Sumit Joshi

Microeconomic Theory

Risk Aversion

1 Introduction
Risk can be defined as the variability in a monetary outcome. It is for
this reason that risk is generally associated with the variance (or standard
deviation) of the probability distribution on the set of monetary outcomes.
In this part we will analyze the behavior towards risk of a consumer with
expected utility preferences. We will take the probability distributions over
outcomes (the lotteries) as fixed and analyze restrictions on the Bernoulli
utility function under which a consumer prefers one distribution over an-
other. We will see that the attitude towards risk is governed by the second
order concavity or convexity property of the Bernoulli utility function.

2 Defining Risk Aversion

An individual’s attitude towards risk can be examined by considering the
individual’s preference between a lottery L, and a degenerate lottery giving
the expected value of L. Recall that a lottery is formally a distribution
function, F , on monetary outcomes. When the individual has expected
utility preferences, then in the discrete case:
X
U (F ) = u(x)p(x) (1)
x: p(x)>0

where u is the Bernoulli utility function over monetary outcomes, U is

the von Neumann - Morgenstern
P utility function over lotteries (or distri-
bution functions), and x: p(x)>0 p(x) = 1. We will always assume that the
Bernoulli utility function, u, is an increasing function, i.e. u0 > 0, implying
that marginal utility from wealth is strictly positive. In the continuous case:
Z +∞ Z +∞
U (F ) = EF [u] = u(x)f (x)dx = u(x)F 0 (x)dx (2)
−∞ −∞
R +∞ R +∞
where −∞ f (x)dx = −∞ F 0 (x)dx = F (+∞) − F (−∞) = 1 − 0 = 1.

1
Definition 1 A consumer is risk averse if, when faced with a choice be-
tween a lottery F and a degenerate lottery which offers the expected value of
F , the consumer always prefers the latter. Given expected utility preferences,
a consumer is risk averse if and only if for any lottery F :
⎛ ⎞
X X
u(x)p(x) ≤ u ⎝ xp(x)⎠ (3)
x: p(x)>0 x:p(x)>0

when X is discrete and:

Z +∞ µZ +∞ ¶
0 0
u(x)F (x)dx ≤ u xF (x)dx (4)
−∞ −∞

when X is continuous.

A risk averse consumer would prefer to get the expected value of F for certain
rather than be exposed to the risk inherent in the lottery F . Inequalities (3)
and (4), for the discrete and continuous case respectively, are referred to as
Jensen’s Inequality and are the defining conditions for concave functions,
u. Therefore, a consumer with expected utility preferences is risk averse if
and only if the consumer’s Bernoulli utility function is concave on the set of
monetary outcomes.

Definition 2 A consumer is said to be strictly risk averse if the consumer

is risk-averse and indifference between a lottery F and receiving the expected
value of F for certain holds only when the lottery F is degenerate. With
expected utility preferences, a consumer is strictly risk averse if and only if
the Bernoulli utility function is strictly concave.

While most economic applications concern themselves with risk averse eco-
nomic agents, in some cases they also consider risk neutral and risk loving
economic agents. These can now be defined for the case of expected utility
preferences. We define it for the continuous case since the discrete case is
similar:

Definition 3 A consumer is risk neutral if and only if for any lottery F :

Z +∞ µZ +∞ ¶
0 0
u(x)F (x)dx = u xF (x)dx (5)
−∞ −∞

2
The above equality holds only for linear real-valued functions. Therefore,
a consumer with expected utility preferences is risk neutral if and only if
the consumer’s Bernoulli utility function is linear on the set of monetary
outcomes.

Definition 4 A consumer is risk loving if and only if for any lottery F :

Z +∞ µZ +∞ ¶
0 0
u(x)F (x)dx ≥ u xF (x)dx (6)
−∞ −∞

The above two inequalities hold only for a convex real-valued function.
Therefore, a consumer with expected utility preferences is risk loving if and
only if the consumer’s Bernoulli utility function is convex on the set of
monetary outcomes.

