Decision Making Under Uncertainty: Decision and Risk Analysis at Du Pont

C H A P T E R 10
Decision
Making
Under
Uncertainty
DECISION AND RISK ANALYSIS AT DU PONT

ormal decision analysis in the face of uncertainty frequently occurs
F at the most strategic levels of a company’s planning process and

typically involves teams of high-level managers from all areas of the
company. This is certainly the case with Du Pont, as reported by two
internal decision analysis experts, Krumm and Rolle (1992), in their article
“Management and Application of Decision and Risk Analysis in Du Pont.”
Du Pont’s formal use of decision analysis began in the 1960s, but because
of a lack of computing power and distrust of the method by senior-level
management, it never really got a foothold. However, by the mid-1980s
things had changed considerably. The company was involved in a
faster-moving, more uncertain environment, more people throughout the
company were empowered to make decisions, and these decisions had to
be made more quickly. In addition, the computing power had arrived to
493
make large-scale quantitative analysis feasible. Since that time, Du Pont has embraced
formal decision-making analysis in all its businesses, and the trend is almost certain to
continue.
The article describes a typical example of decision analysis within the company.
One of Du Pont’s businesses, Business Z (so-called for reasons of confidentiality), was
stagnating. It was not set up to respond quickly to changing customer demands, and
its financial position was declining due to lower prices and market share. A decision
board and a project team were empowered to turn things around. The project team
developed a detailed timetable to accomplish three basic steps: frame the problem,
assess uncertainties and perform the analysis, and implement the recommended deci-
sion. The first step involved setting up a “strategy table” to list the possible strategies
and the factors that would affect or be affected by them. The three basic strategies
were (1) a base-case strategy (continue operating as is), (2) a product differentiation
strategy (develop new products), and (3) a cost leadership strategy (shut down the
plant and streamline the product line).
In the second step, the team asked a variety of experts throughout the company for
their assessments of the likelihood of key uncertain events. In the analysis step they then
used all of the information gained to determine the strategy with the largest expected
net present value. Two important aspects of this analysis step were the extensive use of
sensitivity analysis (many what-if questions) and the emergence of new “hybrid” strat-
egies that dominated the strategies that had been considered to that point. In partic-
ular, the team finally decided on a product differentiation strategy that also decreased
costs by shutting down some facilities in each plant.
By the time of the third step, implementation, the decision board needed little
convincing. Since all of the key people had been given the opportunity to provide input
to the process, everyone was convinced that the right strategy had been selected. All
that was left was to put the plan in motion and monitor its results. The results were
impressive. Business Z made a complete turnaround, and its net present value in-
creased by close to $200 million. Besides this tangible benefit, there were definite
intangible benefits from the overall process. As Du Pont’s vice president for finance
said, “The D&RA [decision and risk analysis] process improved communication within
the business team as well as between the team and corporate management, resulting in
rapid approval and execution. As a decision maker, I highly value such a clear and
logical approach to making choices under uncertainty and will continue to use D&RA
whenever possible.” ■
10.1 INTRODUCTION
n this chapter we will provide a formal framework for analyzing decision prob-
I lems that involve uncertainty. We will discuss the most frequently used criteria for
choosing among alternative decisions, how probabilities are used in the decision-
making process, how decisions made at an early stage affect decisions made at a later
stage, how a decision maker can quantify the value of information, and how attitudes
toward risk can affect the analysis. Throughout, we will employ a powerful graphi-
cal tool—decision trees—to guide the analysis. A decision tree enables the decision
maker to view all important aspects of the problem at once: the decision alternatives,
the uncertain outcomes and their probabilities, the economic consequences, and the
chronological order of events. We will show how to implement decision trees in Ex-
cel by taking advantage of a very powerful and flexible add-in from Palisade called
PrecisionTree.
494 Chapter 10 Decision Making Under Uncertainty

Many examples of decision making under uncertainty exist in the business world.
Here are several examples.
■ Companies routinely place bids for contracts to complete a certain project within
a fixed time frame. Often these are sealed bids, where each of several companies
presents in a sealed envelope a bid for completing the project; then the envelopes
are opened, and the low bidder is awarded the bid amount to complete the project.
Any particular company in the bidding competition must deal with the possible
uncertainty of its actual cost of completing the project (should it win the bid), as
well as the uncertainty involved in what the other companies will bid. The trade-off
is between bidding low in order to win the bid and bidding high in order to make a
profit.
■ Whenever a company contemplates introducing a new product into the market,
there are a number of uncertainties that affect the decision, probably the most
important being the customers’ reaction to this product. If the product generates
high customer demand, then the company will make a large profit. But if demand is
low (and, after all, the vast majority of new products do poorly), then the company
might not even recoup its development costs. Because the level of customer demand
is critical, the company might try to gauge this level by test marketing the product
in one region of the country. If this test market is a success, the company can then
be more optimistic that a full-scale national marketing of the product will also be
successful. But if the test market is a failure, the company can cut its losses by
abandoning the product.
■ Borison (1995) describes an application of formal decision analysis by Oglethorpe
Power Corporation (OPC), a Georgia-based electricity supplier. The basic decision
OPC faced was whether to build a new transmission line to supply large amounts
of electricity to parts of Florida and, if they decided to build it, how to finance this
project. OPC had to deal with several sources of uncertainty: the cost of building
new facilities, the demand for power in Florida, and various market conditions,
such as the spot price of electricity.
■ Ulvila (1987) describes the decision analysis performed by the U.S. Postal Service
regarding the purchase of automation equipment. One of the investment decisions
was which type of OCR (optical character recognition) equipment the Postal Service
should purchase (or convert) for reading single- and/or multiple-line addresses on
packages. An important factor in this decision was the level of use by businesses
of the “zip+4” (nine-digit zip codes). Zip+4 usage had been recommended for
some time but was used only sporadically. The Postal Service was uncertain about
the future level of business zip+4 usage. If businesses used the nine-digit codes
heavily in the future, then a certain type of (expensive) OCR equipment would be
most economical. If business use of zip+4 did not increase, then purchasing this
equipment would be a waste of money. The decision was an extremely important
one, given the expense of the proposed equipment and the fact that the Postal
Service would have to live with whatever equipment it purchased for a number of
years.
■ Utility companies must make many decisions that have significant environmental
and economic consequences. [Balson et al. (1992) provide a good discussion of
such consequences.] For these companies it is not necessarily enough to conform
to federal or state environmental regulations. Recent court decisions have found
companies liable—for huge settlements—when accidents occurred, even though
the companies followed all existing regulations. Therefore, when utility companies
decide, say, whether to replace equipment or mitigate the effects of environmental
10.1 Introduction 495

pollution, they must take into account the possible environmental consequences
(such as injuries to people) as well as economic consequences (such as lawsuits).
An aspect of these situations that makes decision analysis particularly difficult is that
the potential “disasters” are often extremely improbable; hence, their likelihoods
are very difficult to assess accurately.
10.2 ELEMENTS OF A DECISION ANALYSIS

lthough decision making under uncertainty occurs in a wide variety of con-
A texts, all problems have three elements in common: (1) the set of decisions (or
strategies) available to the decision maker, (2) the set of possible outcomes
and the probabilities of these outcomes, and (3) a value model that prescribes results,
usually monetary values, for the various combinations of decisions and outcomes.
Once these elements are known, the decision maker can find an “optimal” decision,
depending on the optimality criterion chosen. Rather than discussing these elements in
the abstract, we introduce them in the context of the following extended example.
E XAMPLE 10.1
BIDDING FOR A GOVERNMENT CONTRACT
AT SCITOOLS
SciTools Incorporated, a company that specializes in scientific instruments, has been
invited to make a bid on a government contract. The contract calls for a specific number
of these instruments to be delivered during the coming year. The bids must be sealed (so
that no company knows what the others are bidding), and the low bid wins the contract.
SciTools estimates that it will cost $5000 to prepare a bid and $95,000 to supply the
instruments if it wins the contract. On the basis of past contracts of this type, SciTools
believes that the possible low bids from the competition, if there is any competition,
and the associated probabilities are those shown in Table 10.1. In addition, SciTools
believes there is a 30% chance that there will be no competing bids.
TABLE 10.1 Data for Bidding Example

Low Bid Probability
Less than $115,000 0.2
Between $115,000 and $120,000 0.4
Between $120,000 and $125,000 0.3
Greater than $125,000 0.1
Solution
Let’s discuss the three elements of SciTools’ problem. First, SciTools has two basic
strategies: submit a bid or do not submit a bid. If SciTools submits a bid, then it must
decide how much to bid. Based on SciTools’ cost to prepare the bid and its cost to
supply the instruments, there is obviously no point in bidding less than $100,000—
SciTools wouldn’t make a profit even if it won the bid. Although any bid amount over

$100,000 might be considered, the data in Table 10.1 might persuade SciTools to limit
its choices to $115,000, $120,000, and $125,000.1
The next element of the problem involves the uncertain outcomes and their proba-
bilities. We have assumed that SciTools knows exactly how much it will cost to prepare
a bid and how much it will cost to supply the instruments if it wins the bid. (In re-
ality, these are probably estimates of the actual costs.) Therefore, the only source of
uncertainty is the behavior of the competitors—will they bid, and if so, how much?
From SciTools’ standpoint, this is difficult information to obtain. The behavior of the
competitors depends on (1) how many competitors are likely to bid and (2) how the
competitors assess their costs of supplying the instruments. However, we will assume
that SciTools has been involved in similar bidding contests in the past and can, there-
fore, predict competitor behavior from past competitor behavior. The result of such
prediction is the assessed probability distribution in Table 10.1 and the 30% estimate
of the probability of no competing bids.
The last element of the problem is the value model that transforms decisions
and outcomes into monetary values for SciTools. The value model is straightforward
in this example, but it can become quite complex in other applications, especially
when the time value of money is involved and some quantities (such as the costs of
environmental pollution) are difficult to quantify. If SciTools decides right now not to
bid, then its monetary value is $0—no gain, no loss. If it makes a bid and is underbid by
a competitor, then it loses $5000, the cost of preparing the bid. If it bids B dollars and
wins the contract, then it makes a profit of B − $100,000, that is, B dollars for winning
the bid, less $5000 for preparing the bid, less $95,000 for supplying the instruments.
For example, if it bids $115,000 and the lowest competing bid, if any, is greater than
$115,000, then SciTools makes a profit of $15,000.
It is often convenient to list the monetary outcomes in a payoff table, as shown
in Table 10.2. For each possible decision and each possible outcome, the payoff table
lists the monetary value to SciTools, where a positive value represents a profit and a
negative value represents a loss. At the bottom of the table, we list the probabilities of
the various outcomes. For example, the probability that the competitors’ low bid is less
than $115,000 is 0.7 (the probability of at least one competing bid) multiplied by 0.2
(the probability that the lowest competing bid is less than $115,000, given that there is
at least one competing bid).
It is sometimes possible to simplify payoff tables to better understand the essence
of the problem. In the present example, if SciTools bids, then the only necessary
information about the competitors’ bid is whether it is lower or higher than SciTools’
TABLE 10.2 Payoff Table for SciTools Bidding Example

Competitors’ Low Bid ($1000s)
No Bid <115 >115, <120 >120, <125 >125

SciTools’ No Bid 0 0 0 0 0
Bid 115 15 −5 15 15 15
($1000s) 120 20 −5 −5 20 20
125 25 −5 −5 −5 25
Probability 0.3 0.7(0.2) 0.7(0.4) 0.7(0.3) 0.7(0.1)
1
The problem with a bid such as $117,000 is that the data in Table 10.1 make it impossible to calculate the
probability of SciTools winning the contract if it bids this amount. Other than this, however, there is
nothing that rules out such an “in-between” bid.
10.2 Elements of a Decision Analysis 497

bid. That is, SciTools cares only whether it wins the contract or not. Therefore, an
alternative way of presenting the payoff table is shown in Table 10.3.
The third and fourth columns of this table indicate the payoffs to SciTools, depend-
ing on whether it wins or loses the bid. The rightmost column shows the probability
that SciTools wins the bid for each possible decision. For example, if SciTools bids
$120,000, then it wins the bid if there are no competing bids (probability 0.3) or
if there are competing bids but the lowest of these is greater than $120,000 (proba-
bility 0.7(0.3 + 0.1)). In this case the total probability that SciTools wins the bid is
0.3 + 0.28 = 0.58.
TABLE 10.3 Alternative Payoff Table for SciTools Bidding Example

Monetary value
Probability That
SciTools Wins SciTools Loses SciTools Wins
SciTools’ No bid NA 0 0.00
Bid 115 15 −5 0.86
($1000s) 120 20 −5 0.58
125 25 −5 0.37
Risk Profiles From Table 10.3 we can obtain risk profiles for each of SciTools’ deci-
sions. A risk profile simply lists all possible monetary values and their corresponding
probabilities. For example, if SciTools bids $120,000, there are two monetary values
possible, a profit of $20,000 or a loss of $5000, and their probabilities are 0.58 and
0.42, respectively. On the other hand, if SciTools decides not to bid, there is a sure
monetary value of $0—no profit, no loss.
A risk profile can also be illustrated graphically as a bar chart. There is a bar above
each possible monetary value with height proportional to the probability of that value.
For example, the risk profile for a $120,000 bid decision is a bar chart with two bars,
one above −$5000 with height 0.42 and one above $20,000 with height 0.58. The risk
profile for the “no bid” decision is even simpler. It has a single bar above $0 with
height 1. We have not shown these bar charts for this example because they are so
simple, but in more complex examples they can provide very useful information.
Expected Monetary Value (EMV) From the information we have discussed so far,
it is not at all obvious which decision SciTools should make. The “no bid” decision
is certainly safe, but it is certain to make zero profit. If SciTools decides to bid, the
probability that it will lose $5000 is smallest with the $115,000 bid, but this bid has the
smallest potential profit. Of course, if SciTools knew what the competitors were going
to do, its decision would be easy. However, this uncertainty is the defining aspect of the
problems in this chapter. The decision must be made before the uncertainty is resolved.
The most common way to make the choice is to calculate the expected monetary
value (EMV) of each alternative and then choose the alternative with the largest EMV.
The EMV is a weighted average of the possible monetary values, weighted by their
probabilities. Formally, if vi is the monetary value corresponding to outcome i and pi
is its probability, then EMV is defined as

EMV = vi pi
In words, EMV is the mean of the probability distribution of possible monetary out-
comes.

TABLE 10.4 EMVs for SciTools Bidding Example
Alternative EMV Calculation EMV
No bid 0(1) $0
Bid $115,000 15,000(0.86) + (−5000)(0.14) $12,200
Bid $120,000 20,000(0.58) + (−5000)(0.42) $9,500
Bid $125,000 25,000(0.37) + (−5000)(0.63) $6,100
The EMVs for SciTools’ problem are listed in Table 10.4. They indicate that if
SciTools uses the EMV criterion for making its decision, it should bid $115,000, as
this yields the largest EMV.
It is very important to understand what an EMV implies and what it does not imply.
If SciTools bids $115,000, then its EMV is $12,200. However, SciTools will certainly
not earn a profit of $12,200. It will earn $15,000 or it will lose $5000. So what does the
EMV of $12,200 really mean? It means that if SciTools could enter many “gambles”
like this, where on each gamble it would win $15,000 with probability 0.86 or lose
$5000 with probability 0.14, then on average it would win $12,200 per gamble. In
other words, the EMV can be interpreted as a long-term average.
It might seem peculiar, then, to base a one-time decision on EMV, which represents
a long-term average. There are two ways to explain this apparent inconsistency. First,
most companies make frequent decisions under uncertainty. Although each decision
might have its own unique characteristics, it seems reasonable that if the company
plans to make many such decisions, it should be willing to “play the averages,” as
it does when it uses EMV as the decision criterion. Second, even if this is the only
such decision the company is ever going to make, decision theorists have proven that
under certain conditions, maximizing EMV is a rational basis for making this decision.
These “certain conditions” relate to the decision maker’s attitude toward risk. As we
will discuss later in this chapter, if the decision maker is risk averse and the possible
monetary payoffs or losses are large relative to her wealth, then EMV is not the
appropriate decision criterion to use. However, the EMV criterion has proved useful in
the vast majority of decision-making applications, so we will use it throughout most
of this chapter.
Decision Trees By now, we have gone through most of the steps of solving SciTools’
problem. We have listed the decision alternatives, the uncertain outcomes and their
probabilities, and the profits and losses from all combinations of decisions and out-
comes. We have then calculated the EMV for each alternative and have chosen the
alternative with the largest EMV. All of this can be done efficiently using a graphical
tool called a decision tree. The decision tree that corresponds to SciTools’ problem
appears in Figure 10.1 (page 500). (This figure is actually part of an Excel spreadsheet
and was created with the PrecisionTree add-in. We will explain how it was created
shortly.)
Decision Tree Conventions To understand Figure 10.1, we need to know the following
conventions that have been established for decision trees.
1. Decision trees are composed of nodes (circles, squares, and triangles) and branches
(lines).
2. The nodes represent points in time. A decision node (a square) is a time when the
decision maker makes a decision. A probability node (a circle) is a time when the
result of an uncertain event becomes known. An end node (a triangle) indicates

