SUPPLEMENT

SUPPLEMENT
TO TO

Chapter 5
Decision Analysis LEARNING
OBJECTIVES
After completing this
supplement, you should
be able to:

LO 1 Describe the decision analysis

SUPPLEMENT OUTLINE process and influence
diagrams, model a single-
Introduction stage decision problem as
Decision Trees and Using DPL Software a payoff table, and use the
Value of Information expected value approach to
Sensitivity Analysis solve it.
Modelling Preferences/Utilities LO 2 Construct a decision tree and
Summary use it to analyze a multistage
Key Terms problem, discuss the risk
Solved Problems profile chart, and use the DPL
Discussion and Review Questions software.
Internet Exercises LO 3 Calculate expected value of
Problems perfect information (EVPI) and
Mini-Case: Southern Company expected value of sample
Mini-Case: ICI Canada (imperfect) information (EVSI)
using Bayes’ rule.
LO 4 Conduct sensitivity analysis to
probability changes in a
simple decision problem, and
understand sensitivity to input
changes and a tornado chart.
LO 5 Describe how risk aversion
can be modelled by a utility
function, and how multiple
objectives can be handled.

ste2687X_ch05s_001-032.indd 1 12/02/21 4:36 PM

 5S-2 Supplement to Chapter 5 Decision Analysis

LO 1 Introduction
Decision analysis is a scientific and structured method for making important difficult decisions. It
decision analysis was promoted in the 1960s by Professor Ron Howard of Stanford University. Decision analysis
A scientific and structured involves modelling the decision problem (alternatives, random variables, consequences/outcomes
method for making important and their value) and then determining the best decision alternative that, for example, maximizes the
difficult decisions. expected value.
A structured approach to decision making is important because the intuitive quick approach
does not consider all the relevant information and possible alternatives, and is biased:
• What is suggested first to the decision maker influences the decision.
• Most people choose to do nothing (i.e., they prefer the status quo).
• There is a tendency to waste good money after bad in order to avoid admitting to having made
a bad decision.
• People are risk averse when a random variable (gamble) is posed in terms of gains, but risk seek-
ing when it is posed in terms of avoiding a loss, even though the two gambles are equivalent.
• People are overconfident in estimating the range of a random variable (i.e., they choose a small
estimation range). Also, they base their estimate on a dramatic unusual past event that they recall.
Decision analysis is suitable for a range of complex decisions, especially when uncertainties
(i.e., random variables) are present. Among them are capacity planning, new product development,
location planning, project portfolio selection, contract bidding, granting bank loans, and medical
treatment. Most large companies—especially oil and gas companies that have to make the right
decisions for exploration, capital-intensive chemical companies, and R&D-intensive pharmaceutical
companies—use decision analysis. See, for example, the “Chevron” OM in Action.

OM in ACTION http://www.chevron.com

Chevron
Chevron is the third largest North American oil and gas
company. Chevron exploration managers have to decide
which prospects to explore and whether to partner with
other companies to explore any prospects. These deci-
sions have costly consequences and have to be made
under uncertainty (amount of hydrocarbon deposit is
unknown). Chevron requires a structured approach to
decision making.
Decision analysis was introduced to Chevron in the
early 1990s and its use in making capital investment
decisions valued at over $50 million became mandatory
Paul Andrew Lawrence/Alamy Stock Photo
in 2000. One of the early uses of decision analysis was
to evaluate a $2 billion proposal to purchase a flexicoker
for a refinery that needed upgrading. This equipment make decisions. These managers get approval for their
would have processed a wider range of crude oil types. projects from decision review boards. Chevron uses the
The decision analysis team considered all the uncer- dialog decision process (see the “GM” OM in Action
tainties, such as future crude and gasoline prices, and below), which puts projects through review and feed-
constructed the risk profile of this project. The downside back. Thousands of managers have been trained to use
of the project became clear to the senior management, decision analysis. There is an annual best-practice shar-
who rejected it for smaller improvements in the refinery. ing meeting of the analysts.
Chevron has more than 100 decision analysts, spread It is generally agreed that the use of decision analysis
among its business units, who help project managers to is one of the main contributors to Chevron’s success.

ste2687X_ch05s_001-032.indd 2 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-3

Decisions that lend themselves to decision analysis tend to have the following characteristics:
• There is at least one random variable with a set of possible values (each is called a state of
nature) that will have a bearing on the result of the decision. state of nature An outcome
• There is a decision with a list of decision alternatives for the decision maker to choose from. over which the decision maker
has no control (e.g., lottery,
• There is an outcome/consequence with a known payoff (or value) for each decision alternative
coin-toss, whether it will rain
under each state of nature. today, state of the economy).
Decision analysis recommends the following decision process:

1. Framing: identify the decision and objective(s)

2. Structure: identify the alternatives, uncertainties, and

preferences/utility, and model the problem

3.3.Solve:
Solve:identify
identifythe
thebest
bestalternative,
alternative,and
andperform
perform
sensitivityanalysis
Sensitivity analysis

Should model Yes

be changed?

No
4. Recommendation

1. Usually the decision maker facing a complex important decision refers the problem to a deci-
sion team which includes a decision analyst. The decision team identifies the scope of the deci-
sion, plans the process of modelling and information gathering, and identifies the decision
maker’s objective(s). An objective may be monetary such as maximizing profit or net present
value, or non-monetary such as quality-adjusted life expectancy. An objective has to be clear
and measurable.
2. A list of possible alternatives for the decision are developed using brainstorming and other
creative methods in order to achieve the objective(s). The decision alternatives should be crea-
tive but realistic.
Then, random variables influencing the outcomes of decision alternatives are identified,
and discrete values for each, called states of nature (e.g., low, medium, and high demand),
should be determined. Note that decision analysis requires discrete random variables. The
probability of each state of nature has to be estimated by interviewing an expert or experts. A
random variable has to be clearly defined to avoid confusion of the expert(s). The interview
should be structured to reduce experts’ biases. influence diagram
Next, the payoff (or value) associated with the outcome of each alternative under each state A compact graphical
of nature has to be determined. When values are small, the decision maker can be assumed to be representation of a decision
risk neutral and the payoffs are used directly to determine the best alternative. However, for large problem.
payoffs, especially when a large loss is possible, empirical studies have shown that most people
are risk averse (need to be paid to take a gamble). In this case, the utility to the decision maker of decision tree A tree-like
each payoff has to be measured and used to determine the alternative with the best expected utility. graphical representation of the
Utilities are described in the last section. decisions, random variables and
Finally, the structure of the model (the relationship between the decision alternatives, random their probabilities, and the
variables, and payoffs) should be determined and represented as an influence diagram and/or payoffs.
a decision tree.

ste2687X_ch05s_001-032.indd 3 12/02/21 4:36 PM

 5S-4 Supplement to Chapter 5 Decision Analysis

3. The best alternative is determined, using, for example, the expected value method. If the deci-
risk profile The probability sion maker is risk averse, the expected utility is used. If there is more than one objective, all of
distribution of the payoff of the them (e.g., a weighted average, or pros and cons of alternatives) are used to select the best
best alternative. alternative. We will focus on the single-objective case until the last section.
Then, sensitivity of the best decision alternative to probabilities of random variables
should be determined. (As the probability of a state of nature changes, would the best alterna-
tive change?) If the best alternative is sensitive to small changes in probabilities of a random
variable, more information may be gathered (if the value of information is more than its cost)
dialog decision process in order to reduce the uncertainty. In addition, the sensitivity of the best alternative to pos-
(DDP) A decision process sible changes in inputs such as price and cost, which are initially estimated as constants, are
which requires communication determined and displayed as a tornado chart (illustrated later). An input that greatly influences
between the support team and the payoff of best decision alternative should be estimated more accurately or be represented
the review board after each by a random variable. After a major change in data or structure, the model should be re-solved.
major stage of decision analysis 4. In the final step, the team presents the best decision alternative and the reasons behind it (pay-
(as opposed to just at the off, risk, etc.) to the decision maker(s). In addition, usually the probability distribution of the
beginning and end). payoff of the best alternative (called the risk profile) is also determined in order to display the
variation of the payoff if the best alternative is chosen.

OM in ACTION www.gm.ca

GM The new general manager of corporate strategy and

knowledge development, who had previously used deci-
GM has been using decision analysis for decades in the sion analysis at Kodak, changed the way decisions are
areas of product planning and development, marketing, made at GM. He employed decision analysts in decision
technology choices, IT, strategy, resource allocation, and support teams and developed the dialog decision pro-
finance. The decision makers are various review boards, cess (DDP), which requires communication between the
many cross-functional, consisting of senior GM managers. decision support team and the review board after each
The review boards are assisted by decision support teams. major stage of decision analysis (as opposed to just at
Before 1990, a decision support team researched the beginning and at the end).
and analyzed the problem and advocated its solution The DDP has improved the quality, cost, and time
to the review board in the form of a business case. The of making complex decisions, and has led to a more
review board either accepted or rejected the recommen- open and cooperative culture with improved learning
dation, or sent it back for rework to the decision support and knowledge sharing. Thousands of new managers
team. Because the objective of a decision support team have participated in decision analysis training and hun-
was to get approval from a review board, over time they dreds of DDPs have been conducted. Recently, internal
increasingly minimized the downside of their recommen- decision consulting has been replaced with outsourced
dation and exaggerated the upside. The result was deci- consultants.
sions made by the review board without knowing all the
risks and opportunities involved.

Influence Diagrams
An influence diagram is a compact graphical representation of a decision problem. The
following example represents the decision of whether or not to introduce a new product,
and at what price. A random variable is represented by a blue circle (called a chance
node), a decision is represented by a cream rectangle (called a decision node), and a
payoff is represented by a green rounded square (called a payoff node). Actually, profit is
the direct payoff, and revenue and cost are the indirect payoffs (called calculation nodes
in brown ovals) that determine the direct payoff. An arrow into a chance node affects its
probability distribution, into a decision node determines its sequence and the information
known before this decision is made, and into a payoff node determines its value (using a
formula). Naturally, the nonexistence of an arrow indicates lack of influence.

ste2687X_ch05s_001-032.indd 4 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-5

Revenue

Introduce a Quantity
Price? Cost Profit
product? sold

Fixed Variable
cost cost

Source: Adapted from K. Chelst, “Can’t See the Forest Because of the Decision Trees: A Critique of
Decision Analysis in Survey Texts,” Interfaces 28, 2, pp. 80–98. Figure 1, p. 85.

Constructing and validating an influence diagram improves communication and consensus

building. Unfortunately, there is no procedure or guideline for drawing an influence diagram. Also,
the context of every decision is different. Furthermore, not all the information needed to make a
decision is displayed in an influence diagram. For example, the above example does not show the
decision alternatives and the states of nature of each random variable. Therefore, a decision tree is
usually drawn based on the influence diagram and used to determine the best alternative.

Payoff Tables
The data for a one variable problem can be summarized in a payoff table, which shows
the payoff for each decision alternative under each state of nature of the random variable. payoff table A table showing
These tables are helpful in choosing among decision alternatives because they facilitate the payoff for each decision
comparison of the alternatives. Consider the following payoff table, which illustrates a alternative under each state of
capacity planning problem. nature of one random variable.

