Decison Theory
Review Lecture
Issues in decision making
Michael Goldstein
1 Decision support
Decision theory tells us how to make decisions when all of the probabilities
and payoffs have been specified.
However, there will always be further considerations involved in mak-
ing decisions. In this way, the theory may be viewed as providing decision
support to guide us in our decision making rather than simply identifying
the best decision to use. We have seen a few of these considerations in the
assessment of risk profiles and EVPI calculations.
One of the most important considerations is that the numbers we input
into our decision analysis are rarely exact. So, one of the key features for
decision support is to assess how sensitive our decisions are to these imprecise
specifications, in order to show us which judgements may require further
consideration.
2 Sensitivity analysis
For example, let us go back to the simple decision problem we studied first.
bad sales good sales
E1 E2
new mach d1 260 440
overtime d2 300 420
probability 0.4 0.6
1
How sensitive is our decision choice to our assessment of the probability
of good sales?
3 Calculating sensitivity
Call p(G) = p. We will assess how critical to our choice of decision was our
choice p = 0.6.
We calculate the EMV’s as a function of p.
EM V (d1 ) = 440p + (1 − p)260 = 180p + 260
EM V (d2 ) = 420p + (1 − p)300 = 120p + 300
Therefore,
EM V (d2 ) > EM V (d1 )
if and only if
120p + 300 > 180p + 260
or equivalently, if and only if
2
p<
3
So if the probability of good sales is less than two thirds, we should use
overtime, otherwise we should buy new machinery.
We assessed p = 0.6. Therefore a 10% change in p would not change our
decision. (Whether that is a large change depends on our confidence in our
probability judgements.)
2
4 Incompatible preferences
Choice 1
Which of the following gambles would you prefer, and why?
g1 = £500, 000
g2 = 0.1£2, 500, 000 ⊕ 0.89£500, 000 ⊕ 0.01£0
Answer
Your preferences between gambles are (clearly!) up to you. The only thing
you could do wrong would be to simply find the EMV for each gamble and
to use this as your utility—money value, for large amounts, is not a plausible
utility for amounts of money.
Choice 2
Which of the following gambles would you prefer, and why?
g3 = 0.11£500, 000 ⊕ 0.89£0
g4 = 0.1£2, 500, 000 ⊕ 0.9£0
Answer
Same answer as above - don’t rely on EMV unless you really are risk neutral.
Question
Suppose that you prefer g1 to g2 and also prefers g4 to g3 . Are your preferences
compatible with the theory of utility?
3
Answer
Many people prefer gamble 1 to gamble 2 (because in gamble 1 you cannot
lose) and also prefer gamble 4 to gamble 3 (because the payoff for 4 is better).
This is inconsistent with any utility function for money, as we’ll now show.
Without loss of generality, let’s assign U (0) = 0 and U (2, 500, 000) = 1.
Then for any assignment U (500, 000) = x for x ∈ [0, 1] we have the following
utilities:
U (g1 ) = x
U (g2 ) = 0.1U (2, 500, 000) + 0.89U (500, 000) + 0.01U (0) = 0.1 + 0.89x
U (g3 ) = 0.11U (500, 000) + 0.89U (0) = 0.11x
U (g4 ) = 0.1U (2, 500, 000) + 0.9U (0) = 0.1
Hence, g1 > ∗g2 means that
0.11x > 0.1
but g4 >∗ g3 means that
0.1 > 0.11x .
Thus we have arrived at a contradiction.
[This example is often termed the Allais paradox.]
What assumptions about preferences are broken by the given preferences?
There are different ways to answer this. Here’s one.
For any gambles s1 , s2 , r
if s1 <∗ s2 , then ps1 ⊕ (1 − p)r <∗ ps2 ⊕ (1 − p)r
(This is substitutability combined with coherence)
Let s1 = £500, 000 for sure.
Let s2 = 10
11
1
2, 500, 000 ⊕ 11 £0
Let r = £500, 000 for sure, t = £0 for sure.
You can check that
g1 ∗ > g2 and g3 <∗ g4 correspond to
0.89r ⊕ 0.11s1 ∗ > 0.89r ⊕ 0.11s2
0.89t ⊕ 0.11s1 <∗ 0.89t ⊕ 0.11s2
This breaks substitutability.
4
5 Logarithmic utility
Consider the utility function Uc (£x) = log(x + c).
It has the following properties
[1] Uc is a concave, increasing function on positive values of x.
[2] The local risk aversion is
1
rc (x) =
x+c
So this utility function is risk averse, but with decreasing risk aversion
with increasing x, where c controls the rate of decrease of risk aversion.
6 St Petersburg Paradox
This function was originally suggested by Bernoulli to resolve the St. Peters-
burg paradox which we discussed earlier.
The example was as follows.
Toss a fair coin until first tail
If n tosses, win £2n
The EMV of the gamble is infinite.
How much would you pay to play?
7 St Petersburg paradox with log utility
Suppose U (£x) = log2 (x);
g = 12 £2 ⊕ 212 £22 ⊕ · · · ⊕ 21n £2n ⊕ · · ·
So, the expected utility of the gamble is
∞ ∞ ∞
X 1 n
X 1 n
X n
n
U (£2 ) = n
log2 (2 ) = =2
n=1
2 n=1
2 n=1
2n
(Note: for any x such that |x| < 1 it holds that ∞
P n d
P∞ n
n=1 nx = x dx n=1 x =
d x x
x dx 1−x = (1−x)2 .)
Hence you would pay £xg where U (£xg ) = 2
Therefore, the certainty equivalent is given by
log2 (xg ) = 2 =⇒ xg = £4
5
8 Discussion
In this problem, the reason for the choice of the logarithmic function to
express utility is motivated in the same paper in which Bernoulli discusses
the St. Petersburg Paradox, where he introduces the principle of Decreasing
Marginal Utility.
This principle states that marginal utility (the extra utility obtained from
consuming a good) decreases as the quantity consumed increases; in other
words, that each additional good consumed is less satisfying than the previous
one.
For example, £1000 is worth more to a person whose total wealth is
£1 than to a person whose total wealth is £1, 000, 000, and each additional
pound is less valuable than the previous pound.
Modern utility theory generalises this idea from the logarithmic form,
while preserving the general insight of Bernoulli’s solution.