FIGURE 10.1 Decision Tree for SciTools Bidding Example
that the problem is completed—all decisions have been made, all uncertainty has
been resolved, and all payoffs/costs have been incurred.
3. Time proceeds from left to right. This means that any branches leading into a node
(from the left) have already occurred. Any branches leading out of a node (to the
right) have not yet occurred.
4. Branches leading out of a decision node represent the possible decisions; the de-
cision maker can choose the preferred branch. Branches leading out of probability
nodes represent the possible outcomes of uncertain events; the decision maker has
no control over which of these will occur.
5. Probabilities are listed on probability branches. These probabilities are conditional
on the events that have already been observed (those to the left). Furthermore, the
probabilities on branches leading out of any particular probability node must sum
to 1.
6. Individual monetary values are shown on the branches where they occur, and
cumulative monetary values are shown to the right of the end nodes. (Actually,
PrecisionTree shows two values to the right of each end node. The top one is
the probability of getting to that end node, and the bottom one is the associated
monetary value.)
The decision tree in Figure 10.1 illustrates these conventions for a single-stage
decision problem, the simplest type of decision problem. In a single-stage problem
all decisions are made first, and then all uncertainty is resolved. Later in this chapter

we will see multistage decision problems, where decisions and outcomes alternate.
That is, a decision maker makes a decision, then some uncertainty is resolved, then the
decision maker makes a second decision, then some further uncertainty is resolved,
and so on. Because these multistage decisions problems are inherently more complex,
we will focus initially on single-stage problems.
Once a decision tree has been drawn and labeled with probabilities and monetary
values, it can be solved easily. The solution for the decision tree in Figure 10.1 is
shown in Figure 10.2. Among other things, it shows that the decision to bid $115,000 is
optimal (follow the decision branches with “True” above them), with a corresponding
EMV of $12,200 (the value under “Bid?” at the left of the tree). This is consistent with
what we saw earlier for this example.
Folding Back Procedure The solution procedure used to develop Figure 10.2 is called
folding back on the tree. Starting at the right of the tree and working back to the left,
the procedure consists of two types of calculations.
1. At each probability node, we calculate the EMV (sum of monetary values times
probabilities) and write it below the name of the node. For example, consider the
node (top right) after SciTools’ decision to bid $115,000 and after it learns that
FIGURE 10.2 Result of Folding Back to Obtain Optimal Decision

there will be a competing bid. From that point, SciTools will either win $15,000
with probability 0.8 or lose $5000 with probability 0.2. The corresponding EMV is
0.8(15,000) + 0.2(−5000) = 11,000
and this value is entered below the node name “Win bid?”.
Now, back up a step and consider the preceding probability node (the one to
the left of the “Win bid?” node). At this point, SciTools has bid $115,000 and is
about to discover whether there will be a competing bid. If there is none, with
probability 0.3, then SciTools will win $15,000. But if there is a competing bid,
with probability 0.7, the EMV from that point on is the $11,000 we just calculated.
Essentially, this $11,000 summarizes the consequences of being at the “Win bid?”
node, and SciTools acts the same as if it were going to receive $11,000 for certain.
Therefore, the EMV for the “Any competing bid?” node is
0.3(15,000) + 0.7(11,000) = 12,200
This EMV is written below the node name.
2. Decision nodes are much easier. At each decision node we find the maximum of
the EMVs and write it below the node name. PrecisionTree indicates the winner by
placing “True” on the decision branch with the maximum EMV and “False” on all
other branches emanating from this node. For example, consider the node where
SciTools is deciding how much to bid (after already having decided to place a bid).
The EMVs under the three succeeding probability nodes are $12,200, $9500, and
$6100. Since the maximum of these is $12,200, the EMV for the “How much to
bid” node is $12,200 and is written below the node name.
After the folding-back process is completed—that is, after we have calculated
EMVs for all nodes—we can trace the “True” labels from left to right to see the optimal
strategy. In this case SciTools should place a bid, and it should be for $115,000. The
EMV written below the leftmost decision node, $12,200, indicates SciTools’ EMV
for this strategy. If SciTools is truly willing to use the EMV criterion, that is, if it is
willing to play the averages, then the company should be indifferent between receiving
$12,200 for certain and bidding $115,000—with the associated risk of winning $15,000
or losing $5000.
The PrecisionTree Add-In Decision trees present a challenge for Excel. We must
somehow take advantage of Excel’s calculating capabilities (to calculate EMVs, for
example) and its graphical capabilities (to depict the decision tree). Fortunately, there is
now a powerful add-in, PrecisionTree developed by Palisade Corporation, that makes
the process relatively straightforward.2 This add-in not only enables us to build and
label a decision tree, but it performs the folding-back procedure automatically and then
allows us to perform sensitivity analysis on key input parameters.
The first thing you must do to use PrecisionTree is to “add it in.” You do this
in two steps. First, you must install the Palisade Decision Tools suite (or at least the
PrecisionTree program) with the Setup program on the CD-ROM accompanying this
book. Of course, you need to do this only once. Then to run PrecisionTree, there are
three options:
2
The educational version of PrecisionTree included with this book is slightly scaled down from Palisade’s
commercial version of PrecisionTree. The difference you are most likely to notice is that the educational
version permits only 50 nodes—of all types combined—in a decision tree.

■ If Excel is not currently running, you can launch Excel and PrecisionTree by
clicking on the Windows Start button and selecting the PrecisionTree item from the
Palisade Decision Tools group of the Programs group.
■ If Excel is currently running, the procedure in the previous bullet will launch
PrecisionTree on top of Excel.
■ If Excel is already running and the Decision Tools toolbar in Figure 10.3 is showing,
you can start PrecisionTree by clicking on its icon (the third from the left).
FIGURE 10.3
Palisade Decision
Tools Toolbar
You will know that PrecisionTree is ready for use when you see its toolbar (shown
in Figure 10.4) and a PrecisionTree menu to the left of the Help menu. By the way, if you
want to unload PrecisionTree without closing Excel, use the PrecisionTree/Help/About
menu item and click on Unload. It’s a bit unconventional, but it works.
Using PrecisionTree PrecisionTree is quite easy to use—at least its most basic items
are—but it can be confusing at first. We will lead you through the steps for the SciTools
example. (The file SCITOOLS.XLS shows the results of this procedure, but you should
work through the steps on your own, starting with a blank spreadsheet.)
1. Inputs. Enter the inputs shown in columns A and B of Figure 10.5.
2. New tree. Click on the new tree button (the far left button) on the PrecisionTree
toolbar, and then click on any cell (say, cell A14) below the input section to start
a new tree. Click on the name box of this new tree (it probably says “tree #1”)
to open a dialog box. Type in a descriptive name for the tree, such as SciTools
Bidding, and click on OK. You should now see the beginnings of a tree, as shown
in Figure 10.6 (page 504).
3. Decision nodes and branches. From here on, keep the finished tree in Figure 10.2
in mind. This is the finished product we eventually want. To obtain decision nodes
FIGURE 10.4
PrecisionTree
Toolbar
FIGURE 10.5
Inputs for SciTools
Bidding Example

FIGURE 10.6
Beginnings of a New
Tree
and branches, click on the (only) triangle end node to open the dialog box in Figure
10.7. Click on the green square to indicate that this is a decision node, and fill in
the dialog box as shown. We’re calling this decision “Bid?” and specifying that
there are two possible decisions. The tree expands as shown in Figure 10.8. The
boxes that say “branch” show the default labels for these branches. Click on either
of them to open another dialog box where you can provide a more descriptive
name for the branch. Do this to label the two branches “No” and “Yes.” Also, you
can enter the immediate payoff/cost for either branch right below it. Since there is
a $5000 cost of bidding, enter the formula
=-BidCost
right below the “Yes” branch in cell B19. (It is negative to reflect a cost.) The tree
should now appear as in Figure 10.9.
4. More decision branches. The top branch is completed; if SciTools does not bid,
there is nothing left to do. So click on the bottom end node, following SciTools’
decision to bid, and proceed as in the previous step to add and label the decision
FIGURE 10.7
Dialog Box for
Adding a New
Decision Node and
Branches
FIGURE 10.8
Tree with Initial
Decision Node and
Branches
FIGURE 10.9
Decision Tree with
Decision Branches
Labeled

node and three decision branches for the amount to bid. (Refer to Figure 10.2.) The
tree to this point should appear as in Figure 10.10. Note that there are no monetary
values below these decision branches because no immediate payoffs or costs are
associated with the bid amount decision.
5. Probability nodes and branches. We now need a probability node and branches
from the rightmost end nodes to capture whether the competition bids. Click on
the top one of these end nodes to bring up the same dialog box as in Figure 10.7.
Now, however, click on the red circle box to indicate that this is a probability
node. Label it “Any competing bid?”, specify two branches, and click on OK.
Then label the two branches “No” and “Yes.” Next, repeat this procedure to form
another probability node (with two branches) following the “Yes” branch, call it
“Win bid?”, and label its branches as shown in Figure 10.11.
6. Copying probability nodes and branches. You could now repeat the same pro-
cedure from the previous step to build probability nodes and branches following
the other bid amount decisions, but because they’re structurally equivalent, you
can save a lot of work by using PrecisionTree’s copy and paste feature. Click on
the leftmost probability node to open a dialog box and click on Copy. Then click
on either end node to bring up the same dialog box and click on Paste. Do this
again with the other end node. Decision trees can get very “bushy,” but this copy
and paste feature can make them much less tedious to construct.
FIGURE 10.10
Tree with All
Decision Nodes and
Branches
FIGURE 10.11 Decision Tree with One Set of Probability Nodes and Branches

7. Labeling probability branches. You should now have the decision tree shown in
Figure 10.12. It is structurally the same as the completed tree in Figure 10.2, but
the probabilities and monetary values on the probability branches are not correct.
Note that each probability branch has a value above and below the branch. The
value above is the probability (the default values make the branches equally likely),
and the value below is the monetary value (the default values are 0). We can enter
any values or formulas in these cells, exactly as we do in typical Excel worksheets.
As usual, it is a good practice to refer to input cells in these formulas whenever
possible. We’ll get you started with the probability branches following the decision
to bid $115,000. First, enter the probability of no competing bid in cell D18 with
the formula
=PrNoBid
and enter its complement in cell D24 with the formula
=1-D18
Next, enter the probability that SciTools wins the bid in cell E22 with the formula
=SUM(B10:B12)
and enter its complement in cell E26 with the formula
=1-E22
FIGURE 10.12 Structure of Completed Tree

(Remember that SciTools wins the bid only if the competitor bids higher, and in
this part of the tree, SciTools is bidding $115,000.) For the monetary values, enter
the formula
=115000-ProdCost
in the two cells, D19 and E23, where SciTools wins the contract. Note that we
already subtracted the cost of the bid (cell B29), so we shouldn’t do so again. This
would be double-counting, and it should always be avoided in decision problems.
8. Enter the other formulas on probability branches. Using the previous step and
Figure 10.2 as a guide, enter formulas for the probabilities and monetary values on
the other probability branches, that is, those following the decision to bid $120,000
or $125,000.
We’re finished! The completed tree in Figure 10.2 shows the best strategy and
its associated EMV, as we discussed earlier. Note that we never have to perform the
folding-back procedure manually. PrecisionTree does it for us. In fact, the tree is
completed as soon as we finish entering the relevant inputs. In addition, if we change
any of the inputs, the tree reacts automatically. For example, try changing the bid cost
in cell B4 from $5000 to some large value such as $20,000. You’ll see that the tree
calculations update automatically, and the best decision is then not to bid, with an
associated EMV of $0.
Risk Profile of Optimal Strategy Once the decision tree is completed, PrecisionTree
has several tools we can use to gain more information about the decision analysis.
First, we can see a risk profile and other information about the optimal decision. To
do so, click on the fourth button from the left on the PrecisionTree toolbar (it looks
like a staircase) and fill in the resulting dialog box as shown in Figure 10.13. (You can
experiment with other options.) The Policy Suggestion option allows us to see only
that part of the tree that corresponds to the best decision, as shown in Figure 10.14
(page 508).
The Risk Profile option allows us to see a graphical risk profile of the optimal
decision. (If we checked the Statistics Report box, we would also see this information
numerically.) As the risk profile in Figure 10.15 (page 508) shows, there are only two
possible monetary outcomes if SciTools bids $115,000. It either wins $15,000 or loses
$5000, and the former is much more likely. (The associated probabilities are 0.86 and
0.14.) This graphical information is even more useful when there are a larger number
of possible monetary outcomes. We can see what they are and how likely they are.
FIGURE 10.13
Dialog Box for
Information About
Optimal Decision

FIGURE 10.14
Subtree for Optimal
Decision
FIGURE 10.15
Risk Profile of
Optimal Decision
Sensitivity Analysis We have already stressed the importance of a follow-up sensi-

tivity analysis for any decision problem, and PrecisionTree makes this relatively easy
to perform. First, we can enter any values into the input cells and watch how the tree
changes. But we can get more systematic information by clicking on PrecisionTree’s
sensitivity button, the fifth from the left on the toolbar (it looks like a tornado). This
brings up the dialog box in Figure 10.16. It requires an EMV cell (and an optional
descriptive name) to analyze at the top and one or more input cells in the middle. The
specifications for these input cells are actually entered at the bottom of the dialog box.
The cell to analyze (at the top) is usually the EMV cell at the far left of the decision
tree—this is the cell shown in the figure—but it can be any EMV cell. For example, if
we assume SciTools will prepare a bid and we want to see how sensitive the EMV from
that point on is to inputs, we could select cell C29 (refer to Figure 10.2) to analyze.
Next, for any input cell such as the production cost cell (B5), we enter a minimum
value, a maximum value, a base value (probably the original value in the model), and
a step size. For example, to specify these for the production cost, we clicked on the
Suggest Values button. This default setting varies the production cost by as much as
10% from the original value in either direction in a series of 10 steps. We can also enter
our own desired values. We did so for the probability of no competing bids, varying its
value from 0 to 0.6 in a sequence of 12 steps.
When we click on Run Analysis, PrecisionTree varies each of the specified inputs
(one at a time if we select the One Way option) and presents the results in several ways
in a new Excel file with Sensitivity, Tornado, and Spider Graph sheets. The Sensitivity
sheet includes several charts, a typical one of which appears in Figure 10.17. This
shows how the EMV varies with the production cost for both of the original decisions

FIGURE 10.16
Sensitivity Analysis
Dialog Box
FIGURE 10.17
EMV versus
Production Cost for
Each of Two
Decisions
(bid or don’t bid). This type of graph is useful for seeing whether the optimal decision
changes over the range of the input variable. It does so only if the two lines cross. In
this particular graph it is clear that the “Bid” decision dominates the “No bid” decision
over the production cost range we selected.
The Tornado sheet shows how sensitive the EMV of the optimal decision is to
each of the selected inputs over the ranges selected. (See Figure 10.18 (page 510).)
The length of each bar shows the percentage change in the EMV in either direction,
so the longer the bar, the more sensitive this EMV is to the particular input. The bars
are always arranged from longest on top to shortest on the bottom—hence the name
tornado chart. Here we see that production cost has the largest effect on EMV, and bid
cost has the smallest effect.
Finally, the Spider Chart sheet contains the chart in Figure 10.19. It shows how
much the optimal EMV varies in magnitude for various percentage changes in the input
variables. The steeper the slope of the line, the more the EMV is affected by a particular
input. We again see that the production cost has a relatively large effect, whereas the
other two inputs have relatively small effects.
Each time we click on the sensitivity button, we can run a different sensitivity
analysis. An interesting option is to run a two-way analysis (by clicking on the Two

FIGURE 10.18
Tornado Chart for
SciTools Example
FIGURE 10.19
Spider Chart for
SciTools Example
Way button in Figure 10.16). Then we see how the selected EMV varies as each pair
of inputs vary simultaneously. We analyzed the EMV in cell C29 with this option,
using the same inputs as before. A typical result is shown in Figure 10.20. For each
of the possible values of production cost and the probability of no competitor bid, this
chart indicates which bid amount is optimal. (By choosing cell C29, we are assuming
SciTools will bid; the question is only how much.) As we see, the optimal bid amount
remains $115,000 unless the production cost and the probability of no competing bid
are both large. Then it becomes optimal to bid $125,000. This makes sense intuitively.
As the chance of no competing bid increases and a larger production cost must be
recovered, it seems reasonable that SciTools should increase its bid.
FIGURE 10.20
Two-Way Sensitivity
Analysis

We reiterate that a sensitivity analysis is always an important aspect in real decision
analyses. If we had to construct decision trees by hand—with paper and pencil—a
sensitivity analysis would be virtually out of the question. We would have to recompute
everything each time through. Therefore, one of the most valuable features of the
PrecisionTree add-in is that it enables us to perform sensitivity analyses in a matter of
seconds. ■
PROBLEMS
Skill-Building Problems TABLE 10.5 Distribution of Possible

1. The SweetTooth Candy Company knows it will need Sugar Prices
10 tons of sugar 6 months from now to implement its Possible Sugar Prices
production plans. Jean Dobson, SweetTooth’s in 6 Months ($/pound) Probability
purchasing manager, has essentially two options for
acquiring the needed sugar. She can either buy the 0.078 0.05
sugar at the going market price when she needs it, 6 0.083 0.25
months from now, or she can buy a futures contract 0.087 0.35
now. The contract guarantees delivery of the sugar in 0.091 0.20
6 months but the cost of purchasing it will be based 0.096 0.15
on today’s market price. Assume that possible sugar
futures contracts available for purchase are for 5 tons
or 10 tons only. No futures contracts can be purchased summarize your findings. In response to which
or sold in the intervening months. Thus, SweetTooth’s model inputs is the expected cost value more
possible decisions are: (1) purchase a futures contract sensitive?
for 10 tons of sugar now, (2) purchase a futures c. Generate a risk profile for SweetTooth’s optimal
contract for 5 tons of sugar now and purchase 5 tons decision.
of sugar in 6 months, or (3) purchase all 10 tons of 2. Carlisle Tire and Rubber, Inc. is considering
needed sugar in 6 months. The price of sugar bought expanding production to meet potential increases in
now for delivery in 6 months is $0.0851 per pound. the demand for one of its tire products. Carlisle’s
The transaction costs for 5-ton and 10-ton futures alternatives are to construct a new plant, expand the
contracts are $65 and $110, respectively. Finally, Ms. existing plant, or do nothing in the short run. The
Dobson has assessed the probability distribution for market for this particular tire product may expand,
the possible prices of sugar 6 months from now (in remain stable, or contract. Carlisle’s marketing
dollars per pound). Table 10.5 contains these possible department estimates the probabilities of these
prices and their corresponding probabilities. market outcomes as 0.25, 0.35, and 0.40, respectively.
a. Given that SweetTooth wants to acquire the Table 10.6 contains Carlisle’s estimated payoff (in
needed sugar in the least-cost way, formulate a dollars) table.
payoff table that specifies the cost (in dollars) a. Use the PrecisionTree add-in to identify the
associated with each possible decision and strategy that maximizes this tire manufacturer’s
possible sugar price in the future. expected profit. Also, perform sensitivity analysis
b. Use the PrecisionTree add-in to identify the on the optimal decision and summarize your
strategy that minimizes SweetTooth’s expected findings. In response to which model inputs is the
cost of meeting its sugar demand. Also, perform expected profit value most sensitive?
sensitivity analysis on the optimal decision and
TABLE 10.6 Payoff Table for Carlisle’s Decision Problem