(Future) Demand
Decision Alternatives Low Moderate High
Small facility $10* $10 $10
Medium facility 7 12 12
Large facility (4) 2 16
* Net present value (in $ millions).

The payoffs, in the body of the table, are net present values, which represent equivalent current
dollar values of estimated future revenues less expenses. This is a convenient measure because it
places all decision alternatives on a comparable basis. If a small facility is built, the payoff will be
$10 million for all three possible states of nature. For a medium facility, low demand will have net
present value of $7 million, etc.
The problem for the decision maker is to select one of the decision alternatives, taking the
payoffs into account. In order to do so, the decision analyst needs to determine the probability of
occurrence of each state of nature (unless one of the alternatives dominates the rest and would have
the highest payoff under each state of nature). Because the states of nature for a random variable
should be mutually exclusive and collectively exhaustive, these probabilities must add up to 1.00.
As mentioned earlier, the probabilities of states of nature are usually subjective opinions of
experts. The decision analyst interviewing an expert should be careful not to ask leading questions
and should be aware of psychological biases such as overconfidence or intentional under- or overes-
timation. For more information on obtaining subjective probabilities, see M.W. Merkhofer, “Quan-
tifying Judgmental Uncertainty: Methodology, Experiences, and Insights,” IEEE Transactions on
Systems, Man, and Cybernetics, 17, 5, 741–751. One method for estimation of a random variable
probability distribution appears in the “Bayer Pharma” OM in Action.

ste2687X_ch05s_001-032.indd 5 12/02/21 4:36 PM

 5S-6 Supplement to Chapter 5 Decision Analysis

OM in ACTION http://pharma.bayer.com

Bayer Pharma such as size of the market were uncertain. The prob-
ability distributions for these data were continuous.
Bayer Pharma is a division of Bayer and makes drugs. These had to be discretized in order to be able to use
Bayer Pharma started to use decision analysis in the late decision analysis. Experts were asked to provide the
1980s. A business unit within Bayer Pharma is the biolog- 10th, 50th, and 90th percentiles for continuous random
ical products unit that makes drugs from living/biologi- variables. Probabilities of 0.3, 0.4, and 0.3 were used
cal cells. The first application of decision analysis in the for these percentiles, respectively. Similarly, subjective
biological products unit was in 1999 when the leadership probabilities for success in each stage were solicited
committee asked the strategic planning department, from experts. When these values were deemed too
which included decision analysts, to commercially evalu- optimistic, they were replaced with historical company
ate a new blood-clot busting drug. data.
A cross-functional team of Bayer Pharma employees After the decision model was built, it was first
was formed (featuring four scientists, a clinician, a regu- evaluated using the 50th percentiles for all the vari-
lator, a production engineer, an industrial engineer, a ables (i.e., assuming certainty). Then, those variables
marketer, and a decision analyst). The team had to deter- that significantly affected the NPV (e.g., market share)
mine the net present value (NPV) of this drug (which were identified using a Tornado chart (described
was expected to have a life of 15 years) by estimating its later). Next these were used as random variables in
demand (size of market, market share), cost of develop- the model (the others were set at their 50th percen-
ment, manufacturing, and marketing, and probability of tile), and the model was resolved. Finally the prob-
success in each stage of development. ability distribution of NPV (the risk profile, described
In order to be approved by the government, a drug later) was determined. The results of the final decision
has to go through many trial stages. These include model and the rigour of decision analysis convinced
testing it on animals in labs (called pre-clinical), then, the Bayer Healthcare board to approve the drug for
if approved, on people, which itself has three stages/ pre-clinical trial.
phases to determine its safety, dosage, and side effects.
Members of the team were asked to estimate the Source: J.S. Stonebraker, “How Bayer Makes Decisions to Develop New
parameters/variables related to their area. Several data Drugs,” INFORMS Journal on Applied Analytics, 2002, 32(6), pp. 77–90.

Expected Value Method Given the probabilities for each state of nature, a widely used approach
expected monetary value for determining the best decision alternative is to use the expected value (also called expected
(EMV) The sum over the states of monetary value (EMV)). The EMV for a decision alternative is the sum over the states of nature
nature of the payoff multiplied by of the payoff multiplied by probability of each state of nature. The EMV is calculated for each
probability of each state of nature. decision alternative and that alternative with the largest expected value (if payoff is profit or NPV)
is selected.

EXAMPLE 5S-1 SOLUTION

Using the expected values, identify the best alternative for Calculate the expected value of each alternative by multiplying
the following payoff table, given these probabilities for future the probability of occurrence of each state of nature by the pay-
demand: low = 0.30, moderate = 0.50, and high = 0.20. off of the alternative for that state of nature and summing them:
​   EV​ small​​ = 0.30($10) + 0.50($10) + 0.20($10) = $10
(Future) Demand ​E
​​ V​  medium​​ = 0.30($7) + 0.50($12) + 0.20($12) = $10.5
     
       ​​​
Alternatives Low Moderate High ​    EV​ large​​ = 0.30(− $4 ) + 0.50($2) + 0.20($16) = $3
Small facility $10* $10 $10 Hence, choose the medium facility because it has the largest
Medium facility 7 12 12 expected net present value.
Large facility (4) 2 16
*Net present value (in $ millions).

ste2687X_ch05s_001-032.indd 6 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-7

As mentioned earlier, expected value is appropriate when a decision maker is neither risk averse
nor risk seeking, but is risk neutral. Note that the expected value amount (e.g., $10.5 million for the
medium facility in Example 5S-1) is not an actual payoff but an expected or average amount that
would be achieved if a large number of identical decisions were to be made. If the decision maker is
risk averse, then he tends to consider the downside risk of payoff. For a risk averse decision maker,
the minimum payoff (if the medium facility is built) of $7 million may still be acceptable. If it is not,
he/she may seek more accurate information about demand or choose to build a small facility that has
a higher minimum payoff ($10 million). As mentioned earlier, a formal approach to deal with risk
averse decision makers is to use utilities (see the last section).
More complex decision problems will have more than one decision variable (e.g., a multistage
decision problem) and/or more than one random variable. In this case, the decision problem cannot
be represented by a payoff table. Instead, decision trees are used to graphically model and solve the
decision problem.

Decision Trees and Using DPL Software LO 2

A decision tree is a tree-like graphical representation of the decisions, random variables and
their probabilities, and the payoffs. The nodes in a decision tree are arranged according to
their occurrence in time (from left to right). The term comes from the tree-like appearance
of the diagram (see Figure 5S-1). Decision trees are particularly useful for analyzing situa-
tions that involve sequential or multistage decisions. For instance, a manager may initially
decide to build a small facility but he/she has to allow for the possibility that demand may
be higher than anticipated. In this case, the manager may plan to make a subsequent deci-
sion on whether to expand or build an additional facility. Note that the second decision usu-
ally influences the initial decision and therefore should be included in the initial analysis.

re 1 Payoff 1
of natu FIGURE 5S-1
State
Format of a decision
e A'1
State
of na Choos Payoff 2 tree.
eA ture
1
s 2 2
oo
Ch Choos
e A'2 Payoff 3
Possible
Initial
1 second
decision decision
e A'3
Ch Choos Payoff 4
oo
se 2
ure 1
A o f nat
State
2
Choos Payoff 5
e A'4
State
of na
ture
2 Payoff 6

Decision Chance event

A decision tree is composed of a number of nodes that have branches emanating from them to
their right (see Figure 5S-1). Square nodes denote decision points and circular nodes denote chance
events (random variables). Branches leaving decision nodes (on their right) represent alternatives
(the decision maker can choose the preferred branch); branches leaving chance events (on their
right) represent the states of nature (the decision maker has no control over these). A branch from
node i to node j means that node i has already occurred and the information about the decision or
state of nature determined at node i is available at node j.
After a decision tree has been drawn and the necessary data collected, it is evaluated from right folding back A method for
to left; that is, it is “rolled” or “folded” back. The folding back procedure is as follows. For each calculating EMVs in a decision
decision, choose the alternative that will yield the greatest (expected) payoff. For each chance node, tree and identifying the best
calculate the expected value of the payoffs of its states of nature. Continue folding back (to the left) alternative.
until the best alternative for all decisions is determined.

ste2687X_ch05s_001-032.indd 7 12/02/21 4:36 PM

 5S-8 Supplement to Chapter 5 Decision Analysis

A manager must decide the size of a video arcade to construct. The manager has narrowed the
EXAMPLE 5S-2 choices to two: small or large. Information has been collected on payoffs and the following
decision tree has been constructed. Determine which alternative (build small or build large)
should be chosen in order to maximize the expected value.

)
d (0.4 $40
e man
Lo wd

g $40
High ot hin
dem
and Don
m all (0.6
ds )
il
Bu
Exp
and
$55
g ($10)
hin
not
Do
Bu )
ild (0.4
lar
ge and Red
dem uce
Low pric
es $50
High
dem
and
(0.6 $70
)

SOLUTION demand, there are two choices: do nothing or reduce

prices. You would choose reduce prices because it has
The dollar amounts at the branch ends (on the right) indicate higher payoff ($50 > −$10), so either replace this second
the estimated payoffs if particular sequences of decision alter- decision by its payoff of $50 and reduce the tree, or place
natives and states of nature occur. For example, if the initial a double slash on the other branch (do nothing). We have
decision is to build a small facility and it turns out that demand chosen the former here:
is low, the payoff will be $40 (thousand). Similarly, if a small
facility is built and demand turns out high, and a later decision )
(0.4 $40
is made to expand, the payoff will be $55. A payoff in paren- and
dem
theses indicates a loss. Low
The figures in parentheses on branches leaving the chance
nodes indicate the probabilities of those states of nature. High
dem
Hence, the probability of low demand is 0.4 and the probability ll and
ma (0.6 $55
uilds )
of high demand is 0.6. Note that these two probabilities add up B
to 1.0. Also note that the initial decision has to be made before
demand is known and the second decision is made after
demand is known.
1. Determine which alternative would be selected for each
possible second decision. For a small facility with high Bu
ild .4) $50
demand, there are two choices: do nothing, or expand. lar
ge an d (0
dem
Because expand has higher payoff ($55 > $40), you Low
would choose it. You can either replace this second deci-
sion with a payoff of $55 and reduce the tree, or block the High
dem
dominated alternative (do nothing) by placing a double and
(0.6 $70
slash through it. Similarly, for a large facility with low )

ste2687X_ch05s_001-032.indd 8 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-9

2. Now, continue with the rollback using the above reduced $49
tree. Determine the product of the state of nature proba-
bilities and their respective payoffs: all
sm
ild
Chance node after Build small Bu

Low demand 0.4($40) = $16

High demand 0.6($55) = $33
Chance node after Build large
Bu
Low demand 0.4($50) = $20 ild
lar
ge
High demand 0.6($70) = $42
$62
Therefore, the expected values of these chance nodes are:
Chance node after Build small: $16 + $33 = $49
3. Continue with rollback. Clearly, the choice for the first
Chance node after Build large: $20 + $42 = $62 decision should be to build a large facility because it has
Replace these chance nodes with their expected value: higher expected value than the small facility ($62 > $49).

