9

Risk Aversion s5
9.1 Introduction
Expected utility provides a framework for the analysis of agents attitudes
toward risk. In this chapter we present a formal denition of risk aversion
and introduce measures of the intensity of risk aversion such as the Arrow
Pratt measures and risk compensation. The main result of this chapter, the
Pratt Theorem, establishes the equivalence of these dierent measures of
risk aversion.
Agents preferences over risky consumption plans are assumed to have an
expected utility representation with continuous von NeumannMorgenstern
utility functions. The consumption plans in the domain of an expected
utility function may be dened either narrowly or broadly. The axioms of
expected utility imply that any consumption plan can be viewed as a random
variable on the set S of states equipped with an agents subjective probability
measure. Thus, if the objects of choice are specied as the consumption plans
that emerge from the axioms of expected utility, these are appropriately
dened narrowly as random variables that can take S values with given
probabilities. However, the analysis of this chapter applies equally well if
consumption plans are broadly interpreted as arbitrary random variables
(that is, as random variables with an arbitrary number of realizations and
arbitrary probabilities). The choice between these interpretations is a matter
of taste.
Except in Section 9.11, it is assumed that date-0 consumption does not
enter the utility functions, and throughout it is assumed that there are at
least two states at date 1, S 2.
92
9.2 Risk Aversion and Risk Neutrality 93
9.2 Risk Aversion and Risk Neutrality
An agents attitude toward risk is characterized by his or her preference
between a risky consumption plan and the deterministic consumption plan
equal to the expectation of the risky plan.
An agent with von NeumannMorgenstern utility function v : R R is
risk averse if he or she prefers the expectation of any consumption plan to
the consumption plan itself; that is,
E[v(c)] v(E(c)) (9.1)
for every consumption plan c.
An agent is risk neutral if
E[v(c)] = v(E(c)) (9.2)
for every consumption plan c.
An agent is strictly risk averse if
E[v(c)] < v(E(c)) (9.3)
for every nondeterministic consumption plan c.
Our term risk aversion means weak risk aversion because only weak
preference is required in inequality (9.1). Note that in this usage risk neu-
trality is a special case of risk aversion.
An agent may be neither risk averse nor risk neutral nor strictly risk
averse and may prefer some nondeterministic consumption plans to the ex-
pectations of the plans. Also, an agent may be risk averse but neither risk
neutral nor strictly risk averse; he or she may strictly prefer the expectation
of some nondeterministic consumption plans to the plans themselves but be
indierent for others.
9.3 Risk Aversion and Concavity
Risk aversion, risk neutrality, and strict risk aversion can be characterized by,
respectively, concavity, linearity, and strict concavity of the von Neumann
Morgenstern utility function:
Theorem 9.3.1
(i) An agent is risk averse i his or her von NeumannMorgenstern utility func-
tion v is concave.
(ii) An agent is risk neutral i his or her von NeumannMorgenstern utility func-
tion v is linear.
94 Risk Aversion s5
Figure 9.1 Construction of a consumption plan c such that v(E(c))<E[v(c)] if v is
not concave.
(iii) An agent is strictly risk averse i his or her von NeumannMorgenstern utility
function v is strictly concave.
Proof: (i) If v is concave, then inequality (9.1) holds (it is Jensens inequal-
ity), and the agent is risk averse. The proof of the converse is straightforward
if consumption plans in denition (9.1) are arbitrary random variables. S:
NEW. I AM BEING MORE EXPLICIT ABOUT THIS. This is so because
satisfaction of Eq. (9.1) for all random variables is the denition of concav-
ity.
We present a proof that applies when consumption plans are restricted
to take S values with given probabilities. S: CHANGES FOLLOW, AS IN
ERRATA Suppose that the agent is risk averse but v is not concave. Then
there exist y
1
, y
2
and

satisfying 0 <

< 1 such that

v(

y
2
+ (1

)y
1
) <

v(y
2
) + (1

)v(y
1
) (9.4)
(Figure 9.1). Consider the set of satisfying 0

and
v(y
2
+ (1 )y
1
) = v(y
2
) + (1 )v(y
1
). (9.5)
This set is nonempty ( = 0 is an element) and closed because v is con-
tinuous. Therefore there exists a maximum, denoted by . Similarly, there
exists a minimum of the set of satisfying