Decision/Market Outcome Market Expands Market Stable Market Contracts
Construct a new plant 400,000 −100,000 −200,000
Expand existing plant 250,000 −50,000 −75,000
Do nothing 50,000 0 −30,000

b. Generate a risk profile for Carlisle’s optimal Table 10.7. The probabilities of observing a strong,
decision. fair, and weak trend in the national economy in the
3. A local energy provider offers a landowner $180,000 coming year are 0.30, 0.50, and 0.20, respectively.
for the exploration rights to natural gas on a certain a. Formulate a payoff table that specifies Techware’s
site and the option for future development. This net revenue (in dollars) for each possible decision
option, if exercised, is worth an additional $1,800,000 and each outcome with respect to the trend in the
to the landowner, but this will occur only if natural national economy.
gas is discovered during the exploration phase. The b. Use the PrecisionTree add-in to identify the
landowner, believing that the energy company’s strategy that maximizes Techware’s expected net
interest in the site is a good indication that gas is revenue from the given marketing opportunities.
present, is tempted to develop the field herself. To do Also, perform sensitivity analysis on the optimal
so, she must contract with local experts in natural gas decision and summarize your findings. In
exploration and development. The initial cost for response to which model inputs is the expected net
such a contract is $300,000, which is lost forever if revenue value most sensitive?
no gas is found on the site. If gas is discovered, c. Generate a risk profile for Techware’s optimal
however, the landowner expects to earn a net profit of decision.
$6,000,000. Finally, the landowner estimates the 5. Consider an investor with $10,000 available to invest.
probability of finding gas on this site to be 60%. He has the following options regarding the allocation
a. Formulate a payoff table that specifies the of his available funds: (1) he can invest in a risk-free
landowner’s payoff (in dollars) associated with savings account with a guaranteed 3% annual rate of
each possible decision and each outcome with return; (2) he can invest in a fairly safe stock, where
respect to finding natural gas on the site. the possible annual rates of return are 6%, 8%, or
b. Use the PrecisionTree add-in to identify the 10%; or (3) he can invest in a more risky stock where
strategy that maximizes the landowner’s expected the possible annual rates of return are 1%, 9%, or
net earnings from this opportunity. Also, perform 17%. Note that the investor can place all of his
sensitivity analysis on the optimal decision and available funds in any one of these options, or he can
summarize your findings. In response to which split his $10,000 into two $5000 investments in
model inputs is the expected profit value most any two of these options. The joint probability
sensitive? distribution of the possible return rates for the two
c. Generate a risk profile for landowner’s optimal stocks is given in Table 10.8.
decision. a. Formulate a payoff table that specifies this
4. Techware Incorporated is considering the investor’s return (in dollars) in one year for each
introduction of two new software products to the possible decision and each outcome with respect
market. In particular, the company has four options to the two stock returns.
regarding these two proposed products: introduce b. Use the PrecisionTree add-in to identify the
neither product, introduce product 1 only, introduce strategy that maximizes the investor’s expected
product 2 only, or introduce both products. Research earnings in one year from the given investment
and development costs for products 1 and 2 are opportunities. Also, perform sensitivity analysis
$180,000 and $150,000, respectively. Note that the on the optimal decision and summarize your
first option entails no costs because research and findings. In response to which model inputs is the
development efforts have not yet begun. The success expected earnings value most sensitive?
of these software products depends on the trend of c. Generate a risk profile for this investor’s optimal
the national economy in the coming year and decision.
on the consumers’ reaction to these products. 6. A buyer for a large department store chain must place
The company’s revenues earned by introducing orders with an athletic shoe manufacturer 6 months
product 1 only, product 2 only, or both products in prior to the time the shoes will be sold in the
various states of the national economy are given in department stores. In particular, the buyer must
TABLE 10.7 Revenue Table for Techware’s Decision Problem

Decision/Trend in National Economy Strong Fair Weak
Introduce neither product $0 $0 $0
Introduce product 1 only $500,000 $260,000 $120,000
Introduce product 2 only $420,000 $230,000 $110,000
Introduce both products $820,000 $390,000 $200,000

TABLE 10.8 Joint Probability Distribution of Safe and Risky
Stock Return Rates
Safe Stock Return Rates (S)/
Risky Stock Return Rates (R) R = 1% R = 9% R = 17%
S = 6% 0.10 0.05 0.10
S = 8% 0.25 0.05 0.20
S = 10% 0.10 0.05 0.10
decide on November 1 how many pairs of the subsequently selling pairs of the new tennis shoes.
manufacturer’s newest model of tennis shoes to order Also, perform sensitivity analysis on the optimal
for sale during the upcoming summer season. decision and summarize your findings. In
Assume that each pair of this new brand of tennis response to which model inputs is the expected
shoes costs the department store chain $45 per pair. earnings value most sensitive?
Furthermore, assume that each pair of these shoes can c. Generate a risk profile for the buyer’s optimal
then be sold to the chain’s customers for $70 per pair. decision.
Any pairs of these shoes remaining unsold at the end
of the summer season will be sold in a closeout sale Skill-Extending Problems
next fall for $35 each. The probability distribution of
7. In designing a new space vehicle, NASA needs to
consumer demand for these tennis shoes (in hundreds
decide whether to provide 0, 1, or 2 backup systems
of pairs) during the upcoming summer season has
for a critical component of the vehicle. The first
been assessed by market research specialists and is
backup system, if included, comes into use only if the
provided in Table 10.9. Finally, assume that the
original system fails. The second backup system, if
department store chain must purchase these tennis
included, comes into use only if the original system
shoes from the manufacturer in lots of 100 pairs.
and the first backup system both fail. NASA
engineers claim that each system, independently of
TABLE 10.9 Distribution of the others, has a 1% chance of failing if called into
Consumer Demand use. Each backup system costs $70,000 to produce
for Tennis Shoes and install within the vehicle. Once the vehicle is in
flight, the mission will be scrubbed only if the
Consumer
original system and all backups fail. The cost of a
Demand Probability
scrubbed mission, in addition to production costs, is
1 0.025 assessed to be $8,000,000.
2 0.050 a. Use the PrecisionTree add-in to identify the
3 0.075 strategy that minimizes NASA’s expected total
cost. Also, perform sensitivity analysis on the
4 0.100
optimal decision and summarize your findings. In
5 0.150 response to which model inputs is the expected
6 0.200 earnings value most sensitive?
7 0.175 b. Generate a risk profile for NASA’s optimal
8 0.100 decision.
8. Mr. Maloy has just bought a new $30,000 sport utility
9 0.075
vehicle. As a reasonably safe driver, he believes that
10 0.050 there is only about a 5% chance of being in an
accident in the forthcoming year. If he is involved in
an accident, the damage to his new vehicle depends
a. Formulate a payoff table that specifies the on the severity of the accident. The probability
contribution to profit (in dollars) from the sale of distribution for the range of possible accidents and
the tennis shoes by this department store chain for the corresponding damage amounts (in dollars) are
each possible purchase decision (in hundreds of given in Table 10.10 (page 514). Mr. Maloy is trying
pairs) and each outcome with respect to consumer to decide whether he is willing to pay $170 each year
demand. for collision insurance with a $300 deductible. Note
b. Use the PrecisionTree add-in to identify the that with this type of insurance, he pays the first $300
strategy that maximizes the department store in damages if he causes an accident and the insurance
chain’s expected profit earned by purchasing and company pays the remainder.

TABLE 10.10 Distribution of Accident TABLE 10.11 Distribution of Defective
Types and Corresponding Components in a Lot
Damage Amounts Proportion of Defective
Type of Conditional Damage to Components Probability
Accident Probability Vehicle
0.05 0.50
Minor 0.60 $200 0.10 0.25
Moderate 0.20 $1,000 0.25 0.15
Serious 0.10 $4,000 0.50 0.10
Catastrophic 0.10 $30,000
each, the purchasing agent is intrigued to learn that

this supplier will now assume the cost of replacing
a. Formulate a payoff table that specifies the cost (in
defective components in excess of the first 100 faulty
dollars) associated with each possible decision
items found in a given lot. This guarantee may be
and type of accident.
purchased by the microcomputer manufacturer prior
b. Use the PrecisionTree add-in to identify the
to the receipt of a given lot at a cost of $1000 per lot.
strategy that minimizes Mr. Maloy’s annual
The purchasing agent is interested in determining
expected total cost. Also, perform sensitivity
whether it is worthwhile for her company to purchase
analysis on the optimal decision and summarize
the supplier’s guarantee policy.
your findings. In response to which model inputs
a. Formulate a payoff table that specifies the
is the expected earnings value most sensitive?
microcomputer manufacturer’s total cost (in
c. Generate a risk profile for Mr. Maloy’s optimal
dollars) of purchasing and repairing (if necessary)
decision.
a complete lot of components for each possible
9. The purchasing agent for a microcomputer manu-
decision and each outcome with respect to the
facturer is currently negotiating a purchase agreement
proportion of defective items.
for a particular electronic component with a given
b. Use the PrecisionTree add-in to identify the
supplier. This component is produced in lots of
strategy that minimizes the expected total cost
1000, and the cost of purchasing a lot is $30,000.
of achieving a complete lot of satisfactory
Unfortunately, past experience indicates that this
microcomputer components. Also, perform
supplier has occasionally shipped defective com-
sensitivity analysis on the optimal decision and
ponents to its customers. Specifically, the proportion
summarize your findings. In response to which
of defective components supplied by this supplier is
model inputs is the expected earnings value most
described by the probability distribution given in
sensitive?
Table 10.11. While the microcomputer manufacturer
c. Generate a risk profile for the purchasing agent’s
can repair a defective component at a cost of $20
optimal decision.
10.3 MORE SINGLE-STAGE EXAMPLES

ll applications of decision making under uncertainty follow the procedures
A discussed so far. We first identify the possible decision alternatives, assess

relevant probabilities, and calculate monetary values. Then we use a decision
tree (or influence diagram) to identify the alternative with the largest EMV and follow
this up with a thorough sensitivity analysis. We can also examine the risk profiles
for the various alternatives. This is particularly useful if criteria other than EMV
maximization are considered, as we will discuss in Section 7.8. In this section we will
illustrate the process with several single-stage examples, where the decision maker
makes one decision and then learns which of several uncertain outcomes occurs. In
the next section we will examine multistage examples, where two or more sequential
decisions must be made.
The following example illustrates a decision problem most of us face on an annual
basis, although most of us probably do not go to the trouble of analyzing it formally.

E XAMPLE 10.2
SELECTING HEALTH CARE PLANS
AT STATE UNIVERSITY
Each year employees at State University are asked to decide on one of three health care
plans. The terms of these are as follows:3
Plan 1: The monthly cost is $24. There is a $500 deductible. The participant pays
all expenses until payments for the year equal $500. After that, 90% of remaining
expenses are paid by the insurer.
Plan 2: This is the same as plan 1, except that the monthly cost is $1 and the
deductible amount is $1000.
Plan 3: The monthly cost is $20. There is no deductible. The employee pays 30%
of all medical expenses. The rest is paid by the insurer.
Which of these three plans should an employee choose?
Solution
Clearly, the solution will vary from one employee to another, depending on the assessed
probability distribution of medical expenses. To illustrate, however, we will consider
an employee who assesses the distribution of yearly medical expenses shown in Table
10.12. These expenses include hospital visits, surgery, office visits, and prescriptions,
all of which are covered under the terms of the plans. As in the previous example, this
distribution is only an approximation of the real distribution, which would contain a
continuum of expenses. However, it is probably adequate for making a decision among
the three plans.
TABLE 10.12 Distribution of Medical Expenses

for Insurance Example
Total Medical Expense Probability
$200 0.30
$600 0.50
$1000 0.15
$5000 0.03
$15,000 0.02
The next step is to determine the employee’s cost for each plan and each outcome.
For example, suppose that the employee chooses plan 1 and incurs $600 in expenses.
Then the total cost is the cost of the insurance plus the full amount of the first $500 in
expenses plus 10% of the last $100 in expenses, that is,
24(12) + 500 + 0.1(100) = $798
However, if this employee’s medical expenses are only $200, then the cost is
24(12) + 200 = $488
3
We assume that these terms apply only to the employee; that is, these are not family plans.
10.3 More Single-Stage Examples 515

The costs for the other plans and other outcomes can be calculated in a similar manner.
We list all of the costs in Table 10.13.
The choice is certainly not clear from this table. The plan with the lowest premium,
plan 2, looks good if the year’s medical expenses are low. This is also true for the no-
deductible plan, plan 3, although its cost is quite large in case of a disaster. For moderate
medical expenses, plan 1 is obviously inferior, but it is the best for guarding against a
disaster. These trade-offs could be illustrated by risk profiles, which you might want to
examine. Instead, we turn directly to the decision tree.
TABLE 10.13 Employee Cost Table for

Insurance Example
Medical Expense Plan 1 Plan 2 Plan 3
$200 $488 $212 $300
$600 $798 $612 $420
$1000 $838 $1012 $540
$5000 $1238 $1412 $1740
$15,000 $2238 $2412 $4740
PRECISION
USING PRECISIONTREE
TREE
The decision tree can be formed with the following steps.
●
1 Inputs. Enter the inputs for the three plans and the probabilities from Table 10.12
in the top left portion of the spreadsheet (down to row 15). (See Figure 10.21 and the
file MEDICAL.XLS.)
●2 Cost table. For later use in the decision tree, calculate the costs to the employee
(not counting insurance premiums) in the range B19:D23. To do this, enter the formula
=IF($A19<=B$6,$A19,B$6+B$7*($A19-B$6))
FIGURE 10.21
Inputs and Cost
Table for Medical
Example

in cell B19 and copy this to the range B19:D23. This IF function says that if the medical
expense is less than the deductible, the employee pays it all. Otherwise, the employee
pays the deductible amount plus a percentage of the remainder.
●3 Decision tree. Use PrecisionTree to create the decision tree shown in Figure
10.22. Here are some tips. First, create the decision node and decision branches, and
enter formulas for their values as 12 times the relevant monthly premiums. Then create
a single probability node and its branches, label the branches, and enter formulas for
the probabilities with absolute references. For example, enter the formula
=$C$11
for the probability of the top branch. Next, copy the probability node to the end nodes
below it. (Do you see the effect of the absolute references?) Finally, link the values for
all of the probability branches to the cells in the cost table. (We know of no quick way
to do this. We entered 15 separate formulas, one for each branch. However, it is much
easier to create a cost table and link branch formulas to it than to create the branch
formulas directly from input values.)
●4 Minimize costs. If we quit here, we would mistakenly choose the worst of the
three plans. This is because PrecisionTree maximizes EMV by default, and in this
problem we want to minimize the EMV of the costs. However, this is simple to change.
Click on the name box at the far left in the decision tree. This brings up a dialog box
(not shown here) where we can select the Minimize option.
FIGURE 10.22
Decision Tree for
Medical Insurance
Example

As we see from Figure 10.22, the optimal plan is plan 3. Its EMV—an expected
cost—is $528. The EMVs for plans 1 and 2 are $753 and $612. Evidently, this em-
ployee’s chances of large medical expenses where plan 3 is at its worst are not large
enough to outweigh plan 3’s no-deductible benefit. However, we might want to exper-
iment with various inputs, either the properties of the plans or the employee’s medical
expense distribution, to see whether plan 3 continues to be the preferred plan. For exam-
ple, if the probabilities in Table 10.12 change to 0.30, 0.40, 0.15, 0.10, and 0.05, so that
large expenses are much more probable, the EMVs for the three plans become $827,
$722, and $750. Now plan 2 is preferred, although the difference in EMV between
plans 2 and 3 is quite small.
We can use this insurance example to illustrate one nonmonetary aspect of decision
problems that is difficult to incorporate into a decision tree. At the university where
we teach, there is another insurance plan in addition to the types in the example.
Its premiums are low, and there are no copayments—the insurer pays all medical
expenses. This plan is clearly the cheapest of all plans offered, but it is not chosen
by many employees. Why? The plan is through an HMO, where all employees must
go to a specified set of physicians; otherwise, the plan does not pay their expenses.
Evidently, many employees believe that the “cost” of having to go to physicians they
would not choose otherwise outweighs the dollar savings from the plan. ■
The following example illustrates one method for using a continuous probability
distribution in a decision tree model.
E XAMPLE 10.3
PURCHASING LIGHTBULBS AT FRESHWAY
SUPERMARKETS
FreshWay, a chain of supermarkets, requires 24,000 fluorescent lightbulbs for its stores.
There are two suppliers of these lightbulbs. Supplies A offers them at $4.00 per bulb
and will replace the first 900 defective bulbs with guaranteed good ones for $3.00 each.
It will replace all defectives after the first 900 for nothing. Supplier B is similar. It will
charge $4.15 per bulb, replace the first 1200 defectives for $1.00 each, and replace all
defectives after the first 1200 for nothing. FreshWay plans to sell these lightbulbs for
$4.40 apiece and charge its customers nothing for replacement of defectives. The only
uncertainty is the number of defective bulbs from either supplier. Based on historical
data from each supplier, FreshWay believes that the percentage of defectives is normally
distributed with mean 4% and standard deviation 1% from supplier A, and mean 4.2%
and standard deviation 1.2% from supplier B. Which supplier should be chosen to
maximize FreshWay’s EMV?
Solution
Let p be the percentage of lightbulbs that are defective. Then the profit to FreshWay
from buying from supplier A is

24,000(4.40 − 4.00) − (24,000 p)(3.00) if p ≤ 900/24,000
Profit =
24,000(4.40 − 4.00) − (900)(3.00) if p > 900/24,000
A similar expression holds for supplier B. The only random quantity in this expression
is p, which is normally distributed. The question is how we can model the continuous
distribution of p in a discrete decision tree—that is, a tree with a discrete number

of probability branches. The method usually used is to approximate the continuous
normal distribution by a discrete distribution with a relatively small number, say 5, of
equally likely values.
The idea is to divide the normal distribution into an equal number of equal proba-
bility regions and take the midpoint (in a probability sense) of each region as a value for
the decision tree. For example, if we use five points, then each region has probability
0.2. The probability halfway between 0 and 0.2 is 0.1, so the first point on the tree
is the 10th percentile of the normal distribution. Similarly, the next point is the 30th
percentile, the next is the 50th, the next is the 70th, and the last is the 90th.
Figure 10.23 illustrates the calculations. (See the file LIGHTBULB.XLS.) Through
row 13 we enter the given inputs for the problem. Then in rows 17–26 we enter the
information we’ll use in the decision tree regarding the percentage defective for each
supplier. This information is based on the five-point approximation to the normal
distribution. For example, the 10th percentile of the normal distribution for supplier A
is found in cell C17 with the formula
=NORMINV(B17,$B$12,$C$12)
and this is copied down to cell C21. Then the cost to FreshWay from defectives,
assuming the value in C17 is the percentage of defectives, is calculated in cell D17
with the formula
=$C$7*IF(C17<=$D$7/Quantity,Quantity*C17,$D$7)
and it is copied down to cell D21. Similar formulas are used for supplier B.
FIGURE 10.23
Inputs and
Calculations for
Lightbulb Example
PRECISION
USING PRECISIONTREE
TREE
It is now straightforward to construct the decision tree shown in Figure 10.24 (page 520).
We enter the revenue from selling the bulbs and the cost of purchasing them in cells
B33 and B47. For example, the formula in cell B33 is
=Quantity*(SellingPrice-B7)

