskaversion -.9990 .1025 .3619 6.9153
(2.2029) (.3095) (.6697) (3.1489)
Constant 29.3199 3.7897 9.8846 1.0199
(4.6181) (.6502) (1.4045) (6.6011)
R2 .0808 .0103 .0080 .0168

NOTE: The table reportsregressionsof the probability of unemploymentp, the coefficientof variationof earningsCV(y),the coeffi-
cient of variationof income CV(x),and the indexof asymmetryAS(x),all multipliedby 100. CV(y)is definedas CV(y)= Sd(y)/E(y),
where Sd(y) is the standarddeviationand E(y) is the mean of earnings.CV(x)is definedas CV(x)= Sd(x)/E(x),whereSd(x) is the
standarddeviationand E(x) the mean of income. AS(x) is definedas [E(x)- M(x)]/Sd(x),where M(x) is the medianof the income
The regressionsfor CV(y),CV(x),and AS(x)are estimatedby least absolutedeviations.The regressionforp is estimated
distribution.
by ordinaryleast squares. The sample excludes the unemployedand observationswith missing values for risk aversion.Omitted
characteristicsare: residentin the center,fairhealth,and low riskaversion.The numberof observationsis 1556. Standarderrorsare
reportedin parentheses.

heads and other observations are lost because of missing or
zero currentearnings, the sample is restrictedto 1556 indi-
viduals. If individuals self-select into jobs according to their
attitudes toward risk, their risk aversion should help predict
the probabilityof unemployment,the coefficient of variation,
and the index of asymmetry,after controlling for individual
characteristics.In general we find the expectation variables
difficult to predict on the basis of observable characteristics,
as witnessed by the low R2 reported in Table 5. This may
reflect large error in measuring income riskiness or the fact
that income riskiness has a large individualcomponentthat is
not associated with standarddemographicvariables.However,
some interestingpatternsemerge from the data.
In column (1) we focus on the probability of unemploy-
ment. As in the descriptive analysis of Table 2, we find that
p has a strong inverse correlation with education. An addi-
tional year of educationreduces the probabilityof unemploy-
ment by 1.2 percentage points (in this sample the mean is
14%). Residence in SouthernItaly, where the unemployment
rate is about 3 times the national average, has the expected
effect: the thought experimentof moving an individual from
the North to the South would raise the probabilityof facing
unemploymentby roughly 10 percentage points. The coeffi-
cient of age is negative but poorly measured,and so are other
demographiccharacteristics.The risk-aversionindicators do
not explain unemploymentrisk.
In column (2) we relate CV(y) to the same set of vari-
ables [recall that CV(y) refers to the earnings distribution
if employed]. Age and education reduce earnings variability,
while being male increases it. The coefficients of the other
demographicvariables and of the dummies for risk aversion
are not statisticallydifferentfrom zero.
The third column of Table 5 refers to CV(x). The gen-
eral patternof results is somewhat similar to that of CV(y),
with one notable exception: the coefficient of the dummy for
high risk aversion in this regressionis negative and precisely
estimated (a t-statistic of -2.4). The coefficient suggests that
more risk-averseindividualschoose occupationswhere CV(x)
is 1.3% lower than for the less risk-averse (the median of
CV(x) in this subsampleis 4.22, the mean 22). This implies
that the self-selection effect cannot be easily dismissed in
empirical studies.
It is worth noting that risk aversion can be measuredcor-
rectly only if income is non random.For instance,if the lottery
is negatively correlatedwith income, the risk-aversemay be
willing to pay the lottery more than its fair price of 5 million
lire. In our case, the hypothetical question about risk aver-
sion implies that the lottery is independentof income. Even
if the two risks are independent, however, the Arrow-Pratt
index of risk aversion depends on income risk. If preferences
are proper in the sense of Pratt and Zeckhauser(1987), the
riskier is income, the less willing is the individualto accept
an additionalindependentrisk, the lower the price that he is
willing to pay for the lottery,and the higher the index of risk
aversion. This "backgroundrisk effect" can therefore atten-
uate the "self-selection effect." Empirically we find that risk
aversion is negatively correlatedwith CV(x), suggesting that
the self-selection effect dominates.Clearly,if the background
risk effect is also present, than the true self-selection is even
strongerthan it appearsfrom our regressions.
Guiso, Jappelli,and Pistaferri:An EmpiricalAnalysisof Earningsand EmploymentRisk
Another importantcharacteristicof the individual income
distribution is asymmetry; for instance, Caballero (1990)
shows that the asymmetry of the distributionprompts pre-
cautionary saving behavior. We measure the asymmetry of
the income distribution with the index AS(x) = [M(x) -
E(x)]/Sd(x), which ranges from -1 to 1. If AS(x) > 0 the
distributionis skewed to the right, implying that very unfavor-
able events receive more weight than favorable events. Intu-
itively, individualswho dislike negative income shocks should
select themselves in occupation with positive AS(x). In col-
umn (4) of Table 5 this intuition is confirmed. We find that
risk aversionis associated with a distributionthat is skewed to
the right. This implies thatrisk-averseindividualsselect them-
selves into occupations where large negative income events
