http://www.owlnet.rice.edu/~cr15/courses/eco501/notes/501_10_08.

pdf

Rice University
ECO 501: Microeconomic Theory I
———————————————–
Lecture 10: Risk
———————————————–
Christian Roessler
November 5, 2008

1 Money Lotteries
A lottery over continuous amounts of money x 2 R can be described
most generally in terms of its cumulative distribution function F : R !
[0; 1]. (The more direct approach would be to use density functions
f ( ), but these do not always exists and exclude
Rx the case of discrete
outcomes. If f ( ) does exist, then F (x) = 1 f (t) dt.)

We now take the lottery space L to be the set of distribution func-

tions on R. The continuous version of the expected utility theorem
guarantees that a continuous preference % on L, that satis…es the in-
dependence axiom, can be represented by
Z
U (F ) = u (x) dF (x) ;

R(Note that f (x)R dF (x) =dx, so that dF (x) = f (x) dx, and thus
u (x) dF (x) = u (x) f (x) dx whenever the density exists.)

The function u ( ), which records the values of degenerate lotteries, i.e.

certain amounts of money, is called Bernoulli utility function. Assume

1
 that u ( ) is increasing, continuous and bounded. (If it were unbounded,
small probability events could make a lottery in…nitely desirable - the
St. Petersburg paradox.)

Exercise (MWG 6.C.2). Suppose an individual’s Bernoulli utility function

u ( ) has the quadratic form u (x) = x2 + x. Show that utility from a
distribution is determined by the mean and variance of the distribution, and
only these moments. (No need to do part b.)

2 Risk Attitude
A risk-averse agent is someone who rejects fair gambles (that have
neither an expected gain nor loss).

De…nition. An agent is risk-averse if she prefers the expected money value

of a lottery to the lottery itself: 8F 2 L,
Z
u xdF (x) U (F ) :

(Strictly risk-averse if this is an equality only if F ( ) is degenerate.) The

agent is risk-neutral if always indi¤erent between the expected value of a
lottery and the lottery itself, i.e. the above is an equality.

The criterion for risk aversion implies, when U ( ) has the expected
utility form, Z Z
u xdF (x) u (x) dF (x) :

This is Jensen’s inequality that characterizes a concave function u ( ).

(If you think of the integral over the probability distribution F ( ) as a
weighted sum, the inequality relates to the basic de…nition of a concave
function u ( x + (1 ) x0 ) u (x) + (1 ) u (x0 ).)

Hence the Bernoulli utility function of a risk-averse agent is concave,

and anyone with a concave Bernoulli function is risk-averse. This fact
has a straightforward explanation: the risk-averse agent’s Bernoulli

2
Figure 1: Bernoulli utility functions for risk-averter (left) and risk-seeker
(right)

utility increases more slowly with gain than it decreases with a loss.
Since the agent has, in utility terms, more to lose than gain from a
lottery that is fair in money terms, she declines the lottery unless the
odds are strictly favorable.

In Figure 1, the expected utility of a fair lottery, i.e. random vari-

able X : fx1 ; x2 g ! [0; 1] with equal probabilities, is labeled U (X).
The Bernoulli utility of certain amount E (X) = (1=2) x1 + (1=2) x2
is labeled u (E (X)). The individual in the left panel is risk-averse,
and U (X) u (E (X)). (Observe how x1 and x2 along the x-axis are
equidistant from E (X), while the corresponding utility value u (E (X))
is closer to the utility of the better state, u (x2 ).)

On the right panel, we have the contrasting case of a risk-seeker, whose

Bernoulli utility is convex.

De…nition. The certainty equivalent c (F; u) of lottery F ( ) is its money

value to an agent whose preference is represented by Bernoulli utility function

3
Figure 2: Certainty equivalent for risk-averter (left) and risk-seeker (right)

u ( ): Z
u (c (F; u)) = u (x) dF (x) :

It is intuitive that a risk-averse person has a certainty equivalent less

than the expected value of the lottery, i.e. she only values the lottery
the same if it produces on average a gain relative to the certainty equiv-
alent. Indeed, concavity of the (increasing) Bernoulli utility function
implies:
Z Z Z
u (x) dF (x) u xdF (x) () u (c (F; u)) u xdF (x)
Z
() c (F; u) xdF (x) :

Figure 2 shows the certainty equivalent for risk-averse and risk-seeking

agents.