1 and Eq. (9.5). Let

denote that minimum. We have <

<

and
v(y
2
+ (1 )y
1
) < v(y
2
) + (1 )v(y
1
) (9.6)
for every < <

.
9.4 Risk Aversion and Expected Utility 95
Let y = y
2
+(1)y
1
and y =

y
2
+(1

)y
1
. It follows from inequality
(9.6) that
v( y + (1 )y) < v( y) + (1 )v(y) (9.7)
for every 0 < < 1.
Consider consumption plan c that takes value y in some (but not all)
states and value y in the remaining states. Note that the deterministic
consumption plan E(c) lies in the interval (y, y). Using inequality (9.7), we
obtain
v[E(c)] < E[v(c)], (9.8)
which contradicts the assumption of risk aversion.
(ii) If v is of the linear form v(y) = ay + b, then Eq. (9.2) holds, and the
agent is risk neutral. The proof of the converse is very similar to the proof
in part (i). The only dierence is that the assumption that v is nonlinear
implies that either inequality (9.4) holds or the opposite strict inequality
holds. Both cases lead to a contradiction of risk neutrality.
(iii) If v is strictly concave, then inequality (9.3) holds (it is Jensens
strict inequality), and the agent is strictly risk averse. To show the converse,
suppose that the agent is strictly risk averse but v is not strictly concave. If
v is linear on some interval [y
1
, y
2
] with y
1
< y
2
, then it follows from part (ii)
that v[E(c)] = E[v(c)] for any consumption plan that takes values in that
interval. This contradicts strict risk aversion. Otherwise, if v is not linear
on any nondegenerate interval in its domain, then the strict inequality (9.4)
must hold. The proof in part (i) leads to a contradiction with strict risk
aversion in this case.
S: NEW SECTION FOLLOWS
9.4 Risk Aversion and Expected Utility
In the preceding chapter we presented derivations of an expected utility
representation of preferences from assumptions on preferences. Of these as-
sumptions, the independence axiom is the most important. Here we present
a much simpler proof of the existence of a concave function v such that pref-
erences are representable as E(v(c)) from the independence axiom and the
assumption that the agent is risk averse. Note that we are proving less here
than in the preceding chapter because the denition of risk aversion pre-
sumes the existence of probabilities, S: NEW whereas the probabilities are
derived under the axiomatic development in the preceding chapter. Despite
96 Risk Aversion s5
this limitation, presentation of this result is justied because the derivation
is very easy.
Suppose that S 3. We have
Theorem 9.4.1 There exists a concave function v such that an agents pref-
erences are represented by E(v(c)) i the agent is risk averse and his pref-
erences obey the independence axiom.
Proof: Necessity is obvious. To prove suciency, observe that S 3 plus
the independence axiom implies the existence of functions v
s
such that pref-
erences are represented by

s
v
s
(c
s
). Now consider the problem
max
c

s
v
s
(c
s
) (9.9)
subject to E(c) = x, for some x. Risk aversion implies that (x, ..., x) is a
solution to Eq. (9.9). The rst-order condition for a maximum is
v

s
(x) =

s

1
v

1
(x). (9.10)
Integrating, there results
v
s
(x) =

s

1
v
1
(x) + C, (9.11)
where C is the constant of integration. This implies that

s
v
s
(c
s
) =

s

s
v
1
(c
s
), up to an ane transformation.
9.5 ArrowPratt Measures of Absolute Risk Aversion
Risk aversion aects agents portfolio choices and equilibriumsecurity prices.
It is useful to have a measure of the intensity of risk aversion. In light of
Theorem 9.3.1, the candidate that comes to mind is the second derivative
v

of the von NeumannMorgenstern utility function. However, the sec-

ond derivative is not invariant to ane transformations of v. As noted in
Chapter 8, a strictly increasing ane transformation of the von Neumann
Morgenstern utility function does not change preferences. Therefore, such a
transformation should not change the measure of risk aversion. The Arrow
Pratt measure of absolute risk aversion is dened by
A(y)
v