Risk Profile
Because the decision maker is usually concerned with the downside risk of payoff, he/she will be inter-
ested in the probability distribution of the payoff of the best alternative. This is called the risk profile.
To determine the risk profile of the best alternative, the probability of each payoff that can be
reached from the best alternative has to be determined. If there is more than one state of nature on
the path from the best alternative to the payoff, their probabilities should be multiplied. Then, prob-
abilities of identical payoffs should be added. If there is only one random variable, the risk profile
is simply the probability distribution of the random variable and its payoffs. For Example 5S-2, the
risk profile for Build large is displayed below:
Prob

0.6

0.4
0.2

$10 $30 $50 $70 Payoff

In this case the minimum payoff ($50) of Build large is larger than the minimum payoff of
Build small ($40), therefore, Build large is better in terms of both average value and risk. In fact, if
the cumulative probability distributions (CPDs) of both large and small alternatives are drawn on
the same chart, the CPD of Build large will be to the right of CPD of Build small (this is called sto-
chastic dominance), implying that any person with any form of utility/risk attitude will prefer Build
large. In general, the downside risk of payoff may be reduced by identifying the random variable
that is causing it and gathering more information on it.
The above problem was entered in the DPL software from Syncopation Software (http://www
.syncopation.com) and solved. Using DPL is described in the following tutorial.

How to Use the DPL Software

Download a free trial version of DPL 9 Professional from http://www.syncopation.com
/dpl-professional-trial and install it.
To start a new model, click the File tab and then New. The DPL model window contains two
panes: the Influence Diagram pane, and the Decision Tree pane. When you start a new model, by

ste2687X_ch05s_001-032.indd 9 12/02/21 4:36 PM

 5S-10 Supplement to Chapter 5 Decision Analysis

default you will be in the Decision Tree mode. You can, however show both panes by working in the
split mode. To achieve the halfway split mode, click the splitter bar just above the top of the Deci-
sion Tree pane and drag down the splitter.
Click the Add button to add a new decision node to the Influence Diagram. Place the decision
node in the Influence Diagram pane. Once placed, the Node Definition Build is opened with the
General tab active. Name the node and the alternatives as follows:

Source: Syncopation Software, Inc., www.Syncopation.com.

Repeat this process, creating two versions of the second decision, each with its own alterna-
tives. Name one of the decision nodes “React to high demand” with alternatives of “Do nothing”
and “Expand”, and the other “React to low demand” with alternatives “Do nothing” and “Reduce
prices”. Next, add a discrete chance node (under the Add button) to the Influence Diagram for
demand. Once placed, name it “Demand”. By default DPL creates three outcome chances node, but
this node only requires two outcomes: Low and High. Delete the Nominal outcome, as shown below:

Source: Syncopation Software, Inc., www.Syncopation.com.

ste2687X_ch05s_001-032.indd 10 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-11

Select the Data tab of the Node Definition dialog for the Demand node and enter the probabili-
ties as follows. Leave the value inputs blank:

Source: Syncopation Software, Inc., www.Syncopation.com.

Close the Node Definition dialog by clicking Ok. Notice that DPL has been building a sym-
metric, default Decision Tree for you as you added nodes to the Influence Diagram. You will now
make edits to this default symmetric decision tree and take over controls. Within the Decision Tree,
select the Build decision node and check the box labeled “Asymmetric” in the grouping section.
Now each alternative of the decision has its own endpoint (blue triangle) and the rest of the subtree
has been disconnected from the node. All of the nodes in the model need to be asymmetric, so repeat
this process for all of the nodes in the model.
Now we need to rebuild the decision tree to match our decision problem. Click and drag the
Demand chance node and drop it on the blue triangle attached to the end of the “Small” branch. Add
a second instance of the Demand chance node to the blue triangle at the end of the “Large” branch
by dragging it down from the Influence Diagram. Again, you’ll need to make this node asymmetric
following the above instructions. Drag the “React to high demand” node you created and drop it at
the end of the High outcome of the Demand node connected to Build | Small. Drag the “React to low
demand” node and drop it at the end of the Low outcome branch of of the Demand node connected
to Build | Large. Double-click each branch ending in a payoff and enter the payoff in Objective
Function text box (highlighted magenta). For example:

Source: Syncopation Software, Inc., www.Syncopation.com.

ste2687X_ch05s_001-032.indd 11 12/02/21 4:36 PM

 5S-12 Supplement to Chapter 5 Decision Analysis

Now your model should look like the following:

Source: Syncopation Software, Inc., www.Syncopation.com.

To run a Decision Analysis and generate results, click the Home tab. Make sure that the Policy
Tree checkbox is checked. The default evaluation method is Fast Sequence, so click the “Decision
Analysis” button. DPL will run the model and generate the selected output. The DPL Policy Tree
is displayed. The expected value solution of Example 5S-2 in DPL is determined and is displayed
at the head of the Policy Tree, as shown below. Note that the optimal alternatives are shown with
bolded lines and the non-optimal part of the tree (Build | Small) is not shown. Also, EMV for any
node is shown on its top left in square brackets. For example, the EMV of Demand at the end of the
Large alternative is 62.0.

Source: Syncopation Software, Inc., www.Syncopation.com.

ste2687X_ch05s_001-032.indd 12 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-13

OM in ACTION www.unilever.ca

Unilever Unilever uses Palisade’s DecisionTools Suite for

DMUU. This is a set of Excel add-ins. The suite includes
Unilever is a large British–Dutch multinational consumer @RISK and PrecisionTree for simulation and decision
goods company with over 172,000 employees world- analysis, respectively. @Risk can model continuous
wide. It manufactures over 1,000 brands, including food probability distributions such as triangular and Beta dis-
(e.g., Becel, Hellmann’s, Knorr, and Breyers ice cream), tributions. However, this capability comes at the cost of
beverages (e.g., Lipton tea), cleaning agents (e.g., Sun- higher computational needs. Hence @Risk is used for
light), and personal care products (e.g., Dove). less complex decision problems such as a capital invest-
Since 1999, Unilever has had a decision analysis ment. On the other hand, PrecisionTree is used for com-
group in the finance department that acts as internal con- plex multistage strategic decision problems such as a
sultant for decision making under uncertainty (DMUU). new product introduction. Over 400 finance managers
DMUU involves the use of simulation and decision analy- have been trained in the use of DMUU and Decision-
sis to make structured decisions based on probabilistic Tools. There are also a library of best practices, user
analysis. DMUU is required for all complex projects that guides, and e-learnings.
have a large number of alternatives and random vari- The DecisionTools Suite tools integrate seamlessly
ables, and involve different company functions with dif- with MS Excel (see the screenshot below).
ferent attitudes toward risk.

Source: Palisade, screenshot of DecisionTools Suite, https://www.palisade.com/decisiontools_suite/

ste2687X_ch05s_001-032.indd 13 12/02/21 4:36 PM

 5S-14 Supplement to Chapter 5 Decision Analysis

LO 3 Value of Information
In certain situations, it is possible to buy information about the states of nature of a random variable
in order to more accurately estimate their probabilities and hence make a better decision. For
instance, one can perform a market survey about customer attitude toward a new product. Or infor-
mation about technical potential of a product could come from product testing.
expected value of perfect If the information is perfect (i.e., it is 100 percent accurate and will pinpoint the state of nature
information (EVPI) The of a random variable), then the expected gain in payoff is called the expected value of perfect
difference between the expected information (EVPI), which is the difference between the expected payoff with perfect information
payoff with perfect information and the expected payoff without the information.
and the expected payoff without
the information. Expected value of perfect information = E
 xpected payoff with perfect information
(5S-1)
               − Expected payoff without the information

Then combine these by weighting each payoff by the prob-

EXAMPLE 5S-3
ability of that state of nature and adding the amounts. Thus,
the best payoff under low demand is $10, the best payoff
Using the data from Example 5S-1, reproduced below, determine under moderate demand is $12, and the best payoff under high
the expected value of perfect information using Formula 5S-1. demand is $16. The expected payoff with perfect information
(Future) Demand is, then:
Alternatives Low Moderate High
Small facility $10* $10 $10
0.30($10) + 0.50($12) + 0.20($16) = $12.2
Medium facility 7 12 12 The expected payoff without perfect information, as calculated
Large facility (4) 2 16 in Example 5S-1, is $10.5. The EVPI is the difference between
*Net present value (in $ millions). these:

SOLUTION EVPI = $12.2 − $10.5 = $1.7

First, calculate the expected payoff with perfect information.
To do this, identify the best payoff under each state of nature.

EVPI indicates the maximum the decision maker should be willing to spend to obtain informa-
tion. Thus, if the cost exceeds EVPI, the decision maker would be better off not to spend additional
money and simply to go with the alternative that has the highest expected payoff.
In most cases, the information that can be obtained about a random variable is not perfect.
In these cases, a statistical result called Bayes’ rule can be used to update the (prior) probability
distribution of the states of nature of a random variable. The difference between the expected payoff
with imperfect (sample) information and the expected payoff without sample information is called
the expected value of sample information (EVSI).
expected value of sample
information (EVSI) The Bayes’ Rule
difference between the expected Bayes’ rule is a mathematical formula for determining conditional probability named after 18th-cen-
payoff with sample (imperfect) tury British mathematician Thomas Bayes. Let A be an “information” event, and let B1 and B2 be mutu-
information and the expected ally exclusive and exhaustive states of nature. Let P(B1) and P(B2) be the prior probabilities of the states
payoff without sample of nature, and P(A | B1) and P(A | B2) be the likelihood of observing A given B1 and B2, respectively (i.e.,
information. the conditional probabilities). Then, the posterior probabilities P(B1 | A) and P(B2 | A) are:
Bayes’ rule A mathematical
​P (​B1​  ​​​|​​  A) = ​   
______________________
   | 1
P​(​​A​​​ ​B​  ​​​)​​P​(​​ ​B​  ​​​)​​
1
 ​​ (5S-2)
formula for determining ( | 1) ( 1) ( | 2) ( 2)
P​ ​​A​​​ ​B​  ​​​ ​​P​ ​​ ​B​  ​​​ ​​  + P​ ​​A​​​ ​B​  ​​​ ​​P​ ​​ ​B​  ​​​ ​​
conditional probability, named
after 18th-century British ​P​(​​ ​B2​  ​​​|​​A​)​​ = 1 − P​(​​ ​B1​  ​​​|​​A​)​​​ (5S-3)
mathematician Thomas Bayes.
​P​(​​A​)​​  = P​(​​A​|​​ ​B1​  ​​​)​​P​(​​ ​B1​  ​​​)​​  + P​(​​A​|​​ ​B2​  ​​​)​​P​(​​ ​B2​  ​​​)​​​ (5S-4)

ste2687X_ch05s_001-032.indd 14 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-15