4
 Beyond these formal characterizations, risk-averse behavior is evident
in the propensity to buy insurance, even when premia are not "actuar-
ially fair" (i.e. the expected payout is less than the premium).

Example. A strictly risk averse agent with initial wealth w faces a possible
damage of $D with probability . The agent is o¤ered insurance at a fair
premium per dollar-payout in the event of a loss. If dollars of insurance
(i.e. conditional payout) are purchased at this premium, the agent’s wealth
will be either w (if no damage occurs) or w D+ = w+
(1 ) D (if there is damage). Expected utility from a choice of ,
which induces a lottery over w + (1 ) D and w with probabilities
( ;1 ), is then

U ( ) = u (w + (1 ) D) + (1 ) u (w ):

The …rst-order condition with respect to ,

u0 (w + (1 ) D) = u0 (w );

can be solved for w + (1 ) D=w , i.e. = D because u0 ( )

is strictly decreasing from strict concavity. Thus, a risk-averse agent insures
fully if the premium is actuarially fair.

Example. Suppose an amount of wealth can be invested in a safe asset,

which yields $1, Rand a risky asset with earnings distribution F (z) such that
expected return zdF (z) is greater than 1. The portfolio choice problem is
to determine the optimal shares and to invest in these assets, such that
+ = 1. Since the random return, given and , is + z, utility is a
random variable u ( + z) = u (1 + z). An expected utility maximizer
solves: Z
max u (1 + (z 1)) dF (z) ;
2[0;1]

which has …rst-order condition, with respect to ,

Z
(z 1) u0 (1 + (z 1)) dF (z) = 0

(at anRinterior solution). Since the left side is greater than zero at = 0,
given zdF (z) > 1 and that u ( ) is increasing everywhere, we must have

5
 > 0, whether or not the individual is risk-averse. The general principle
is that an agent will always invest some share of wealth in an actuarially
favorable asset.

Exercise (MWG 6.C.19).

Risk aversion can be quanti…ed and compared by means of the absolute

and relative coe¢ cients of risk aversion.

De…nition. The Arrow-Pratt coe¢ cient of absolute risk aversion is, for a
twice di¤erentiable Bernoulli utility function u ( ) at x,
u00 (x)
rA (x; u) = :
u0 (x)

This is essentially a measure of the concave curvature of the Bernoulli

utility function (for an increasing concave function, u0 (x) > 0 and
u00 (x) < 0). Note that we cannot compare the degree of concavity based
on u00 (x) alone, since this derivative can be scaled by a positive linear
transformation (which yields another utility function that represents
the preference). But such a transformation would also scale u0 (x), so
it cannot a¤ect rA (x; u).

For a risk-neutral agent, u0 (x) is constant and u00 (x) = 0, so rA (x; u) =

0.

The more risk-averse of two agents has a lower certainty equivalent

for any given lottery F ( ). Moreover, 2 is more risk-averse than 1 in
the sense that rA (x; u2 ) rA (x; u1 ) if and only if u2 ( ) is a concave
transformation of u1 ( ). I.e. there exists an increasing concave function
( ) such that 8x, u2 (x) = (u1 (x)).

Exercise (MWG 6.C.6, C.7).

Example. It is possible to recover the preference from the Arrow-Pratt coef-

…cient. Suppose rA (x; u) = a for all x. Integrating u00 (x) = au0 (x) on both

6
sides, we have u0 (x) =u (x) = @ ln u (x) =@x = a, and integrating once more
on both sides, u (x) = e ax , i.e. the utility function is exponential when the
Arrow-Pratt coe¢ cient is constant. I have constructed one particular utility
function, assuming the integration constants are zero, but others are still
exponential (solve the di¤erential equation to see this). Exponential utility
functions therefore constitute the CARA (constant absolute risk aversion)
class.

The DARA class of Bernoulli utility functions has the plausible prop-
erty that wealthier people tend to be less risk-averse.

De…nition. Bernoulli utility function u ( ) exhibits decreasing absolute risk

aversion (DARA) if rA (x; u) is a decreasing function of x, i.e. 8x, rA (x; u) >
rA (x0 ; u) whenever x < x0 .