3 Equivalent Formulation of Risk Aversion

An equivalent definition of risk aversion can be offered in terms of the cer-
tainty equivalent of a lottery. Given a lottery F , the certainty equivalent
of F , denoted by c(F ), is the monetary amount that will make a consumer
indifferent between the lottery F and the certain amount c(F ). Mathemat-
ically, in the discrete and continuous cases respectively:
X
u(x)p(x) = u(c(F ))
x: p(x)>0
Z +∞
u(x)F 0 (x)dx = u(c(F )) (7)
−∞

We will now show that a consumer is risk averse if and only if the certainty
equivalent of any lottery F is less than or equal to the expected value of F ,
that is: ( P
x: p(x)>0 xp(x),
R +∞
X discrete
c(F ) ≤ 0 (8)
−∞ xF (x)dx, X continuous

3
Consider the continuous case. Using equality (7) and inequality (4) respec-
tively:
Z +∞
u(c(F )) = u(x)F 0 (x)dx
−∞
µZ +∞ ¶
≤ u xF 0 (x)dx
−∞

and the result follows from the fact that u is an increasing function.

We can characterize risk aversion in yet another way using the notion of a
risk premium, π(F ). It is defined as the maximum amount a consumer is
willing to pay (from the expected value of F ) in order to replace the lottery
F with a certain return. Mathematically, the risk premium is defined by the
following condition:
µZ +∞ ¶ Z +∞
0
u xF (x)dx − π(F ) = u(x)F 0 (x)dx (9)
−∞ −∞
| {z } | {z }
Sure Return Expected Utility from F

Recalling equality (7), it follows that:

Z +∞
π(F ) = xF 0 (x)dx − c(F ) (10)
−∞

Therefore, a consumer is risk averse if and only if the risk premium associated
with any lottery is non-negative.

4 Absolute Measure of Risk Aversion

In many economic applications, we would like to have a measure of the
consumer’s risk aversion. We could then meaningfully investigate the change
in risk aversion due to a change in some parameter such as wealth. Since
risk aversion is equivalent to the concavity of the Bernoulli utility function,
a natural candidate to measure the degree of risk aversion at x seems to be
the curvature of the utility function as measured by u00 (x). However, the
second derivative by itself is not a valid measure of risk aversion. Recall
that a consumer’s expected utility preferences remain the same under any
affine transformation of utility. Therefore, the Bernoulli utility function,
v(x) = au(x) + b, a > 0, represents the same preferences between any two

4
lotteries as u. But, v 00 (x) = au00 (x), and if a 6= 1 then v 00 (x) 6= u00 (x).
But this contradicts the fact that v should exhibit the same degree of risk
aversion at x as u.

In order to have a measure of the degree of risk aversion which is invariant

to affine transformations of the Bernoulli utility function, we define it at x
as follows:
u00 (x)
rA (x) = − 0 (11)
u (x)
It is called the Arrow-Pratt measure of absolute risk aversion. It is
easy to verify that all Bernoulli utility functions which are affine transfor-
mations of each other will have the same Arrow-Pratt measure of absolute
risk aversion.

The Arrow Pratt measure can be interpreted in a different way by looking at

the risk premium that an individual with wealth w is willing to pay in order
to avoid a fair lottery. Consider a lottery F and assume that the lottery is
fair in the sense that its expected value is equal to zero, that is:
Z +∞
E(X) = xF 0 (x)dx = 0 (12)
−∞

By the definition of the risk premium:

u(w − π(F )) = E[u(w + X)] (13)

We will expand both sides of the above equation using the Taylor series
approximation. Since the risk premium is a constant amount, we take a
first order approximation of the left hand side:

u(w − π(F )) = u(w) − π(F )u0 (w)

We take a second order approximation of the right hand side:

∙ ¸
0 1 2 00
E[u(w + X)] = E u(w) + Xu (w) + X u (w)
2

Note that EX = 0. Therefore, EX 2 = σ 2 , the variance of the lottery.

Equating the two sides and simplifying we get:

2π(F )
rA (w) = (14)
σ2

5
Therefore, the measure of absolute risk aversion is proportional to the risk
premium a risk averse consumer is willing to pay to avoid a small but fair
lottery. In particular, absolute risk aversion is equal to twice the risk pre-
mium the consumer is willing to pay to avoid a unit of variance for very
small risks.

