Study Manual (Studiewijzer)

Decision Theory (Besliskunde)

2007 2008

Silja Renooij

Version 6, 24122008
General information

This study manual was designed to help guide your self studies. As such, it does not include
reading material that is not already present in the course book, or the transparencies. Hope-
fully, it will help you take a deeper dive into the subjects and enhance your understanding.

The Decision Theory (or: Decision Analysis) course is a 7.5 ECTS course, which means that
an average student is expected to require 210 hours of work to complete and pass the course.

The course is graded based on a practical assignment and a written exam. You are expected
to have knowledge of, insight in, and, most importantly, understanding of the different course
subjects. This also means that you have practiced with applying different procedures and
algorithms and can correctly do so in little time. The exercises in this study manual can help
you practice.

Ideal textbooks do not exist. We have not been able to find a textbook that covers all subjects
that we would like to treat at the required level and in an understandable way. We therefore
chose to base this course on a textbook which covers a lot of subjects, is easy to read, and
includes numerous exercises. Several versions of the book are available ([C96] with the green
cover and [CR01], or its 2004 update, with the blue cover), which are the same as far as the
theoretical contents relevant for this course is concerned; the 2001 edition, however, includes
software (updated in 2004) and a number of exercises using the software. Unfortunately, the
theory in the book is rather superficial, which is why we have supplemented the course book
with transparencies. The transparencies provide more details about the subjects presented in
the book, and cover those subjects not discussed in the book. The transparencies therefore
consist of essential supplementary material.

This manual lists, for each of the subjects treated in the course, the corresponding sections
in the textbook, relevant exercises1 , and sometimes a number of questions that can help to
guide your self-study. Note that each chapter in the textbook includes a summary, a number
of problems, questions and case studies, as well as a list of references for those interested in
further reading.

The exercises listed in this manual enable you to practice with what you have learned and give
you an idea of what you should be able to do if you understand the material. The exercises
mainly cover the application of methods and require less insight in the course material than
will be required during the exam. The listed questions and exercises should give you some
indication of how you are doing thus far.

The last section of this manual contains answers to selected numerical exercises (for the
starred exercises, answers can be found between Appendix F and the Author index of the
textbook).

1
A large number of exercises was constructed or collected by L.C. van der Gaag in 2003.

2
1 Introduction

Decision theory can be considered to be a branch of mathematics, providing a precise and

systematic study of the formal or abstract properties of decision-making scenarios. Tradition-
ally, decision theory, as opposed to for example game theory, considers only the decisions of
a single individual.

The aim of this course is to acquire knowledge and understanding of different models and
decision analysis techniques for (rational) decision making with and without uncertainty or
risk. Roughly, the following subjects are discussed:

decision criteria
decision trees
utility theory
multi attribute decision making
analyses of uncertainties

The first part of the course focuses on how to actually perform a decision analysis for a person
or company facing a difficult decision. The remainder of the course is of a more theoretical
nature and focuses on the theory behind different ingredients of decision theory. The former
is discussed at length in the textbook; the latter is only touched upon briefly by the book, so
we mostly rely on material on the transparencies.

When considering what the best decision, or sequence of decisions, would be in a given
situation, we first of all need to identify what the actual decision problem or context is (for
example: where to go for lunch). Each decision comes with a number of possible mutually
exclusive options or alternatives to choose between; from among the alternatives, the decision
maker can make one and only one choice. Decisions may depend on or result in the resolution
of uncertain events. The different things that can happen as a result of resolving an uncertain
event, typically beyond the control of the decision maker, are called outcomes of the event.
An uncertain event is relevant to the problem at hand if its outcome has an impact on at least
one of your (fundamental) objectives. Objectives are the reasons for making the decision, that
is, they describe the goal(s) we want to achieve (for example: we want to still our appetite
and we dont want to spend too much money). Objectives are measured by attributes (for
example: money spent in dollars); the two terms are, however, often used interchangeably.
Finally, what happens with respect to each objective after the last decision is made and all
uncertain events are resolved is called the consequence of your decision(s). Consequences
may be valued differently by different people; the value of a consequence can be specified by
a monetary reward, or more generally, by a utility. This basically summarises the different
elements of a decision problem.

To read/study:
Transparencies: Introduction
The M.C. Airport problem
Textbook: Ch. 1 all, except in [CR01]: Where does the
software fit in?
Ch. 2 all except The time value of money
Ch. 3 up to and including Getting the decision context
right(pages 41 50 in [C96], or 43 52 [CR01])

3
Different decision analysis techniques exist, depending on the number of (conflicting) objec-
tives or attributes you want to consider, and depending on whether or not you want to take
the uncertainty about or risk involved in the consequences of your decision into account.
If consequences are uncertain, but no probabilities are assigned to them, then the decision
situation is called decision under uncertainty. If probabilities are assigned, then the situation
is called decision under risk. This is a basic distinction in decision theory, although the terms
are used inconsistently. Examples of techniques for the different combinations are summarised
in the following table and are treated, to some extent, in this course.
Certainty Uncertainty Risk
1 - Maximin - Bayes criterion
attribute - Maximax - Dominance (prob.)
- Minimax regret - Utility theory (UT)
- Dominance (det.)
n2 - Analytic Hierarchy - Multi-attribute UT
attributes Process (AHP) (MAUT)

Questions

Why is it important to structure a decision problem?

What is the difference between a fundamental and a means objective? And why is this
difference important?

What is the difference between a chance variable and a decision variable?

What are general properties of the values of a chance variable?

Can a chance variable have more than one probability distribution associated with it?

What would a reward function look like if we have more than one decision variable and
more than one chance variable?

Exercises

1.1 (Adapted from [W94]) Pizza King and Noble Greek are two competing restaurants.
Each must decide, simultaneously, whether to undertake small, medium, or large advertising
campaigns. Pizza King believes that it is equally likely that Noble Greek will undertake a
small, a medium, or a large advertising campaign. Given the actions chosen by each restau-
rant, Pizza Kings profits are as follows:

Pizza King Noble Greek campaign

campaign small medium large
small e 6000 e 5000 e 2000
medium e 5000 e 6000 e 1000
large e 9000 e 6000 e0

a. Phrase Pizza Kings decision problem.

b. Specify the various elements of Pizza Kings decision problem.

4
Suppose that Pizza King and Noble Greek stop advertising but must decide upon the price
they will charge for each pizza sold. Pizza King believes that Noble Greeks price is a chance
variable P with the following probability distribution:
Pr(P = e 6) = 0.25; Pr(P = e 8) = 0.50; Pr(P = e 10) = 0.25
If Pizza King charges a price p1 and Noble Greek charges a price p2 , Pizza King will sell
100 + 25 (p2 p1 ) pizzas. It costs Pizza King e 4 to make a pizza. Pizza King is considering
charging e 5, e 6, e 7, e 8, or e 9 for a pizza.

c. Specify the various elements of Pizza Kings new decision problem.

1.2 Solve the following exercises from the textbook (starting on page 90 in [C96], or page
98 in [CR01]):

a. Exercise 3.1
b. Exercise 3.15
c. Exercise 3.18

2 Decision criteria

This subject is concerned with single attribute decision making under uncertainty. That is,
there is uncertainty with respect to the consequences of our decision, but this uncertainty
is considered to be independent of the consequence; basically this amounts to assuming a
uniform distribution.