Organizers of an international athletics competition are consid- EXAMPLE 5S-4

ering mandatory testing. However, the test is not 100 percent
accurate. The conditional probabilities are as follows: If a per-
son uses drugs, then the test will be positive 93 percent of the Note that if the drug test is positive, the probability that
time, whereas if he does not, the test will be negative 97 per- the employee tested is a non-user (i.e., a false positive)
cent of the time. Suppose we know that generally 5 percent of is not small (38 percent).
athletes in this field use drugs. Similarly P(A−|B1) = 1 − 0.93 = 0.07, P(A−|B2) = 0.97,
a. What are the posterior probabilities—that is, the probabil- P(B1|A−) = (0.07(0.05))/(0.07(0.05) + 0.97(0.95)) =
ities that a drug user and a non-user are correctly identi- 0.0038, and P(B2|A−) = 1 − 0.0038 = 0.9962.
fied after drug testing? Note that if the drug test is negative, then the probability
b. If the cost of not identifying (and dismissing) a drug user that the employee tested is a user (i.e., a false negative) is
(i.e., a false negative test result) is $1,000, cost of falsely very low (0.0038).
accusing a non-user (i.e., a false positive test result) is b. EVSI can be determined from the following decision tree
$200, and other costs are zero, what is the maximum the obtained by DPL. Note that negative payoffs are costs and
manager should be willing to pay for a drug test (expected the probabilities obtained in part a are used for outcomes
value of sample information)? related to drug testing. For example, the probability that
an employee is a drug user, given a positive test result, is
SOLUTION
0.62 or 62%.
a. Consider a random employee. Let B1 = The event that he is
a drug user, B2 = The event that he is not a drug user. We
have P(B1) = 0.05 and hence P(B2) = 1 − 0.05 = 0.95. Let
A+ = The event that the test result is positive. We have
P(A+ | B1) = 0.93 and P(A+ | B2) = 1 − 0.97 = 0.03. Then,

​​​P​(​​ ​B1|A+​)​​  = ​(​​0.93​(​​0.05​)​​​)​​  / ​(​​0.93​(​​0.05​)

           + 0.03​ (0.95​
​​ )​​​)​​ = 0.62​​
​P​(​​ ​B2​​  |A+​)​​  = 1 − 0.62 = 0.38.​​
Also, P​(​​A+)​​  = P​(​​A+|B1​​​)​​P​(​​ ​B1​​​)​​ + P​(​​A+|​B2​​​)​​
     P​(​​ ​
B2​​​)​​  = 0.93​(​​0.05​)​​ + 0.03​(​​0.95​)​​  = 0.075,​​ 
​Also, P​(​​A−)​​  =   1 − P​(​​A+)​​  = 1 − 0.075 = 0.925​ ​​
Employee 1 User [0.000]
Positive [–76.000] 62.00% 0.000
Test result 7.50% Non-user [–200.000]
Yes [–9.215] 38.00% –200.000
Test Employee 2 User [–1000.000]
[–9.215] Negative [–3.800] 0.38% –1000.000
92.50% Non-user [0.000]
99.62% 0.000
Employee 3 User [–1000.000]
No [–50.000] 5.00% –1000.000
Non-user [0.000]
95.00% 0.000
The maximum that the manager should pay for a drug test for
an employee (i.e., EVSI) is $50 − $9.215 = $40.785.

Sensitivity Analysis LO 4

Generally speaking, both the payoffs (and their components such as price and cost, and other inputs)
and the probabilities in a decision problem are estimated values. Consequently, it can be useful for
the decision maker to know how sensitive the choice or payoff of the best alternative is to changes in
one or more of these values. Unfortunately, it is impossible to consider all possible combinations of
every change in a typical problem. Nevertheless, we will consider two types of sensitivity: (1) sensi-
tivity to probability in a one-variable problem, and (2) sensitivity to an input.

ste2687X_ch05s_001-032.indd 15 12/02/21 4:36 PM

 5S-16 Supplement to Chapter 5 Decision Analysis

Sensitivity to a Probability
sensitivity to a
probability Range of a Sensitivity to a probability provides the range of probability over which an alternative has the best
probability for which an alternative expected payoff. The approach illustrated here is useful when there are two states of nature. It
has the best expected payoff. involves constructing a graph and then using algebra to determine the range of probability for which
a given alternative is best. Example 5S-5 illustrates the procedure.

To find exact values of the ranges for P2, determine where

EXAMPLE 5S-5
the lines intersect. Note that at the intersections, the two alter-
natives represented by the lines would be equivalent in terms
Given the following payoff table, determine the range of prob- of expected value. To determine the intersections, you must
ability for state of nature #2, P2, for which each alternative is obtain the equation of each line. Because these are straight
optimal using the expected values. lines, they have the form y = a + bx, where a is the y-intercept
State of Nature value at the left axis, b is the slope of the line, and x is P2.
#1 #2 Slope is defined as the change in y for a one-unit change in x.
A 4 12 In this case, the distance between the two vertical axes is 1.0.
Alternative B 16 2 Consequently, the slope of each line is equal to the right-hand
C 12 8 value minus the left-hand value. The slopes and equations are:

#1 #2 Slope Equation
A 4 12 12 − 4 = +8 4 + 8P2
SOLUTION
B 16 2 2 − 16 = −14 16 − 14P2
First, plot the expected payoff of each alternative relative to C 12 8 8 − 12 = −4 12 − 4P2
P2. To do this, plot the #1 payoff (i.e., the payoff if the state
of nature #1 occurs) on the left side of the following graph From the graph, we can see that alternative B is best from the
and the #2 payoff on the right side. For instance, for alterna- point P2 = 0 to the point where the alternative B line intersects
tive A, plot 4 on the left side of the graph and 12 on the right the alternative C line. To find that point, solve for the value of P2
side. Then, connect these two points with a straight line. The at their intersection. This requires setting the two equations
three alternatives are plotted on the graph as shown below. equal to each other and solving for P2. Thus,
16 − 14P2 = 12 − 4P2
16 16
Rearranging terms yields
14 B 14 4 = 10P2
12 12 Solving yields P2 = 0.40. Thus, alternative
A
C B is best from P2 = 0 up to P2 = 0.40. Alterna-
10 10
tives B and C are equivalent at P2 = 0.40.
#1 #2
8 8 Payoff Alternative C is best from that point
Payoff
until its line intersects alternative A’s line.
6 6 To find that intersection, set those two
4 4 equations equal and solve for P2. Thus,
4 + 8P2 = 12 − 4P2
2 2
B best C best A best
Rearranging terms results in
0 0.2 0.4 0.6 0.8 1.0 12P2 = 8
P (2) Solving yields P2 = 0.67. Thus, alternative
C is best from P2 > 0.40 up to P2 = 0.67,
The graph shows the range of values of P2 over which each where alternatives A and C are equivalent. For values of P2 greater
alternative is optimal (i.e., has the largest expected value). than 0.67 up to P2 = 1.0, alternative A is best.
Thus, for low values of P2 (and thus high values of P1, Note: If a problem calls for ranges with respect to P 1,
because P1 + P2 = 1.0), alternative B will have the highest you could find the P 2 ranges as above, and then subtract
expected value; for intermediate values of P2, alternative C is each P 2 from 1.00 (e.g., 0.40 becomes 0.60, and 0.67
best; and for higher values of P2, alternative A is best. becomes 0.33).

ste2687X_ch05s_001-032.indd 16 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-17

Sensitivity to an Input
Sensitivity to an input is the range of the expected payoff of the best alternative as a result of vary-
ing an input. An input such as price, cost, or the quantity of high demand is initially estimated by one sensitivity to an input The
value (i.e., it is a constant). This best estimate can be called the nominal value. However, the nominal range of the expected payoff of
value is just an estimate, and a range from low to high can be determined for this input. It would be the best alternative as a result of
difficult to vary all inputs simultaneously to see their effect on the best alternative. So, this perturba- varying an input.
tion is done one input at a time, leaving the other inputs at their nominal value. The resulting range
of expected payoff for the best alternative, as a result of varying each input, is displayed in order of
importance in a chart called a tornado chart.

Tornado Chart A tornado chart is a chart that shows the range of expected payoff for the best
alternative, as a result of varying each input, in order of importance of input, most important in the tornado chart A chart that
top. It is named tornado chart because the more important inputs that have longer bars are placed on shows the range of expected
the top, thus making the chart look like a tornado. payoff for the best alternative, as
An influential input, especially one with large downside expected payoff, should in fact be a result of varying each input, in
considered a random variable and its values and probabilities properly estimated. order of importance of input,
most important in the top.

Reconsider Example 5S-2, the video arcade problem. Sup- EXAMPLE 5S-6
pose the payoff (profit) is calculated as follows: Payoff =
Revenue − Cost. Revenue is number of customers times rev-
enue per customer. Cost is amortized building cost plus oper- SOLUTION
ating costs. The nominal values are as follows: High demand The low and high of expected payoffs for each of these input
= 30,000 per year, Low demand = 15,000 per year, Revenue ranges, while keeping the other inputs at their nominal value,
per customer = $5, Reduced-prices revenue per customer = are as follows:
$4.5, Increased demand as a result of reduced prices =
28,889, and Cost per year of Build large = $80,000. Using the
nominal values, the payoff (profit per year) for Build large High demand:
and high demand is 30,000($5) − $80,000 = $70,000, and for Build large and High demand payoff = [25,000($5) −
low demand/reduced prices/increased demand is 28,889($4.5) $80,000, 35,000($5) − $80,000] = [$45,000, $95,000]
− $80,000 = $50,000. These are what were given in Example Expected payoff of Build large = [0.4($50,000) + 0.6($45,000),
5S-2. The solution using the nominal values is reproduced 0.4($50,000) + 0.6($95,000)] = [$47,000, $77,000]
below for convenience. Increased demand as a result of reduced prices:
Build large and Low (but increased as
a result of reduced prices) demand
payoff = [15,000($4.5) − $80,000,
30,000($4.5) − $80,000] =
[−$12,500, $55,000]
 xpected payoff of Build large =
E
[0.4(−$12,500) + 0.6($70,000),
0.4($55,000) + 0.6($70,000)] =
[$37,000, $64,000]
Now, the analyst would like to investigate the sensitivity of Cost per year of Build Large:
the Build large alternative and its expected payoff ($62,000) to Build large and High demand payoff = [30,000($5) −
the values of some of the inputs. The experts are asked for a $90,000, 30,000($5) − $70,000] = [$60,000, $80,000]
range of values for each input and the following are their
 uild large and Low (but increased as a result of reduced
B
answers [low, high]: High demand = [25,000, 35,000],
prices) demand payoff = [28,889($4.5) − $90,000,
Increased demand as a result of reduced prices = [15,000,
28,889($4.5) − $70,000] = [$40,000, $60,000]
30,000], Cost per year of Build large = [$70,000, $90,000].
Determine and draw the tornado chart.  xpected payoff of Build large = [0.4($40,000) + 0.6($60,000),
E
0.4($60,000) + 0.6($80,000)] = [$52,000, $72,000]
(continued )

ste2687X_ch05s_001-032.indd 17 12/02/21 4:36 PM

 5S-18 Supplement to Chapter 5 Decision Analysis

The ranges of expected payoff for changes in the above Even though “Increased demand as a result of reduced prices”
inputs are $30,000, $27,000, and $20,000, respectively, by is not at the top of the tornado chart, its expected downside payoff
chance in decreasing order of size. Therefore, the tornado chart may be of concern to the decision maker. Perhaps this can be fur-
is as follows: ther investigated and made into a random variable.

