Economics of Education Review 21 (2002) 635–640

www.elsevier.com/locate/econedurev

Absolute risk aversion and the returns to education

Giorgio Brunello *
Department of Economics, University of Padua, via del Santo 33, 35100 Padova, Italy

Received 25 July 2000; accepted 14 June 2001

Abstract

Individual absolute risk aversion is measured in a sample of 1583 male house-hold heads, using the data drawn from
the 1995 wave of the Survey on the Income and Wealth of Italian households. This measure, conditional on household
financial wealth, is used as an instrument for attained education in a standard log earnings equation. In line with most
of the literature, I find that the gap between the IV and the OLS estimates of the returns to education is large.
 2002 Elsevier Science Ltd. All rights reserved.

JEL classification: J24; J31

Keywords: Returns to education; Risk aversion

1. Introduction
It is well known that the estimation of the returns to
education is difficult because of the presence of measure-
ment errors and because unobserved ability can affect
both educational choice and the returns to education. One
of the strategies used to deal with this problems consists
of selecting instrumental variables, that are correlated
with schooling but not with earnings (conditional on
schooling). The typical instruments used in the literature
are school reforms, family background variables and
smoking. An alternative is to use data on twins. See Card
(1999) for a review of the existing evidence. In this note
I add to the current list an additional candidate, the absol-
ute degree of risk aversion. I start by showing in a simple
static model that risk aversion affects in a natural way
educational choice by influencing the marginal utility of
schooling. Perhaps one reason why this variable has not
been used so far is that it is difficult to measure risk
aversion in survey data (see Barsky, Juster, Kimball, &
Shapiro (1997) and Hartog, Ferrer i Carbonell, & Jonker
(2000) for an attempt). I use the 1995 wave of the Survey
on the Income and Wealth of Italian households and pre-
vious work on these data by Guiso and Paiella (2000) to
measure individual absolute risk aversion in a sample of
1583 married Italian male household heads. This variable
is then used as an instrument, for education in a standard
Mincerian earnings function. In line with most of the
current literature, I find that the gap between IV and OLS
estimates is substantial.
2. Schooling choice
Following Card (1999) I assume that an individual
chooses S, the years of schooling, by maximizing the
following objective function1
U(y)⫺f(S) (1)
where y is (hourly) earnings, U is a concave function of
y and f is a convex function of S. Hourly earnings are
related to S by the following function

* Corresponding author. Tel.: +39-049-8274210; fax: +39-

1
049-8274221. This assumption ignores the importance of parental choice
E-mail address: brunello@decon.unipd.it (G. Brunello). in the education of children.