(y)
v

(y)
(9.12)
for a scalar variable y such that v

(y) = 0. It is invariant to strictly increasing

ane transformations of the utility function v.
9.6 Risk Compensation 97
Figure 9.2 Here z takes on values of plus or minus 1 with equal porbability. The
expected utility of y +z equals the utility of y , and thus is the compensation
for risk z.
S: NOTE CHANGE If the reciprocal of the ArrowPratt measure of ab-
solute risk aversion is nonzero, its reciprocal is
T(y)
1
A(y)
(9.13)
can be used as a measure of risk tolerance.
9.6 Risk Compensation
Another measure of risk aversion that is closely related to the ArrowPratt
measure of absolute risk aversion is risk compensation. We dene risk com-
pensation as the amount of deterministic consumption one would have to
charge an agent in exchange for relieving him or her of a risk (Figure 9.2,
where the risk has payo plus or minus 1 with equal probability). In non-
nance applications of the theory of choice under uncertainty, this variable
is almost always referred to as the risk premium. Here and in other nance
applications, however, the term risk premium refers to the expected return
on a security less the risk-free return.
The risk compensation for the additional consumption plan (risk) z at
deterministic initial consumption y is the value (y, z) that satises
E[v(y + z)] = v(y (y, z)), (9.14)
and thus the deterministic consumption y(y, z) is the certainty equivalent
of risky consumption y + z.
Note that an agent is risk averse i risk compensation (y, z) is positive
(that is, strictly positive or zero) for every y and every risk z with E(z) = 0.
98 Risk Aversion s5
An agent is risk neutral i risk compensation is zero for all risks z with
E(z) = 0.
For small risk z, risk compensation (y, z) equals approximately half the
product of the variance
2
z
of z and the ArrowPratt measure of absolute
risk aversion at y.
Theorem 9.6.1 For small z with E(z) = 0,
(y, z)

=
A(y)
2
z
2
. (9.15)
Proof: The quadratic approximation of v(y + z) is
v(y + z)

= v(y) + v

(y)z + v

(y)
z
2
2
. (9.16)
Taking expectations, we obtain
E[v(y + z)]

= v(y) + v

(y)

2
z
2
. (9.17)
Similarly, a linear expansion of the right-hand side of Eq. (9.14) yields
v(y (y, z))

= v(y) v

(y)(y, z). (9.18)

If the left-hand sides of Eqs. (9.17) and (9.18) are set equal and the denition
of the measure of absolute risk aversion A is used, Eq. (9.12) results.
The forms of approximation used in Eqs. (9.16) and (9.18) reveal the
meaning of small in the statement of Theorem 9.6.1. For random variable
z, small means that the variance is of rst-order signicance. Approxi-
mations (9.16) and (9.18) take into account only the rst-order signicant
terms.
9.7 The Pratt Theorem
The two measures of risk aversion the ArrowPratt measure and risk
compensation can be used to compare the risk aversion of two agents. An
important theorem says that comparisons using the ArrowPratt measure
and risk compensation always give the same result. Further, one agent
is more risk averse than another if the von NeumannMorgenstern utility
function of the rst is a concave transformation of that of the second.
Let v
1
and v
2
be two von NeumannMorgenstern utility functions on
R, and let
i
and A
i
denote the risk compensation and the ArrowPratt
measure of absolute risk aversion, respectively, of v
i
for i = 1, 2.
We have
9.7 The Pratt Theorem 99
Theorem 9.7.1 Suppose that utility functions v
1
and v
2
are twice dieren-
tiable with continuous second derivatives and strictly increasing. Then, the
following conditions are equivalent:
(i) A
1
(y) A
2
(y) for every y.
(ii)
1
(y, z)
2
(y, z) for every y and every random variable z.
(iii) v
1
is a concave transformation of v
2
; that is, v
1
= f v
2
for f concave and
strictly increasing.
Proof: We rst show that (i) implies (iii). Because v
2
is strictly increasing,
the inverse function v
1
2
exists, and the function f of (iii) is dened by
f(t) = v
1
(v
1
2
(t)).
We have to show now that f is strictly increasing and concave. The
derivative of f is
f

(t) =
v

1
_
v
1
2
(t)
_
v

2
_
v
1
2
(t)
_ (9.19)
and is strictly positive because v

i
> 0 for i=1, 2. Calculation of the second
derivative of f yields
f

(t) =
v

1
(y) (v

2
(y)v

1
(y))/v

2
(y)
[v

2
(y)]
2
, (9.20)
where y = v
1
2
(t). This can be rewritten as
f

(t) = (A
2
(y) A
1
(y))
v

1
(y)
[v

2
(y)]
2
. (9.21)
Thus f

0, and hence f is concave.