Exercise (MWG 6.C.8).

De…nition. The coe¢ cient of relative risk aversion is, for a twice di¤eren-
tiable Bernoulli utility function u ( ) at x,
u00 (x)
rR (x; u) = x :
u0 (x)

If rR (x; u) = xrA (x; u) is decreasing in x, then clearly rA (x; u) must

be decreasing in x. The converse is not true. Therefore, non-increasing
relative risk reversion is a stronger property than decreasing absolute
risk aversion.

Example. Consider the Bernoulli utility function u (x) = x1 = (1 ),

where 2 (0; 1). Since rA (x; u) = =x is decreasing in x, this function is in
the DARA class. But rR (x; u) = xrR (x; u) = is constant, so DARA does
not imply DRRA.

Exercise (MWG 6.C.18).

Exercise (MWG 6.C.12).

7
3 Stochastic Dominance
Up to now, we have compared agents in terms of the risk aversion
exhibited by their utility functions. Now we are interested in compar-
ing lotteries and …nding criteria by which they can be ranked, given
properties of preference, such as risk attitude.
If distribution F yields a higher expected utility than lottery G, re-
gardless of risk attitude (i.e. the speci…c form of the Bernoulli utility
function), F is said to …rst-order stochastically dominate G.

De…nition. Distribution F ( ) …rst-order stochastically dominates G ( ), writ-

ten F %F OSD G, if
Z Z
u (x) dF (x) u (x) dG (x)

for any non-decreasing function u : R ! R.

If distribution F yields a higher expected utility than lottery G for a

risk-averse agent (i.e. a concave Bernoulli utility function), then F is
said to second-order stochastically dominate G.

De…nition. Distribution F ( ) second-order stochastically dominates G ( ),

written
R F %SOSD
R G, if F ( ) and G ( ) have the same expectation of x, i.e.
xdF (x) = xdG (x), and
Z Z
u (x) dF (x) u (x) dG (x)

for any non-decreasing concave function u : R+ ! R.

F %F OSD G is equivalent to the property: for any x, getting more than

x is more likely under F than under G.
A related statement can be made for second-order dominance. F %SOSD
G is equivalent to the property: for any x, probability mass accumulates
faster toward x under G than under F . (I.e. G spreads the probability
mass more evenly and gives more weight to the extremes. I will give a
precise characterization in a moment.)

8
The next results make use of some integral relationships that may be
found through "integration by parts." The technique is based on the
product rule applied to u (x) F (x):
d
u (x) F (x) = u0 (x) F (x) + u (x) f (x)
dx
implies
d
u (x) f (x) = u (x) F (x) u0 (x) F (x)
dx
and, integrating both sides,
Z Z
u (x) f (x) dx = u0 (x) F (x) dx;

where is a constant (since du (x) F (x) =dx integrated on ( 1; 1) is

u (1) F (1) u ( 1) F ( 1), and any distribution function satis…es
F ( 1) = 0 and F (1) = 1).
Rx
Applied to u0 (x) 1 F (t) dt, the product rule gives
Z x Z x
d 0 00
u (x) F (t) dt = u (x) F (t) dt + 2u0 (x) F (x)
dx 1 1
Rx Rx
(since the derivative of 1 F (t) dt with respect to x is 1 f (t) dt +
F (x) = 2F (x) by Leibniz’rule). Thus
Z x Z x
0 1 d 0 1 00
u (x) F (x) = u (x) F (t) dx u (x) F (t) dt
2 dx 1 2 1

and Z Z Z x
0 1 00
u (x) F (x) dx = u (x) F (t) dt dx:
2 1
R1
(where is a constant, since 21 u0 (1) 1
F (t) dt = 12 u0 (1)).
Hence
Z Z
u (x) dF (x) = u0 (x) F (x) dx
Z Z x
1 00
= + u (x) F (t) dt dx:
2 1

9
Proposition. The payo¤ distribution F ( ) …rst-order stochastically domi-
nates G ( ) if and only if 8x, F (x) G (x).
Proof.R (If) Let F (x) GR(x) for all x. From Rintegration by parts,
we
R 0 have u (x) dF (x) = u0 (x) F (x) dx and u (x) dG (x) =
u (x) G (x) dx, so
Z Z Z Z
0 0
u (x) F (x) dx u (x) G (x) dx () u (x) dF (x) u (x) dG (x) :