To read/study:
Transp.: Decision criteria
Textbook: Ch. 4 Risk profiles through Dominance: An alternative. . .
(pages 118 127 in [C96], or 128 137 in [CR01])

For those interested in further reading, we suggest chapter 13 of [W94]:

W.L. Winston Operations Research: Applications an Algorithms, Duxbury Press, 1994.

Questions

What is the difference between deterministic dominance and probabilistic dominance?

What is the difference between deterministic dominance as defined in the transparencies
and as defined in the text book?
What is the difference between probabilistic (stochastic, or statistical) dominance as defined
in the transparencies and as defined in the text book?
What is the use of the different decision criteria?
Why are dominance properties interesting?
What is a risk-profile?
For a simple decision problem, we can draw risk-profiles per decision alternative. For what
do we construct a risk-profile in more complex decision problems?

5
Exercises

2.1 (Adapted from [W94]) Reconsider Pizza King and Noble Greek from Exercise 1.1

a. For the maximin, the maximax, and the minimax regret decision criteria, determine
Pizza Kings best choice of advertising campaign.

b. For the maximin, the maximax, and the minimax regret decision criteria, determine the
price per pizza that Pizza King should charge.

c. Which price should Pizza King charge based on Bayes decision criterion?

2.2 Reconsider Aries Straatnieuws problem from the first lectures. Arie purchases copies
of the Straatnieuws at e 1.00 each and sells them at e 1.50. The daily demand for copies is
between 16 and 20, where each number is equally likely.

a. Show that the decision to purchase 21 copies is deterministically dominated by the

decision to purchase 20 copies.

b. Is the decision to purchase 21 copies also probabilistically dominated by the decision to

purchase 20 copies?

c. Now suppose that the probability distribution over the daily demand for copies C is:

Pr(C = 16) = 0.1; Pr(C = 17) = 0.2; Pr(C = 18) = 0.6;

Pr(C = 19) = 0.05; Pr(C = 20) = 0.05

1. Is the decision to purchase 21 copies still deterministically dominated by the deci-

sion to purchase 20 copies?
2. Is the decision to purchase 21 copies probabilistically dominated by the decision
to purchase 20 copies?

2.3 Solve the following exercises from the textbook (starting on page 141 in [C96], or page
159 in [CR01]):

a. Exercise 4.2

b. Exercise 4.9

c. Exercise 4.11

d. Exercise 4.16

6
3 Decision trees

On how to model a decision problem under risk in a decision tree and how to determine the
best decision (foldback-analysis).

To read/study:
Transp.: Decision trees
Evaluating decision trees
Textbook: Ch. 3 Structuring decisions: decision trees through Decision details: . . .
(pages 67 76 in [C96], or 69 78 in [CR01])
Ch. 4 up to and including Decision trees and expected monetary value
(pages 101 108 in [C96], or 111 118 in [CR01])

Questions

What is the difference between a chance tree and a decision tree?

What is the use of a decision tree?

How is a chance variable represented in a chance/decision tree?

How is a decision variable represented in a decision tree?

What does inverting a chance tree represent in probability theory?

What is the purpose of inverting a chance tree in practice?

Is it possible to invert a decision tree?

What is a scenario in a decision tree?

What is a strategy in a decision tree? Can you have strategies in chance trees?

How do you determine the strategies in a decision tree?

If the value of a chance node is known before a decision is to be made, is it allowed to

model the chance node behind the decision node in a decision tree?

How do you reduce a decision tree?

What is the idea behind the fold-back algorithm?

Does the fold-back algorithm require the use of a decision tree?

Does the fold-back algorithm require the use of a reduced decision tree?

7
Exercises

3.1 For a medical decision problem, the following chance tree has been constructed:

Pr(s | m) = 0.52

Pr(m) = 0.51

Pr(s | m) = 0.48

Pr(s | m) = 0.09

Pr(m) = 0.49

Pr(s | m) = 0.91

In the tree, the chance variable M denotes whether or not a patient is suffering from a
myocardial infarction. The chance variable S indicates the SGOT-level found in the patient;
the value of the chance variable S can be determined by means of a blood test. Invert the
chance tree.

3.2 (Adapted from [W94]) Oilco must determine whether or not to drill for oil in the
South China Sea. It costs e 100 000 to drill, and if oil is found, the value is estimated to
be e 600 000. At present, Oilco believes there is a 45% chance that the field contains oil.
Before drilling, Oilco can hire (at e 10 000) a geologist to obtain more information about the
likelihood that the field will contain oil. There is a 50% chance that the geologist will issue
a favourable report and a 50% chance of an unfavourable report. Given a favourable report,
there is an 80% chance that the field indeed contains oil. Given an unfavourable report, there
is a 10% chance that the field contains oil nonetheless.

a. Specify the various elements of Oilcos decision problem.

b. Construct a decision tree for Oilcos decision problem.

c. Perform a fold-back analysis of your decision tree.

3.3 (Adapted from [W94]) Farmer Jones must determine whether to plant corn or wheat.
If he plants corn and the weather is warm, he earns e 8 000; if he plants corn and the weather
is cold, he earns e 5 000. If he plants wheat and the weather is warm, he earns e 7 000; if
he plants wheat and the weather is cold, he earns e 6 500. In the past, 40% of all years have
been cold and 60% have been warm. Before planting, Jones can pay e 600 for an expert
weather forecast. If the year will actually be cold, there is a 90% chance that the forecaster
will predict a cold year. If the year is actually going to be warm, there is a 80% chance that
the forecaster will predict a warm year.

a. Specify the various elements of Jones decision problem.

b. Construct a decision tree for Jones decision problem.

c. Solve Jones decision problem.

8
3.4 Solve the following exercises from the textbook (starting on page 141 in [C96], or page
159 in [CR01]):

a. Exercise 4.3

b. Exercise 4.4*

c. Exercise 4.6*

d. Exercise 4.7

e. Exercise 4.8*

f. Exercise 4.14

3.5 Solve the following exercises from the textbook (starting on page 251 in [C96], or page
281 in [CR01]):

a. Exercise 7.6*

b. Exercise 7.22*

c. Exercise 7.34

4 Probability theory and assessment

The following subjects are roughly discussed: probability basics (you should already be famil-
iar with this), subjective probability assessments, heuristics, biases, and continuous variables.
Probabilities and their assessment are important for quantifying a decision tree.

To read/study:
Transp.: Probability theory and assessment
Textbook: Ch. 7 all except Decision-analysis software . . . and Optional sections
Ch. 8 all except Coherence and the Dutch book, and
in [CR01]: Constructing distributions using RISKView

Questions

What properties does a statistical/stochastic/random variable adhere to?

What are the properties of a probability distribution?

What is a conditional probability (distribution) and what notation is used to describe one?

What is a cumulative probability distribution and what notation(s) is/are used to describe
one?

What is Bayes rule/theorem?

What is (probabilistic/stochastic/statistical) independence?

Why is it important to know about subjective probability assessment?

9
 What is the difference between a betting model and a lottery model?
How can you construct a cumulative probability distribution?
What is the use of the Pearson-Tukey and bracket-median methods?
What types of distribution are best approximated by the Pearson-Tukey and bracket-
median methods?