Next we show that (iii) implies (ii). By the denition of risk compensation
we have
E[v
1
(y + z)] = v
1
(y
1
(y, z)). (9.22)
Because v
1
= f v
2
and f is concave, application of Jensens inequality
yields
E[v
1
(y + z)] = E{f[v
2
(y + z)]} f{E[v
2
(y + z)]}. (9.23)
The right-hand side of inequality (9.23) equals f[v
2
(y
2
(y, z))]. Combining
inequality (9.23) with Eq. (9.22) yields
v
1
(y
1
(y, z)) v
1
(y
2
(y, z)). (9.24)
Because v
1
is strictly increasing, inequality (9.24) implies
1
(y, z)
2
(y, z).
100 Risk Aversion s5
Finally, we show that (ii) implies (i). Suppose that
A
1
(y

) < A
2
(y

) (9.25)
for some y

. Because A
1
and A
2
are continuous, there is an interval around
y

such that A
1
(y) < A
2
(y) for every y in this interval. Using the arguments
of the proofs above with interchanged roles of v
1
and v
2
, it can be shown
that
1
(y, z) <
2
(y, z) whenever y + z takes values in that interval. This
contradicts (ii).
We emphasize again that the set of random variables z in Theorem 9.7.1
(condition (ii)) can be either the set of all random variables on the set of
states S with given probabilities or the set of all arbitrary random variables.
Note also that no restriction on consumption has been imposed in Theorem
9.7.1. Therefore, the theorem is valid as stated only for utility functions
dened on the entire real line. However, the same equivalence holds for
utility functions dened only for positive (strictly positive) consumption
when risk z in (ii) is such that y + z is positive (strictly positive).
There is also a strict version of Theorem 9.7.1. The equivalence of con-
ditions (i), (ii), and (iii) remains valid if the inequalities in (i) and (ii) are
strict, and the transformation f in (iii) is strictly concave as well as strictly
increasing.
Further, there is an equality version of Theorem 9.7.1: conditions (i), (ii),
and (iii) remain equivalent with equalities in (i) and (ii) and strictly increas-
ing ane transformation f in (iii). This version is a simple corollary to 9.7.1.
It implies that if two utility functions have equal ArrowPratt measures of
risk aversion, then each is a strictly increasing ane transformation of the
other. For instance, the only constant absolute risk aversion utility function
is (up to a strictly increasing ane transformation) the negative exponen-
tial function. Because a strictly increasing ane transformation of a utility
function describes the same expected utility preferences, the ArrowPratt
measure completely characterizes preferences.
9.8 Decreasing, Constant, and Increasing Risk Aversion
If absolute risk aversion A(y) of an agent is decreasing in y, then he or she
has decreasing absolute risk aversion. If A(y) is constant (increasing) in y,
the agent has constant (increasing) absolute risk aversion.
The Pratt theorem implies that an equivalent expression of decreasing
(constant, increasing) absolute risk aversion is that risk compensation (y, z)
is decreasing (constant, increasing) in y for every z.
9.9 Relative Risk Aversion 101
Corollary 9.8.1 For a strictly increasing and twice-dierentiable (with con-
tinuous second derivative) utility function v,
(i) (y, z) is increasing in y for every z i A(y) is increasing in y.
(ii) (y, z) is constant in y for every z i A(y) is constant in y.
(iii) (y, z) is decreasing in y for every z i A(y) is decreasing in y.
Proof: Let us dene utility function v
1
by v
1
(y) v(y + y) for some
y 0. The ArrowPratt measure of absolute risk aversion and the risk
compensation of v
1
are A
1
(y) = A(y + y) and
1
(y, z) = (y + y, z).
Applying Pratts theorem 9.7.1 to v
1
and v yields that A(y + y) A(y)
i (y + y, z) (y, z). Because y is arbitrary, (i) follows.
The proofs of (ii) and (iii) are similar.
9.9 Relative Risk Aversion
Sometimes it is of interest to measure risk aversion relative to the initial
consumption. There are two measures of relative risk aversion the Arrow
Pratt measure of relative risk aversion and relative risk compensation.
The ArrowPratt measure of relative risk aversion is dened by
R(y)
v