5 Comparative Statics: Comparison Across Wealth

The Arrow-Pratt measure is a local measure of risk aversion because it evalu-
ates the derivatives of the Bernoulli utility function at a given level of wealth.
Therefore, it will not be the same at every level of wealth. An interesting
question is whether an individual’s risk aversion increases or decreases with
the level of wealth. We can classify utility functions in terms of constant,
increasing or decreasing absolute risk aversion over some subset of wealth if
rA (x) remains constant, increases or decreases respectively with wealth over
this subset.
drA (x)
Constant absolute risk aversion (CARA) dx =0
drA (x)
Increasing absolute risk aversion (IARA) dx >0
drA (x)
Decreasing absolute risk aversion (DARA) dx <0

Example: The Bernoulli utility function:

u(x) = −e−ax + b, a>0

displays constant absolute risk aversion (CARA). The utility function:

b
u(x) = a + bx − cx2 , a, b, c > 0, 0<x<
2c
displays increasing absolute risk aversion (IARA). The utility function:

u(x) = log x

displays decreasing absolute risk aversion (DARA).

Of the three possibilities, DARA is the most realistic restriction on the

preferences of a risk averse consumer. It is consistent with the observation
that individuals will be less averse to risk at higher levels of wealth.

6
6 Comparative Statics: Comparison Across Con-
sumers
The Arrow-Pratt measure can also be used to compare risk aversion across
consumers. Consider consumers 1 and 2 with Bernoulli utility functions
u1 and u2 over monetary outcomes respectively. In many economic applica-
tions, we are interested in comparing the optimal decisions of two consumers
where one, say consumer 1, is always more risk averse than consumer 2. We,
therefore, need a definition of what it means for consumer 1 to be always
more risk-averse than consumer 2. Given different Bernoulli utility functions
for the two individuals, the most obvious way is to compare the Arrow-Pratt
measure of absolute risk aversion for the two individuals at the same level
of wealth:

Definition 5 Given two consumers, 1 and 2, with strictly increasing and

concave Bernoulli utility functions u1 and u2 respectively, consumer 1 is
more risk averse than consumer 2 if and only if for every x:

1 u001 (x) u00 (x) 2

rA (x) = − 0 ≥ − 20 = rA (x) (15)
u1 (x) u2 (x)

Using the above definition, we can obtain other equivalent characterization

results for comparing risk aversion across consumers. Note that since both u1
and u2 are strictly increasing functions, we can find some strictly increasing
real-valued function φ such that:

u1 (x) = φ(u2 (x)) (16)

Differentiating (16) yields:

u01 (x) = φ0 (u2 (x))u02 (x)

Differentiating the above a second time yields:

£ ¤2
u001 (x) = φ0 (u2 (x))u002 (x) + φ00 (u2 (x)) u02 (x)

Dividing the above expression by u01 and simplifying yields:

∙ 00 ¸
1 2 0 φ (u2 (x))
rA (x) = rA (x) − u2 (x) 0 (17)
φ (u2 (x))

7
It follows from the above expression that rA 1 (x) ≥ r 2 (x) is equivalent to
A
00
φ ≤ 0, i.e. φ is concave. This gives us the following characterization result:
consumer 1 is more risk averse than consumer 2 if and only if the Bernoulli
utility function of consumer 1 is more concave than that of consumer 2 in
the sense that u1 is a concave transformation of u2 .

Using the above result we can obtain the following equivalent characteri-
zation of comparative risk aversion: consumer 1 is more risk averse than
consumer 2 if and only if for every lottery F , the certainty equivalent of
consumer 1, c1 (F ), is less than or equal to the certainty equivalent of con-
sumer 2, c2 (F ). To see this, note from the definition of certainty equivalent
that:
Z +∞
u1 (c1 (F )) = u1 (x)F 0 (x)dx
−∞
Z +∞
u2 (c2 (F )) = u2 (x)F 0 (x)dx
−∞

Using our result that u1 is a concave transformation of u2 :

Z +∞
u1 (c1 (F )) = φ(u2 (x))F 0 (x)dx
−∞
µZ +∞ ¶
≤ φ u2 (x)F 0 (x)dx
−∞
= φ(u2 (c2 (F )))
= u1 (c2 (F ))

and the result follows since u1 is an increasing function.