FIGURE 10.24
Decision Tree for
Lightbulb Example
Then we link the monetary values below the probability branches to the relevant cells
in the D17:D26 range.
The EMVs for suppliers A and B are $7088 and $5027, so supplier A is the clear
choice. Evidently, the higher price charged by supplier B and its slightly higher mean
percentage of defects outweigh its better deal on replacing defectives. Of course, if
supplier B really wants to get FreshWay’s business, it could attempt to sweeten its deal
in a number of ways. Sensitivity analysis is useful to see how the EMV for supplier
B (in cell C47) is affected by the various input parameters. We tried this, varying the
inputs in cells B8, C8, D8, and B13 by PrecisionTree’s default values (10% in either
direction) and keeping track of the change in the EMV for supplier B. The tornado chart
in Figure 10.25 makes it very clear that the most important input is the unit purchase
cost. The effects of the other three inputs are practically negligible in comparison. If
supplier B wants FreshWay’s business, it will have to lower its unit purchase cost.
FIGURE 10.25
Tornado Chart to
Analyze the EMV
for Supplier B

MODELING ISSUES
The discrete approximation used in Example 10.3 can be used in any decision tree
with continuous probability distributions, regardless of whether they are normal. We
first need to decide how many values to have in the discrete approximation. The usual
choices are 5 or 3. (Surprisingly, a three-point approximation does an adequate job
in many situations.) Then we need to use the “inverse” function—in the previous
example it was the NORMINV function—to find the values to use in the decision tree.
The appropriate inverse function is available in Excel for a number of widely used
continuous distributions. ■
PROBLEMS
Skill-Building Problems TABLE 10.15 Distribution of Possible

10. Each day the manager of a local bookstore must Competing Bids for
decide how many copies of the community Construction Project
newspaper to order for sale in her shop. She must Buffalo Valley
pay the newspaper’s publisher $0.40 for each Construction’s Bid Probability
copy and sells the newspapers to local residents
for $0.50 each. Newspapers that are unsold at $160,000 0.40
the end of day are considered worthless. The $165,000 0.30
probability distribution of the number of copies of $170,000 0.20
the newspaper purchased daily at her shop is $175,000 0.10
provided in Table 10.14. Employ a decision tree to
find the bookstore manager’s profit-maximizing
daily order quantity. within the local community, Fine Line Homes
believes that it will likely be awarded the project in
the event that it and Buffalo Valley Construction
TABLE 10.14 Distribution of Daily Local submit exactly the same bids. Employ a decision
Newspaper Demand tree to identify Fine Line Homes’ profit-maximizing
bid for the new community center building.
Daily Demand for
12. Suppose that you have sued your employer for
Local Newspaper Probability
damages suffered when you recently slipped and fell
10 0.10 on an icy surface that should have been treated by
11 0.15 your company’s physical plant department.
12 0.30 Specifically, your injury resulting from this accident
was sufficiently serious that you, in consultation
13 0.20 with your attorney, decided to sue your company for
14 0.15 $500,000. Your company’s insurance provider has
15 0.10 offered to settle this suit with you out of court. If
you decide to reject the settlement and go to court,
your attorney is confident that you will win the case
11. Two construction companies are bidding against one but is uncertain about the amount the court will
another for the right to construct a new community award you in damages. He has provided his
center building in Lewisburg, Pennsylvania. The first assessment of the probability distribution of the
construction company, Fine Line Homes, believes court’s award to you in Table 10.16 (page 522). Let
that its competitor, Buffalo Valley Construction, will S be the insurance provider’s proposed out-of-court
place a bid for this project according to the settlement (in dollars). For which values of S will
distribution shown in Table 10.15. Furthermore, you decide to accept the settlement? For which
Fine Line Homes estimates that it will cost $160,000 values of S will you choose to take your chances in
for its own company to construct this building. court? Of course, you are seeking to maximize the
Given its fine reputation and long-standing service expected payoff from this litigation.

TABLE 10.16 Distribution of Possible b. Suppose now that your colleague can place all of
Court Award Amounts his available funds in one of these three
investments as before, or he can invest $1000 in
Amount of
one alternative and $1000 in another. Assuming
Court Award Probability
that he seeks to maximize his expected total
$0 0.025 earnings in one year, how should he allocate his
$50,000 0.075 $2000?
$100,000 0.100
Skill-Extending Problems
$200,000 0.125
$300,000 0.175 14. A home appliance company is interested in
marketing an innovative new product. The company
$400,000 0.200
must decide whether to manufacture this product
$500,000 0.300 essentially on its own or employ a subcontractor to
manufacture it. Table 10.19 contains the estimated
probability distribution of the cost of manufacturing
13. Suppose that one of your colleagues has $2000
1 unit of this new product (in dollars) under the
available to invest. Assume that all of this money
alternative that the home appliance company
must be placed in one of three investments: a
produces the item on its own. Table 10.20 contains
particular money market fund, a stock, or gold. Each
the estimated probability distribution of the cost of
dollar your colleague invests in the money market
purchasing 1 unit of this new product (in dollars)
fund earns a virtually guaranteed 12% annual return.
under the alternative that the home appliance
Each dollar he invests in the stock earns an annual
company commissions a subcontractor to produce
return characterized by the probability distribution
the item.
provided in Table 10.17. Finally, each dollar he
a. Assuming that the home appliance company
invests in gold earns an annual return characterized
seeks to minimize the expected unit cost of
by the probability distribution given in Table 10.18.
manufacturing or buying the new product, should
a. If your colleague must place all of his available the company make the new product or buy it
funds in a single investment, which investment from a subcontractor?
should he choose to maximize his expected b. Perform sensitivity analysis on the optimal
earnings over the next year? expected cost. Under what conditions, if any,
TABLE 10.17 Distribution of Annual TABLE 10.19 Distribution of Unit

Returns for Given Stock Production Cost under
Annual Returns for "Make" Alternative
Given Stock Probability Cost Per Unit Probability
0% 0.10 $50 0.20
6% 0.20 $53 0.25
12% 0.40 $55 0.30
18% 0.20 $57 0.20
24% 0.10 $59 0.05
TABLE 10.18 Distribution of Annual TABLE 10.20 Distribution of Unit

Returns for Gold Production Cost under
Annual Returns "Buy" Alternative
for Gold Probability Cost Per Unit Probability
−36% 0.10 $50 0.10
−12% 0.20 $53 0.20
12% 0.40 $55 0.40
36% 0.20 $57 0.20
60% 0.10 $59 0.10

would the home appliance company select an a. Formulate a payoff table that specifies the
alternative different from the one you identified contribution to profit (in dollars) from the sale of
in part a? the tennis shoes by this department store chain
15. A grapefruit farmer in central Florida is trying to for each possible purchase decision (in hundreds
decide whether to take protective action to limit of pairs) and each outcome with respect to
damage to his crop in the event that the overnight consumer demand. Use an appropriate discrete
temperature falls to a level well below freezing. He approximation of the given normal demand
is concerned that if the temperature falls sufficiently distribution.
low and he fails to make an effort to protect his b. Construct a decision tree to identify the buyer’s
grapefruit trees, he runs the risk of losing his entire course of action that maximizes the expected
crop, which is worth approximately $75,000. Based profit (in dollars) earned by the department store
on the latest forecast issued by the National Weather chain from the purchase and subsequent sale of
Service, the farmer estimates that there is a 60% tennis shoes in the coming year.
chance that he will lose his entire crop if it is left 17. Consider again the purchasing agent’s decision
unprotected. Alternatively, the farmer can insulate problem described in Problem 9. Assume now that
his fruit by spraying water on all of the trees in his the proportion of defective components supplied by
orchards. This action, which would likely cost the this supplier is well described by the triangular
farmer C dollars, would prevent total devastation distribution with parameters 0, 0, and 1. (This is
but might not completely protect the grapefruit trees called the right triangular distribution with
from incurring some damage as a result of the range 1.)
unusually cold overnight temperatures. Table 10.21 a. Formulate a payoff table that specifies the
contains the assessed distribution of possible microcomputer manufacturer’s total cost
damages (in dollars) to the insulated fruit in light of (in dollars) of purchasing and repairing (if
the cold weather forecast. Of course, this farmer necessary) a complete lot of components for each
seeks to minimize the expected total cost of coping possible decision and each outcome with respect
with the threatening weather. to the proportion of defective items. Use an
appropriate discrete approximation of the given
triangular distribution for the proportion of
TABLE 10.21 Distribution of Damages to defective items.
Insulated Grapefruit Crop b. Construct a decision tree to identify the
Damage to purchasing agent’s course of action that
Grapefruit Crop Probability minimizes the expected total cost (in dollars)
$0 0.30 of achieving a complete lot of satisfactory
components.
$5000 0.15
18. A retired partner from Goldman Sachs has 1 million
$10,000 0.10 dollars available to invest in particular stocks or
$15,000 0.15 bonds. Each investment’s annual rate of return
$20,000 0.30 depends on the state of the economy in the
forthcoming year. Table 10.22 (page 524) contains
the distribution of returns for these stocks and bonds
a. Find the maximum value of C below which the as a function of the economy’s state in the coming
farmer will choose to insulate his crop in hopes year. This investor wants to allocate her $1 million to
of limiting damage as result of the unusually maximize her expected total return 1 year from now.
cold weather. a. If X = Y = 15%, find the optimal investment
b. Set C equal to the value identified in part a. strategy for this investor.
Perform sensitivity analysis to determine under b. For which values of X (where 10% < X < 20%)
what conditions, if any, the farmer might be and Y (where 12.5% < Y < 17.5%), if any, will
better off not spraying his grapefruit trees and this investor prefer to place all of her available
taking his chances in spite of the threat to his funds in the given stocks to maximize her
crop. expected total return one year from now?
16. Consider again the department store buyer’s c. For which values of X (where 10% < X < 20%)
decision problem described in Problem 6. Assume and Y (where 12.5% < Y < 17.5%), if any, will
now that consumer demand for the new tennis shoes this investor prefer to place all of her available
model (in hundreds of pairs) during the upcoming funds in the given bonds to maximize her
summer season is normally distributed with mean 6 expected total return one year from now?
and standard deviation 1.5.

TABLE 10.22 Distribution of Annual Returns for Given Stocks
and Bonds
State of Annual Returns Annual Returns
the Economy Probability for Given Stocks for Given Bonds
Very strong 0.20 25% 20%
Moderately strong 0.40 20% 17.5%
Fair 0.25 X% Y%
Moderately weak 0.10 10% 12.5%
Very weak 0.05 5% 10%
10.4 MULTISTAGE DECISION PROBLEMS

o far, all of the examples have required a single decision. We now examine
S a problem where the decision maker must make at least two decisions that
are separated in time, such as when a company must decide whether to buy
information that will help it make a second decision. The following example illustrates
the typical situation.
E XAMPLE 10.4
MARKETING A NEW PRODUCT AT ACME
The Acme Company is trying to decide whether to market a new product. As in
many new-product situations, there is considerable uncertainty about whether the new
product will eventually “catch on.” Acme believes that it might be prudent to introduce
the product in a regional test market before introducing it nationally. Therefore, the
company’s first decision is whether to conduct the test market. Acme estimates that the
fixed cost of the test market is $3 million. If it decides to conduct the test market, it
must then wait for the results. Based on the results of the test market, it can then decide
whether to market the product nationally, in which case it will incur a fixed cost of
$90 million. On the other hand, if the original decision is not to run a test market, then
the final decision—whether to market the product nationally—can be made without
further delay. Acme’s unit margin, the difference between its selling price and its unit
variable cost, is $18 (in the test market and in the national market).
Acme classifies the results in either the test market or the national market as great,
fair, or awful. Each of these is accompanied by a forecast of total units sold. These sales
volumes (in 1000s of units) are 200, 100, and 30 for the test market and 6000, 3000,
and 900 for the national market. Based on previous test markets for similar products,
Acme estimates that probabilities of the three test market outcomes are 0.3, 0.6, and
0.1. Then, based on historical data from previous products that were test marketed
and eventually marketed nationally, it assesses the probabilities of the national market
outcomes given each possible test market outcome. If the test market is great, the
probabilities for the national market outcomes are 0.8, 0.15, and 0.05. If the test market
is fair, these probabilities are 0.3, 0.5, and 0.2. If the test market is awful, they are
0.05, 0.25, and 0.7. (Note how the probabilities of the national market outcomes tend
to mirror the test market outcomes.)
The company wants to use a decision tree approach to find the best strategy.

Solution
We begin by discussing the three basic elements of this decision problem: the possible
strategies, the possible outcomes and their probabilities, and the value model. The
possible strategies are clear. Acme must first decide whether to conduct a test market.
Then it must decide whether to introduce the product nationally. However, it is important
to realize that if Acme decides to conduct a test market, it can base the national
market decision on the results of the test market. In this case its final strategy will
be a contingency plan, where it conducts the test market, then introduces the product
nationally if it receives sufficiently positive test market results and abandons the product
if it receives sufficiently negative test market results. The optimal strategies from many
multistage decision problems involve similar contingency plans.
Regarding the uncertain outcomes and their probabilities, we note that the given
probabilities—probabilities of test market outcomes and conditional probabilities of
national market outcomes given test market outcomes—are exactly the ones we need in
the decision tree. This is because the test market outcome is known before the national
market outcome will occur. However, suppose Acme decides not to run a test market
and then decides to market nationally. Then what are the probabilities of the national
market outcomes?
It is important to realize that we cannot simply assess three new probabilities for
this situation. These probabilities are implied by the given probabilities. This follows
from the rules of conditional probability. If we let T1 , T2 , and T3 be the test market
outcomes, and N be any of the national market outcomes, then by the addition rule for
probability and the conditional probability formula,
P(N ) = P(N and T1 ) + P(N and T2 ) + P(N and T3 ) (10.1)
= P(N |T1 )P(T1 ) + P(N |T2 )P(T2 ) + P(N |T3 )P(T3 ) (10.2)
(This is sometimes called the law of total probability.) For example, if N1 represents
a great national market, then from equation (10.1),
P(N1 ) = (0.8)(0.3) + (0.3)(0.6) + (0.05)(0.1) = 0.425
Similarly, we find that P(N2 ) = 0.37 and P(N3 ) = 0.205. These are the probabilities
we need to use for the probability branches when no test market is used.
Finally, the monetary values in the tree are straightforward. There are fixed costs
of test marketing or marketing nationally, and these are incurred as soon as these “go
ahead” decisions are made. From that point, we observe the sales volumes and multiply
them by the unit margin to obtain the profits.
PRECISION
USING PRECISIONTREE
TREE
The inputs for the decision tree appear in Figure 10.26 (page 526). (See file ACME.
XLS.) The only calculated values in this part of the spreadsheet are in row 28, which
follow from equation (10.1). Specifically, the formula in cell B28 is
=SUMPRODUCT(B22:B24,$B$16:$B$18)
which we copy across row 28. The tree is then straightforward to build and label, as
shown in Figure 10.27 (page 527). Note how the fixed costs of test marketing and
marketing nationally appear on the decision branches where they occur, so that only
the selling profits need to be placed on the probability branches. Also, the probabilities
on the various probability branches are exactly those listed in Figure 10.26.
The interpretation of this tree is fairly straightforward if we realize that each value
just below each node name is an EMV. For example, the 807 in cell B43 is the EMV for
10.4 Multistage Decision Problems 525

FIGURE 10.26
Inputs for Acme
Marketing Example
the entire decision problem. It means that Acme’s best EMV is $807,000. As another
example, the 5910 in cell D47 means that if Acme ever gets to that point—the test
market has been conducted and it has been great—the EMV for ACME is $5,910,000.
Each of these EMVs has been calculated by the folding-back procedure we discussed
earlier, starting from the right and working back toward the left. PrecisionTree takes
EMVs at probability nodes and maximums at decision nodes.
We can also see Acme’s optimal strategy by following the “TRUE” branches from
left to right. Acme should first run a test market. If the test market results are great,
then the product should be marketed nationally. However, if the test market results are
only fair or awful, the product should be abandoned. In these cases the prospects from
a national market look bleak, so Acme should cut its losses. (And there are losses. In
these latter two cases, Acme has spent $3,000,000 on the test market and has recouped
only $1,800,000 or $540,000 on test market sales.)
The risk profile from the optimal strategy appears in Figure 10.28 (page 528). It
is based on the data in Figure 10.29 (page 528). (These were obtained by clicking on
PrecisionTree’s “staircase” button and selecting the Statistics and Risk Profile options.)
We see that there is a small chance of two possible large losses (approximately $73
million and $35 million), there is a 70% chance of a moderate loss of about $1 or $2
million, and there is a 24% chance of an $18.6 million profit. Of course, the net effect
is an EMV of $807,000.
You might argue that the large potential losses and the slightly higher than 70%
chance of some loss should persuade Acme to abandon the product right away—without
a test market. However, this is what “playing the averages” with EMV is all about.
Since the EMV of this optimal strategy is greater than 0, the EMV of abandoning the
product right away, Acme should go ahead with this optimal strategy if the company is
indeed an EMV maximizer. In Section 10.8 we will see how this reasoning can change
if Acme is a risk-averse decision maker—as it might be with multimillion dollar losses
looming in the future!