Exercises

4.1 Suppose that an infectious disease expert gave you the following advice about various
treatment alternatives for one of your patients:

The likelihood that the infection will respond to antibiotic A is greater than the likeli-
hood that the infection will respond to antibiotic B.
The likelihood that the infection will not respond to antibiotic C is greater than the
likelihood that the infection will not respond to antibiotic B.
The likelihood that the infection will respond to antibiotic C is greater than the likeli-
hood that the infection will respond to antibiotic A.

a. Is this advice sound (i.e. no contradictions)?

The expert further provides you with the following probability assessments:

The probability that the patient will die sometime during the next seven days is 0.50.
The probability that the patient will die after day 7, yet on or before day 14 is 0.40.
The probability that the patient will survive beyond day 14 is 0.30.

b. Are these assessment coherent?

4.2
a. A neurologist has been studying the relationship of stroke to elevated blood pressure
in elderly patients and has obtained the following statistical information from previous
investigations and through a review of the literature on cerebral vascular incidents: for
persons over age 70, 10 percent will have a stroke within the next 5 years; of all patients
aged 70+ who have had a stroke, 40 percent demonstrate elevated blood pressure; 20%
of people over 70 have elevated blood pressure. A patient aged 70+ in the neurologists
practice has elevated blood pressure. What is the probability that this patient will have
a stroke within the next 5 years?
b. Suppose that the neurologist has further assessed the probability that within the next
5 years neither a stroke nor a myocardial infarction will occur in another patient aged
85, at 0.10. He has assessed the probability that only a stroke will occur at 0.30 and the
probability that only a myocardial infarction will occur at 0.20. Do you have enough
information to determine the probability that both incidents will occur in this patient?
c. Can the probability that both a stroke and a myocardial infarction will occur in a
patient within the next 5 years be greater than the probability that a stroke will occur
in a patient in whom a myocardial infarction has occurred?

10
4.3 Solve the following exercises from the textbook (starting on page 254 in [C96], or page
284 in [CR01]):

a. Exercise 7.21
b. Exercise 7.25
c. Exercise 7.26

4.4 Solve the following exercises from the textbook (starting on page 299 in [C96], or page
336 in [CR01]):

a. Exercise 8.3
b. Exercise 8.15c and 8.15d
c. Exercise 8.13*
d. Exercise 8.27

5 Introduction to utility theory and assessment

To solve a decision problem under risk it is important to know how the decision maker values
different outcomes or consequences. The value of a consequence is given by its utility. Here
we cover some basics from utility theory (Lotteries and the utility axioms) and discuss how
utilities can be assessed from decision makers.

To read/study:
Transp.: Introduction to utility theory and assessment
Assessing utilities
Textbook: Ch. 3 Defining measurement scales for fundamental objectives
(pages 77 82 in [C96], or 79 83 in [CR01])
Ch. 13 Utility function assessment
(pages 473 477 in [C96], or 539 543 in [CR01])
Some caveats
(page 488 in [C96], or 556 in [CR01])
Ch. 14

For those interested in further reading, we suggest chapter 4 of [KR93]:

R.L. Keeney & H. Raiffa (1993). Decisions with Multiple Objectives, Cambridge University
Press.

Questions

What is the difference between a utility and a reward?

What is the relation between a consequence and a utility?
Why is the Reduction of compound uncertain events (see Clemen) not really an axiom of
utility theory?
How does the main theorem of utility theory follow from the axioms?

11
Exercises

5.1 (Adapted from [W94]) Consider the following four lotteries:

0.10
e 1 000 000
1.0
L1 : e 1 000 000 L3 :
0.90
e0

0.50 0.05
e 3 000 000 e 3 000 000
L2 : L4 :
0.50 0.95
e0 e0

a. Most people prefer L1 to L2 and L4 to L3 . Explain why.

b. Suppose that Peter subscribes to the Von Neumann-Morgenstern axioms and prefers
L1 to L2 . Show that he must also prefer L3 to L4 .

5.2 (Adapted from [W94]) Patricia is trying to determine which of two courses to take. If
she would take the decision analysis course, she believes that she has a 10% chance of receiving
an A, a 40% chance for a B, and a 50% chance for a C. If she would take the statistics course,
she believes to have a 70% chance for a B, a 25% chance for a C, and a 5% chance for a D.
Patricia is indifferent between the following two lotteries:
0.25
A
1.0
C and
0.75
D
She is also indifferent between the following two lotteries:
0.70
A
1.0
B and
0.30
D

a. Specify a preference ordering over the four possible grades that respects Patricias in-
differences. Are there any other orderings that satisfy Patricias indifferences?
b. Construct a utility function over the four possible grades, given the preference ordering
that you have specified in part a. of this exercise.
c. If Patricia wants to take the course that maximises the expected utility of her final
grade, given the utility function that you have constructed in part b. of this exercise,
which course should she take?

5.3 The Dutch company Multilever is going to invest e 1 000 000. Two investments can
be made: in washing powder and in canned soup. If the e 1 000 000 is invested in washing
powder, Multilever is certain to end up with a net profit of e 1 296 000. If they invest in
canned soup, there is a 0.75 chance that they will end up with a net profit of e 400 000 and
a 0.25 chance that they will end up with a net profit ofe 10 000 000. For e x profit, their

utility function is given by u(x) = x. Should Multilever invest in washing powder or in
canned soup?

12
5.4 Reconsider the gangrene problem from the transparencies, with the four consequences:
A = amputated above knee
B = amputated below knee
C = cured
D = death

a. Suppose that u(A) = 0.95, u(C) = 1.00, and u(D) = 0. For which range of u(B) would
delaying surgery be the preferred treatment alternative?

b. Now suppose that u(A) = 0.95, u(B) = 0.99, u(C) = 1.00, and u(D) = 0. Would this
patient prefer to lose his leg below the knee for certain or take a chance consisting of a
0.75 probability of total cure and a 0.25 probability of loss of the leg above the knee?

5.5 (Adapted from [W94]) Carly has been chosen to appear on Illusive Millionaire ! The
rules of the well-known show are as follows: there are four hidden cards. One says STOP and
the other three have euro amounts of 150 000, 200 000 and 1 000 000. Carly gets to choose a
card. If the card says STOP, she wins no money. At any time, she may decide to quit and
keep the largest amount of money that has appeared on any card she has chosen, or continue.
If she continues and chooses the STOP card, however, she wins no money. As a example, she
may first choose the e 150 000 card, then the e 200 000 card, and then she may decide to
quit and receive e 200 000.

a. If Carlys goal is to maximise expected payoff, which strategy should she follow?

b. Now suppose that Carlys utility function for cash satisfies u(0) = 0, u(e 40 000) = 0.25,
u(e 120 000) = 0.50, u(e 400 000) = 0.75, and u(e 1 000 000) = 1.0. Determine a
strategy for Carly that maximises her expected utility.

5.6 Solve the following exercises from the textbook (starting on page 521 in [C96], or page
589 in [CR01]):

a. Exercise 14.7

b. Exercise 14.16 a, b

6 Risk attitudes

Whether or not a decision maker is prepared to take some risks will influence his/her utility
function. Here we discuss how attitudes towards risk affect utility functions.