0272-7757/02/$ - see front matter  2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 2 7 2 - 7 7 5 7 ( 0 1 ) 0 0 0 6 2 - 0
636 G. Brunello / Economics of Education Review 21 (2002) 635–640
y ⫽ g(S) ⫽ elS
In this setup, additional schooling increases earnings
and utility U at the price of higher investment costs f.
The first order condition associated to the maximization
of Eq. (1) is
U⬘g⬘(S) ⫽ f⬘(S)
where the prime if for the first order derivative.2
Using a first order Taylor approximation of U⬘ around
y=0, Eq. (3) can be re-written as follows
U⬘(0)[1⫺ARAg(S)]g⬘(S) ⫽ f⬘(S)
U⬙(y)
where ARA ⫽ ⫺ is the Arrow Pratt coefficient of
U⬘(y)
absolute risk aversion (see Laffont (1990)).
I can now establish the following
Claim 1 The selected years of schooling S decrease
when absolute risk aversion ARA increases.
Proof. Differentiation of Eq. (4) with respect to ARA
and S yields
再 冎
f⬙(S)
g⬙(S)[1⫺ARAg(S)]⫺ARAg⬘(S)2⫺ ∂S
U⬘(0)
⫽ g(S)g⬘∂ARA
The expression within curly brackets on the left hand
side is negative because of the second order conditions
for a maximum. The right hand side is positive.
Let the marginal cost of an additional year of school-
ing, f⬘(S), be equal to r1 ⫹ r2S. Next, assume that the
utility function U(y) belong to the CARA (constant
absolute risk aversion) class
1 these instruments include school reforms and features of
U(y) ⫽ ⫺ exp(⫺sy)
s
In this case the coefficient ARA is equal to s,
U⬘(0) ⫽ 1 and optimal schooling S* is given by3
S∗ ⫽ S(l,r1,r2,s)
Individual differences in educational attainment can be
explained in this simple model both by differences in
marginal returns l and marginal costs r and by differ-
ences in the absolute degree of risk aversion.
The expected variation of earnings should also matter.
In a slight complication of the model, I can add to Eq.
(2) the stochastic term n and assume that log earnings
are normally distributed with mean lS and variance sν2.
2
The sufficient condition for an interior maximum is
U⬙g⬘(S)2 ⫹ U⬘g⬙(S)⫺f⬙(S) ⬍ 0 where the two primes are for
the second derivative.
3 4
Card uses the CRRA (constant relative risk aversion) util-
ity function U(y) ⫽ ln y.
(2) If college wages varied more than, say, high school
wages, there is an additional reason why risk aversion
can influence educational attainment, because more risk
averse individuals will require higher expected earnings
to invest in college education.
(3)
3. The empirical model
Consider the standard regression model
(4) ln y ⫽ X⬘d ⫹ aS ⫹ e (7)
S ⫽ X⬘b ⫹ Z⬘g ⫹ h (8)
where X is a vector of controls that affect the marginal
benefits of education, e and h are error terms, Eq. (7) is
the Mincerian earnings function and Eq. (8) is the attain-
ment function, that depends on X and on variables that
affect marginal costs and capture individual prefer-
ences (Z).
Ordinary least squares (OLS) estimates of Eq. (7)
yield a consistent estimate of a only when e and h are
uncorrelated. Unobserved ability and measurement errors
are two well known factors that affect both schooling S
and earnings y conditional on schooling, thereby
inducing correlation between the error terms. This prob-
lem can be addressed if one can identify variables that
affect schooling but not (conditional) earnings. These
variables can be used as instruments (IV) to generate
consistent estimates of the returns to education.
Card (1999) presents a detailed review of previous
studies based on instrumental variables and discusses the
validity of the instruments used in each study. Briefly,
(5) the school system, family background and the use of
samples of twins. Another instrument recently used but
not discussed by Card is smoking. The argument here is
that smoking habits are likely to be highly correlated
with the discount rate, but do not influence earnings
(6) directly.4 Therefore, they can be used as a valid instru-
ment for schooling S.
The simple model presented in the previous section
suggests that a measure of individual absolute risk aver-
sion ARA is another potential candidate. In the model,
the variable ARA affects the schooling decision because
it affects the marginal utility of income (and
consumption), but does not affect the marginal returns
to schooling l.
See Fersterer and Winter-Ebmer (2000) for a recent dis-
cussion.
 G. Brunello / Economics of Education Review 21 (2002) 635–640
4. The measure of risk aversion
Following Laffont (1990), the coefficient of absolute
risk aversion at a given level of wealth W is twice the
risk premium per unit of variance for small risk. The risk
premium is the maximum amount that an agent is willing
to pay to have the sure return rather than the expected
return from a lottery ticket. According to this definition,
risk aversion is not easy to measure and this perhaps