FIGURE 10.27 Decision Tree for Acme Marketing Example
Expected Value of Sample Information The role of the test market in the Acme
marketing example is to provide information in the form of more accurate probabilities
of national market results. Information usually costs something, as it does in Acme’s
problem. Currently, the fixed cost of the test market is $3 million, which is evidently not
too much to pay because Acme’s best strategy is to conduct the test market. However,
we might ask how much this test market is worth. This is easy to answer. From the
decision tree in Figure 10.27, we see that the EMV from test marketing is $807,000
better than the decision not to test market (and then abandon the product). Therefore, if
the fixed cost of test marketing were any more than $807,000 above its current value,

FIGURE 10.28
Risk Profile of
Optimal Strategy
FIGURE 10.29
Distribution of
Profit/Loss from the
Optimal Strategy
Acme would be better not to run a test market. Equivalently, the most Acme would be
willing to pay for the test market (as a fixed cost) is $3.807 million.
This value is called the expected value of sample information, or EVSI. In
general, we can write the following expression for EVSI:
EVSI = EMV with free information − EMV without information
In Acme’s problem, the EMV with free information is $3.807 million (just don’t charge
for the test market fixed cost), and the EMV without any test market information is $0
(because Acme abandons the product when there is no test market available). Therefore,
EVSI = $3.807 − $0 = $3.807 million
Expected Value of Perfect Information The reason for the term sample is that the
information does not remove all uncertainty about the future. That is, even after the
test market results are in, there is still uncertainty about the national market results.
Therefore, we might go one step further and ask how much perfect information is
worth. We can imagine perfect information as an envelope that contains the true final
outcome (of the national market). That is, either “the national market will be great,”
“the national market will be fair,” or “the national market will be awful” is written
inside the envelope. Admittedly, no such envelope exists, but if it did, how much would
Acme be willing to pay for it?
We can answer this question with the simple decision tree in Figure 10.30. Now
the probability node on the left corresponds to opening the envelope. Its probabilities
are the same as before (when there is no test market available). Note the reasoning
here. Acme doesn’t know what the contents of the envelope will be, so we need a
probability node. However, once the envelope is opened, the true national market
outcome will be revealed. At that point Acme’s decision is fairly obvious. If it learns
that a national market will be great, it knows the product will be profitable and will
market it. Otherwise, if it learns that the national market will be fair or poor, it knows

FIGURE 10.30
EVPI for Acme
Marketing Example
that there will be a loss from marketing nationally, so it will abandon the product.
Folding back in the usual way produces an EMV of $7.65 million.
Now compare this $7.65 million with the EMV in the top part of Figure 10.27 that
results from no test market, namely, $0. The difference, $7.65 million, is called the
expected value of perfect information, or EVPI. It represents the maximum amount
the company would pay for perfect information about the final outcome (of the national
market). In general, the expression for EVPI is
EVPI = EMV with free perfect information − EMV with no information
In Acme’s case this expression becomes
EVPI = $7.65 − $0 = $7.65 million
The EVPI may appear to be an irrelevant concept since perfect information is almost
never available—at any price. However, it is often useful because it represents an upper
bound on the EVSI for any potential sample information. That is, no sample information
can ever be worth more than the EVPI. For example, if Acme is contemplating an
expensive test market with an anticipated fixed cost of more than $8 million, then
there is really no point in pursuing it any further. The information gained from this test
market, no matter how reliable it is, cannot possibly justify its cost because its cost is
greater than the EVPI. ■

PROBLEMS
Skill-Building Problems 40% chance that its annual maintenance cost will be
$300. The CPA’s office manager believes that the
19. The senior executives of an oil company are trying repairperson will issue a satisfactory report on the
to decide whether to drill for oil in a particular field first machine with probability 0.50.
in the Gulf of Mexico. It costs the company a. Provided that the CPA wishes to minimize
$300,000 to drill in the selected field. Company the expected total cost of purchasing and
executives believe that if oil is found in this field its maintaining one of these two machines for a
estimated value will be $1,800,000. At present, this 1-year period, which machine should she
oil company believes that there is a 50% chance that purchase? When, if ever, would it be worthwhile
the selected field actually contains oil. Before for the CPA to obtain the repairperson’s review
drilling, the company can hire a geologist at a cost of the first machine?
of $30,000 to prepare a report that contains a b. Compute and interpret the expected value of
recommendation regarding drilling in the selected sample information (EVSI) in this decision
field. There is a 55% chance that the geologist problem.
will issue a favorable recommendation and a c. Compute and interpret the expected value of
45% chance that the geologist will issue an perfect information (EVPI) in this decision
unfavorable recommendation. Given a favorable problem.
recommendation from the geologist, there is a 75% 21. FineHair is developing a new product to promote
chance that the field actually contains oil. Given an hair growth in cases of male pattern baldness. If
unfavorable recommendation from the geologist, FineHair markets the new product and it is
there is a 15% chance that the field actually successful, the company will earn $500,000 in
contains oil. additional profit. If the marketing of this new
a. Assuming that this oil company wishes to product proves to be unsuccessful, the company will
maximize its expected net earnings, determine its lose $350,000 in development and marketing
optimal strategy through the use of a decision costs. In the past, similar products have been
tree. successful 60% of the time. At a cost of $50,000, the
b. Compute and interpret the expected value of effectiveness of the new restoration product can be
sample information (EVSI) in this decision thoroughly tested. If the results of such testing are
problem. favorable, there is an 80% chance that the marketing
c. Compute and interpret the expected value of efforts of this new product will be successful. If the
perfect information (EVPI) in this decision results of such testing are not favorable, there is a
problem. mere 30% chance that the marketing efforts of this
20. A local certified public accountant must decide new product will be successful. FineHair currently
which of two copying machines to purchase for her believes that the probability of receiving favorable
expanding business. The cost of purchasing the first test results is 0.60.
machine is $4500, and the cost of maintaining the a. Identify the strategy that maximizes FineHair’s
first machine each year is uncertain. The CPA’s expected net earnings in this situation.
office manager believes that the annual maintenance b. Compute and interpret the expected value of
cost for the first machine will be $0, $150, or $300 sample information (EVSI) in this decision
with probabilities 0.35, 0.35, and 0.30, respectively. problem.
The cost of purchasing the second machine is c. Compute and interpret the expected value of
$3000, and the cost of maintaining the second perfect information (EVPI) in this decision
machine through a guaranteed maintenance problem.
agreement is $225 per year. 22. Hank is considering placing a bet on the upcoming
Before the purchase decision is made, the showdown between the Penn State and Michigan
CPA can hire an experienced copying machine football teams in State College. The winner of this
repairperson to evaluate the quality of the first contest will represent the Big Ten Conference in the
machine. Such an evaluation would cost the CPA Rose Bowl on New Year’s Day. Without any
$60. If the repairperson believes that the first additional information, Hank believes that each team
machine is satisfactory, there is a 65% chance that has an equal chance of winning this big game. If he
its annual maintenance cost will be $0 and a 35% wins the bet, he will win $500; if he loses the bet, he
chance that its annual maintenance cost will be will lose $550. Before placing his bet, he may decide
$150. If, however, the repairperson believes that the to pay his friend Al, who happens to be a football
first machine is unsatisfactory, there is a 60% chance sportswriter for the Philadelphia Enquirer, $50 for
that its annual maintenance cost will be $150 and a Al’s expert prediction on the game. Assume that Al

predicts that Penn State will win similar games 55% recommendation with probability 0.50. Given a
of the time, and that Michigan will win similar positive recommendation to market the new product,
games 45% of the time. Furthermore, Hank knows the new brand will eventually succeed in the
that when Al predicts that Penn State will win, there marketplace with probability 0.75. Given a negative
is a 70% chance that Penn State will indeed win the recommendation regarding the marketing of the new
football game. Finally, when Al predicts that product, the new brand will eventually succeed in
Michigan will win, there is a 20% chance that Penn the marketplace with probability 0.25.
State will proceed to win the upcoming game. a. In order to maximize expected profit in this case,
a. In order to maximize his expected profit from what course of action should the C&B product
this betting opportunity, how should Hank manager take?
proceed in this situation? b. Compute and interpret the expected value of
b. Compute and interpret the expected value of sample information (EVSI) in this decision
sample information (EVSI) in this decision problem.
problem. c. Compute and interpret the expected value of
c. Compute and interpret the expected value of perfect information (EVPI) in this decision
perfect information (EVPI) in this decision problem.
problem.
23. A product manager at Clean & Brite seeks to Skill-Extending Problems
determine whether her company should market a
24. A publishing company is trying to decide whether to
new brand of toothpaste. If this new product
publish a new business law textbook. Based on a
succeeds in the marketplace, C&B estimates that it
careful reading of the latest draft of the manuscript,
could earn $1,800,000 in future profits from the sale
the publisher’s senior editor in the business textbook
of the new toothpaste. If this new product fails,
division assesses the distribution of possible payoffs
however, the company expects that it could lose
earned by publishing this new book. Table 10.23
approximately $750,000. If C&B chooses not to
contains this probability distribution. Before making
market this new brand, the product manager believes
a final decision regarding the publication of the
that there would be little, if any, impact on the
book, the editor can learn more about the text’s
profits earned through sales of C&B’s other
potential for success by thoroughly surveying
products. The manager has estimated that the new
business law instructors teaching at universities
toothpaste brand will succeed with probability
across the country. Historical frequencies based on
0.55. Before making her decision regarding this
similar surveys administered in the past are provided
toothpaste product, the manager can spend $75,000
in Table 10.24.
on a market research study. Such a study of
a. Find the strategy that maximizes the publisher’s
consumer preferences will yield either a positive
expected payoff (in dollars).
recommendation with probability 0.50 or a negative
TABLE 10.23 Distribution of Payoffs for New

Business Law Textbook
Textbook Estimated
Performance Probability Payoff (if published)
Very strong 0.20 $100,000
Moderately strong 0.20 $50,000
Fair 0.20 $0
Poor 0.40 −$50,000
TABLE 10.24 Historical Frequencies of Combinations of

Past Survey Results and Actual Outcomes
Survey Indication/ Very Moderately
Actual Performance Strong Strong Fair Poor
Very strong 13 12 2 3
Moderately strong 10 20 6 4
Fair 5 12 15 8
Poor 1 3 9 22

b. What is the most (in dollars) that the publisher a. For which values of p, if any, does Sharp Outfits
should be willing to pay to conduct a new survey minimize its expected total cost by choosing to
of business law instructors? postpone shipping its customer orders via UPS?
c. If the actual cost of conducting the given survey b. Suppose now that, at a cost of $1000, Sharp
is less than the amount identified in part a, what Outfits can purchase information regarding the
should the publisher do? likelihood of a UPS strike in the near future.
d. Assuming that a survey could be constructed that Based on similar strike threats in the past, the
provides “perfect information” to the publisher, probability that this information indicates the
how much should the company be willing to pay occurrence of a UPS strike is 27.5%. If the
to acquire and implement such a survey? purchased information indicates the occurrence
25. Sharp Outfits is trying to decide whether to ship of a UPS strike, the chance of a strike actually
some customer orders now via UPS or wait until occurring is 0.105/0.275. If the purchased
after the threat of another UPS strike is over. If information does not indicate the occurrence of a
Sharp Outfits decides to ship the requested UPS strike, the chance of a strike actually
merchandise now and the UPS strike takes place, the occurring is 0.680/0.725. Provided that
company will incur $60,000 in delay and shipping p = 0.15, what strategy should Sharp Outfits
costs. If Sharp Outfits decides to ship the customer pursue to minimize its expected total cost?
orders via UPS and no strike occurs, the company c. Continuing part b, compute and interpret the
will incur $4000 in shipping costs. If Sharp Outfits expected value of sample information (EVSI)
decides to postpone shipping its customer orders via when p = 0.15.
UPS, the company will incur $10,000 in delay costs d. Continuing part b, compute and interpret the
regardless of whether or not UPS goes on strike. Let expected value of perfect information (EVPI)
p represent the probability that UPS will go on when p = 0.15.
strike and impact Sharp Outfits’s shipments.
10.5 BAYES’ RULE

n multistage decision problems we typically have alternating sets of decision nodes
I and probability nodes. The decision maker makes a decision, some uncertain out-
comes are observed, the decision maker makes another decision, more uncertain
outcomes are observed, and so on. In the resulting decision tree, all probability branches
at the right of the tree are conditional on outcomes that have occurred earlier, to their
left. Therefore, the probabilities on these branches are of the form P( A|B), where B
is an event that occurs before event A in time. However, it is sometimes more natural
to assess conditional probabilities in the opposite order, that is, P(B| A). Whenever
this is the case, we require Bayes’ rule to obtain the probabilities we need on the tree.
Essentially, Bayes’ rule is a mechanism for updating probabilities as new information
becomes available. We illustrate the mechanics of Bayes’ rule in the following example.
[See Feinstein (1990) for a real application of this example.]
E XAMPLE 10.5
DRUG TESTING COLLEGE ATHLETES
If an athlete is tested for a certain type of drug usage (steroids, say), then the test result
will be either positive or negative. However, these tests are never perfect. Some athletes
who are drug free test positive, and some who are drug users test negative. The former
are called false positives; the latter are called false negatives. We will assume that 5%

of all athletes use drugs, 3% of all tests on drug-free athletes yield false positives, and
7% of all tests on drug users yield false negatives. The question then is what we can
conclude from a positive or negative test result.
Solution
Let D and N D denote that a randomly chosen athlete is or is not a drug user, and
let T + and T − indicate a positive or negative test result. We are given the following
probabilities. First, since 5% of all athletes are drug users, we know that P(D) =
0.05 and P(N D) = 0.95. These are called prior probabilities because they represent
the chance that an athlete is or is not a drug user prior to the results of a drug
test. Second, from the information on drug test accuracy, we know the conditional
probabilities P(T + |N D) = 0.03 and P(T − |D) = 0.07. But a drug-free athlete tests
either positive or negative, and the same is true for a drug user. Therefore, we also have
the probabilities P(T − |N D) = 0.97 and P(T + |D) = 0.93. These four conditional
probabilities of test results given drug user status are often called the likelihoods of
the test results.
Given these priors and likelihoods, we want posterior probabilities such as
P(D|T +), the probability that an athlete who tested positive is a drug user, or
P(N D|T −), the probability that an athlete who tested negative is drug free. They
are called posterior probabilities because they are assessed after the drug test results.
This is where Bayes’ rule enters. We will develop Bayes’ rule in some generality and
then apply it to the present example.
Let A be any “information” event, such as the result of a drug test, and let
B1 , B2 , . . . , Bn be any mutually exclusive and exhaustive set of events. That is, exactly
one of the Bi ’s must occur. To apply Bayes’ rule, we assume that the prior probabilities
P(B1 ), P(B2 ), . . . , P(Bn ) are given, as are the likelihoods P( A|Bi ) for each i. Then
we want the posterior probabilities P(Bi | A) for each i. Bayes’ rule shows how to find
these. For any i, we have
P( A|Bi )P(Bi )
P(Bi | A) = Bayes’ rule
P( A|B1 )P(B1 ) + · · · + P( A|Bn )P(Bn )
This formula says that a typical posterior probability is a ratio. The numerator is
a likelihood times a prior, and the denominator is the sum of likelihoods times
priors.
Before illustrating Bayes’ rule numerically, we make two other observations about
the terms in Bayes’ rule. First, we can use the multiplication rule of probability to write
any product of a likelihood and a prior as
P( A|Bi )P(Bi ) = P( A and Bi )
The probability on the right, that both A and Bi occur, is called a joint probability.
Second, we can use the definition of conditional probability directly to write
P( A and Bi )
P(Bi | A) =
P( A)
Therefore, the probability in the denominator of Bayes’ rule is really just the probability
of A:
P( A) = P( A|B1 )P(B1 ) + · · · + P( A|Bn )P(Bn )
As we will see shortly, this natural by-product of Bayes’ rule will come in very handy
in decision trees.
10.5 Bayes’ Rule 533

It is fairly easy to implement Bayes’ rule in a spreadsheet, as illustrated in Figure
10.31 for the drug example. Here A corresponds to either test result, and B1 and B2
correspond to D and N D. (See the file DRUGBAYES.XLS.4) In words, we want to
see how the chances of D and N D change after seeing the results of the drug test.
The given priors and likelihoods are listed in the ranges B5:C5 and B9:C10. We
then calculate the products of likelihoods and priors in the range B15:C16. The formula
in cell B15 is
=B$5*B9
and this is copied to the rest of the B15:C16 range. Their row sums are calculated in
the range D15:D16. These represent the unconditional probabilities of the two possible
results. They are also (as we saw above) the denominators of Bayes’ rule. Finally, we
calculate the posterior probabilities in the range B21:C22. The formula in cell B21 is
=B15/$D15
and this is copied to the rest of the B21:C22 range. The various 1’s in the margins of
Figure 10.31 are row sums or column sums that must equal 1. We show them only as
checks of our logic.
FIGURE 10.31
Bayes’ Rule for
Drug-Testing
Example
Note that a negative test result leaves little doubt that the athlete is drug free.
The posterior probability that the athlete is drug free, given a negative test result, is
0.996. However, there is still a lot of doubt about an athlete who tests positive. The
posterior probability that the athlete uses drugs, given a positive test result, is only
0.620. This asymmetry occurs because of the prior probabilities. We are fairly certain
that a randomly selected athlete is drug free because only 5% of all athletes use drugs.
It takes a lot of evidence to convince us otherwise. This initial bias, plus the fact that
the test produces a few false positives, means that athletes with positive test results still
have a decent chance (probability 0.380) of being drug free. Is this a valid argument
4
The Bayes2 sheet in this file illustrates how Bayes’ rule can be used when there are more than two
possible test results and/or drug user categories.

for not requiring drug testing of athletes? We explore this question in the following
continuation of the drug-testing example. It all depends on the “costs.” (It might also
depend on whether there is a second type of test that could help confirm the findings
of the first test. However, we won’t consider such a test.) ■
E XAMPLE 10.5 (continued)

DRUG TESTING COLLEGE ATHLETES
The administrators at State University are trying to decide whether to institute manda-
tory drug testing for the athletes. They have the same information about priors and
likelihoods as in the previous example, but now they want to use a decision tree
approach to see whether the benefits outweigh the costs.5
Solution
We have already discussed the uncertain outcomes and their probabilities. Now we need
to discuss the decision alternatives and the monetary values—the other two elements
of a decision analysis. We will assume that there are only two alternatives: perform
drug testing on all athletes or don’t perform any drug testing. In the former case we
assume that if an athlete tests positive, this athlete is barred from sports.
The “monetary” values are more difficult to assess. They include
■ the benefit B from correctly identifying a drug user and barring him or her from
sports
■ the cost C1 of the test itself for a single athlete (materials and labor)
■ the cost C2 of falsely accusing a nonuser (and barring him or her from sports)
■ the cost C3 of not identifying a drug user (either by not testing at all or by obtaining
a false negative)
■ the cost C4 of violating a nonuser’s privacy by performing the test
It is clear that only C1 is a direct monetary cost that is easy to measure. However,
the other “costs” and the benefit B are real, and they must be compared on some scale
to enable administrators to make a rational decision. We will do so by comparing
everything to the cost C1 , to which we will assign value 1. (This does not mean that the
cost of testing an athlete is necessarily $1; it just means that we will express all other
costs as multiples of C1 .) Clearly, there is a lot of subjectivity involved in making these
comparisons, so sensitivity analysis on the final decision tree is a must.
Before developing this decision tree, it is useful to form a benefit–cost table for
both alternatives and all possible outcomes. Because we will eventually maximize
expected net benefit, all benefits in this table have a positive sign and all costs have
a negative sign. These net benefits appear in Table 10.25 (page 536). The first two
columns are relevant if no tests are performed; the last four are relevant when testing
is performed. For example, if a positive test is obtained for a nonuser, there are three
5
Again, see Feinstein (1990) for an enlightening discussion of this drug-testing problem at a real
university.