To read/study:
Transp.: Risk attitudes
Textbook: Ch. 2 The time value of money
(pages 28 31 in [C96], or 30 33 in [CR01])
Ch. 13 all except Risk tolerance & sensitivity analysis
(all pages except 480 482 in [C96], or 546 551 in [CR01])

13
Questions

What is the relation/difference between a consequence, a utility and an attribute?

What are the product and chain rules for computing derivatives?

What are the first and second derivatives of

a 3x
axn , , ln x, ex , 3x 5x2 ,
x+b x+1

What, in general, is the interpretation of a first and second derivative?

Consider a decision problem with monetary rewards. Under what conditions is the utility
of the expected reward the same as the expected utility? Does this have consequences for
a fold-back analysis?

Is the risk-premium a property of a utility function or of a lottery?

Is it possible to determine a risk-premium without knowing a utility function?

What is the relation between the risk-aversion function and risk tolerance (see Clemen).

Exercises

6.1 (Adapted from [W94]) Suppose my utility function for asset position x is given by
u(x) = ln x.

a. Am I risk-averse, risk-neutral, or risk-prone?

b. I now have e 20.000 and am considering the following two (unfortunate) lotteries:
L1 : with probability 1, I lose e 1000.
L2 : with probability 0.9, I gain e 0,
and with probability 0.1, I lose e 10,000
Determine which lottery I prefer. In addition, determine the risk premium of L2 .

6.2 (Adapted from [W94]) Show that a decision maker who has a linear utility function
will rank two lotteries according to their expected (monetary) value.

6.3 (Adapted from

p [W94]) A decision maker has a utility function for monetary gains x
given by u(x) = (x + 10.000).

a. Is the decision maker risk-averse, risk-neutral, or risk-prone?

b. Give a lottery that models the decision makers status quo.

c. Show that the person is indifferent between the status quo and lottery L:
L: with probability 31 , he or she gains e 80.000
with probability 32 , he or she loses e 10.000

d. There is a 10% chance that a painting valued at e 10.000 will be stolen during the next
year.

14
 1. What is the risk-premium of this lottery?
2. What is the most (per year) that the decision maker would be willing to pay for
insurance covering the loss of the painting?
3. Explain the difference between your previous two answers.

6.4 (Adapted from [W94]) We now have e 5000 in assets and are given a choice between
investment 1 and investment 2. With investment 1, 80% of the time we increase our asset
position by e 295.000 and 20% of the time we increase our asset position by e 95.000. With
investment 2, 50% of the time we increase our asset position by e 595.000 and 50% of the time
we increase our asset position by e 5.000. Our utility function for final asset position x is u(x).
We are given the following values for u(x): u(0) = 0, u(90.000) = 0.30, u(490.000) = 0.7,
u(640.000) = 0.80, u(810.000) = 0.90, u(1.000.000) = 1.

a. Are we risk-averse, risk-neutral, or risk-prone? Explain.

b. Will we prefer investment 1 or investment 2?

c. Give the risk premium of both investments.

6.5 Solve two of the following exercises: Show that

a. u(x) = x, x > 0, describes risk neutrality

b. u(x) = e0.2x , x > 0, describes constant risk aversion

c. u(x) = 1 + 2x x2 , x < 1, describes increasing risk aversion

d. u(x) = log(x + 30), x > 30, describes decreasing risk aversion

e. u(x) = e0.4x , x > 0, describes constant risk proneness

f. u(x) = x2 , x > 0, describes decreasing risk proneness

2
g. u(x) = ex , x > 1, describes increasing risk proneness

6.6 Solve the following exercises from the textbook (starting on page 489 in [C96], or page
557 in [CR01]):

a. Exercise 13.1

b. Exercise 13.3

c. Exercise 13.4

d. Exercise 13.6*

e. Exercise 13.7*

f. Exercise 13.9

g. Exercise 13.10

h. Exercise 13.11

15
 i. Exercise 13.12

j. Exercise 13.14

k. Exercise 13.17

l. Exercise 13.19

m. Exercise 13.20

n. Exercise 13.21

o. Exercise 13.26

p. Exercise 13.29

7 Multi-attribute utility theory

Up till now we have learned the structure of a decision tree, how to assess probabilities and
utilities and to be aware of risk-attitudes and how they affect a persons utility function. In
the process, we considered only very simple decision problems.

Often, when solving a decision problem, there is more than one goal to obtain. These goals are
typically conflicting. Multi-attribute utility theory is concerned with modelling such two- or
more-dimensional problems. To simplify this multi-dimensional problem, often independence
assumptions are made and attributes, or combinations thereof, are weighed separately.

To read/study:
Transp.: Multi-attribute utility theory
Multi-attribute utilities in practice
Textbook: Ch. 4 pages 127 135 in [C96], or
pages 137 146 in [CR01]
Ch. 15
Ch. 16

For those interested in further reading about multi-attribute utility theory, we recommend
[KR93] (chapter 6) and [W94] (chapter 14).

Questions

What is the relation between consequences and attributes?

What is the relation between attributes and utilities?

What are the domain and range of a utility function?

When are you allowed to use weight-assessment methods such as Pricing out and Swing
weighting?

Why is decomposed assessment only possible under certain assumptions?

Argue why Pricing out and Swing weighting cannot be used to assess the scaling constants
of the Additive utility function - Ia. How should these scalingconstants be assessed?

16
 Argue why utility independence implies preferential independence.

In practice, is additive independence likely to hold? What about utility independence?

What about preferential independence?

Additive independence can be established on its own or can be derived from mutual utility
independence. Explain the difference in properties that should hold to conclude additive
independence.

What is the use of an iso-preference curve?

What is the difference between a utility function and an iso-preference curve.

Is utility independence a requirement for the use of an iso-preference curve?

Do the methods for decision making under uncertainty used in medical practice build on
the same or different principles as/than discussed in this course?

What is the relation between a survival curve and life-expectancy? Can a survival curve
be interpreted as a utility function? Can life-expectancy be interpreted as a utility value?

Why do we need the DEALE method?

What is the difference between a QALE and a QALY and what is their relation to a
multi-attribute utility?

Exercises

7.1 Consider two attributes X and Y with values x1  . . .  xn , n 1, and y1  . . . 

ym , m 1. Prove that if X is utility independent of Y , then for each value yl of Y , there
exist real functions gl > 0 and hl , such that for all xi and yj we have that

u(xi , yj ) = gl (yj ) u(xi , yl ) + hl (yj )

7.2 (Adapted from [W94]) Public service Indiana (PSI) is considering two sites for a
nuclear power plant. The following two attributes will influence its determination about
where to build the plant:

Attribute C Cost of the plant (in millions of euros)

Attribute A Acres of land damaged by building the plant

Assume that PSIs multiattribute utility function for any value c of C and any value a of A
is given by

u(c, a) = 0.70 uC (c) + 0.20 uA (a) + 0.10 uC (c) uA (a), where

uC (c) = 0.1 + e0.1c and uA (a) = 2.5 2.5 e0.0006a0.48

Two locations for the power plant are under consideration:
Location 1 is equivalent to the lottery [0.50, (50, 300); 0.50, (30, 400)];
Location 2 is equivalent to the lottery [0.50, (60, 200); 0.50, (15, 600)].
Which location should be chosen?