occur with relatively low probability.This channel adds to the
selection effect due to the income uncertaintydescribedabove.
8. CONCLUSIONS
In this articlewe propose a new set of indicatorsof expected
income and subjective income risk using the 1995 Bank of
Italy Survey of Household Income and Wealth. Their main
advantageis that they are derived from simple yet powerful
questions. With suitable assumptions, these questions allow
estimation of moments of the distributionof future income,
taking into account the probability of unemployment and
the distributionof unemploymentcompensation.We can thus
examine the entire conditional distributionof income, rather
than focusing on just one aspect, like most of the empirical
literature.We point out that variationsin the perceived proba-
bility of unemploymentexplain a large part of the differences
in income prospects. This suggests that one should account
separatelyfor employmentand income risk in tests of house-
holds' behavior such as consumptionand portfolio choice.
The second and thirdmomentof the income distributionand
the perceived probabilityof unemploymentare then relatedto
a set of demographicand preference characteristics.We find
that demographiccharacteristicssuch as age, education, and
geographicallocation affect both unemploymentrisk and the
varianceof the subjectivedistributionof future earnings.
So far there is very limited evidence on the evolution of
individual income risk over the life cycle and on the effect
of risk aversion on income risk (the self-selection problem).
Despite some theoreticalwork suggesting that asymmetrymay
be importantfor precautionarysaving (Caballero 1990), there
has been no attempt at measuring higher moments of the
income distribution.We provideevidence on these issues. First
of all, we find strong evidence that the profiles of income and
unemploymentrisk decline with age differentlyfor individuals
in differenteducation groups, a finding that is consistent with
models of the labor marketin which asymmetricinformation
and on-the-job-learningplay an importantrole. Second, con-
trolling for demographicvariables, we find that risk aversion
is a predictor of income risk. This correlation suggests that
the more risk-averse select themselves into occupations with
low-income risk. This findingis also consistent with the claim
that the concern for security is a major factor in the traditio-
nally long queue of Italiansseeking civil service jobs. Finally,
we find that the risk-aversetend to self-select in jobs with low
probabilityof low-income realizations.
251
ACKNOWLEDGMENTS
We thank two anonymousreferees, the Editor,Franco Per-
acchi, and seminarparticipantsin Rome and Tilburgfor com-
ments on a preliminarydraft of the article. This article is part
of a research project on "StructuralAnalysis of Household
Savings and WealthPositions over the Life Cycle."The Train-
ing and Mobility of Researchers Network Program (TMR)
of the European Commission DGXII, the Italian National
Research Council (CNR), the Ministryof Universityand Sci-
entific Research (MURST), and the "TaubeFaculty Research
Fund" at the Stanford Institute for Economics and Policy
Research have provided financial support.
APPENDIX
A: THE1995 SURVEYOF HOUSEHOLD
INCOMEANDWEALTH
The 1995 Surveyof HouseholdIncome and Wealthhas data
on income, consumption,financial wealth, real estate wealth,
and several demographicvariablesfor a representativesample
of 8135 Italian households (including 14,699 income recipi-
ents). Interviews were conducted between May and October
of 1996. Balance-sheet items are end-1995 values. Income
and flow variables refer to 1995. The survey also covers job
search activity, hours of work, and labor force participation.
Brandolini and Cannari(1994) describe the main features of
the SHIW, its sample design, interviewing procedure, and
response rates. Details about the 1995 sample can be found in
Banca d'Italia (1997). The datasetcan be obtainedby writing
to: Banca d'Italia, Research Department,Via Nazionale 91,
00196 Rome, Italy.
APPENDIXB: THEQUESTIONSON
INCOMEEXPECTATIONS
The 1995 SHIW contains two special sections, respectively
on subjectiveincome expectationsand past labor marketexpe-
rience. To reduce overall interview time, half of the sample is
asked questions from the first section and the other half ques-
tions from the second section. Allocation of households to the
two sections is randomand based on the year of birth of the
head (odd or even).
Four questions on income expectations were thus asked to
half of the overall sample after excluding the currentlyretired
and people not in the labor force (a total of 4799 individu-
als). Both the employed, the unemployed,and the job seekers
are asked to state, on a scale from 0 to 100, their chances of
having a job in the 12 months following the interview. Each
individual assigning a positive probabilityto being employed
is then asked to report the minimum (ym) and the maximum
(yM) incomes he or she expects to earn if employed, and the
probabilityof earningless than the midpointof the supportof
the distribution,Prob(y < (ym+ YM)/2) = 7r. The exact word-
ing of the survey questions is providedbelow. All respondents
are first asked:
(a) Do you expect to voluntarilyretire or stop working in
the next 12 months?
If the answer is "Yes"the interviewergoes on to the next
survey section. If the answeris "No" each respondentis asked
252 Journalof Business & EconomicStatistics,April2002