TABLE 10.25 Net Benefit for Drug-Testing Example
Don’t Test Perform Test
D ND D and T + N D and T + D and T − N D and T −

−C3 0 B − C1 −(C1 + C2 + C4 ) −(C1 + C3 ) −(C1 + C4 )
costs: the cost of the test (C1 ), the cost of falsely accusing the athlete (C2 ), and the cost
of violating the nonuser’s privacy (C4 ). The other entries are obtained similarly.
The solution with PrecisionTree shown in Figure 10.32 is now fairly straightfor-
ward. (See the file DRUG.XLS.) We first enter all of the benefits and costs in an input
section. These, together with the Bayes’ rule calculations from before, appear at the
top of the spreadsheet. Then we use PrecisionTree in the usual way to build the tree
and enter the links to the values and probabilities.
FIGURE 10.32 Decision Tree for Drug-Testing Example

Before we interpret this solution, we discuss the timing (from left to right). If drug
testing is performed, the result of the drug test is observed first (a probability node).
Each test result leads to an action (bar from sports or don’t), and then the eventual
benefit or cost depends on whether the athlete uses drugs (again a probability node).
You might argue that the university never knows for certain whether the athlete uses
drugs, but we must include this information in the tree to get the benefits and costs
correct. If no drug testing is performed, then there is no intermediate test result node
or branches.
Now to the interpretation. First, we discuss the benefits and costs shown in Fig-
ure 10.32. These were chosen fairly arbitrarily, but with some hope of reflecting reality.
They say that the largest cost is falsely accusing (and barring) a nonuser. This is 50
times as large as the cost of the test. The benefit of identifying a drug user is only
half this large, and the cost of not identifying a user is 40% as large as barring a
nonuser. The violation of privacy of a nonuser is twice as large as the cost of the
test. Based on these values, the decision tree implies that drug testing should not be
performed. The EMVs for testing and for not testing are both negative, indicating that
the costs outweigh the benefits for each, but the EMV for not testing is slightly less
negative.6
What would it take to change this decision? We’ll start with the assumption,
probably accepted by most people in our society, that the cost of falsely accusing a
nonuser (C2 ) ought to be the largest of the benefits or costs in the range B4:B10. In
fact, because of possible legal costs, we might argue that C2 should be more than 50
times the cost of the test. But if we increase C2 , the scales are tipped even farther
in the direction of not testing. On the other hand, if the benefit B from identifying a
user and/or the cost C3 for not identifying a user increase, then testing might be the
preferred alternative. We tried this, keeping C2 constant at 50. When B and C3 both
had value 45, no testing was still optimal, but when they both increased to 50—the
same magnitude as C2 —then testing won out by a small margin. However, it would be
difficult to argue that B and C3 should be of the same magnitude as C2 .
Other than the benefits and costs, the only other thing we might vary is the accuracy
of the test, measured by the error probabilities in cells B14 and B15. Presumably, if the
test makes fewer false positives and false negatives, testing might be a more attractive
alternative. We tried this, keeping the benefits and costs the same as those shown in
Figure 10.32 but changing the error probabilities. Even when each error probability
was decreased to 0.01, however, the no-testing alternative was still optimal—by a fairly
wide margin.
In summary, based on a number of reasonable assumptions and parameter settings,
this example has shown that it is difficult to make a case for mandatory drug testing. ■
6
The university in the Feinstein (1990) study came to the same conclusion.

PROBLEMS
Skill-Building Problems that the company will gain $500,000. If she decides
to fire this vice-president but he is not the informer,
26. Consider a population of 2000 individuals, 800 of
the company will lose his expertise and still have an
whom are women. Assume that 300 of the women
informer within the staff; the CEO estimates that
in this population earn at least $60,000 per year, and
this outcome would cost her company about $2.5
200 of the men earn at least $60,000 per year.
million. If she decides not to fire this vice-president,
a. What is the probability that a randomly selected
she estimates that the firm will lose $1.5 million
individual from this population earns less than
whether or not he actually is the informer (since in
$60,000 per year?
either case the informer is still with the company).
b. If a randomly selected individual is observed to
Before deciding whether to fire the
earn less than $60,000 per year, what is the
vice-president for finance, the CEO could order lie
probability that this person is a man?
detector tests. To avoid possible lawsuits, the lie
c. If a randomly selected individual is observed to
detector tests would have to be administered to all
earn at least $60,000 per year, what is the
company employees, at a total cost of $150,000.
probability that this person is a woman?
Another problem she must consider is that the
27. Yearly automobile inspections are required for
available lie detector tests are not perfectly reliable.
residents of the state of Pennsylvania. Suppose that
In particular, if a person is lying, the test will reveal
18% of all inspected cars in Pennsylvania have
that the person is lying 95% of the time. Moreover,
problems that need to be corrected. Unfortunately,
if a person is not lying, the test will indicate that the
Pennsylvania state inspections fail to detect these
person is not lying 85% of the time.
problems 12% of the time. Consider a car that is
a. In order to minimize the expected total cost of
inspected and is found to be free of problems. What
managing this difficult situation, what strategy
is the probability that there is indeed something
should the CEO adopt?
wrong that the inspection has failed to uncover?
b. Should the CEO order the lie detector tests for all
28. Consider again the landowner’s decision problem
of her employees? Explain why or why not.
described in Problem 3. Suppose now that, at a cost
c. Determine the maximum amount of money that
of $90,000, the landowner can request that a
the CEO should be willing to pay to administer
soundings test be performed on the site where
lie detector tests.
natural gas is believed to be present. The company
that conducts the soundings concedes that 30% of 30. A customer has approached a bank for a $10,000
the time the test will indicate that no gas is present one-year loan at a 12% interest rate. If the bank does
when it actually is. When natural gas is not present not approve this loan application, the $10,000 will
in a particular site, the soundings test is accurate be invested in bonds that earn a 6% annual return.
90% of the time. Without additional information, the bank believes
a. Given that the landowner pays for the soundings that there is a 4% chance that this customer will
test and the test indicates that gas is present, what default on the loan, assuming that the loan is
is the landowner’s revised estimate of the approved. If the customer defaults on the loan, the
probability of finding gas on this site? bank will lose $10,000.
b. Given that the landowner pays for the soundings At a cost of $100, the bank can thoroughly
test and the test indicates that gas is not present, investigate the customer’s credit record and supply a
what is the landowner’s revised estimate of the favorable or unfavorable recommendation. Past
probability of not finding gas on this site? experience indicates that in cases where the
c. Should the landowner request the given customer did not default on the approved loan, the
soundings test at a cost of $90,000? Explain why probability of receiving a favorable recommendation
or why not. If not, when (if ever) would the on the basis of the credit investigation was 77/96.
landowner choose to obtain the soundings test? Furthermore, in cases where the customer defaulted
29. The chief executive officer of a firm in a highly on the approved loan, the probability of receiving a
competitive industry believes that one of her key favorable recommendation on the basis of the credit
employees is providing confidential information to investigation was 1/4.
the competition. She is 90% certain that this a. What course of action should the bank take to
informer is the vice-president of finance, whose maximize its expected profit?
contacts have been extremely valuable in obtaining b. Compute and interpret the expected value of
financing for the company. If she decides to fire this sample information (EVSI) in this decision
vice-president and he is the informer, she estimates problem.

c. Compute and interpret the expected value of the disease test negative. The government’s goal is
perfect information (EVPI) in this decision to maximize the expected net financial benefits per
problem. potential immigrant.
31. A company is considering whether to market a new a. Let x = 10 (i.e., 10%). What is the largest value
product. Assume, for simplicity, that if this product of T at which the U.S. government will choose
is marketed, there are only two possible outcomes: to test potential immigrants for the disease?
success or failure. The company assesses that the b. How does your answer to the question in part a
probabilities of these two outcomes are p and change when x increases to 15?
1 − p, respectively. If the product is marketed and it c. Let x = 10 and T = $100. Find the
proves to be a failure, the company will lose government’s optimal strategy in this case.
$450,000. If the product is marketed and it proves to d. Let x = 10 and T = $100. Compute and
be a success, the company will gain $750,000. interpret the expected value of perfect
Choosing not to market the product results in no information (EVPI) in this decision problem.
gain or loss for the company.
The company is also considering whether to
survey prospective buyers of this new product. The Skill-Extending Problems
results of the consumer survey can be classified as 33. A city in Ohio is considering replacing its fleet of
favorable, neutral, or unfavorable. In similar cases gasoline-powered automobiles with electric cars.
where proposed products proved to be market The manufacturer of the electric cars claims that this
successes, the likelihoods that the survey results municipality will experience significant cost savings
were favorable, neutral, and unfavorable were 0.6, over the life of the fleet if it chooses to pursue the
0.3, and 0.1, respectively. In similar cases where conversion. If the manufacturer is correct, the city
proposed products proved to be market failures, the will save about $1.5 million dollars. If the new
likelihoods that the survey results were favorable, technology employed within the electric cars is
neutral, and unfavorable were 0.1, 0.2, and 0.7, faulty, as some critics suggest, the conversion to
respectively. The total cost of administering this electric cars will cost the city $675,000. A third
survey is C dollars. possibility is that less serious problems will arise
a. Let p = 0.4. For which values of C, if any, and the city will break even with the conversion. A
would this company choose to conduct the consultant hired by the city estimates that the
consumer survey? probabilities of these three outcomes are 0.30, 0.30,
b. Let p = 0.4. What is the largest amount that this and 0.40, respectively.
company would be willing to pay for perfect The city has an opportunity to implement a
information about the potential success or failure pilot program that would indicate the potential cost
of the new product? or savings resulting from a switch to electric cars.
c. Let p = 0.5 and C = $15,000. Find the strategy The pilot program involves renting a small number
that maximizes the company’s expected earnings of electric cars for 3 months and running them under
in this situation. Does the optimal strategy typical conditions. This program would cost the city
involve conducting the consumer survey? $75,000. The city’s consultant believes that the
Explain why or why not. results of the pilot program would be significant but
32. The U.S. government is attempting to determine not conclusive; she submits Table 10.26 (page 398),
whether immigrants should be tested for a a compilation of probabilities based on the
contagious disease. Let’s assume that the decision experience of other cities, to support her contention.
will be made on a financial basis. Furthermore, For example, the first row of her table indicates that
assume that each immigrant who is allowed to enter given that a conversion to electric cars actually
the United States and has the disease costs the results in a savings of $1.5 million, the conditional
country $100,000. Also, each immigrant who is probabilities that the pilot program will indicate that
allowed to enter the United States and does not have the city saves money, loses money, and breaks even
the disease will contribute $10,000 to the national are 0.6, 0.1, and 0.3, respectively.
economy. Finally, assume that x percent of all a. What actions should this city take to maximize
potential immigrants have the disease. The U.S. the expected savings?
government can choose to admit all immigrants, b. Should the city implement the pilot program at a
admit no immigrants, or test immigrants for the cost of $75,000?
disease before determining whether they should be c. Compute and interpret the expected value of
admitted. It costs T dollars to test a person for the sample information (EVSI) in this decision
disease; the test result is either positive or negative. problem.
A person who does not have the disease always tests
negative. However, 20% of all people who do have

TABLE 10.26 Likelihoods of Pilot Program Outcomes Given
Actual Conversion Outcomes
Actual Outcome of Conversion/
Pilot Program Indication Savings Loss Break Even
Savings 0.6 0.1 0.3
Loss 0.1 0.4 0.5
Break Even 0.4 0.2 0.4
34. A manufacturer must decide whether to extend b. Should the manufacturer routinely obtain credit
credit to a retailer who would like to open an rating reports on those retailers who seek credit
account with the firm. Past experience with new approval? Why or why not?
accounts indicates that 45% are high-risk customers, c. Compute and interpret the expected value of
35% are moderate-risk customers, and 20% are sample information (EVSI) in this decision
low-risk customers. If credit is extended, the problem.
manufacturer can expect to lose $60,000 with 35. A television network earns an average of $1.6
a high-risk customer, make $50,000 with a million each season from a hit program and loses an
moderate-risk customer, and make $100,000 with a average of $400,000 each season on a program that
low-risk customer. If the manufacturer decides not turns out to be a flop. Of all programs picked up by
to extend credit to a customer, the manufacturer this network in recent years, 25% turn out to be hits
neither makes nor loses any money. and 75% turn out to be flops. At a cost of C dollars,
Prior to making a credit extension decision, the a market research firm will analyze a pilot episode
manufacturer can obtain a credit rating report on the of a prospective program and issue a report
retailer at a cost of $2000. The credit agency predicting whether the given program will end up
concedes that its rating procedure is not completely being a hit. If the program is actually going to be
reliable. In particular, the credit rating procedure a hit, there is a 90% chance that the market
will rate a low-risk customer as a moderate-risk researchers will predict the program to be a hit. If
customer with probability 0.10 and as a high-risk the program is actually going to be a flop, there is a
customer with probability 0.05. Furthermore, the 20% chance that the market researchers will predict
given rating procedure will rate a moderate-risk the program to be a hit.
customer as a low-risk customer with probability a. Assuming that C = $160,000, identify the
0.06 and as a high-risk customer with probability strategy that maximizes this television network’s
0.07. Finally, the rating procedure will rate a expected profit in responding to a newly
high-risk customer as a low-risk customer with proposed television program.
probability 0.01 and as a moderate-risk customer b. What is the maximum value of C that this
with probability 0.05. television network should be willing to incur in
a. Find the strategy that maximizes the choosing to hire the market research firm?
manufacturer’s expected net earnings. c. Compute and interpret the expected value of
perfect information (EVPI) in this decision
problem.
10.6 INCORPORATING ATTITUDES TOWARD RISK

ational decision makers are sometimes willing to violate the EMV maxi-
R mization criterion when large amounts of money are at stake. These decision
makers are willing to sacrifice some EMV to reduce risk. Are you ever willing
to do so personally? Consider the following scenarios.
1. You have a chance to enter a lottery where you will win $100,000 with probability
0.1 or win nothing with probability 0.9. Alternatively, you can receive $5000 for
certain. How many of you—truthfully—would take the certain $5000, even though

the EMV of the lottery is $10,000? Or change the $100,000 to $1,000,000 and the
$5000 to $50,000 and ask yourself whether you’d prefer the sure $50,000!
2. You can either buy collision insurance on your expensive new car or not buy it,
where the insurance costs a certain premium and carries some deductible provision.
If you decide to pay the premium, then you are essentially paying a certain amount
to avoid a gamble—the possibility of wrecking your car and not having it insured.
You can be sure that the premium is greater than the expected cost of damage;
otherwise, the insurance company would not stay in business. Therefore, from an
EMV standpoint you should not purchase the insurance. But how many of you
drive without this type of insurance?
These examples, the second of which is certainly realistic, illustrate situations
where rational people do not behave as EMV maximizers. Then how do they act? This
question has been studied extensively by many researchers, both mathematically and
behaviorally. Although the answer is still not agreed upon universally, most researchers
believe that if certain basic behavioral assumptions hold, people are expected utility
maximizers—that is, they choose the alternative with the largest expected utility. Al-
though we will not go deeply into the subject of expected utility maximization, the
discussion in this section will acquaint you with the main ideas.
Utility Functions
We begin by discussing an individual’s utility function. This is a mathematical function
that transforms monetary values—payoffs and costs—into utility values. Essentially,
an individual’s utility function specifies the individual’s preferences for various mon-
etary payoffs and costs and, in doing so, it automatically encodes the individual’s
attitudes toward risk. Most individuals are risk averse, which means intuitively that
they are willing to sacrifice some EMV to avoid risky gambles. In terms of the utility
function, this means that every extra dollar of payoff is worth slightly less to the in-
dividual than the previous dollar, and every extra dollar of cost is considered slightly
more costly (in terms of utility) than the previous dollar. The resulting utility functions
are shaped as shown in Figure 10.33. Mathematically, these functions are said to be
increasing and concave. The increasing part means that they go uphill—everyone
prefers more money to less money. The concave part means that they increase at a
decreasing rate. This is the risk-averse behavior.
There are two problems involved in implementing utility maximization in a real
decision analysis. The first is obtaining an individual’s (or company’s) utility function;
we will discuss this below. The second is using the resulting utility function to find the
best decision. This second step is actually quite straightforward. We simply substitute
FIGURE 10.33 Utility

Risk-Averse Utility
Function
Monetary value
10.6 Incorporating Attitudes Toward Risk 541

utility values for monetary values in the decision tree and then fold back as usual.
That is, we calculate expected utilities at probability branches and take maximums (of
expected utilities) at decision branches. We will look at a numerical example later in
this section. So the real work involves finding an individual’s (or company’s) utility
function in the first place.
Assessing a Utility Function