explains why it has never been used as an instrument
for attained education. A survey that contains detailed
information on individual attitudes towards risk is the
Italian Survey on Household Income and Wealth
(SHIW), conducted every two years by the Bank of Italy.
The survey is very useful for my purposes because it
includes information on earnings, educational attain-
ment, household wealth and attitudes towards risk for a
nationally representative sample of households.
In the survey, each household head is offered an hypo-
thetical lottery and is asked to report the maximum price
that he would be willing to pay to participate.5 The exact
question is
“We would like to ask you a hypothetical question
that we would like you to answer as if the situation
was a real one. You are offered the opportunity of
acquiring a security permitting you, with the same
probability 1/2, to either gain 10 million lire or to
gain nothing. What is the most that you are prepared
to pay for this security?”
Ten million lire corresponds to just over Euros 5,000.
Guiso and Paiella (2000) explain that the interviews were
conducted personally by professional interviewers. In
order to help the respondent understand the question,
they were supposed to show an illustrative card and to
provide explanations. The respondent could answer in
one of three ways: a) declare the maximum amount he
is willing to pay to participate, denoted here by M,
known as the compensating certainty equivalent; b) don’t
know; c) unwilling to answer.6
Using the information provided by the answers to this
question I can measure the Arrow–Pratt index of absol-
ute risk aversion for each household head. Let W denote
the non-random household endowment and let the ran-
dom prize of the lottery be 10 million lire and 0 with
equal probability. The maximum entry price is given by:
5
This section draws extensively from Guiso and Paiella
(2000).
6
Guiso and Paiella (2000) find that the sample selection
effects induced by missing answers are small and unlikely to
constitute a problem.
637
1 1
U(W) ⫽ U(W ⫹ 10⫺M) ⫹ U(W⫺M) (9)
2 2
Taking a second-order Taylor expansion of the right-
hand side of Eq. (9) around W yields
ARA ⫽ 4(5⫺M) / [102 ⫹ 2M2⫺20M] (10)
that uniquely defines the Arrow–Pratt measure of absol-
ute risk aversion in terms of the parameters of the lottery
in the survey.
Three problems remain to be discussed before using
ARA as a valid instrument of educational attainment in
the Mincerian earnings function. First, the measure of
absolute risk aversion in Eq. (10) can vary with individ-
ual wealth. Laffont (1990) argues that
“%it is difficult to obtain sufficient information about
an agent’s preferences in order to know whether his
absolute risk aversion increases or decreases%[with
wealth]. However, %since we must assume that
absolute risk aversion decreases with wealth to obtain
results that accord with both intuition and obser-
vations of rational behavior%we can infer that agents
must satisfy this assumption in general%” (p. 24).
Clearly, if absolute risk aversion varies with house-
hold wealth, and wealth is correlated with hourly (net)
earnings, the variable ARA fails to meet the fundamental
requirement for an instrumental variable and cannot be
used as a valid instrument for schooling S. The avail-
ability of detailed information on household real and
financial wealth in the SHIW dataset allow me to regress
individual ARA on these measures of wealth and to use
the residuals of this regression as instruments for school-
ing. Define this generated variable as RISK. By construc-
tion, RISK is orthogonal to household wealth and reflect
both individual differences in characteristics (age, edu-
cation and region of birth) and innate differences in tast-
es.
A second and more difficult problem is that absolute
risk aversion can affect the log earnings of individuals
with the same educational attainment by influencing their
occupational choice. In this case, the generated variable
RISK is not a valid instrument. I evaluate in detail this
possibility in the next section.
A third problem is that educational choice depends on
absolute risk aversion at the time of the choice, not on
current risk aversion. Therefore, my measure of risk
aversion is meaningful only if the time invariant compo-
nent of risk is important. Empirical evidence in support
of the importance of innate preferences is provided by
Guiso and Paiella (2000), who find that the main predic-
tor of absolute risk aversion in the SHIW sample is
region of birth.
Finally, there is no particular reason to expect that risk
aversion, conditional on household wealth, be correlated
638 G. Brunello / Economics of Education Review 21 (2002) 635–640