questions (b) through(e) below:
(b) What are the chances that in the next 12 months you
will keep your job or find one (or start a new activity)? In
other words, if you were to assign a score between 0 and 100
to the chance of keeping your job or of finding one (or of
startinga new business), what score would you assign? ("0" if
you are certainnot to work, "100" if you are certainto work).
The following table is shown to the respondent:
* for those working in firms with over 50 employees and
earning a gross monthly salary above 1291 Euro, unemploy-
ment benefits are set at 775 Euro a month and are received
for 12 months following the layoff;
* for those working in firms with over 50 employees and
earning a gross monthly salary below 1291 Euro, benefits
are set either at 646 Euro monthly or 80% of gross salary,
whicheveris the less (durationis again 12 months);
* for those working in small firms (under 50 employees),

0
I....I....I....I....I....I....I....I....I....
S....
10 20 30 40 50 60 70 80 90 100
benefits are set at 30% of gross monthly income and are
received for 6 months.

I am sure that I am sure that Finally, we set unemploymentbenefits equal to minimum

I will not I will earnings (ym) when the former exceeds the latter.
have a job have a job
APPENDIXD: THETRIANGULAR
DISTRIBUTION
(c) Suppose you will keep your job or that in the next Assume that the earningsdistributionis triangularover the
12 months you will find one. What is the minimum annual
two intervals [ym, (Ym + YM)/2] and ((ym + YM)/2, YM]. The
income, net of taxes and contributions,thatyou expect to earn
from this job? probabilitymass to the left of the midpoint(ym+YM)/2 is con-
strainedto be equal to 7r as in Figure 2. The density function
(d) Again suppose you will keep yourjob or thatin the next of the distributionis
12 months you will find one. What is the maximum annual
income, net of taxes and contributions,that you expect to earn "T)(Y-m) ) <y (Ym+YM)
from this job? (y)=8 (YM- ym)2 2
(e) What are the chances that you will earn less than X +8(1- -
)(ym Y) 1 (ym+ YM) }Y
(where X is computedby the intervieweras [(c) + (d)]/2)? In (YM - Ym)2 -2
other words, if you were to assign a score between 0 and 100
to the chance of earning less than X, what score would you and the mean and varianceof earningsare
assign? ("0" if you are certain to earn more than X, "100" if "T + 1 - 7T
E(y) = -(2ym YM)+ 3 (Ym+ 2yM)
you are certainto earn less than X). "3
The following table is shown to the respondent:
Var(y) = 24
I....I I I I I
0 10 ...20 ...30 ...40 ...50 60
I I 100I
I ...
70 80 90....
m +j10YYM+3y2)
7T(11y2
1- 7
+ + 11y) )- [E(y)]2.
24 (3y + 10YmYM
I am sure that I am sure
I will earn I will earn E(x) and Var(x) are then computed using Equations(6) and
more than X less than X
(7) in the text. The median of the distributionof earnings
M(y) is given by