We will outline a method that can be used to estimate a person’s utility function. There
are two things we must understand about this method. First, it asks the person to make
a series of trade-offs. Because each of us has different attitudes toward risk, we will
not all make the trade-offs in the same way. Therefore, each of us will obtain our
own utility function. Second, even a particular person’s utility function is not unique.
If U (x) represents a person’s utility function, then it turns out that aU (x) + b also
describes that person’s utility function, for any constants a and b with a > 0. They are
equivalent in the sense that they lead to exactly the same decisions.
We take advantage of this nonuniqueness by specifying two points on the utility
function. Specifically, we begin by asking the person for two monetary values that
represent the worst possible loss and the best possible gain imaginable. Let’s say these
values are − A and B. Then we arbitrarily assign utility values 0 and 1 to these two
monetary values, that is, U (− A) = 0 and U (B) = 1. Don’t worry about the absolute
magnitudes, 0 and 1, we’ve assigned—we could assign any other values, such as 14
and 320. The important thing is to use these as “anchors” and then obtain other utility
values in terms of them.
The procedure is as follows. Given any two known utility values, say, U (x) and
U (y), where x and y are monetary values, we present the person with a choice between
the following two options:
■ Option 1: Obtain a certain payoff of z.
■ Option 2: Obtain a payoff of either x or y, depending on the flip of a fair coin.
Then we ask the person to select the monetary value z in option 1 so that he or she
is indifferent between the two options. If the person is indifferent, then the expected
utilities from the two options must be equal. We will call the resulting value of z the
indifference value. This leads to the equation for U (z):
U (z) = 0.5U (x) + 0.5U (y) (10.3)
In words, we have generated a new utility value from two known utility values. This
process continues until we have enough utility values to approximate a utility curve.
(Note that if any of x, y, and z are negative, then “payoff” really means “cost.”) We
will illustrate this procedure with the following example.
E XAMPLE 10.6
ASSESSING THE UTILITY FUNCTION FOR A
SMALL BUSINESS
John Jacobs owns his own business. Because he is about to make an important decision
where large losses or large gains are at stake, he wants to use the expected utility
criterion to make his decision. He knows that he must first assess his own utility

function, so he hires a decision analysis expert, Susan Schilling, to help him out. How
might the session between John and Susan proceed?
Solution
Susan first asks John for the largest loss and largest gain he can imagine. He answers
with the values $200,000 and $300,000, so she assigns utility values U (−200,000) = 0
and U (300,000) = 1 as anchors for the utility function. Now she presents John with
the choice between two options:
■ Option 1: Obtain a payoff of z (really a loss if z is negative).
■ Option 2: Obtain a loss of $200,000 or a payoff of $300,000, depending on the flip
of a fair coin.
Susan reminds John that the EMV of option 2 is $50,000 (halfway between
−$200,000 and $300,000). He realizes this, but because he is quite risk averse, he
would far rather have $50,000 for certain than take the gamble in option 2. Therefore,
the indifference value of z must be less than $50,000. Susan then poses several values
of z to John. Would he rather have $10,000 for sure or take option 2? He says he’d
rather take the $10,000. Would he rather pay $5000 for sure or take the gamble in op-
tion 2? (This is like an insurance premium.) He says he’d rather take option 2. By this
time, we know the indifference value of z must be less than $10,000 and greater than
−$5000. With a few more questions of this type, John finally decides on z = $5000 as
his indifference value. He is indifferent between obtaining $5000 for sure and taking
the gamble in option 2. We can substitute these values into equation (10.3):
U (5000) = 0.5U (−200,000) + 0.5U (300,000) = 0.5(0) + 0.5(1) = 0.5
Note that John is giving up $45,000 in EMV because of his risk aversion. The EMV of
the gamble in option 2 is $50,000, and he is willing to accept a sure $5000 in its place.
The process would then continue. For example, since she now knows U (5000) and
U (300,000), Susan could ask John to choose between these options:
■ Option 1: Obtain a payoff of z.
■ Option 2: Obtain a payoff of $5000 or a payoff of $300,000, depending on the flip
of a fair coin.
If John decides that his indifference value is now z = $130,000, then with equa-
tion (10.3) we know that
U (130,000) = 0.5U (5000) + 0.5U (300,000) = 0.5(0.5) + 0.5(1) = 0.75
Note that John is now giving up $22,500 in EMV because the EMV of the gamble in
option 2 is $152,500. By continuing in this manner, Susan can help John assess enough
utility values to approximate a continuous utility curve. ■
As this example illustrates, utility assessment is tedious. Even in the best of

circumstances, when a trained consultant attempts to assess the utility function of a
single person, the process requires the person to make a series of choices between
hypothetical alternatives involving uncertain outcomes. Unless the person has some
training in probability, these choices can be difficult to understand, let alone make,
and it is unlikely that the person will answer consistently as the questioning proceeds.
The process is even more difficult when a company’s utility function is being assessed.
Because company executives involved typically have different attitudes toward risk, it
is difficult for these people to reach a consensus on a common utility function.

Exponential Utility
For these reasons there are classes of “ready-made” utility functions that have been
developed. One important class is called exponential utility and has been used in many
financial investment analyses. An exponential utility function has only one adjustable
numerical parameter, and there are straightforward ways to discover the most appropri-
ate value of this parameter for a particular individual or company. So the advantage of
using an exponential utility function is that it is relatively easy to assess. The drawback
is that exponential utility functions do not capture all types of attitudes toward risk.
Nevertheless, their ease of use has made them popular.
An exponential utility function has the following form:
U (x) = 1 − e−x/R (10.4)
Here x is a monetary value (a payoff if positive, a cost if negative), U (x) is the utility of
this value, and R > 0 is an adjustable parameter called the risk tolerance. Basically,
the risk tolerance measures how much risk the decision maker will tolerate. The larger
the value of R, the less risk averse the decision maker is. That is, a person with a large
value of R is more willing to take risks than a person with a small value of R.
To assess a person’s (or company’s) exponential utility function, we need only to
assess the value of R. There are a couple of tips for doing this. First, it has been shown
that the risk tolerance is approximately equal to that dollar amount R such that the
decision maker is indifferent between the following two options:
■ Option 1: Obtain no payoff at all.
■ Option 2: Obtain a payoff of R dollars or a loss of R/2 dollars, depending on the
flip of a fair coin.
For example, if you are indifferent between a bet where you win $1000 or lose
$500, with probability 0.5 each, and not betting at all, then your R is approximately
$1000. From this criterion it certainly makes intuitive sense that a wealthier person (or
company) ought to have a larger value of R. This has been found in practice.
A second tip for finding R is based on empirical evidence found by Ronald
Howard, a prominent decision analyst. Through his consulting experience with sev-
eral large companies, he discovered tentative relationships between risk tolerance and
several financial variables—net sales, net income, and equity. [See Howard (1992).]
Specifically, he found that R was approximately 6.4% of net sales, 124% of net in-
come, and 15.7% of equity for the companies he studied. For example, according to
this prescription, a company with net sales of $30 million should have a risk tolerance
of approximately $1.92 million. Howard admits that these percentages are only guide-
lines. However, they do indicate that larger and more profitable companies tend to have
larger values of R, which means that they are more willing to take risks involving given
dollar amounts.
We illustrate the use of the expected utility criterion, and exponential utility in
particular, with the following example.
E XAMPLE 10.7
DECIDING WHETHER TO ENTER RISKY
VENTURES AT VENTURE LIMITED
Venture Limited is a company with net sales of $30 million. The company currently
must decide whether to enter one of two risky ventures or do nothing. The possible

outcomes of the less risky venture are a $0.5 million loss, a $0.1 million gain, and a $1
million gain. The probabilities of these outcomes are 0.25, 0.50, and 0.25. The possible
outcomes of the more risky venture are a $1 million loss, a $1 million gain, and a $3
million gain. The probabilities of these outcomes are 0.35, 0.60, and 0.05. If Venture
Limited can enter at most one of the two risky ventures, what should it do?
Solution
We will assume that Venture Limited has an exponential utility function. Also, based
on Howard’s guidelines, we will assume that the company’s risk tolerance is 6.4% of
its net sales, or $1.92 million. (We’ll do a sensitivity analysis on this parameter later
on.) We can substitute into equation (10.4) to find the utility of any monetary outcome.
For example, the gain from doing nothing is $0, and its utility is
U (0) = 1 − e−0/1.92 = 1 − 1 = 0
As another example, the utility of a $1 million loss is
U (−1) = 1 − e−(−1)/1.92 = 1 − 1.683 = −0.683
These are the values we use (instead of monetary values) in the decision tree.
PRECISION
USING PRECISIONTREE
TREE
Fortunately, PrecisionTree takes care of all the details. After we build a decision tree
and label it (with monetary values) in the usual way, we click on the name of the tree
(the box on the far left of the tree) to open the dialog box in Figure 10.34. We then fill
in the utility function information as shown in the upper right section of the dialog box.
This says to use an exponential utility function with risk tolerance 1.92. It also indicates
that we want expected utilities (as opposed to EMVs) to appear in the decision tree.
FIGURE 10.34
Dialog Box for
Specifying the
Exponential Utility
Criterion
The completed tree for this example appears in Figure 10.35 (page 546). (See
the file VENTURE.XLS.) We build it in exactly the same way as usual and link
probabilities and monetary values to its branches in the usual way. For example, there
is a link in cell C22 to the monetary value in cell A10. However, the expected values
shown in the tree (those shown in color on your screen) are expected utilities, and the
optimal decision is the one with the largest expected utility. In this case the expected
utilities for doing nothing, investing in the less risky venture, and investing in the more

FIGURE 10.35
Decision Tree for
Risky Venture
Example
risky venture are 0, 0.0525, and 0.0439. Therefore, the optimal decision is to invest in
the less risky venture.
Note that the EMVs of the three decisions are $0, $0.175 million, and $0.4 million.
The latter two of these are calculated in row 14 as the usual “sumproduct” of monetary
values and probabilities. So from an EMV point of view, the more risky venture is defi-
nitely best. However, Venture Limited is sufficiently risk averse, and the monetary val-
ues are sufficiently large, that the company is willing to sacrifice EMV to reduce its risk.
How sensitive is the optimal decision to the key parameter, the risk tolerance? We
can answer this by changing the risk tolerance (through the dialog box in Figure 10.34)
and watching how the decision tree changes.7 You can check that when the company
becomes more risk tolerant, the more risky venture eventually becomes optimal. In
fact, this occurs when the risk tolerance increases to approximately $2.075 million. In
the other direction, when the company becomes less risk tolerant, the “do nothing”
decision eventually becomes optimal. This occurs when the risk tolerance decreases
to approximately $0.715 million. So the “optimal” decision depends heavily on the
attitudes toward risk of Venture Limited’s top management.
Certainty Equivalents Now suppose that Venture Limited has only two options. It
can either enter the less risky venture or receive a certain dollar amount x and avoid
7
We show the risk tolerance in cell B5, but the values in the decision tree are not linked to that cell. We
need to go through the dialog box to change the risk tolerance.

the gamble altogether. We want to find the dollar amount x such that the company is
indifferent between these two options. If it enters the risky venture, its expected utility
is 0.0525, calculated above. If it receives x dollars for certain, its (expected) utility is
U (x) = 1 − e−x/1.92
To find the value x where it is indifferent between the two options, we set 1 − e−x/1.92
equal to 0.0525, or e−x/1.92 = 0.9475, and solve for x. Taking natural logarithms of
both sides and multiplying by −1.92, we obtain
x = −1.92 ln(0.9475) ≈ $0.104 million
This value is called the certainty equivalent of the risky venture. The company is
indifferent between entering the less risky venture and receiving $0.104 million to
avoid it. Although the EMV of the less risky venture is $0.175 million, the company
acts as if it is equivalent to a sure $0.104 million. In this sense, the company is willing
to give up the difference in EMV, $71,000, to avoid a gamble.
By a similar calculation, the certainty equivalent of the more risky venture is
approximately $0.086 million. That is, the company acts as if this more risky venture
is equivalent to a sure $0.086 million, when in fact its EMV is a hefty $0.4 million!
So in this case it is willing to give up the difference in EMV, $314,000, to avoid this
particular gamble. Again, the reason is that the company dislikes risk. We can see these
certainty equivalents in PrecisionTree by adjusting the Display box in Figure 10.34
to show Certainty Equivalent. The tree then looks as in Figure 10.36. The certainty
equivalents we just discussed appear in cells C24 and C32.
FIGURE 10.36
Decision Tree with
Certainty Equivalents
Is Expected Utility Maximization Used?

The above discussion indicates that utility maximization is a fairly involved task. The
bottom line, then, is whether the difficulty is worth the trouble. Theoretically, expected
utility maximization might be interesting to researchers, but is it really used? The
answer appears to be: not very often. For example, one recent article on the practice
of decision making [see Kirkwood (1992)] quotes Ronald Howard—the same person
we quoted earlier—as having found risk aversion to be of practical concern in only

5% to 10% of business decision analyses. This same article quotes the president of
a Fortune 500 company as saying, “Most of the decisions we analyze are for a few
million dollars. It is adequate to use expected value (EMV) for these.”
With these comments in mind, it is clear that knowledge of expected utility maxi-
mization is an important requirement for anyone intending to specialize in the field. In
some of the greatest success stories, expected utility maximization was indeed imple-
mented. For nonspecialists, however, a passing knowledge of the concepts is sufficient.
PROBLEMS
Skill-Building Problems 40. Consider again Techware’s decision problem

described in Problem 4. Suppose now that
36. Suppose that a decision maker’s utility as a function Techware’s utility function of net revenue r
of his wealth, x, is given by U (x) = ln x (the (measured in dollars), earned from the given
natural logarithm of x). marketing opportunities, is U (r ) = 1 − e−r/350,000 .
a. Is this decision maker risk averse? Explain why a. Find the course of action that maximizes
or why not. Techware’s expected utility. How does this
b. The decision maker now has $10,000 and two optimal decision compare to the optimal decision
possible decisions. For decision 1, he loses $500 with an EMV criterion? Explain any difference
for certain. For decision 2, he loses $0 with in the two optimal decisions.
probability 0.9 and loses $5000 with probability b. Repeat part a when Techware’s utility function is
0.1. Which decision maximizes the expected U (r ) = 1 − e−r/50,000 .
utility of his net wealth? 41. Consider again the bank’s customer loan decision
37. An investor has $10,000 in assets and can choose problem in Problem 30. Suppose now that the
between two different investments. If she invests in bank’s utility function of profit π (in dollars) is
the first investment opportunity, there is an 80% U (π ) = 1 − e−π/10,000 . Find the strategy that
chance that she will increase her assets by $590,000 maximizes the bank’s expected utility in this case.
and a 20% chance that she will increase her assets How does this optimal strategy compare to the
by $190,000. If she invests in the second investment optimal decision with an EMV criterion? Explain
opportunity, there is a 50% chance that she will any difference in two optimal strategies.
increase her assets by $1.19 million and a 50%
chance that she will increase her assets by $1000. Skill-Extending Problems
This investor has an exponential utility function for
final assets with a risk tolerance parameter equal to 42. Suppose that a decision maker has a utility
$600,000. Which investment opportunity will she function for monetary gains x given by
prefer? U (x) = (x + 10,000)0.5 .
38. Consider again FreshWay’s decision problem a. Show that this decision maker is indifferent
described in Example 10.3. Suppose now that between gaining nothing (i.e., $0) and entering a
FreshWay’s utility function of profit π, earned from risky situation where she gains $80,000
the acquisition and sale of the 24,000 fluorescent with probability 1/3 and loses $10,000 with
lightbulbs, is U (π ) = ln(π ). Find the course of probability 2/3.
action that maximizes FreshWay’s expected utility. b. If there is a 10% chance that one of the decision
How does this optimal decision compare to the maker’s family heirlooms, valued at $5000, will
optimal decision with an EMV criterion? Explain be stolen during the next year, what is the most
any difference in the two decisions. that she would be willing to pay each year for an
39. Consider again the landowner’s decision problem insurance policy that completely covers the
described in Problem 3. Suppose now that the potential loss of her cherished item?
landowner’s utility function of financial gain x 43. A decision maker is going to invest $2000 for a
is U (x) = x 2 . Find the course of action that period of 6 months. Two potential investments are
maximizes the landowner’s expected utility. How available to him: U.S. Treasury bills and gold. If this
does this optimal decision compare to the optimal decision maker invests the $2000 in T-bills, he is
decision with an EMV criterion? Explain any sure to end the 6-month period with $2592. If this
difference in the two decisions. decision maker invests in gold, there is a 75%
chance that he will end the 6-month period with

$800 and a 25% chance that he will end up with this case his gain or loss from either investment
$20,000. The decision maker’s utility√function of is reduced proportionally. For example, if he
ending up with x dollars is U (x) = x. invests half of his money in gold, he will either
a. Should this decision maker invest in gold or lose $600 with probability 0.75 or gain $9000
T-bills? with probability 0.25.
√ Given the same utility
b. Suppose the decision maker invests a proportion function U (x) = x, find the investor’s optimal
y of his $2000 in T-bills and the remaining choice of y.
fraction (1 − y) of his available funds in gold. In
10.7 CONCLUSION
n this chapter we have discussed methods that can be used in decision-making
I problems in which future uncertainty is a key element. Perhaps the most important
skill we can gain from this chapter is the ability to approach decision problems
that include uncertainty in a systematic manner. This systematic approach requires the
decision maker to list all possible decisions or strategies, list all possible uncertain
outcomes, assess the probabilities of these outcomes (possibly with the aid of Bayes’
rule), calculate all necessary monetary values, and finally do the calculations necessary
to obtain the best decision. If large dollar amounts are at stake, it might also be necessary
to perform a utility analysis, where the decision maker’s feelings toward risk are taken
into account. Once the basic analysis has been completed, using “best guesses” for the
various parameters of the problem, a sensitivity analysis should be conducted to see
whether the best decision continues to be best within a range of problem parameters.
PROBLEMS
TABLE 10.27 Distribution of Annual