(y)
v

(y)
y, (9.26)
and thus R(y) = yA(y).
The relative risk compensation for the relative risk z at deterministic initial
consumption y is the value
r
(y, z) that satises
E[v(y + yz)] = v(y y
r
(y, z)). (9.27)
Relative risk compensation
r
is related to (absolute) risk compensation
via

r
(y, z) =
(y, yz)
y
. (9.28)
For small relative risk z with E(z) = 0, it follows from Theorem 9.6.1 that

r
(y, z)

=
R(y)
2
z
2
. (9.29)
The parallel forms of approximations (9.15) and (9.29) provide a motivation
for denition (9.26) of the measure R of relative risk aversion.
A version of Pratts theorem holds for relative risk aversion: comparisons
of relative risk aversion of two agents using the ArrowPratt measure and
102 Risk Aversion s5
the relative risk compensation always give the same result. A reference is
given in the notes.
S: I REPARAMETRIZED POWER UTILITY FUNCTION SO THAT
ALPHA = ABSOLUTE RISK AVERSION IF GAMMA IS INFINITE, AND
GAMMA IS RELATIVE RISK AVERSION IF ALPHA = 0.
9.10 Utility Functions with Linear Risk Tolerance
The functions most often used as von NeumannMorgenstern utility func-
tions in applied work and as examples are linear utility and the following
utility functions:

Negative exponential utility. The utility function

v(y) = e
y
, (9.30)
for > 0, has absolute risk aversion that is constant and equal to .

Logarithmic utility. The utility function

v(y) = ln(y +
1

), y + 1/ > 0, (9.31)
has absolute risk aversion that is decreasing and equal to 1/(y +1/). If equals
, relative risk aversion equals 1.

Quadratic utility. The utility function

v(y) =
1
2
_
1

y
_
2
, 1/ y > 0, (9.32)
has absolute risk aversion that is increasing and equal to 1/(1/ y).

Power utility. The utility function

v(y) =

1
_
1

+
y

_
1
,
1

+
y

> 0, (9.33)
for = 0 and = 1, has absolute risk aversion equal to 1/(1/+y/). If > 0,
absolute risk aversion is decreasing. Otherwise, if < 0, it is increasing. If
equals , relative risk aversion equals .
S: NOTE THAT WE CANT HAVE ALPHA IN THE USUAL FORM IN
BOTH NEGATIVE EXPONENTIAL AND QUADRATIC SINCE BOTH
ARE SPECIAL CASES OF POWER. I CHOSE NEGATIVE EXPONEN-
TIAL. AGREE?
Negative exponential, logarithmic and quadratic utility can be viewed as
limiting cases of power utility when approaches , 1 or 1, respectively
. If the power utility function is written as
v(y) =

1
_
1 +
y

_
1
, (9.34)
9.11 Risk Aversion with Two-Date Consumption 103
which is an ane transformation of (9.33), where > 0, then v(y) converges
to e
y
as approaches zero.
If a dierent ane transformation of (9.33) is considered,
v(y) =