17
7.3 (Adapted from [W94]) National Express Carriers is interested in two attributes :

Attribute C The average cost of delivering a letter (known to be between e 1 and e 5)

Attribute T Percentage of all letters reaching their destination on time (known to be

between 70 % and 100 %)

a. Would Nationals multiattribute utility function exhibit mutual utility independence?

b. Would Nationals utility function be additive?

c. Assume that attributes C and T exhibit mutual utility independence. Suppose National
expresses the following indifferences:

[1.00, (e 1, 70%)] [0.30, (e 1, 100%); 0.70, (e 5, 70%)]

[1.00, (e 5, 100%)] [0.50, (e 1, 100%); 0.50, (e 5, 70%)]

Find Nationals multiattribute utility function. Express Nationals multiattribute utility
function in terms of uC (C) and uT (T ).

7.4 (Adapted from [W94]) A blood bank must determine at the beginning of each week
how many pints of blood should be ordered. Any blood left over at the end of the week spoils
(it is outdated). For the blood bank, two attributes of interest are as follows:

Attribute S Number of pints of blood by which ordered blood falls short of the weeks
demand (the weekly shortage). The weekly shortage is known to be always between 0
and 10 pints.

Attribute R Number of pints of blood that are outdated (known to be always between
0 and 10 pints).

Assume that attributes S and R exhibit mutual utility independence.

Suppose the blood bank expresses the following indifferences:

[1.00, (0, 10)] [0.40, (0, 0); 0.60, (10, 10)]

[1.00, (10, 0)] [0.50, (0, 0); 0.50, (10, 10)]

Let s denote any value of attribute S, and r any value of attribute R and suppose that

r2
uS (s) = 0.58 e(1 10 ) 0.58
s
and uR (r) = 1
100

a. Determine the blood banks multiattribute utility function.

b. Suppose that each week there is a 0.5 chance that the demand for blood will be 25 pints
and a 0.5 chance it will be 35 pints. Would the blood bank be better off ordering 28
pints, 30 pints, or 32 pints?

18
7.5 (Adapted from [W94]) Gotham City is trying to determine how many ambulances
it should have and how to staff them. Each ambulance may be staffed with paramedics
or emergency medical technicians. Paramedics are considered to provide better service and
are paid higher salaries. Budgetary limitations have forced the city to choose between the
following two alternatives:

Alternative 1 Four ambulances, two staffed with emergency medical technicians and two
staffed with paramedics

Alternative 2 Three ambulances,all staffed with paramedics

The city fathers believe that the following two attributes determine the citys satisfaction
with ambulance service:

Attribute T Time until an ambulance reaches a patient

Attribute C Percentage of ambulance calls handled by paramedics.

Assume that Gotham Citys multiattribute utility function u(T, C) exhibits mutual utility
independence and that for any value t of T and any value c of C

t2 c2
uT (t) = 1 and uC (c) =
900 10, 000
The time for an ambulance to reach a patient is always between 0 and 30 minutes. The city
fathers express the following indifferences

[1.00, (30, 100%)] [0.40, (0, 100%); 0.60, (30, 0%)]

[1.00, (0, 0%)] [0.80, (0, 100%); 0.20, (30, 0%)]

Assume that if an ambulance is available when a call comes in, then the ambulance will arrive
in 5 minutes; if an ambulance is not available when a call comes in, it will arrive in 20 minutes.
With three ambulances, one will be immediately available 60% of the time, and with four
ambulances, one will be immediately available 80% of the time.

a. Determine the city fathers multiattribute utility function.

b. Which alternative should they choose?

7.6 (Adapted from [W94]) Consider attributes X and Y . Let x1 , x2 be values of X

such that x1 x2 and y1 , y2 be values of Y such that y1 y2 . Consider the following four
consequences:
A = (x1 , y1 ); B = (x1 , y2 ); C = (x2 , y2 ); D = (x2 , y1 )
Assume that a decision makers utility function exhibits mutual utility independence. Con-
sider the following two lotteries:

L1 = [0.50, A; 0.50, C] and L2 = [0.50, B; 0.50, D]

a. Show that if kXY > 0, then L1  L2 .

b. Show that if kXY < 0, then L2  L1 .

19
 c. Show that if the decision maker exhibits additive independence (kXY = 0), then L1
L2 .

d. Let attribute E be the performance of an army on its eastern front, and attribute W be
the performance of the army on the western front. A high level of an attribute means
that the army did well, and a low level of an attribute means that the army did poorly.
Suppose that the army will suffer defeat if it performs poorly on either front. If these
attributes exhibit mutual utility independence, what would be the sign of kEW ?

e. General Motors has domestic and international divisions. Let attribute D be the profits
in the domestic division and attribute I the profits in the international division. Suppose
General Motors is reasonably happy if at least one division has a good year, but is very
unhappy if both divisions have a bad year. If these attributes exhibit mutual utility
independence, what would be the sign of kDI ?

7.7 Consider attributes X and Y . Prove that additive independence of X and Y implies
mutual utility independence of X and Y .

7.8 Consider the following additive utility function:

u(X, Y ) = kX uX (X) + kY uY (Y )

over attributes X and Y with values x1  . . .  xn and y1  . . .  ym , respectively. Assume

that u(X, Y ) is normalised such that u(x1 , y1 ) = 0 and u(xn , ym ) = 1, uX (X) is normalised
such that uX (x1 ) = 0 and uX (xn ) = 1, and uY (Y ) is normalised such that uY (y1 ) = 0 and
uY (yn ) = 1.

Show that kX and kY are positive scaling constants summing to 1. (Hint: exploit that also
u(X, Y ) = u(X, y1 ) + u(x1 , Y ) with u(x1 , y1 ) = 0.)

7.9 Explain the differences and similarities between mutual utility independence as defined
for two attributes and mutual utility independence as defined for n (n > 2) attributes.

7.10 Let X and Y be two attributes with values x1  . . .  xs and y1  . . .  yt , respec-

tively. Consider the two-attribute multilinear utility function (form IIa) as defined on the
transparencies.

a. Show that this function is a special case of the n-attribute multilinear utility function
defined on the transparencies;

b. Show that this function is a special case of the n-attribute multiplicative utility function
defined on the transparencies.

Now consider the two-attribute additive utility function (form Ia) as defined on the trans-
parencies.

c. Show that this function is a special case of the n-attribute additive utility function as
defined on the transparencies.

20
7.11 Consider Exercise 16.15 from the textbook (page 603 in [C96], or page 671 in [CR01]).

a. Solve the exercise;

b. Show that the other two solutions that follow from the given equation are invalid (see
the constraints for the n-attribute multiplicative utility function on the transparencies).

7.12

a. A surgeon considers an operative procedure for one of his patients. For this particular
patient, he assesses the life expectancy without treatment at 5 years. If the patient
would survive the surgery, his life expectancy would have increased to 7 years. The
surgical procedure, however, involves a 10% risk of operative death. Now suppose that
the surgeon decides, in close consultation with his patient, to perform the procedure.
What is the patients life expectancy?

b. For one of his other patients, the surgeon considers operative removal of a primary
tumour. This surgical procedure involves a 20% risk of operative death. If the patient
would survive the surgery, she would stand a 50% chance of dying within the first year
following the surgery. If she lives through the first year, she has a 20% chance of dying
within the second year. What chance does she have of surviving beyond the second year
following the surgery?

c. For a third patient, the surgeon has to decide between three different treatment alter-
natives:

treatment A has associated an annual death rate of 10%;

treatment B has associated a 5-year survival of 50%;
treatment C has associated a life expectancy of 5.0 years.