High demand

Increased demand

Cost per year

30 35 40 45 50 55 60 65 70 75 80 Expected payoff

OM in ACTION

Oglethorpe Power or high). The payoffs were measured in terms of net

present value.
Oglethorpe Power is an electricity generation and distri- The team estimated the probability of the states
bution cooperative supplying 20 percent of the electricity of nature for each random variable and the payoff for
in Georgia. Most of the remaining electricity in Georgia is each combination of decision alternatives and states
supplied by Georgia Power Co. Oglethorpe and Georgia of nature with the help of the planning staff (“experts”)
Power jointly own most of the power transmission lines within Oglethorpe. Using the DPL software, the alterna-
in Georgia, and supply their excess electricity to Florida. tive for each decision variable with highest expected
Some years ago, Florida Power and Light (FPL) indi- value was line (alone), control (Oglethorpe), and upgrade
cated to Oglethorpe that it was interested in the con- (no). Then, the risk profile (the distribution of NPV) of this
struction of another major transmission line between solution was determined by DPL. Because this showed
Florida and Georgia. Oglethorpe invited the consulting considerable possible negative NPV, the team identified
company Applied Decision Analysis Inc. to assist it in the random variable that affected the downside risk: the
making a decision. competitive situation in Florida. The next step was to col-
A team was formed to investigate the decision vari- lect further information on this random variable in order
ables, random variables, and their payoffs. Brainstorm- to reduce the range of values for the payoff. The team
ing resulted in an influence diagram and identification also performed sensitivity analysis of all the inputs (e.g.,
of three decision variables (alternatives): line (joint the discount rate) and found that the optimal decisions
with Georgia Power, alone, or no line), nature of con- [line (alone), control (Oglethorpe), and upgrade (no)] are
trol (Oglethorpe or FPL), and whether to upgrade the fairly robust. The team presented all its findings to top
associated facilities (joint with Georgia Power, alone, or management (including the risk profile of joint line and
no upgrade). Five random variables (states of nature) no line). With this information, top management started
were identified: construction cost (low, medium, or to negotiate with FPL.
high), competitive situation in Florida (good, fair, or bad), Applied Decision Analysis Inc. is now part of
Florida demand (low, medium, or high), Oglethorpe’s PricewaterhouseCoopers.
share of demand (very low, low, medium, high, or very Source: A. Borison, “Oglethorpe Power Corporation Decides About Invest-
high), and future spot price for electricity (low, medium, ing in a Major Transmission System,” Interfaces 25(2), pp. 25–36.

LO 5 Modelling Preferences/Utilities
Expected monetary payoffs may not reflect the preference and attitude of the decision maker toward
risk. In this case, utilities can be used. A utility is a quantification of a person’s value for various
utility A quantification of a payoffs. Empirical studies have identified that most people are risk averse (this is the reason why
person’s value for various payoffs. most people buy insurance). A common risk averse utility function is the exponential function:
​​U​(x​​ ​)​​ = 1 – ​e​​  – x/R​​​ (5S-5)

ste2687X_ch05s_001-032.indd 18 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-19

where x is the payoff and R is the risk tolerance. A wealthier person or larger organization has a
larger R. R is determined as follows: the decision maker is indifferent between zero payoff and a risk tolerance Measures the
gamble with a 50–50 chance of receiving R dollars or losing R/2 dollars. Professor Ron Howard amount of loss that a decision
found that R was approximately 6 percent of annual sales revenue of the large companies he studied.1 maker is willing to withstand: he/
Exponential utility has a downward sloping shape. For example, exponential utility with​ she is indifferent between zero
R = $100, 000​is as follows: payoff and a gamble with 50–50
chance of receiving R dollars or
1.5 losing R/2 dollars.

0.5

0
–100,000 –50,000 0 50,000 100,000 150,000 200,000 250,000 300,000

–0.5

–1

–1.5

–2 util The unit of a utility.

Consider the video arcade problem of Example 5S-2, but suppose that the decision maker has EXAMPLE 5S-7
exponential utility with risk tolerance R = $100,000. How would the answer change?

SOLUTION

We first need to translate the dollar payoffs in terms of a util (the unit of utility) for the decision maker. Specifically, U($40,000)
= 0.33, U($55,000) = 0.42, U(−$10,000) = −0.105, U($50,000) = 0.39, and U($70,000) = 0.5. Using these in the problem as
payoffs would present the following decision tree:

Demand

React
Build Low to high
demand
0.33
Small Do nothing
High 0.33
Expand

0.42

React
to low
Demand demand

Do nothing
Low –0.105
Large Reduce prices
0.39
High
0.5
(continued )

1
R.A. Howard, “Decision Analysis: Practice and Promise,” Management Science 34(6), pp. 679–695.

ste2687X_ch05s_001-032.indd 19 12/02/21 4:36 PM

 5S-20 Supplement to Chapter 5 Decision Analysis

The alternative that maximizes the expected utility of the decision maker is still Build large (note that DPL rounds the utils to one
decimal digit):
Demand
Small [0.4]
Build
[0.5] React to low demand Do nothing [–0.1]
Demand Low [0.4] –0.1
Large [0.5] 40% Reduce prices [0.4]
0.4
High [0.5]
60% 0.5

More exactly, the utility of Build large is 0.4(0.39) + 0.6(0.5) = 0.456.

As utils don’t have an intuitive interpretation, a util is frequently translated back into dollars.
certainty equivalent The This is called its certainty equivalent, the guaranteed amount of money that an individual would
guaranteed amount of money that view as equally desirable as a risky asset. The equation ​U ​​(x​​ ​)​​​ = 1 − e​ ​​  -x/R​​is used in reverse to do so:​
an individual would view as 1 − U​​(x​​ ​)​​​ = e​ ​​  -x/R​; ​take natural log of both sides :​-x/R = ln ​​(​​1 − U ​​(x​​ ​)​​​​)​​​​ or
equally desirable as a risky asset. x = –R ln(1 – U (x)) (5S-6)
For Example 5S-7, the certainty equivalent to U ​ ​​(x​​ ​)​​​ = 0.456​ is ​x = −$100, 000 ln  ​​(​​1 − 0.456​)​​​ =
$60, 881.​Note that this is only a little smaller than the expected payoff of $62,000. The difference is
called risk premium. In this case, risk premium is ​$62, 000 − $60, 881 = $1, 119.​

Multiple Objectives
Multiple objectives or attributes occur when the value of a consequence/outcome cannot be mea-
sured by one unit (i.e., dollars). These cases mostly occur for government agencies whose objectives
are not limited to dollars, but also include public safety and environmental stewardship. For exam-
ple, BC Hydro studied its strategic values and determined that they consist of economic, environ-
mental, health and safety, and service quality.2
Identifying the fundamental objective may require asking “why” a few times, as done for find-
ing the root cause of a quality problem. For example, the objective of transportation of dangerous
goods may appear to be “minimize distance travelled.” However, asking why is likely answered by
“to minimize accidents.” Again asking why is likely answered by “to minimize health impacts.”
Now the answer to why is simply that health impacts are important. Thus, the fundamental objective
is to minimize health impacts, whereas to minimize distance travelled and to minimize accidents are
not. This becomes clear when one considers that a shortest route may go through a populated area
and that many accidents outside a populated area may be better than one in a populated area.3
If the objectives can be measured, but not in one unit, it may be possible to compose a weighted
sum of the scores of each consequence/outcome and use that as payoff. For example, if a conse-
quence/outcome has $10 million economic cost and results in 5,000 hectares of lost forest, its payoff
may be measured as ​10 ​​(​​0.395​)​​​ + 5000​ ​(​​0.0025​)​​​ ​​(​​0.25​)​​​ = 3.95 + 3.125 = 7.075,​ where 0.395 and
0.25 are weights of economic and environment objectives, respectively, and 0.0025 is the scale used
for a hectare lost (in million dollars; in effect a hectare is worth $2,500).
If the objectives cannot be measured but weights for the objectives can be determined, then each
objective of each consequence/outcome has to be subjectively scored by the decision maker and a weighted
score for each consequence/outcome determined (similar to the factor rating method of Chapter 8).
If weights of objectives cannot be determined, an overall subjective score for each consequence/
outcome has to be determined (similar to a utility) and used. A related method is the analytic hierar-
chy process, which compares a pair of outcomes/consequences at a time and determines the weights
for objectives (factors) and scores of each consequence/outcome.
2
R.L. Keeney and T.L. McDaniels, “Value-Focused Thinking About Strategic Decisions at BC Hydro,” Interfaces,
22, 6, 94–109.
3
W. Edwards et al., Advances in Decision Analysis, Cambridge University Press, 2007, pp. 114–115.

ste2687X_ch05s_001-032.indd 20 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-21

Summary
• Decision analysis is a formal approach to Two concepts that can be used in deciding
decision making. It involves identifying the whether or not to buy information are
decision alternatives, chance events (random expected value of perfect information (EVPI)
variables) and their probabilities, and payoffs and expected value of sample information
(associated with consequences/outcomes). A (EVSI). EVSI uses Bayes’ rule to determine
visual tool useful for modelling the decision conditional probabilities.
problem is an influence diagram. A simple • Graphical sensitivity analysis can be used for
decision problem can be represented as a pay- one-variable problems to determine the effect
off table. A common evaluation method uses of changes in the probability assigned to a
the expected values. state of nature on the best alternative (called
• Complex decision problems are modelled as a sensitivity to probability). Also, sensitivity of
decision tree. In addition to using the expected expected payoff of the best alternative to an
values to determine the best alternative, the input (e.g., cost) can be determined and plot-
probability distribution of payoff for the best ted in the tornado chart (with the most impor-
alternative is determined and displayed (called tant input on the top of the chart).
the risk profile). Downside risk of payoff is a • Utilities are used to measure the preference
concern to the decision maker and may be of a decision maker for a payoff when risk is
reduced by gathering more information on the involved. Most people are risk averse. A
random variable causing it. A popular deci- common utility is the exponential function.
sion analysis software is DPL. Multiple objectives/attributes occur when it
• Information can be used to reduce the vari- is hard to measure the consequence in one
ability of the important random variables. unit.

Key Terms
Bayes’ rule (EVPI) risk tolerance
certainty equivalent expected value of sample information sensitivity to a probability
decision analysis (EVSI) sensitivity to an input
decision tree folding back state of nature
dialog decision process (DDP) influence diagram tornado chart
expected monetary value (EMV) payoff table util
expected value of perfect information risk profile utility

Solved Problems
Solved Problems 1–4 refer to this payoff (profits) table:
New Bridge Built No New Bridge
Alternative capacity A-small 1 14
for new store B-medium 2 10
C-large 4 6

Problem 1
Using graphical sensitivity to probability, determine the probability for “No New Bridge” state of
nature for which each alternative would be optimal.

ste2687X_ch05s_001-032.indd 21 12/02/21 4:36 PM

 5S-22 Supplement to Chapter 5 Decision Analysis

Solution
Plot a straight line for each alternative. Do this by plotting the payoff for “new bridge built” on the left
axis and the payoff for “no new bridge” on the right axis, and then connecting the two points with a
straight line. Each line represents the expected profit for an alternative for the entire range of probabil-
ity of “no new bridge.” Because the lines represent expected profit, the line that is highest for a given
value of P (no new bridge) is optimal. Thus, from the graph, you can see that for low values of this
probability alternative C is best, and for higher values alternative A is best (alternative B is never the
highest line, so it is never optimal; it is dominated).