M(y) = + .51
APPENDIXC: IMPUTATION
OF ym 2jiY7>
UNEMPLOYMENT BENEFITS
Under the Italian welfare programs in place in 1995-96,
unemployment benefits depend on labor market status and
individual characteristics.We separate the sample into four
main groups: first-job seekers, long-term unemployed, cur-
rently employed, and self-employed. In principle, first-job while the median of the income distributionM(x) is
seekers, long-termunemployed,and the self-employed are not b if p .5
entitled to benefits. However,they may be eligible for special (YM - y1Tm)
welfare programs offering part-timejobs. For individuals in YM (Y-- ) if 7r< .5 and p<.5
these groups we impute, by region, an average benefit equal - p)7
2 V2(1
to the ratio between 1995 public expenditureon these special
welfare programsand the numberof first-jobseekers.Data are
drawnfrom Casotti and Gheido (1997). Y+Yif r = .5 and p < .5
Only the currentlyemployed receive an explicit compensa-
tion in case of temporarylayoff. Unemploymentcompensation if p=0.
M(y)
depends on gross earnings at the time of layoff and on firm
size. Benefit durationvaries by firm size. Following current
legislation, we use the following values: [Received September1999. Revised February2001.]
Guiso, Jappelli,and Pistaferri:An EmpiricalAnalysisof Earningsand EmploymentRisk 253

REFERENCES Dominitz, J. (1998), "Earnings Expectations, Revisions, and Realizations,

"Reviewof Economics and Statistics, 374-388.
Banca d'Italia (1997), "I Bilanci delle Famiglie Italiane Nell'Anno 1995," Dominitz, J., and Manski, C. F. (1997a), "Using ExpectationsData to Study
Supplementoal Bollettino Statistico, No. 14, March. SubjectiveIncome Expectations,"Journal of the AmericanStatisticalAsso-
Barski, R. B., Juster,T., Kimball, M. S., and Shapiro,M. D. (1997), "Prefer- ciation, 92, 855-867.
ence Parametersand BehavioralHeterogeneity:An ExperimentalApproach
- (1997b), "Perceptions of Economic Insecurity: Evidence From
in the Health and Retirement Survey,"Quarterly Journal of Economics,
the Survey of Economic Expectations,"Public Opinion Quarterly, 61,
112, 537-580. 261-287.
Bertola, G., and Ichino, A. (1995), "Wage Inequality and Unemployment: K. (1993), "HowPrudentare Consumers?,"Journal of Political Econ-
United States vs. Europe,"NBER Macroeconomics Annual 1995, Cam- Dynan,
omy, 101, 1104-1113.
bridge, MA: MIT Press.
Brandolini,A., and Cannari,L. (1994), "MethodologicalAppendix,"in Saving Flinn, C. J. (in press), "LaborMarketStructureand Inequality:A Comparison
and the Accumulationof Wealth,eds. A. Ando, L. Guiso, and I. Visco, of Italy and the US," Review of Economic Studies.
Cambridge,UK: CambridgeUniversity Press. Guiso, L., Jappelli, T., and Terlizzese, D. (1992), "Earnings Uncer-
Browning, M., Hansen, L., and Heckman,J. (1999), "Microdataand General tainty and PrecautionarySaving," Journal of Monetary Economics, 30,
EquilibriumModels," in Handbook of Macroeconomics,eds. J. B. Taylor 307-337.
and M. Woodford,Amsterdam:North-Holland. Kimball, M. S. (1990), "PrecautionarySaving in the Small and in the Large,"
Caballero,R. (1990), "ConsumptionPuzzles and PrecautionarySaving,"Jour- Econometrica,58, 53-73.
nal of MonetaryEconomics, 25, 113-136. MaCurdy,T. E. (1982), "The Use of Time Series Processes to Model the
Carroll, C. D. (1997), "Buffer-StockSaving and the Life Cycle/Permanent Error Structureof Earnings in a LongitudinalData Analysis," Journal of
Income Hypothesis,"QuarterlyJournal of Economics, 112, 1-56. Econometrics, 18, 82-114.
Casotti, A., and Gheido, M. R. (1997), "LavoriSocialmenteUtili," in Diritto Manski,C. F., and Straub,J. D. (2000), "WorkerPerceptionsof Job Insecurity
e Pratica del Lavoro (Vol. 28), Rome: ISPOA. in the Mid-1990s: Evidence From the Survey of Economic Expectations,"
Das, M., and Donkers, B. (1999), "How Certainare Dutch Households About Journal of HumanResources, 35, 447-480.
FutureIncome? An EmpiricalAnalysis,"Reviewof Income and Wealth,45, Pratt, J. W., and Zeckhauser,R.
(1987), "ProperRisk Aversion,"Economet-
325-338. rica, 55, 143-154.