Which treatment should he choose if he wishes to maximise the patients life expectancy?

7.13

a. Suppose that a recent study of prostate cancer reports a 5-year survival of 25% for
50-year-old men. Assume that the age-specific hazard for these men is 0.009. Compute
the life expectancy for a 55-year-old patient in whom recently prostate cancer has been
established. Assume that the age-specific hazard for men 55 years of age is 0.014.

b. Suppose that a recent study of surgical treatment of colon cancer reports a 5-year
survival of 40% with treatment and of 30% without treatment. The study included 55-
year-old men, with an age-specific hazard of 0.014. Compute the life expectancy, with
and without treatment, for a 65-year-old man suffering from colon cancer. Assume that
the age-specific hazard for 65-year-old men is 0.023.

c. Reconsider the 65-year-old patient from part b. and suppose that the attending physician
considers surgical treatment of the cancer. The physician is concerned, however, that
his patient may face a significant risk of operative death. How high would the risk have
to be for the patient in order for surgery not to increase life expectancy?

21
7.14 Suppose that a patient faces the decision problem captured by the following decision
tree:

operative death
p = 0.01
surgery
live one year
p = 0.20 without symptoms
live 5 years
p = 0.99 p = 0.30 with symptoms
live 5 years
p = 0.50 without symptoms

medication live 5 years

with symptoms
The patient has a cancer that will cause his death in about 5 years time. He can choose either
surgical treatment or medical treatment to control the symptoms of his cancer:

the medical treatment is risk-free; however, it is only partially effective, so that the
patient will continue to have some of the symptoms;
the mortality for the operative procedure is 0.01; the probability that the surgical
treatment will leave the patient free of symptoms is 0.50; the probability that the
patients life will be shortened to one year by the procedure is 0.20; there is a 30%
chance that the operation will not relieve the symptoms.

The patients utilities for a life without symptoms are

number of
symptom-free years utility
0 0
1 0.60
2 0.80
3 0.90
4 0.95
5 1.00

To the patient, living with the symptoms is equivalent to reducing the length of his life by
20%. For example, to the patient, living 5 years with symptoms is equivalent to living 4 years
without symptoms. Should the patient choose medical treatment, or surgery?

7.15 Assume that a patient who is currently undergoing dialysis, faces a life expectancy of
10 years. If he were to undergo a kidney transplant, he would have a 20% chance of dying
during the surgery. If he would survive the operation, he would face a 40% chance of losing
the transplanted kidney within one year; beyond the first year, he will no longer run the risk
of losing the kidney. If after the operation, he would lose the kidney, then he would have to
undergo dialysis for the rest of his life; his total life expectancy (including the first year) then
is 10 years. If he retains the kidney, his life expectancy after the first year is 12.5 years.

The health status of the patient during the remaining years of his life will fall in one of four
quality levels:

22
 T1 : the patient is carrying a successful transplant and has returned to his prior level of
activity;

T2 : the patient is carrying a successful transplant but there are serious limitations on his
activity;

D1 : the patient is undergoing dialysis and is near his prior level of activity;

D2 : the patient is undergoing dialysis and there are serious limitations on his activity.

A patient carrying a transplant has a 60% chance of having a health status of quality T1 and
a 40% chance of having a health status of quality T2 ; a patient undergoing dialysis has a 50%
chance of having a health status of quality D1 and a 50% chance of having a health status of
quality D2 . The attending physician estimates that the patient would be indifferent between

ten years of life with a health status of quality T2 and six years of life of quality T1 ;

eight years of life with a health status of quality D2 and four years of life of quality D1 ;

ten years of life with a health status of quality D1 and eight years of life of quality T1 .

Moreover, one year of quality-adjusted life equals a year spent in quality T1 .

At the time the decision is made:

a. what is the quality-adjusted life expectancy for the patient if he were to keep on under-
going dialysis?

b. what is the quality-adjusted life expectancy for the patient if he were to survive a
transplant operation and retain the kidney?

7.16 Solve the following exercises from the textbook:

a. Exercise 4.12 (page 142 in [C96], or page 161 in [CR01])

b. Exercise 4.18

c. Exercise 15.3 (page 561 in [C96], or page 629 in [CR01])

d. Exercise 15.6

e. Exercise 15.9*

f. Exercise 15.16

g. Exercise 15.26

23
7.17 Solve the following exercises from the textbook (starting on page 599 in [C96], or page
667 in [CR01]):

a. Exercise 16.1 16.4

b. Exercise 16.5

c. Exercise 16.6*

d. Exercise 16.7

e. Exercise 16.8

f. Exercise 16.13

8 Analyses of uncertainties

The utilities and probabilities that are specified in a decision tree are most often assessed
by the decision maker and consequently inaccurate. How sensitive is the outcome of your
decisionproblem to these inaccuracies? And suppose there is some way to obtain (more)
certainty about parts of the tree, that is about the problem itself, is that worth the try? We
discuss different algorithms to answer these questions.

To read/study:
Transp.: Sensitivity analysis
The value of information
Textbook: Ch. 5 all except Sensitivity analysis by computer/TopRank, PrecisionTree
and Optional sections
Ch. 12 all except Value of Information and experts
and Calculating EVPI. . .

Questions

What does a sensitivity analysis amount to?

What does a threshold analysis amount to?

What is the relationship between the tornado diagrams (see textbook) and sensitivity
analysis as discussed in the transparencies?

Can the general form of the sensitivity function as stated in the transparencies be applied
for any type of parameter in your decision problem? (are you sure?!)

Can the general form of the sensitivity function be applied to non-reduced decision trees?

How would you perform a sensitivity analysis in a non-reduced decision tree?

What is the purpose of a value of information analysis?

Consider a decision problem with decision node D1 , chance node C1 , probability distribu-
tions Pr(C | D) and a set of rewards R. We perform a value of information analysis. Which
of the mentioned decision elements will change in the more informed problem?

24
 Can value of information analyses only be performed for problems with monetary rewards,
or for any decision problem with utilities as consequences?
P
The entropy of a (discrete) probability distribution is given by i pi ln pi and is a
measure of uncertainty in the distribution. Explain why gathering information reduces
entropy. What is the entropy corresponding to perfect information?
Why do we consider the inverted decision problem in a value of perfect information analysis?
Is it always possible to do a value of perfect information analysis?
Why do the probabilities under consideration remain unchanged in a value of perfect in-
formation analysis?
Why is EVPI (expected value of perfect information) an upperbound on EVI (expected
value of information2 )?
Why is EVI always greater than or equal to zero?

Exercises

8.1 Reconsider the decision tree for the laryngeal cancer problem, from the transparencies
(the one with qalys).

a. Conduct a one-way sensitivity analysis for the two strategies represented in the tree, of
the probability of recurrence of the primary tumour after radiotherapy. Try to explain
the signs of the derivatives of the two sensitivity functions.
b. Conduct a threshold analysis of the probability of recurrence of the primary tumour
after radiotherapy. What do you conclude from the results of the analysis?