Table 1 ute risk aversion is negatively correlated to financial

Means and standard deviations of the main variables wealth FW and positively correlated to household size
NC, given wealth. On the other hand, house ownership
Mean Std Dvt is positively but not significantly correlated to risk aver-
sion.
log y 2.658 0.34 Notice that the selected measures of household wealth
S 10.502 3.63 absorb only 1.8% of the total variation of absolute risk
A 42.277 6.84 aversion. I use the residuals and the estimated constant
ARA 0.153 0.09
term from the regression in Table 2 to construct the vari-
FWa 30.414 61.03
NC 3.612 1.11 able RISK. The sample average and the sample standard
H 0.647 – deviation of this variable are 0.128 and 0.094 respect-
ively.
a
FW is in million lire. As mentioned above, RISK could affect log earnings,
conditional on schooling, by influencing occupational
with unmeasured ability. The maintained hypothesis choice. I deal with this problem by using three different
used in the model and in the empirical exercise is that definitions of job/occupation. The first definition disting-
the causal relation runs from absolute risk aversion to uishes between jobs in the private and in the public sec-
educational attainment, not vice versa. tor. The second definition allocates individuals in four
occupational categories: blue collar, white collar, school
teacher and managerial employee. The third definition
5. Empirical results combines the previous two and adds information on sec-
tor of activity (industry, services and the public sector)
I estimate Eq. (7) using the sample of married male to obtain 10 occupational categories.8 I use the first
household heads aged between 28 and 55 years with at definition to run a probit regression where the dependent
least primary education, who were employed full-time variable is the probability of joining the public sector
and for the full year in 1995. As a preliminary step, I and the explanatory variables are years of schooling S,
regress ARA on two measures of household wealth, fin- the controls in the vector X (age, age squared, marital
ancial wealth FW, that includes all financial assets held status, regional and urban residence dummies9) and
by the household in 1995, and the dummy H, equal to RISK. Since the p-value of the estimated coefficient asso-
1 if the household head owns the house she lives in.7 I ciated to RISK is 0.43, my filtered measure of risk aver-
also add a control for household size, the number of sion does not significantly affect the choice of public
household members NC. versus private sector jobs. Next, I fit two multinomial
The summary statistics of the key variables used in logits for the more detailed job definitions by using the
this section are in Table 1. The results of the regression same regressors as above, and test whether RISK signifi-
are shown in Table 2. As expected, the measure of absol- cantly affects choice. In both cases, the c2 statistic rejects
significance at the 5 percent level of confidence.10
Table 2 As a final check, I consider the most detailed classi-
OLS regression of ARA on measures of household wealth. fication of occupations and select jobs that are filled by
Dependent variable: ARAa individuals with the same educational attainment. For
each selected job, I show in Table 3 the fitted average
Coefficient P-value log hourly wage (Fy), obtained by adding the residuals
to the constant term in a regression of log y on X, and
FW ⫺0.150 0.002 the average value of RISK. To save space, I choose two
NC 7.255 0.001 jobs for each educational attainment, and require these
H 4.797 0.339 jobs to have similar values of average RISK.
N obs 1583 Consider first individuals with only primary education
R2 0.018 –

a 8
Note: Robust standard errors. All the coefficients are multi-
plied by 1000. The regression includes a constant term.
7
I exclude household net income, because an important part
of it is the household head’s personal net income. The inclusion
of this variable would remove part of the correlation between
risk aversion and log earnings. At the extreme, the instrument
would lose its power completely.
These categories are: blue collars, white collars and mana-
gerial employees in the three sectors and school teachers in the
public sector.
9
I use age rather than potential experience because the for-
mer variable can be treated as exogenous in the earnings
regression. See Harmon and Walker (1995). There are 18
regional dummies and 3 urban residence dummies, depending
on the size of the town of residence.
10
The p-values are 0.14 and 0.28 respectively.
 G. Brunello / Economics of Education Review 21 (2002) 635–640 639