1
_
_
1

+
y

_
1
1
_
, (9.35)
then by using lHopitals rule it can be shown that v(y) converges to ln(y +
1/) as approaches 1.
If = 1, then (9.33) takes the form (9.32).
S: NEW The Pratt theorem, with its implication that a utility function
is completely characterized by its risk aversion function, can be used to
verify the convergence of power utility functions to negative exponential,
logarithmic or quadratic utility as approaches , 1or 1. The power
utility function has absolute risk aversion given by 1/(1/+y/). If = ,
this reduces to , the risk aversion function for negative exponential utility.
If = 1 it reduces to 1/(1/+y), the risk aversion function for logarithmic
utility, while if = 1 it reduces to 1/(1/ y), the risk aversion function
for quadratic utility.
All these utility functions are strictly increasing, strictly concave, and
have risk tolerance that depends linearly on consumption (strictly, the de-
pendence is ane, not linear). For the negative exponential utility function
(9.30), risk tolerance is constant, T(y)=1/; for the logarithmic utility func-
tion (9.31), risk tolerance is T(y) = 1/ + y; for the power utility function
(9.33), risk tolerance is T(y) = 1/+y/. These utility functions are called
linear risk tolerance (LRT) utility functions (alternatively, HARA utility
functions, where HARA stands for hyperbolic absolute risk aversion because
A(y) denes a hyperbola).
The domain of an LRT utility function can be conveniently written as
{y : T(y) > 0}. Note that the parameter (with = 0 for the negative
exponential utility function and = 1 for the logarithmic utility function)
is the slope of the risk tolerance function.
The LRT utility functions have many attractive properties, as will be seen
in Chapters ?? and ??.
9.11 Risk Aversion with Two-Date Consumption
The denitions of risk aversion and risk neutrality can easily be adapted to
the case in which date-0 consumption enters agents utility functions.
An agent with von NeumannMorgenstern utility function v : R
2
R is
104 Risk Aversion s5
risk averse if
E[v(c
0
, c
1
)] v(c
0
, E(c
1
)), (9.36)
for every c
0
and every c
1
, and is risk neutral if
E[v(c
0
, c
1
)] = v(c
0
, E(c
1
)), (9.37)
for every c
0
and every c
1
.
By Theorem 9.3.1 an agent is risk averse i the von NeumannMorgenstern
utility function v(y
0
, y
1
) is concave in y
1
for every y
0
and risk neutral
i v(y
0
, y
1
) is linear in y
1
for every y
0
. For instance, utility functions
v(y
0
, y
1
) = y
0
+ y
1
and v(y
0
, y
1
) = y
0
y
1
imply risk neutrality.
When v is not additively separable over time, the measures of date-1 risk
aversion of Section 9.5 depend on date-0 consumption. Consequently, an
agent can be risk neutral in date-1 consumption for some values of c
0
and
strictly risk averse for others, for example. In the case of time-separable
expected utility (8.20), an agents attitude toward date-1 risk depends only
on the form of the date-1 utility function v
1
; the S: NOTE ADDITION
form of the date-0 utility function and the level of date-0 consumption is
irrelevant.
For a time-separable power utility function (with = 0),
v(y
0
, y
1
) =
1
1
_
(y
0
)
1
1

+ (y
1
)
1
1

_
, (9.38)
where = 0, 1, the measure of absolute (date-1) risk aversion is 1/y
1
and
depends only on y
1
; the measure of relative (date-1) risk aversion is S: NOTE
CHANGE 1/. Note that the marginal rate of substitution between date-
0 consumption and date-1 consumption under this power utility function
is (y
1
/y
0
)
1/
, and the elasticity of substitution is S: NOTE CHANGE .
Thus, the elasticity of substitution depends on the coecient of relative
risk aversion. In general, the intertemporal elasticity of substitution and
the coecient of risk aversion are interdependent under the expected utility
representation.
9.12 Notes
The equivalences proved in Theorem 9.3.1 between risk aversion (strict risk
aversion, risk neutrality) and concavity (strict concavity, linearity) of utility
function are also an implication of the Pratt Theorem (take v
1
= v and
linear v
2
). However, the Pratt Theorem applies only to dierentiable utility
functions, whereas Theorem 9.3.1 applies to all continuous utility functions.
9.12 Notes 105
S: NEW Section 9.4 is based on Werner [?]. Werner also presented a
proof that does not depend on dierentiability of v. For proof that the
independence axiom implies that preferences are separable when S 3, see
Debreu [?]. If S = 2 the result does not obtain because the independence
axiom does not imply separable preferences in that case.
The ArrowPratt measures of absolute and relative risk aversion were
proposed in Arrow [?], [?] and Pratt [?]. The Pratt Theorem is found in
Pratt [?]. A version of the Pratt Theorem for relative risk aversion can also
be found in Pratt [?].
An illuminating discussion of measures of risk aversion can be found in
Yaari [?].
Measures of risk aversion introduced in this chapter are based on the
assumption that a risk-free payo is attainable. More general measures that
apply when a risk-free payo is not attainable have been proposed by Ross
[?]; see also Machina and Neilsen [?]. Cohen [?] discussed concepts of risk
aversion without the expected utility representation of preferences.
Kihlstrom and Mirman [?] addressed problems in extending the Arrow
Pratt theory of risk aversion to multivariate risks (for example, state-contingent
consumption plans with multiple goods).