8.2 Reconsider the decision tree for the laryngeal cancer problem, from the transparencies
(the one with qalys).

a. Conduct a one-way sensitivity analysis for the two strategies represented in the tree, of
the utility of artificial speech.
b. What do you think is a reasonable plausible interval for the utility of artificial speech?

8.3 Reconsider the decision tree for the laryngeal cancer problem, from the transparencies
(the one with qalys).

a. Conduct a two-way sensitivity analysis for the two strategies represented in the tree,
of the probability of recurrence of the primary tumour after a laryngectomy and the
utility of hoarseness. Do the two parameters under study show an interaction effect?
b. Conduct a two-way sensitivity analysis for the two strategies represented in the tree, of
the probability of recurrence of the primary tumour after radiotherapy and the utility
of hoarseness. Do these two parameters show an interaction effect?
c. Conduct a threshold analysis of the two parameters mentioned in part a. of this exercise.
What do you conclude from the results of the analysis?
2
In the textbook EVI is called either EVSI (EV of sample information) or EVII (EV of imperfect informa-
tion)

25
8.4 (Adapted from [W94]) Greta has one die in her left hand and one in her right hand.
One die has six dots painted on each face, and the other has one dot painted on two of the
faces and six dots painted on each of the other four faces. Abdul is to pick a die (either
Gretas left hand or her right hand) and will receive e 10 for each dot painted on the die that
is picked. Before choosing, Abdul may pay Greta to toss the die in her left hand and tell him
how many dots are painted on the face that comes up. How much is Abdul willing to pay for
the resulting information if he is trying to maximise his monetary reward?

8.5 Oilco must determine whether or not to drill for oil in the South China Sea. It costs
e 100 000 to drill, and if oil is found, the value is estimated to be e 600 000. At present,
Oilco believes there is a 45% chance that the field contains oil. Before drilling, Oilco can hire
(at e 10 000) a geologist to obtain more information about the likelihood that the field will
contain oil. There is a 50% chance that the geologist will issue a favourable report and a 50%
chance of an unfavourable report. Given a favourable report, there is an 80% chance that the
field indeed contains oil. Given an unfavourable report, there is a 10% chance that the field
contains oil nonetheless.

a. Construct a decision tree for the core problem of Oilco.

b. Compute the expected value of the information resulting from the decision to hire the
geologist. Should Oilco hire the geologist?

c. Compute the expected value of perfect information for Oilcos decision problem.

8.6 (Adapted from [W94]) Farmer Jones must determine whether to plant corn or wheat.
If he plants corn and the weather is warm, he earns e 8 000; if he plants corn and the weather
is cold, he earns e 5 000. If he plants wheat and the weather is warm, he earns e 7 000; if
he plants wheat and the weather is cold, he earns e 6 500. In the past, 40% of all years have
been cold and 60% have been warm. Before planting, Jones can pay e 600 for an expert
weather forecast. If the year will actually be cold, there is a 90% chance that the forecaster
will predict a cold year. If the year is actually going to be warm, there is a 80% chance that
the forecaster will predict a warm year.

a. Construct a decision tree for the core problem of farmer Jones.

b. Compute the expected value of the information resulting from the decision to consult
the expert weather forecast. Should Jones consult the weather forecast?

c. Compute the expected value of perfect information for farmer Jones decision problem.

8.7 Solve the following exercises from the textbook:

a. Exercise 5.7 (page 179 in [C96], or page 208 in [CR01])

b. Exercise 5.8

c. Exercise 7.35 (page 257 in [C96], or page 287 in [CR01])

d. Exercise 12.2* (page 452 in [C96], or page 518 in [CR01])

e. Exercise 12.7, parts a., c., and d.

26
9 The Analytic Hierarchy Process

We started this course discussing decision criteria for decision making under uncertainty
with one attribute. We then looked at single-attribute decision making under risk and also
discussed multi-attribute decision making under risk. We now close the circle by looking at
multi-attribute decision making under certainty.

The analytic hierarchy process is a method designed to solve multi-attribute decision problems
for which the consequences of the different decision alternatives are known with certainty.

To read/study:
Transp.: Multiattribute decision making under certainty:
the analytic hierarchy process

For those interested in further reading, we recommend chapter 14 of [W94] and chapter 3 of
[KR93].

Questions

What are advantages and drawbacks if the analytic hierarchy process?

What is a potential problem of the transitivity property of a consistent pair-wise comparison

matrix?

Why is it important that a pair-wise comparison matrix is consistent?

Why would you assess an entire pair-wise comparison matrix if a consistent one can be
derived from only one column?

Which properties of a consistent pair-wise comparison matrix does the Checking for con-
sistency method already assume a matrix to have? Why?

If we have a decision problem with n alternatives and m objectives, how many pair-wise
comparison matrices are to be constructed for an AHP analysis?

Exercises

10.1 (Adapted from [W94]) Each professors annual salary increase is determined by per-
formance in three areas: teaching, research, and service to the university. The administration
has come up with the following pairwise comparison matrix for these objectives:

Teaching
Research Service

1
Teaching 1 3 5

Research 3
1 7

1 1
Service 5 7 1
The administration has compared two professors with regard to their teaching, research, and
service over the past year. The pairwise comparison matrices for teaching, research and
service are, respectively

27
 Professor 1 Professor 2
" #
Professor 1 1 4 (teaching)
1
Professor 2 4 1

Professor 1 Professor 2
" #
1 (research)
Professor 1 1 3
Professor 2 3 1

Professor 1 Professor 2
" #
Professor 1 1 6 (service)
1
Professor 2 6 1

a. Which professor should receive a bigger raise?

b. Does the AHP indicate how large a raise each professor should be given?

c. Check the pairwise comparison matrix for consistency.

10.2 (Adapted from [W94]) A business is about to purchase a new personal computer.
Three objectives are important in determining which computer should be purchased: cost,
user-friendliness, and software availability. The pairwise comparison matrix for these objec-
tives is as follows:

Cost
User-friend Softw avail

1 1
Cost 1 4 5

1
User-friend 4
1 2


Softw avail 5 2 1
Three computers are being considered for purchase. The performance of each computer with
regard to each objective is indicated by the following pairwise comparison matrices. For cost
(low cost is good, high cost is bad!), user-friendliness, and software availability, they are,
respectively

Computer
1 Computer 2 Computer 3
Computer 1 1 3 5
(cost)
Computer 2 1 1 2
3
1 1
Computer 3 5 2 1

Computer
1 Computer 2 Computer 3
1 1
Computer 1 1 3 2
(user-friendliness)
Computer 2 3
1 5

1
Computer 3 2 5 1

Computer
1 Computer 2 Computer 3
1 1
Computer 1 1 2 7
(software availability)
1
Computer 2 2
1 5


Computer 3 7 5 1

28
 a. Which computer should be purchased?
b. Check the pairwise comparison matrices for consistency.