If u ( ) is an increasing function, the …rst inequality holds, given that F (x)

G (x) for all x. Hence F …rst-order dominates G.
(Only if) Suppose F (x) > G (x) for some x. To show that F fails to
…rst-order
R dominate G, R we need to …nd a non-decreasing function u ( ) such
that u (x) dF (x) < u (x) dG (x) at some x. Consider u (x) = 1 for x > x
and u (x) = 0 for x x, which is non-decreasing. Then
Z Z 1
u (x) dF (x) = dF (x) = 1 F (x)
x
Z 1 Z
< 1 G (x) = dG (x) = u (x) dF (x) :
x

Exercise (MWG 6.D.1).

Exercise (MWG 6.D.2).

Proposition. The payo¤ distribution F ( ) second-order stochastically dom-

inates G ( ) if and only if 8x,
Z x Z x
G (t) dt F (t) dt:
1 1

Rx Rx
Proof. (If) Let 1 F (t) dt 1
G (t) dt for all x. From integrating
R R Rx
by parts, we have u (x) dF (x) = + 21 u00 (x) 1 F (t) dt dx and
R R Rx
u (x) dG (x) = + 12 u00 (x) 1 G (t) dt dx. If u ( ) is concave, then
u00 (x) < 0, so
Z x Z x Z Z
F (t) dt G (t) dt () u (x) dF (x) u (x) dG (x) :
1 1

10
Hence F second-order Rdominates G. R
x x
(Only if) Suppose 1 F (t) dt > 1 G (t) dt for some x. To show that
F fails to second-order
R dominateR G, we need to …nd a concave function
u ( ) such that u (t) dF (t) < u (t) dG (t) at x. Let u (t) = t, except
R x0 R x0
in an interval [x; x] containing x where 1 F (t) dt > 1 G (t) dt for all
x0 2 [x; x]. On this interval, let u (t) be strictly concave.
From integration by parts,
Z Z Z Z x Z Z x
00 00
u (t) dF (t) > u (t) dG (t) () u (x) F (t) dt dx > u (x) G (t) dt dx:
1 1

Since u00 (x) = 0 for all x 2= [x; x] and u00 (x) < 0 for all x 2 [x; x], where
Rx Rx R
F (t) dt > x G (t) dt, the left inequality cannot hold, so u (x) dF (x) <
R x
u (x) dG (x). This means G SOSD F , a contradiction.

F %SOSD G is also equivalent to the property that G is a "mean-

preserving spread" of F . I.e. if G is obtainable from F by replac-
ing every certain outcome x with a lottery that yields x + z, where z
is
R a zero-mean random variable, distributed according to Hx (z) with
zHx (z) = 0.

Example. Consider two lotteries that reward outcomes of rolling a fair die.
F gives $1 if a number up to 3 is rolled and $2 if the number is greater
than 3. G pays nothing for a 1 and $5 for a 6, and $1 otherwise. These
lotteries have the same mean, 3=2, and probabilities (1=2; 1=2) over ($1; $2),
respectively (1=6; 2=3; 1=6) over ($0; $1; $5). To obtain G from F , replace
the $1 and $2 wins in F with lotteries that give ($0; $1; $5) respectively with
probabilities (1=3; 7=12; 1=12) and (0; 3=4; 1=4). Observe that the expected
values of these lotteries are $1 and $2. The compound lottery over ($0; $1; $5)
that plays (1=3; 7=12; 1=12) and (0; 3=4; 1=4) with equal probability reduces
to (1=6; 2=3; 1=6) = G. So we have constructed G as a mean-preserving
spread of F , i.e. F %SOSD G.

Example. Continuing in the previous scenario, distribution H is called an

"elementary increase in risk" from G if it redistributes all probability mass
to the extreme points in G’s domain (while preserving the mean). I.e. H

11
is the lottery (7=10; 0; 3=10). This is a mean-preserving spread via lotteries
(1; 0; 0), (4=5; 0; 1=5) and (0; 0; 1) in place of the $0, $1 and $5 wins.

Exercise (MWG 6.D.4).

Exercise (MWG 6.D.3).

12