Payoff if
new bridge 14
A Payoff if
no new
bridge

10
B

C 6

2
1 C best A best
0
0 0.27 1.0
P (no new bridge)

The dividing line between the ranges where alternatives C and A are optimal occurs where the two
lines intersect. To find that probability, first formulate the equation for each line. To do this, let the
intersection with the left axis be the y intercept; the slope equals the right-side payoff minus the left-
side payoff. Thus, for C you have ​4 + (​​
​​ 6 − 4​)​​ ​P,​which is ​4 + 2P. ​For ​A, 1 + ​​(​​14 − 1​)​​ ​P,​which is
1 + 13P. Setting these two equal to each other, you can solve for P:

​4 + 2P = 1 + 13P​
Solving, P = 3/11 = 0.27. Therefore, the ranges for P(no new bridge) for each alternative to be best
are:
A: 0.27 ≤ P ≤ 1.00
  B:
   never optimal​ ​​​
​​   
C: 0 ≤ P ≤ 0.27

Problem 2
Using the probabilities of 0.60 for a new bridge and 0.40 for no new bridge,
a. Calculate the expected value of each alternative, and identify the alternative that would be
selected using expected values.
b. Construct a decision tree for the problem showing expected values.

Solution
a. ​​A: 0.60(1) + 0.40(14) = 6.20 [best]​
B: 0.60(2) + 0.40(10) = 5.20​
C: 0.60(4) + 0.40(6) = 4.80 ​​

ste2687X_ch05s_001-032.indd 22 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-23

Expected
values

) $1 0.60(1) = 0.60
0. 60
e(
dg
w bri
Ne
$6.20 (best)

No n
ew b
ridg
e (0. $14 0.40(14) = 5.60
40)
A

.60
) $2 0.60(2) = 1.20
g e (0
w brid
Ne
B $5.20
No
new
brid
ge
(0.4
0) $10 0.40(10) = 4.00
C

0) $4 0.60(4) = 2.40
(0.6
br idge
New

$4.80
No
ne
wb
rid
ge
(0.
40
) $6 0.40(6) = 2.40

Problem 3
Calculate the EVPI using the data from the previous problem.

Solution
Using Formula 5S-1, the EVPI is the expected payoff with perfect information minus the expected
payoff without perfect information (i.e., the maximum expected value). The expected payoff with
perfect information involves multiplying the best payoff in each column by the column probability and
then summing those amounts. The best payoff in the first column is 4, and the best in the second
column is 14. Thus,

Expected payoff with perfect information = 0.60(4) + 0.40(14) = 8.00

The maximum expected value is 6.20 from the previous problem. Therefore,

EVPI = 8.00 − 6.20 = 1.80

Problem 4
Suppose that the values in the payoff table represent costs instead of profits.
a. Using sensitivity to a probability, determine the range of P(no new bridge) for which each alter-
native would be optimal.
b. If P(new bridge) = 0.60 and P(no new bridge) = 0.40, find the alternative to minimize expected
cost.

Solution
a. The graph is identical to that shown in Solved Problem 1. However, the lines now represent
expected costs, so the best alternative for a given value of P(no new bridge) is the lowest line.

ste2687X_ch05s_001-032.indd 23 12/02/21 4:36 PM

 5S-24 Supplement to Chapter 5 Decision Analysis

Hence, for very low values of P(no new bridge), A is best; for intermediate values, B is best; and
for high values, C is best. You can set the equations of A and B, and B and C, equal to each other
in order to determine the values of P(no new bridge) at their intersections.
Thus,
A = B: 1 + 13P = 2 + 8P; solving, P = 1 / 5 = 0.20
B = C: 2 + 8P = 4 + 2P; solving, P = 2 / 6 = 0.33
Hence, the ranges are:
A best: 0 ≤ P ≤ 0.20
B best: 0.20 ≤ P ≤ 0.33
C best: 0.33 ≤ P ≤ 1.00

Cost with
new bridge 14
A Cost with
no new
bridge

10
B

C 6

2 A B
1 best best
C best
0
0 0.20 0.33 1.0
P (no new bridge)

b. Expected value calculations are the same whether the values represent costs or profits. Hence,
the expected payoffs for costs are the same as the expected payoffs for profits that were calcu-
lated in Solved Problem 2. However, now you want the alternative that has the lowest expected
payoff rather than the one with the highest payoff. Consequently, alternative C is the best
because its expected payoff (4.80) is the lowest of the three (6.20, 5.20, 4.80).

Problem 5
A diagnostic test of a certain disease has 95 percent probability that the test will be positive for a
patient with the disease and 96 percent probability that the test will be negative for a healthy person.
Only 1 percent of the population has the disease in question. If the diagnostic test reports that a person
chosen at random tests positive, what is the conditional probability that the person does, in fact, have
the disease?

Solution
P(+test | disease) = 0.95
P(−test | no disease) = 0.96
P(disease) = 0.01
P(disease | +test) = ?
Bayes’ rule:
P(disease | +test)
= (P(+test | disease)P(disease))/(P(+test | disease)P(disease) + P(+test | no disease)P(no disease))
0.95(0.01)
= ________________________
​​   
    ​​ = 0.1935
0.95(0.01) + (1 − 0.96) (1 − 0.01)
or 19.35 percent, a small probability.

ste2687X_ch05s_001-032.indd 24 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-25

Discussion and Review Questions

LO1 1. What is decision analysis? LO4 9. What is sensitivity to a probability and sensitivity to an
input, and how can they be useful to a decision maker?
LO1 2. When should decision analysis be used?
LO4 10. What is a tornado chart?
LO1 3. Why is a structured approach to decision making important?
LO5 11. What is utility?
LO1 4. Under what circumstances is using expected value appro-
priate? When isn’t it appropriate? LO1 12. What is the dialogue decision process?

LO2 5. When using expected value, suppose there is a tie between LO1 13. Why do oil exploration companies need a structured
two alternatives. How should the decision maker make the approach to decision making?
decision?
LO1 14. Briefly describe Bayer Pharma’s decision process for
LO1 & 2 6. What is the difference between a decision tree and an introduction of a new drug.
influence diagram? When should each be used?
LO2 15. What is Unilever’s decision making under uncertainty
LO2 7. What is a risk profile? (DMUU)? What software does it use for DMUU?
LO3 8. Will expected value of perfect information always be greater LO4 16. Briefly describe Oglethorpe Power’s decision process for
than or equal to expected value of sample information? introduction of an electricity line.
Briefly explain.

Internet Exercises
LO1–4 1. Read http://www.decisionprofessionals.com/assets/storage LO1 2. View https://www.youtube.com/watch?v=dFV-lzIqfRA, and
/docs/Processes-and-tools-of-portfolio-mgmt.pdf and sum- identify the six requirements for decision quality.
marize how decision analysis is being used at Baxter
Healthcare.

Problems
LO1–3 1. A small building contractor has recently experienced two LO4 2. Refer to Problem 1. Construct a graph that will enable you
successive years in which work opportunities exceeded to perform sensitivity to a probability on the problem. Over
the firm’s capacity. The contractor must now make a what range of P(high) would each of the alternatives be best?
decision on capacity for next year. Estimated profit for
LO2–5 3. A company that plans to expand its product line must
each alternative under each of the two possible states of
decide whether to build a small or a large facility to produce
nature for next year’s demand is shown to the right.
the new products. If it builds a small facility and demand is
low, the net present value after deducting for building costs
Next Year’s Demand
will be $400,000. If demand is high, the company can either
Alternative Low High maintain the small facility or expand it. Expansion would
Do nothing $50* $60 have a net present value of $450,000, and maintaining the
small facility would have a net present value of $350,000.
Expand 20 80
If a large facility is built and demand is high, the esti-
Subcontract 40 70 mated net present value is $800,000. If demand turns out
*Profit in $ thousands. to be low, the net present value will be −$10,000.
The probability that demand will be high is estimated
Suppose that, after a certain amount of discussion, the to be 0.60, and the probability of low demand is estimated
­contractor is able to subjectively assess the probabilities of to be 0.40.
low and high demand as: P ​ ​​(​​low​)​​​ = 0.3 and P​​(​​high​)​​​ = 0.7.​ a. Analyze this decision problem using a decision tree.
a. Determine the expected profit of each alternative. b. Draw the risk profile (for the better alternative).
Which alternative is best? c. Calculate the EVPI. How could this information be used?
b. Analyze the problem using a decision tree. Show the d. Determine the range of P(demand low) over which
expected profit of each alternative on the tree. each alternative would be better.
c. Calculate the expected value of perfect information. e. Use an exponential utility with risk tolerance R =
How could the contractor use this knowledge? $300,000 to evaluate this problem.

ste2687X_ch05s_001-032.indd 25 12/02/21 4:36 PM

 5S-26 Supplement to Chapter 5 Decision Analysis

LO2 4. Determine the course of action that has the highest

expected payoff for the decision tree below.