Table 3 wise comparisons for intermediate levels of education

Average values of RISK and FW for selected jobsa yield similar results: conditional on education, the
between-jobs variation in the average hourly wage is not
Job S FW RISK accompanied by a similar variation in the average value
of risk aversion, thereby confirming that RISK does not
BC services 5 1.03 0.15 affect log earnings, conditional on schooling S and con-
WC public 5 1.21 0.15
trols X.
BC industry 8 1.16 0.12
WC public 8 1.36 0.12
Overall, I take this evidence as supportive of my selec-
WC industry 13 1.39 0.11 tion of RISK as a valid instrument for schooling in the
ME industry 13 1.54 0.11 earnings regression. An instrument needs also to be cor-
WC services 17 1.50 0.11 related with S, the endogenous variable. Table 4 presents
ME services 17 1.92 0.11 the results of the reduced form schooling equation (with
and without age squared among the controls). As pre-
a
Note: WC: white collar: BC: blue collar: ME: managerial dicted by theory, I find that, conditional on age, years
employee. Services: services in the private sector. of schooling are significantly higher for individuals with
lower RISK.
My estimates of the returns to education are presented
Table 4
OLS estimate of the reduced form schooling equation. Depen-
in Table 5. I find that the estimated marginal return to
dent variable: Sa schooling based on OLS is 0.047. This return rises to
0.078, a 65% increase, when I use the IV procedure with
Coefficient P-value Coefficient P-value RISK as the instrument for schooling. The gap between
(1) (2) the two estimates is substantial. The IV estimate, how-
ever, is less precise than the OLS estimate and I cannot
A 0.162 0.266 ⫺0.052 0.000 reject the hypothesis that the two estimated coefficients
A2 ⫺0.03 0.143 are equal. Similar results are common in the literature
RISK ⫺2.703 0.000 ⫺2.755 0.003 that uses school reforms or family background as instru-
N obs 1583 1583 ments for educational attainment (see Harmon & Walker
R2 0.048 0.046
(1995) for a discussion).
a
The large estimated gap between IV and OLS is diffi-
Note: robust standard errors. The regression includes also a
cult to explain with the presence of measurement errors,
constant, marital status, regional and urban residence dummies.
that according to Card (1999) can account for at most a
10% gap. An alternative explanation, also suggested by
(S=5). The white collar jobs in the public sector filled Card, is that variations in RISK affect to a higher extent
by these individuals pay a significantly higher average the subgroups of the population with lower education,
log hourly wage than the blue collar jobs in services, who have higher marginal returns to schooling. Guiso
but attract individuals with the same average RISK. Next and Paiella show that consumers born in the indus-
consider individuals with a college degree. In this case, trialized North of the country are less risk averse than
managerial jobs in services pay a higher average hourly consumers born in the underdeveloped South, who are
wage than white collar jobs in the same sector, but are typically less educated than average. An exogenous
filled by individuals with the same average RISK. Pair- reduction in risk aversion, induced perhaps by a more

Table 5
OLS estimate of Eq. (7). Dependent variable: ln ya

OLS IV
Coefficient P-value Coefficient P-value

S 0.047 (0.002) 0.000 0.078 (0.029) 0.008

A 0.045 (0.012) 0.000 0.040 (0.014) 0.004
A2 ⫺0.0004 (0.0001) 0.004 ⫺0.0003 (0.0001) 0.064
N obs 1583 1583
R2 0.31 0.21

a
Note: robust standard errors within parentheses. The regression includes also a constant, marital status, regional and urban
residence dummies.
640 G. Brunello / Economics of Education Review 21 (2002) 635–640
effective government control of criminal activities, is
likely to affect more Southern households, who live in
crime intensive regions, are less educated and have
higher marginal returns to education, and to produce as
a consequence a higher IV estimate of the average
returns to schooling.
6. Summary
In this note I have added to the current list of instru-
ments for endogenous schooling in an earnings
regression a new entry, the absolute degree of risk aver-
sion. I have used the 1995 wave of the Survey on the
Income and Wealth of Italian households to measure this
variable in a sample of 1583 married Italian male house-
hold heads. In line with an important part of the relevant
literature, I have found that the IV estimate of the mar-
ginal return to schooling is much higher than the OLS
estimate.
Acknowledgements
I am grateful to two anonymous referees, to Daniele
Checchi, Luigi Guiso, Tullio Jappelli, Claudio Lucifora
and Guglielmo Weber for advice. The usual disclaimer
applies.
References
Barsky, R., Juster, T., Kimball, M., & Shapiro, M. (1997). Pref-
erence parameters and behavioural heterogeneity: an experi-
mental approach in the health and retirement study. The
Quarterly Journal of Economics, 112(2), 537–580.
Card, D. (1999). The causal relationship of education on earn-
ings. In O. Ashenfelter, & D. Card (Eds.), Handbook of
labor economics, vol. 3 (pp. 1801–1863). Amsterdam:
North-Holland.
Fersterer, J., & Winter Ebmer, R. (2000). Smoking, discount
rates and returns to education. IZA Discussion Paper n.
126.
Guiso, L., & Paiella, M. (2000). Risk aversion and financial
market imperfections, mimeo, Ente Einaudi, Rome.
Harmon, C., & Walker, I. (1995). Estimates of the economic
returns to schooling for the United Kingdom. American
Economic Review, 85, 1278–1286.
Hartog, J., Ferrer i Carbonell, A., & Jonker, N. (2000). On a
simple survey measure of individual risk aversion. Univer-
sity of Amsterdam.
Laffont, J. J. (1990). The economics of uncertainty and infor-
mation. Cambridge, MA: MIT Press.