The next characterization follows by recalling the definition of a risk pre-

mium. For any lottery F , since c1 (F ) ≤ c2 (F ):
Z +∞
π 1 (F ) = xF 0 (x)dx − c1 (F )
−∞
Z +∞
≥ xF 0 (x)dx − c2 (F )
−∞
= π 2 (F )

Therefore, for the same lottery F , the relatively more risk-averse consumer
has a greater risk premium.

8
The last equivalent characterization of comparative risk aversion follows
from the above analysis and states that for the same initial wealth, if con-
sumer 1 prefers a lottery F to the sure outcome x, then consumer 2 will also
prefer the lottery F to the sure outcome x. In other words, a less risk averse
consumer is always willing to accept more lotteries than a more risk averse
one.

b.
To see this, suppose consumer 1 prefers a lottery F to the sure outcome x
Then, by definition:
Z +∞
u1 (x)F 0 (x)dx ≥ u1 (b
x)
−∞

Recalling that u1 (x) = φ(u2 (x)), the above can be written equivalently as:
Z +∞
φ(u2 (x))F 0 (x)dx ≥ φ(u2 (b
x))
−∞

Using Jensen’s inequality:

µZ +∞ ¶
0
φ u2 (x)F (x)dx ≥ φ(u2 (b
x))
−∞

and since φ is an increasing function:

Z +∞
u2 (x)F 0 (x)dx ≥ u2 (b
x)
−∞

showing as required that consumer 2 also prefers the lottery F to the sure
b.
outcome x

7 Relative Measure of Risk Aversion

The Arrow-Pratt measure of absolute risk aversion considers those lotter-
ies where the uncertain monetary outcomes are independent of the current
wealth of the consumer. In other words, the random variable refers to ab-
solute changes to the current level of wealth. If the current wealth is w0 ,
then wealth changes according to w0 + δ where δ is a random variable with
distribution function F . However, there are many lotteries whose uncertain

9
monetary outcomes are gains or losses as a percentage or proportion of cur-
rent wealth. In other words, if the current wealth is w0 , then wealth changes
according to δw0 where δ is a random variable with distribution function F .

To measure risk-aversion to such lotteries that affect wealth proportionately

requires an alternative measure referred to as the Arrow-Pratt measure
of relative risk aversion:
xu00 (x)
rR (x) = − = xrA (x) (18)
u0 (x)
Analogous to the absolute risk aversion case, we have the following charac-
terization of relative risk aversion:
drR (x)
Constant relative risk aversion (CRRA) dx =0
drR (x)
Increasing relative risk aversion (IRRA) dx >0
drR (x)
Decreasing relative risk aversion (DRRA) dx <0

Empirically we observe DARA and CRRA in the real world.

Example: Consider the following Bernoulli power utility function:

½ xa
u(x) = a , a < 1 a 6= 0
log x, a = 0

The above function displays decreasing absolute risk aversion but constant
relative risk aversion.

Now consider a lottery which causes proportionate changes in wealth equal

to δw, where δ has distribution function F . It is assumed that the lottery
is fair, i.e. expected value of E[δ] is equal to zero. Further, the variance
of this lottery, E[δ 2 ] = σ 2 . Define the proportional risk premium of F,
π p (F ), as the maximum proportion of wealth w that the consumer is willing
to forego in order to substitute the lottery F for a sure return. Then:

u(w − π p (F )w) = E[u(w + δw)] (19)

Expand both sides of the above equation using a Taylor series approxima-
tion. Since the proportional risk premium is constant, we take a first order
approximation of the LHS:

u(w − π p (F )w) = u(w) − wπ p (F )u0 (w)

10
Taking a second order approximation of the RHS
∙ ¸
0 1 2 2 00
E[u(w + δw)] = E u(w) + δwu (w) + δ w u (w)
2

Noting that E[δ] = 0 and E[δ 2 ] = σ 2 , equating the two sides we get:

2π p (F )
rR (w) = (20)
σ2
Therefore, analogous to the case of absolute risk aversion, the Arrow-Pratt
measure of relative risk aversion is directly related to the proportional risk
premium a risk averse consumer is willing to pay to avoid a small but fair
lottery.

References
[1] J.W. Pratt (1964) “Risk Aversion in the Small and in the Large”, Econo-
metrica 32, 122-136.

11