$1.0*
)
.4
(0 g
an
d othin $1.3
Do n
dem
all
Sm (0.5
)
Expa
em and nd $1.3
edi um d
M
g
Larg othin $1.5
e dem Do n
and Expand
(0.1) $1.6
Build
ct

$1.8
tra
on
bc

g
othin $0.7
Su

Do n
Other use #1
$1.5
(0 .4)
and Othe
r use
em
all d #2 $1.0
Sm
Expand Medium demand (0.5)
$1.6
Lar
ge g
dem othin $1.6
and Do n
(0.1 Subcontract
) $1.5
Build
$1.7
Bu
ild

g
othin ($0.9)
Do n
Other use #1
.4) $1.4
ma nd (0 Othe
ll de r use
Sma #2 $1.0
Med
ium
dem g
and othin $1.0
La (0.5) Do n
rg Other use #1
e $1.1
de
m Othe
an r use
d #2 $0.9
(0
.1)

*Net present value in millions $2.4

LO1–3 5. The lease of an amusement park is about to expire. Man- Suppose that the management of the amusement park
agement must decide whether to renew the lease for has decided that there is a 75 percent probability that the
another 10 years or to relocate near the site of a proposed resort’s application will be approved.
resort. The town planning board is currently debating the a. If the management uses expected values, which alter-
merits of granting approval to the resort. A consultant has native should it choose?
estimated the net present value of amusement park’s two b. Represent this problem in the form of a decision tree
alternatives under each state of nature as shown below. and solve it.
Resort Resort
c. If the management has been offered the option of a
Options Approved Rejected temporary lease while the town planning board consid-
ers the resort’s application, would you advise the man-
Renew $3,000,000 $4,000,000
agement to sign the lease? The temporary lease will
(current location)
cost $200,000.
Relocate $5,000,000 $2,000,000
near resort

ste2687X_ch05s_001-032.indd 26 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-27

LO4 6. Construct a graph that can be used for sensitivity to a s­ ubcontract some production, which would have a net pre-
probability of the preceding problem. sent value of $110,000; the third option is to purchase a
a. How sensitive is the solution to the problem in terms second machine, which would have a net present value of
of the probability estimate of 0.75? $100,000.
b. Suppose that, after consulting with a member of the How many machines should the manager purchase
town planning board, management decides that the initially? Use a decision tree to analyze this problem.
estimate of resort approval should be 0.60. How LO2 9. Determine the course of action that has the highest
­sensitive is the solution to this revised estimate? expected value for the following decision tree.
Explain.
c. Suppose that management is confident of all the esti-
mated payoffs except for $5 million. If the probability 1/3
0
of resort approval is 0.60, for what range of payoff for 1/3
60
relocate/approved will the relocate alternative remain 1/3
90
optimal? 0.30
LO2–4 7. A company must decide whether to construct a small, 40
medium, or large plant. A consultant’s report indicates a 0.50
A 44
0.20 probability that demand will be low and a 0.80 prob- tive
na
ability that demand will be high. A lter 0.20 60
If the company builds a small facility and demand
turns out to be low, the net present value will be $42 mil-
1/3
lion; if demand turns out to be high, the company can (45)
either subcontract some production and realize the net 1/3

Al
45
te
­present value of $43 million or expand for a net present 1/3

value of $48 million. r

na
99
tiv 0.30
The company could build a medium-size facility: if e
B 40
demand turns out to be low, its net present value is esti-
0.50
mated at $22 million; if demand turns out to be high, the 50
company could do nothing and realize a net present value
of $46 million, or it could expand and realize a net present 30
0.20
value of $50 million. 1/2

If the company builds a large facility and demand 40

is low, the net present value will be −$20 million, 1/2
whereas high demand will result in a net present value 50
of $72 million.
a. Analyze this problem using a decision tree. What is
the best alternative? LO1–3 10. The director of social services of a province has learned
b. Calculate the EVPI and interpret it. that a new act has mandated additional information
c. Perform graphical sensitivity to P(high). requirements. This will place an additional burden on
the agency. The director has identified three alterna-
LO2 8. A manager must decide how many machines of a certain tives to handle the increased workload: (1) reassign
type to buy. The machines will be used to manufacture a present staff, (2) hire and train two new workers, and
new gear for which there is increased demand. The man- (3) redesign current practice so that workers can read-
ager has narrowed the decision to two alternatives: buy ily collect the information with little additional effort.
one machine or buy two machines. If only one machine An unknown factor is the caseload for the coming year
is purchased and demand is more than it can handle, a when the new data will be collected on a trial basis. The
­second machine can be purchased at a later time. However, estimated costs for various options and caseloads are
the cost per machine would be lower if the two machines shown in the following table:
were purchased at the same time.
The estimated probability of low demand is 0.30, and Caseload
the estimated probability of high demand is 0.70. (Cost in $ thousands)
The net present value associated with the purchase
Moderate High Very High
of two machines initially is $75,000 if demand is low and
$130,000 if demand is high. Reassign staff $50 60 85
The net present value for one machine and low New staff 60 60 60
demand is $90,000; if demand is high, there are three
Redesign data 40 50 90
options: one option is to do nothing, which would have
collection
a net present value of $90,000; a second option is to

ste2687X_ch05s_001-032.indd 27 12/02/21 4:36 PM

 5S-28 Supplement to Chapter 5 Decision Analysis

The director has decided that reasonable caseload prob- LO4 14. Given the following payoff table:
abilities are 0.10 for moderate, 0.30 for high, and 0.60 for
very high. State of Nature
a. Which alternative will yield minimum expected cost? #1 #2
b. Construct a decision tree for this problem and solve it.
A $120* 20
c. Determine the expected value of perfect information.
d. Suppose the director has the option of hiring an addi- Alternative B 60 40
tional worker if one worker is hired initially and the case- C 10 110
load turns out to be high or very high. Under that plan, D 90 90
the cost if demand is moderate will be $40(000), the cost
*In $ thousands.
if demand is high will be $70(000), and the cost if
demand is very high will also be $70(000). Construct a a. Determine the range of P(#2) for which each alterna-
decision tree that shows this additional alternative and tive would be best, treating the payoffs as profits.
determine which alternative will minimize expected cost. b. Answer part a, treating the payoffs as costs.
LO4 11. A manager has compiled estimated profits for various LO1 & 4 15. A law firm is representing a company that is being sued
capacity alternatives but is reluctant to assign probabilities by a customer. The decision is whether to go through liti-
to the states of nature. The payoff table is: gation or settle the case out of court. The cost of losing
the case is estimated to be $1 million, whereas the cost of
State of Nature settling the case is $200,000. Winning the case in court
#1 #2 would result in no loss. What is the minimum chance of
A $20* 140 winning for which the company should contest the case?
Alternative B 120 80 LO2 16. An 18-year-old youth has just arrived at a hospital com-
C 100 40 plaining of abdominal pain. The medical findings are
*In $ thousands. consistent with appendicitis but not totally typical of
appendicitis. The lab and X-ray test results are not clear.
a. Plot the expected profit lines on a graph as a function The surgeon is wondering whether to operate or wait
of P(#2). 12 hours (in case it is not appendicitis). In this case, if the
b. Is there any alternative that would never be appropriate? pain does not recede, then it must be appendicitis and the
c. For what range of P(#2) would alternative A be the surgeon will operate. The probability that it is a­ ppendicitis
is 56 percent using past experience with similar cases.
best choice?
Also, from similar cases, the probability that the appen-
d. For what range of P(#1) would alternative A be the
dicitis will perforate during the 12 hours of waiting is
best choice? 6 ­percent. The payoffs are measured in terms of death rate.
LO4 12. Repeat all parts of Problem 11, assuming the values in the The death rate of operating when appendicitis is present
payoff table are estimated costs and the goal is to mini- is 0.09 percent; it is 0.04 percent when appendicitis is not
mize expected cost. present. The death rate of operating on a perforated appen-
dicitis is 0.64 percent. Draw the decision tree and deter-
LO4 13. Consider the following payoff table of estimated NPV for
mine the best course of action in this case.4
four alternatives:
LO1 & 4 17. An oil company has some land that may contain oil.5 The
State of State of oil reserve and its probability (based on experience in the
Nature 1 Nature 2 area) are:
#1 $10* −2
Alternative #2 8 3 Oil reserve 500 200 50 0
#3 5 5 (in 1,000 barrels)

#4 0 7 Probability 0.1 0.15 0.25 0.50

*In $ thousands. The company has to decide whether to drill for oil
or lease the land to a rancher. The cost of drilling will
Relative to the probability of state of nature 2, deter-
be $500,000 if it is a dry well or $750,000 if it is oil-­
mine the range of probability for which each of the alter-
producing. The profit per barrel will be $45. The revenue
natives would maximize the expected NPV.
for leasing the land to the rancher will be $3,000,000.

4
R. Clarke, “The Application of Decision Analysis to Clinical Medicine,” Interfaces 17(2), pp. 27–34.
5
F. S. Hillier and G. J. Lieberman, Operations Research, 2nd ed, San Francisco: Holden-Day, p. 598.

ste2687X_ch05s_001-032.indd 28 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-29

a. Determine the best decision for the oil company. LO2 & 3 21. An offshore underwater structure has tested positive
b. If the cost of drilling a dry well can range between for oil by two exploratory wells drilled near the crest of
$400,000 and $600,000, cost of drilling an oil-producing the ­structure.9 Management is now trying to determine
well can range between $600,000 and $900,000, and whether it is wise to make a definite commitment to begin
profit per barrel can range between $30 and $60, draw building a platform, and, if so, what size, or to defer the
decision to allow for the drilling of an additional explora-
the tornado chart.
tory well. The principal uncertainty is the size of the field.
LO2 & 3 18. The board of governors of Santa Clara University is con- If the structure is nearly filled with oil, the area of the
templating mandatory testing of its student athletes.6 The field will be about 15 square miles and it would require
test is not 100 percent accurate. The conditional probabili- installation of a large platform costing $50 million. If the
ties are as follows: If an athlete uses drugs, then the test structure is only partially filled with oil, the area will be
will be positive 94 percent of the time, whereas if he does much smaller and the field could be depleted with fewer
not (i.e., he is a non-user), the test will be negative 98 per- wells and a smaller platform costing $30 million. Based
cent of the time. Suppose the board suspects that 4 percent on the subjective judgment of the company’s geologists,
of the athletes use drugs. the likelihood of the structure being completely filled is
a. Calculate the posterior probabilities (i.e., probability 0.4 (i.e., field is large). If a small platform is built and the
of correct identification after a drug test). field turns out to be large, a second small platform will
b. If a test costs $50, the cost of not identifying (and bar- have to be built at the same cost. If another exploratory
ring) a drug user is $1,000, the cost of falsely accusing well is drilled along the flanks of the structure, there is a
10 percent chance that the result will be erroneous. How-
a non-user is $100, and other costs are zero, should the
ever, management will base its platform size decision on
university test any athlete? Draw and use a decision
the result of the exploratory well. The exploratory well
tree. will cost $2 million.
LO2 & 4 19. An electric utility is trying to decide whether to replace a. Calculate the posterior probabilities given that the
the PCB transformer in a generating station with a new exploratory well indicates a small field.
and safer one.7 To evaluate this decision, the utility needs b. Which decision has the lowest expected cost? Draw
information about the likelihood of an incident, such as a the decision tree.
fire, the cost of such an incident, and the cost of replacing
the transformer. Suppose that the total cost of replacement LO3 22. The management of a printed circuit board (PCB) manu-
is $75,000. If the transformer is replaced, there is virtually facturer is concerned about the significant number of bad
no chance of a fire. However, if the current transformer is (defective) PCBs that are being shipped to the company’s
retained, the probability of a fire is assessed to be 0.0025. If customers.10 The quality control procedure for the PCBs
a fire occurs, then the clean-up cost could be high ($80 mil- is as follows: each PCB is sent to a testing station where
lion) or low ($5 million). The probability of a high clean-up it is tested using an electronic instrument; those that pass
cost, given that a fire occurs, is assessed to be 0.2. are shipped to the customers; those that fail are retested;
a. Should the company replace the transformer? those that pass the second test are shipped to the custom-
ers; those that fail the second time are retested again;
b. Perform sensitivity analysis to the probability of a fire.
those that pass the third test are shipped to customers, and
Does the optimal decision from part a remain optimal
those that fail the third time are discarded. A consultant
for a wide range of values for this probability? is brought in to examine the situation. She studied 1,136
LO3 20. A 59-year-old executive has just completed his annual PCBs. Of these, 1,087 passed the quality test at the first
physical exam.8 The exam included a test on a sample of attempt. The 49 that failed were retested, and 37 passed.
his stool to screen for cancer of the large intestine (colon The remaining PCBs all failed the third time and were
cancer). The test looks for small amounts of blood in the ­discarded. The consultant asked for retesting of the 37 that
stool. The executive is informed that the test showed pres- passed the second time. After several tests, it was discov-
ence of blood in his stool and he is quite anxious. The like- ered that 16 were defective. Also, the consultant asked for
lihood of any 59-year-old man having colon cancer is 1 in retesting of those that passed the first time. A sample of
1,000. The test detects 85 percent of cases of colon c­ ancer, 200 was retested several times and it was discovered that
but 2 percent of patients without cancer have a small 7 were defective.
amount of blood in their stool. What is the probability that a. Calculate the probability of a good (non-defective)
the man has colon cancer? PCB passing the test the first time and the probability
6
C. D. Feinstein, “Deciding Whether to Test Student Athletes for Drug Use,” Interfaces 20(3), pp. 80–87.
7
W. E. Balson et al., “Using Decision Analysis and Risk Analysis to Manage Utility Environmental Risk,”
Interfaces 22(6), pp. 126–139.
8
P. H. Hill et al., Making Decisions, Massachusetts: Addison-Wesley, p. 170.
9
P. D. Newendorp, Decision Analysis for Petroleum Exploration, Tulsa, Oklahoma: PennWell Publishing,
p. 524.
10
P. C. Bell, Management Science/Operations Research: A Strategic Perspective, Cincinnati: South-Western
College Publishing, p. 90.