10.3 (Adapted from [W94]) Woody is ready to select his mate for life and has determined
that beauty, intelligence, and personality are the key factors in selecting a satisfactory mate.
His pairwise comparison matrix for these objectives is as follows:

Beauty
Intelligence Personality

Beauty 1 3 5

Intelligence 1 1 3
3
1 1
Personality 5 3 1
Three women (Melanie Griffith, Meryl Streep, and Molly Ringwald) are begging to be Woodys
mate. His views of these womens beauty, intelligence and personality, respectively, are given
in the following pairwise comparison matrices:

Melanie
Meryl Molly

Melanie 1 5 3
(beauty)
Meryl 1 1 1
5 2
1
Molly 3 2 1

Melanie
Meryl Molly

1 1
Melanie 1 6 4
(intelligence)
Meryl 6
1 2

1
Molly 4 2 1

Melanie
Meryl Molly

1
Melanie 1 4 4
(personality)
Meryl 1 1 1
4 9
Molly 4 9 1

a. Whom should Woody choose as his lifetime mate?

b. Evaluate all pairwise comparison matrices for consistency.

10.4 (Adapted from [W94]) In determining where to invest my money, two objectives
expected rate of return and degree of risk are considered equally important. Two investments
(1 and 2) have the following pairwise comparison matrices for expected return and degree of
risk, respectively:

Investment 1 Investment 2
" #
1 (expected return)
Investment 1 1 2
Investment 2 2 1

Investment 1 Investment 2
" #
Investment 1 1 3 (degree of risk)
1
Investment 2 3 1

29
 a. How should I rank these investments?

b. Now suppose another investment (investment 3) is available. Suppose the pairwise

comparison matrices for these investments are as follows with respect to expected return
and degree of risk, respectively:
Investment
1 Investment 2 Investment
3
1
Investment 1 1 2 4
(expected return)
Investment 2 2
1 8

1 1
Investment 3 4 8 1

Investment
1 Investment 2 Investment
3
1
Investment 1 1 3 2
(degree of risk)
Investment 2 1 1 1
3 6
Investment 3 2 6 1

Observe that the entries in the comparison matrices for investments 1 and 2 have not
changed. How should I now rank the investments? Contrast my ranking of investments
1 and 2 with the answer from part a.

10.5 (Adapted from [W94]) You are trying to determine which MBA program to attend.
You have been accepted at two programs: Indiana and Northwestern. You have chosen three
attributes to use in helping you make your decision:

Attribute 1 Cost

Attribute 2 Starting salary

Attribute 3 Ambience of school (can we party there?!!)

Your pairwise comparison matrix for these attributes is as follows:

Cost
Salary Ambience

1
Cost 1 4 2

Salary 4
1 7

1 1
Ambience 2 7 1
For the attributes cost, starting salary, and ambience, respectively, the pairwise comparison
matrices for Indiana and Northwestern are:

Indiana Northwestern
" #
Indiana 1 6 (cost)
1
Northwestern 6 1

Indiana Northwestern
" #
1 (starting salary)
Indiana 1 3
Northwestern 3 1

30
 Indiana Northwestern
" #
Indiana 1 4 (ambience)
1
Northwestern 4 1

Which MBA program should you attend?

10.6 (Adapted from [W94]) You have been hired to determine which of the following
accounts receivable procedures should be used in audit of the Keating Five and Dime Store:

A. Analytic review, B. Confirmations, C. Test of subsequent collections (receipts)

The three criteria used to distinguish between the procedures are as follows:

1. Reliability, 2. Cost, 3. Validity

The pairwise comparison matrix for the three criteria is as follows:

Reliability
Cost Validity

Reliability 1 5 7

Cost 1 1 2
5
1 1
Validity 7 2 1
For the reliability, cost, and validity criterion, respectively, the pairwise comparison matrices
of the three procedures are as follows:

Review
Confirm Test

1 1
Review 1 6 2
(reliability)
Confirm 6
1 4

1
Test 2 4 1

Review
Confirm Test

Review 1 5 3
(cost)
Confirm 1 1 1
5 2
1
Test 3 2 1

Review
Confirm Test

Review 1 3 2
(validity)
Confirm 1 1 1
3 2
1
Test 2 2 1

Use the AHP to determine which auditing procedure should be used. Also check the first
pairwise comparison matrix for consistency.

10.7 (Adapted from [W94]) You are trying to determine which of two secretarial can-
didates (Jack and Jill) to hire. The three objectives that are important to your decision
are personality, typing ability, and intelligence. You have assessed the following pairwise
comparison matrix:

31
 Personality
Typing Intelligence

1 1
Personality 1 4 3

1
Typing 4
1 2


Intelligence 3 2 1
The score of each employee on each objective is as follows:

Personality Typing Intelligence

" #
Jack 0.4 0.6 0.2
Jill 0.6 0.4 0.8

If you follow the AHP method which employee should be hired?

References

[C96] R.T. Clemen (1996). Making Hard Decisions: An Introduction to Decision Analysis,
Duxbury Press, 2nd edition.

[CR01] R.T. Clemen & T. Reilly (2001). Making Hard Decisions with DecisionTools,
Duxbury Press.

[KR93] R.L. Keeney & H. Raiffa (1993). Decisions with Multiple Objectives, Cambridge Uni-
versity Press.

[W94] W.L. Winston (1994). Operations Research: Applications an Algorithms, Duxbury

Press.

32
 3
Answers to selected exercises

2.1 a) small ; large; large. b) 7,00; 9,00; 8,00. c) 8,00.

2.2 b) yes. c) yes; yes.

3.2 c) 180.000.

3.3 c) 6800 (dont hire, plant either)

4.1 a) no. b) no.

4.2 a) 0,20. b) yes. c) no.

5.3 canned soup

5.4 a) u(B) < 0, 966. b) lose lower leg.

5.5 a) draw until million euro card. b) quit after first draw.

6.1 a) risk-averse. b) L1 ; 339, 34.

6.3 a) risk-averse. b) [1.0, 0]. d) 900, 1900.

6.4 a) risk-averse. b) Investment 1. c) 235.000, 290.000.

6.6 k) (1.00, 0.95, 0.90, 0.80, 0.00).

7.2 Location 2.

7.4 b) 28 pints.

7.5 b) Alternative 2.

7.6 d) kEW > 0. e) kDI < 0.

7.11 b) other solutions: k = 0 and k = 15.224.

3
Answers to starred exercises can be found in the textbook

33
7.12 a) 6,3 b) 0,32 c) A.

7.13 a) 3,5 b) 5,2; 4,0 c) > 0,23.

7.14 medical treatment

7.15 a) 6,0 b) 11,3

7.16 g) u(portalo) = 0.45, u(norushi) = 0.6175, u(standard) = 0.55. kl > 0.67: the
portalo; kl < 0.40: the standard.

7.17 c) kX = 0.48, kY = 0.67. e) kX = 0.83, kY = 0.81. Alternative B.

8.3 a) no b) yes.

8.4 16,80

8.5 b) 20.000; hire c) 55.000

8.6 b) 415,20; dont hire c) 600.

9.1 a) Professor 2. c) Consistent enough.

9.2 a) Computer 3. c) User-friendliness: seriously inconsistent(!); others: consistent enough.

9.3 a) Melanie Griffith is the chosen one. c) All matrices are consistent enough.

9.4 a) Investment 1 Investment 2. c) Investment 2 Investment 3 Investment 1!

9.5 Attend the Northwestern MBA program.

9.6 The Confirm procedure should be used. The first matrix is consistent enough.

9.7 Jill should be hired.

34