ste2687X_ch05s_001-032.indd 29 12/02/21 4:36 PM

 5S-30 Supplement to Chapter 5 Decision Analysis

of a bad (defective) PCB failing the test the first time. the data (number of claims for a specific period) was clas-
Also, calculate the proportion of PCBs that are good sified as follows:
(non-defective) and the proportion that are bad
Claim Does
(defective).
Converts Not Convert
b. Propose a testing procedure that improves the quality
Classified as high risk 3,580 13,396
of PCBs sent to the customers.
Classified as low risk 1,479 140,645
LO2 23. Decision analysis can be used to identify best decisions
for a baseball manager. For example, suppose the team
a. If a worker is classified as high risk, what is the prob-
has a player on first base and no outs. Most manag-
ers choose to bunt, unless the batter has a particularly ability that his/her claim converts?
high batting ­average. Suppose the batter’s average is b. If a worker is classified as low risk, what is the prob-
not high and the manager has the following statistics ability that his/her claim does not convert?
for the average ­number of runs his team scored in the LO3 25. The fecal occult blood test examines stool sample for
inning given a particular base occupied and a particular blood.12 It is used to test for colon and rectal cancer. A
­number of outs: study found that out of 15,000 people over 50 years old
tested, 10 percent had blood in their stool. These 10 per-
Expected cent underwent further testing including colonoscopy.
No. of Runs Bases Occupied Only 2.5 percent of those had cancer. Also, 5 out of 1,000
Outs None First Second who tested negative in the fecal blood test did eventually
develop cancer. What is the probability of a positive result
0 0.46 0.81 1.19
given cancer and the probability of a negative result given
1 0.24 0.50 0.67
no cancer of fecal blood test?
2 0.10 0.22 0.30
LO3 26. A physician has to decide whether to treat a 15-year-old
with sore throat, suspected of being caused by Streptococ-
a. Explain why, in general, bunting with a player on first
cus, with antibiotics.13 Failure to treat strep may lead to
base and no outs does not maximize the expected
rheumatic heart disease RHD (which leads to early death).
number of runs scored in the inning. Assume that the However, taking antibiotics may lead to allergic reaction
batter will be thrown out at first base. (anaphylaxis) leading to death. Probability of anaphylaxis
b. When, in general, would bunting with a player on first is 0.0003. If strep is present and there is no treatment,
base and no outs be a good decision? probability of RHD is 0.0063, but if there is treatment,
c. Assume there is a player on first base and no outs. probability of RHD is 0.0013. The physician estimates
What probability of stealing second base makes steal- the probability of strep as 0.6. The physician uses quality-
ing a good decision? adjusted life expectancy (QALE) as payoff and estimates
QALE as follows: death due to anaphylaxis = 0 year, early
LO3 24. The Workers’ Compensation Board of British Columbia death due to RHD = 25 years, no RHD or no strep = 60
receives claims from injured workers for support.11 Some years. Should antibiotics be prescribed?
of the claims convert from short-term disability (STD) to
long-term disability (LTD), which costs a lot more. So LO3 & 5 27. A drilling prospect has 15 percent chance of success, in
early detection of these cases helps reduce the costs. A which case it is estimated that 6 billion cubic feet (BCF) of
study was done to find a way to classify the claims into natural gas can be recovered.14 No drilling has a payoff of 0.
high risk (of conversion) and low risk (of conversion). The Dry hole has payoff of −$725,000. If successful, the com-
factors identified were the age of worker and the number pany can decide not to drill further, in which case the payoff
of days a worker is on STD. These vary by type of injury. will be $2,450,000. Otherwise, a second well can be drilled
Focusing on sprains and strains, threshold days for STD, that will have a 60 percent chance of increasing the total
given age of worker, were determined based on costs and find to 35 BCF of natural gas for a payoff of $26,500,000. If
historical data. Given the determined classification rule, the second hole is dry, the payoff will be $485,000.

11
E. Urbanovich et al., “Early Detection of High-Risk Claims at the Workers’ Compensation Board of
­British Columbia,” Interfaces, 33(4), 15–26.
12
R.T. Clemen and T. Reilly, Making Hard Decisions, 2nd edition, Duxbury (Brooks/Cole), Pacific Grove,
California, 2001, pp. 288–289.
13
R.F. Nease, JR and D.K. Owens, “Use of Influence Diagrams to Structure Medical Decisions,” Medical
Decision Making, 17, 263–275.
14
M.R. Walls, et al., “Decision Analysis of Exploration Opportunities in the Onshore US at Phillips Petroleum
Company,” Interfaces 25, 6, 39–56.

ste2687X_ch05s_001-032.indd 30 12/02/21 4:36 PM

 Supplement to Chapter 5 Decision Analysis 5S-31

a. Using expected values, determine the best alternatives. inventory count (and actually to decrease it if a sample
b. Assuming exponential utility with risk tolerance of showed high records accuracy) by using the following
$16,667,000, what are the best alternatives? Determine example. Suppose the record accuracy in a warehouse is
the certainty equivalent. 97 percent; that is, 97 percent of the items have an inven-
tory record that matches the actual number of the items
LO3 28. A large distributor with many warehouses had inaccurate in stock. If a record is accurate, then the probability that
inventory records despite performing a physical inven- the count will confirm its accuracy is estimated to be 95
tory count once a year.15 The management wanted to percent, in which case nothing is done; else the accurate
increase the frequency of inventory counts but an ana- record will be set to the inaccurate count. If a record is
lyst’s view was that this would actually make the records inaccurate, then the probability that the count will indi-
more inaccurate. His idea was based on the fact that cate falsely that it is accurate is assumed to be 20 per-
the physical inventory count was itself inaccurate, and cent, in which case nothing is done; else the inaccurate
because its results were used to change the records, it record is fixed. Calculate the accuracy of records after a
actually contributed to record inaccuracy. He convinced physical count. Should a physical count be conducted in
the management not to increase the frequency of physical this case?

Mini-Cases
MINI-CASE

Southern Company tives: (1) do nothing, (2) use oil analysis only, (3) replace oil only,
or (4) replace oil and do oil analysis. For option (1) the probability
Fleet managers have a large pool of cars and trucks to maintain.16 of failure is 0.1, and the cost of failure is $1,200. For option (2), the
One approach to the vehicle maintenance is to use periodic oil probability of failure remains at 0.1. If the unit is about to fail, the
analysis: the oil from the engine and transmission are subjected oil analysis will indicate this with probability 0.7; if the unit is not
periodically to a test. These tests can sometimes signal an about to fail, the oil analysis will indicate this with probability 0.8.
impending failure (for example, iron particles in the oil), and pre- The oil analysis itself costs $20, and if it indicates that failure is
ventive maintenance is then performed (at a relatively low cost), about to occur, the oil will be changed at the cost of $14.80 and
eliminating the risk of failure (failure would result in a relatively preventive maintenance will be performed. The cost of preventive
high cost). However, oil analysis costs money and it is not per- maintenance to restore a unit that is about to fail is $500, whereas
fect—it can indicate that a unit is defective when in fact it is not, the cost of maintenance for a unit that is not about to fail is $250.
and it can indicate that a unit is non-defective when in fact it is. For options (3) and (4), probability of failure decreases from 0.1 to
As a possible substitute for oil analysis, the company could sim- 0.04. Analyze this decision problem.
ply change the oil periodically, thereby reducing the probability
of failure. Question
The fleet manager for the Southern Company, an electrical Using a decision tree, which alternative should the fleet manager
utility based in Atlanta (parent of Georgia Power), has four alterna- for the Southern Company choose?

15
Ido Millet, “A Novena to Saint Anthony, or How to Find Inventory by Not Looking,” Interfaces 24(2),
pp. 69–75.
16
J. M. Mellichamp, “The Southern Company Uses a Probability Model for Cost Justification of Oil Sample
Analysis,” Interfaces 23(3), pp. 118–124.

ste2687X_ch05s_001-032.indd 31 12/02/21 4:36 PM

 5S-32 Supplement to Chapter 5 Decision Analysis

MINI-CASE

ICI Canada The primary economic factors and their expense/gain (in million
dollars) were the following:
ICI’s Canadian subsidiary (now part of AkzoNobel) discovered a
new but unpatentable application for a chemical agent to reduce • The marketing development cost to determine whether
pulp-mill water pollution.17 However, everything was quite there is a significant market, and research expenses to
uncertain, and the management was trying to decide whether to identify a new production process for the product (C1):
go ahead with its R&D or abandon the product. The following $1 ± 25%
questions indicate the primary risks: • The process development costs, including pre-sanction
• Would market tests confirm that there is a significant engineering and commercial development (C2): $3.5 ± 25%
market for the product and could the company develop • The commercial development costs after the board’s
a new process for making this product—that is, is it sanction (C3): $1.0 ± 25%
technically feasible? • The venture value (net present value) if successful (R):
• After a production process is developed, would the $25 ± 50%
company’s board sanction production on a commercial
The plus-or-minus values indicate management’s considerable
scale?
uncertainty about the values.
• Would the venture turn out to be commercially successful?
Question
The management team assumed that each of these questions had
Should management go ahead with R&D for this product? How
a yes or no answer. The probabilities of yes answers are shown
can this question be fully answered?
below. The plus-or-minus values indicate management’s uncer-
tainty about the true probabilities.

Significant market and technical feasibility (P1) 0.36 ± 0.09

Board sanctions plant (P2) 0.8 ± 0.2
Commercial success (P3) 0.8 ± 0.2

17
S. W. Hess, “Swinging on the Branch of a Tree: Project Selection Applications,” Interfaces 23(6),
pp. 5–12.

ste2687X_ch05s_001-032.indd 32 12/02/21 4:36 PM