Skill-Building Problems Demand
44. Ford is going to produce a new vehicle, the Pioneer,
and wants to determine the amount of annual Annual Demand Probability
capacity it should build. Ford’s goal is to maximize 400,000 0.25
the profit from this vehicle over the next 10 years. 900,000 0.50
Each vehicle will sell for $13,000 and incur a
1,300,000 0.25
variable production cost of $10,000. Building 1 unit
of annual capacity will cost $3000. Each unit of
capacity will also cost $1000 per year to maintain,
even if the capacity is unused. Demand for the TABLE 10.28 Distribution of Cash
Pioneer is unknown but marketing estimates the Flow
distribution of annual demand to be as shown in Cash Flow Probability
Table 10.27. Assume that unit sales during a year is
the minimum of capacity and annual demand. −1,000,000 0.35
a. Explain why a capacity of 1,300,000 is not a 500,000 0.60
good choice. 3,000,000 0.05
b. Which capacity level should Ford choose?
45. You are CEO of the venture capital firm D&D. Billy
comes to you with an investment proposition. You a. If you are trying to maximize the expected value
estimate that your distribution of cash flows from of the firm’s cash flows, should you take the
this investment is as shown in Table 10.28. project?
10.7 Conclusion 549

b. Suppose you assess your firm to be risk averse, indicate that there is a 60% chance that the summer
with an exponential utility function. You also use will be sunny (in which case there is an average of 6
the rule of thumb that the firm’s risk tolerance is rainy days during the summer) and a 40% chance
about 6.4% of its annual revenues, which are $30 the summer will be rainy (an average of 30 rainy
million. Determine whether D&D should enter days during the summer).
the venture. Before the summer begins, Adam has the
46. Pizza King (PK) and Noble Greek (NG) are option of purchasing a long-range weather forecast
competitive pizza chains. Pizza King believes there for $1. The forecast predicts a sunny summer 80%
is a 25% chance that NG will charge $6 per pizza, a of the time and a rainy summer 20% of the time. If
50% chance NG will charge $8 per pizza, and a 25% the forecast predicts a sunny summer, there is a 70%
chance that NG will charge $10 per pizza. If PK chance that the summer will actually be sunny. If the
charges price p1 and NG charges price p2 , PK will forecast predicts a rainy summer, there is an 80%
sell 100 + 25( p2 − p1 ) pizzas. It costs PK $4 to chance that the summer will actually be rainy.
make a pizza. PK is considering charging $5, $6, $7, Assuming that Adam’s goal is to minimize his total
$8, or $9 per pizza. In order to maximize its expected cost for the summer, what should he do?
expected profit, what price should PK charge for a Also find the EVSI and the EVPI.
pizza? 50. Erica is going to fly to London on August 5, and
47. Sodaco is considering producing a new product: return home on August 20. It is now July 1. On
Chocovan soda. Sodaco estimates that the annual July 1, she may buy a one-way ticket (for $350) or a
demand for Chocovan, D (in thousands of cases), round-trip ticket (for $660). She may also wait until
has the following probability distribution: August to buy a ticket. On August 1, a one-way
P(D = 30) = 0.30, P(D = 50) = 0.40, ticket will cost $370, and a round-trip ticket will
P(D = 80) = 0.30. Each case of Chocovan sells for cost $730. It is possible that between July 1 and
$5 and incurs a variable cost of $3. It costs $800,000 August 1, her sister (who works for the airline) will
to build a plant to produce Chocovan. Assume that if be able to obtain a free one-way ticket for Erica. The
$1 is received every year (forever), this is equivalent probability that her sister will obtain the free ticket
to receiving $10 at the present time. If Sodaco is 0.30. If Erica has bought a round-trip ticket on
decides to build the plant and produce Chocovan, July 1 and her sister has obtained a free ticket, she
find the expected net present value of its profit. may return “half” of her round trip to the airline. In
48. Many decision problems have the following simple this case, her total cost will be $330 plus a $50
structure. A decision maker has two possible penalty. Use a decision tree approach to determine
decisions, 1 and 2. If decision 1 is made, a sure cost how to minimize Erica’s expected cost of obtaining
of c is incurred. If decision 2 is made, there are two round-trip transportation to London.
possible outcomes, with costs c1 and c2 and prob- 51. A nuclear power company is deciding whether to
abilities p and 1 − p. We assume that c1 < c < c2 . build a nuclear power plant at Diablo Canyon or at
The idea is that decision 1, the riskless decision, has Roy Rogers City. The cost of building the power
a “moderate” cost, whereas decision 2, the risky plant is $10 million at Diablo and $20 million at
decision, has a “low” cost c1 or a “high” cost c2 . Roy Rogers City. If the company builds at Diablo,
a. Find the decision maker’s cost table, that is, the however, and an earthquake occurs at Diablo during
cost for each possible decision and each possible the next 5 years, construction will be terminated and
outcome. the company will lose $10 million (and will still
b. Calculate the expected cost from the risky have to build a power plant at Roy Rogers City).
decision. Without further expert information the company
c. List as many scenarios as you can think of that believes there is a 20% chance that an earthquake
have this structure. (Here’s an example to get you will occur at Diablo during the next 5 years. For $1
started. Think of insurance, where you pay a sure million, a geologist can be hired to analyze the fault
premium to avoid a large possible loss.) structure at Diablo Canyon. She will predict either
49. During the summer, Olympic swimmer Adam that an earthquake will occur or that an earthquake
Johnson swims every day. On sunny summer days will not occur. The geologist’s past record indicates
he goes to an outdoor pool, where he may swim for that she will predict an earthquake on 95% of the
no charge. On rainy days he must go to a domed occasions for which an earthquake will occur and no
pool. At the beginning of the summer, he has the earthquake on 90% of the occasions for which an
option of purchasing a $15 season pass to the domed earthquake will not occur. Should the power
pool, which allows him use for the entire summer. If company hire the geologist? Also find the EVSI and
he doesn’t buy the season pass, he must pay $1 each the EVPI.
time he goes there. Past meteorological records

52. Joan’s utility function for her asset position x 20% of the time. Finally, in response to similar
(for x between
√ 0 and $160,000) is given by threats in the past, weather forecasters have issued
U (x) = x/400. predictions of a major hurricane making landfall
a. Is Joan risk averse? Explain. near a particular coastal location 40% of the time.
b. Currently, Joan’s assets consist of $10,000 in a. Let L be the economic valuation of the loss of
cash and a $90,000 home. During a given year, human life resulting from a coastal strike by the
there is a 0.001 probability that Joan’s home will hurricane. Employ a decision tree to help these
be destroyed by fire or other causes. How much city officials make a decision that minimizes the
should Joan be willing to pay for insurance that expected cost of responding to the threat of the
covers her home completely from this type of impending storm as a function of L . To proceed,
destruction? you might begin by choosing an initial value of
53. My current annual income is $40,000. I believe that L and then perform sensitivity analysis on
I owe $8000 in taxes. For $500, I can hire a CPA to the optimal decision by varying this model
review my tax return. There is a 20% chance she parameter. Summarize your findings.
will save me $4000 in taxes and an 80% chance she b. For which values of L will these city officials
won’t save me anything. If x is my disposable always choose to evacuate the coastal residents,
income for the current
√ year, my utility function is regardless of the Hurricane Center’s prediction?
given by U (x) = x/200. 55. A homeowner wants to decide whether he should
a. Am I risk averse or risk seeking? install an electronic heat pump in his home. Given
b. Should I hire the accountant? that the cost of installing a new heat pump is fairly
large, the homeowner would like to do so only if he
Skill-Extending Problems can count on being able to recover the initial
expense over five consecutive years of cold winter
54. City officials in Ft. Lauderdale, Florida, are trying to
weather. Upon reviewing historical data on the
decide whether to evacuate coastal residents in
operation of heat pumps in various kinds of winter
anticipation of a major hurricane that may make
weather, he computes the expected annual costs of
landfall near their city within the next 48 hours.
heating his home during the winter months with and
Based on previous studies, it is estimated that it will
without a heat pump in operation. These cost figures
cost approximately 1 million dollars to evacuate the
are shown in Table 10.29. The probabilities of
residents living along the coast of this major
experiencing a mild, normal, colder than normal,
metropolitan area. However, if city officials choose
and severe winter are 0.2(1 − x), 0.5(1 − x),
not to evacuate their residents and the storm strikes
0.3(1 − x), and x, respectively.
Fort Lauderdale, there would likely be some deaths
a. Given that x = 0.1, what is the most that the
as a result of the hurricane’s storm surge along the
homeowner is willing to pay for the heat pump?
coast. While city officials are reluctant to place
b. If the heat pump costs $500, how large must x be
an economic value on the loss of human life
before the homeowner decides it is economically
resulting from such a storm, they realize that it may
worthwhile to install the heat pump?
ultimately be necessary to do so to make a sound
c. Given that x = 0.1, compute and interpret the
judgment in this situation. Prior to making the
expected value of perfect information (EVPI)
evacuation decision, city officials consult hurricane
when the heat pump costs $500.
experts at the National Hurricane Center in Coral
d. Repeat part c when x = 0.15.
Gables regarding the accuracy of past predictions.
56. Consider a company that manufactures computer
They learn that in similar past cases, hurricanes that
memory chips in lots of ten chips. From past
were predicted to make landfall near a particular
experience, the company knows that 80% of all lots
coastal location actually did so 60% of the time.
contain 10% defective chips, and 20% of all lots
Moreover, they learn that in past similar cases
contain 50% defective chips. If an acceptable (that
hurricanes that were predicted not to make landfall
is, 10% defective) batch of chips is sent on to the
near a particular coastal location actually did so
next stage of production, processing costs of
TABLE 10.29 Expected Winter Heating Costs for Homeowner’s

Decision Problem
Decision Colder than
Alternatives Mild Normal Normal Severe
Purchase pump $420 $590 $720 $900
Don’t purchase pump $358 $503 $612 $765
10.7 Conclusion 551

$10,000 are incurred. If an unacceptable (that is, in Quebec City into two groups. Two-thirds of the
50% defective) batch is sent on to the next stage of men were asked to be tested for prostate cancer and
production, processing costs of $40,000 are one-third were not asked. Eventually, 8137 men
incurred. This company also has the option of were screened for prostate cancer (PSA plus digital
reworking a batch of chips at a cost of $10,000. A rectal exam) in 1989; 38,056 men were not
reworked batch is guaranteed to be acceptable. screened. By 1997 only 5 of the screened men had
Alternatively, at a cost of $1000, the company can died of prostate cancer while 137 of the men who
test one memory chip from each batch in an were not screened had died of prostate cancer.
attempt to determine whether the given batch is (Source: New York Times May 19,1998)
unacceptable. If a randomly selected chip is found a. Discuss why this study seems to indicate that
to be defective, the batch from which the chip came screening for prostate cancer saves lives.
is acceptable with probability 8/18. If a randomly b. Despite the results of this study, many doctors
selected chip is found not to be defective, the batch are not convinced that early screening for
from which the chip came is acceptable with prostate cancer saves lives. Can you see why they
probability 72/82. doubt the conclusions of the study?
a. Determine how this company can minimize the 59. You have just been chosen to appear on “Hoosier
expected total cost per batch of computer Millionaire”! The rules are as follows: There are
memory chips. four hidden cards. One says “STOP” and the other
b. Compute and interpret the expected value of three have dollar amounts of $150,000, $200,000,
sample information (EVSI) in this decision and $1,000,000. You get to choose a card. If the card
problem. says “STOP,” you win no money. At any time you
c. Compute and interpret the expected value of may quit and keep the largest amount of money that
perfect information (EVPI) in this decision has appeared on any card you have chosen, or you
problem. may continue. If you continue and choose the STOP
d. Suppose now that this manufacturer’s utility card, however, you win no money. As an example,
function of cost c per batch is U (c) = −c0.5 . you might first choose the $150,000 card, then the
Find the strategy that maximizes the $200,000 card, and then choose to quit and receive
manufacturer’s expected utility. How does this $200,000.
optimal strategy compare to the optimal decision a. If your goal is to maximize your expected payoff,
with an EMV criterion? Explain any difference what strategy should you follow?
in two optimal strategies. b. Suppose your utility function for an increase in
57. Patty is trying to determine whether to take cash satisfies U (0) = 0, U ($40,000) = 0.25,
management science or statistics. If she takes U ($120,000) = 0.50, U ($400,000) = 0.75 and
management science, she believes there is a 10% U ($1,000,000) = 1. Are you risk averse?
chance she will receive an A, a 40% chance she will Explain.
receive a B, and a 50% chance she will receive C. If c. After drawing a curve through the points in part
Patty takes statistics, she has a 70% chance of b, determine a strategy that maximizes your
receiving a B, a 25% chance of a C, and a 5% expected utility. (Alternatively, you might want
chance of a D. Patty is indifferent between the to assess and use your actual utility function.)
following two options: 60. You are trying to determine how much money to put
■ Option 1: Receiving a B for certain in your Tax Saver Benefit (TSB) plan. At the
■ Option 2: A 70% chance at an A and a 30% beginning of the calendar year, a TSB allows you
chance at a D to put money into an account. The money in
Patty is also indifferent between the following the account can be used to pay for medical
two options: expenses incurred during the year. Once the TSB is
■ Option 3: Receiving a C for certain exhausted, you must pay the medical expenses out
■ Option 4: A 25% chance at an A and a 75% of pocket. The benefit of the TSB is that money
chance at a D placed in the TSB is not subject to federal taxes. The
In order to maximize the expected utility catch is that any money left in the TSB at the end of
associated with her final grade, which course should the year is lost to you. Suppose the federal tax rate is
Patty take? 40% and your current annual salary is $50,000. You
58. Many men over 50 take the PSA blood test. The believe that it is equally likely that your medical
purpose of the PSA test is to detect prostate cancer expenses during the current year will be $3000,
early. Dr. Rene Labrie of Quebec conducted a study $4000, $5000, $6000, or $7000.
to determine whether the PSA test can actually a. If you are risk neutral and want to maximize
prevent cancer deaths. In 1989, Dr. Labrie randomly your expected disposable income, how much
divided all male registered voters between 45 and 80 should you put in your TSB?

b. Suppose you assess a utility function for cells) and automatically see her optimal EMV
disposable income given by U (x) = and optimal strategy from the tree.
0.000443x 0.713595 . (Who said they all have to b. If p2 = 0.8 and p3 = 0.1, what value of p1
have nice round numbers?) Are you risk averse? makes Chang indifferent between abandoning
How much should you put in the TSB? the project and going ahead with it?
61. Peter is thinking of purchasing an advertising c. How much would Chang be willing to pay the
company from Amanda. At present, only Amanda Olympic organization (now) to guarantee her the
(not Peter) knows the current value of the company. contract in the case where her company is
Peter knows, however, that there is an equal chance successful in developing the contract? (This
that the company is worth 10, 20, 30, 40, 50, 60, 70, guarantee is in force only if she is successful in
80, 90, or 100 million dollars. Amanda will accept developing the product.) Assume p1 = 0.4,
an offer from Peter only if Peter bids at least the p2 = 0.8, and p3 = 0.1.
value of the company. For example, if Amanda d. Suppose now that this a “big” project for Chang.
knows the company is worth $20 million, she will Therefore, she decides to use expected utility as
accept any bid of $20 million or higher. As soon as her criterion, with an exponential utility function.
Peter purchases the company, his reputation as a Using some trial and error, see which risk
skilled businessman immediately increases the tolerance changes her initial decision from “go
actual value of the company by 80%. ahead” to “abandon” when p1 = 0.4, p2 = 0.8,
a. Suppose Peter is risk neutral and is considering and p3 = 0.1.
bidding 10, 20, 30, 40, 50, 60, 70, 80, 90, or 100 63. Suppose an investor has the opportunity to buy the
million dollars. What should he bid? following contract, a stock call option, on March 1.
b. Suppose Peter’s utility function for financial The contract allows him to buy 100 shares of ABC
gains or losses (in millions of dollars) is given by stock at the end of March, April, or May at a
U (x) = ((x + 82)/144)1.7 . Determine whether guaranteed price of $50 per share. He can “exercise”
Peter is risk averse or risk seeking and determine this option at most once. For example, if he
Peter’s optimal decision. purchases the stock at the end of March, he can’t
62. Sarah Chang is the owner of a small electronics purchase more in April or May at the guaranteed
company. In 6 months a proposal is due for an price. The current price of the stock is $50. Each
electronic timing system for the 1998 Olympic month, we assume the stock price either goes up by
Games. For several years, Chang’s company has a dollar (with probability 0.6) or down by a dollar
been developing a new microprocessor, a critical (with probability 0.4). If the investor buys the
component in a timing system that would be superior contract, he is hoping that the stock price will go up.
to any product currently on the market. However, The reasoning is that if he buys the contract, the
progress in research and development has been slow, price goes up to $51, and he buys the stock (that is,
and Chang is unsure about whether her staff can he exercises his option) for $50, he can turn around
produce the microprocessor in time. If they succeed and sell the stock for $51 and make a profit of $1 per
in developing the microprocessor (probability p1 ), share. On the other hand, if the stock price goes
there is an excellent chance (probability p2 ) that down, he doesn’t have to exercise his option; he can
Chang’s company will win the $1 million Olympic just throw the contract away.
contract. If they do not, there is a small chance a. Use a decision tree to find the investor’s optimal
(probability p3 ) that she will still be able to win the strategy (that is, when he should exercise the
same contract with an alternative, inferior timing option), assuming he purchases the contract.
system that has already been developed. b. How much should he be willing to pay for such a
If she continues the project, Chang must invest contract?
$200,000 in research and development. In addition, 64. The Ventron Engineering Company has just been
making a proposal (which she will decide whether awarded a $2 million development contract by the
to do after seeing whether the R&D is successful or U.S. Army Aviation Systems Command to develop
not) requires developing a prototype timing system a blade spar for its Heavy Lift Helicopter program.
at an additional cost of $50,000. Finally, if Chang The blade spar is a metal tube that runs the length of
wins the contract, the finished product will cost an and provides strength to the helicopter blade. Due to
additional $150,000 to produce. the unusual length and size of the Heavy Lift
a. Develop a decision tree that can be used to solve Helicopter blade, Ventron is unable to produce a
Chang’s problem. You can assume in this part single-piece blade spar of the required dimensions,
that she is using EMV (of her net profit) as a using existing extrusion equipment and material.
decision criterion. Build the tree so that she can The engineering department has prepared two
enter any values for p1 , p2 , and p3 (in input alternatives for developing the blade spar:
10.7 Conclusion 553

(1) sectioning or (2) an improved extrusion process. steps of the improved extrusion process? (This
Ventron must decide which process to use. (Backing information would tell Ventron, right now, the
out of the contract at any point is not an option.) The ultimate success/failure outcomes of both steps.)
risk report has been prepared by the engineering 65. Ligature, Inc. is a company that does contract work
department. The information from it is explained for publishing companies. It specializes in writing
below. textbooks for secondary schools. Because states
The sectioning option involves joining several such as Texas and California typically adopt only
shorter lengths of extruded metal into a blade spar of about four to eight textbooks for any given subject
sufficient length. This work will require extensive and grade level (from which individual schools can
testing and rework over a 12-month period at a total choose), the potential for large profits is great.
cost of $1.8 million. While this process will Ligature is currently negotiating a contract with
definitely produce an adequate blade spar, it merely Brockway and Coates (B&C), a large publishing
represents an extension of existing technology. company, to write a social studies series for grades
To improve the extrusion process, on the other 9–12. Actually, the development of the books is
hand, it will be necessary to perform two steps: already well under way, and the only details not yet
(1) improve the material used, at a cost of $300,000, worked out concern the fee B&C will pay Ligature.
and (2) modify the extrusion press, at a cost of Ligature has always operated on a fixed fee basis.
$960,000. The first step will require 6 months of Under this arrangement, B&C would pay Ligature
work, and if this first step is successful, the second its costs, in this case $4.15 million, plus 25%.
step will require another 6 months of work. If both Ligature would receive this payment in 6 months, at
steps are successful, the blade spar will be available the beginning of year 1. Although this is still an
at that time, that is, a year from now. The engineers option, the companies have also been discussing a
estimate that the probabilities of succeeding in royalty arrangement as an alternative.
steps 1 and 2 are 0.9 and 0.75, respectively. Under the royalty plan, B&C would still pay
However, if either step is unsuccessful (which will Ligature its $4.15 million costs at the beginning of
be known only in 6 months for step 1 and in a year year 1, but Ligature would then receive yearly
for step 2), Ventron will have no alternative but to royalty payments at the ends of years 1 through 5.
switch to the sectioning process—and incur the These payments would depend on (1) total sales over
sectioning cost on top of any costs already incurred. the five years, (2) the timing of sales, and (3) the
Development of the blade spar must be negotiated royalty rate, that is, Ligature’s percentage
completed within 18 months to avoid holding up the of each sales dollar. As for timing, both parties agree
rest of the contract. If necessary, the sectioning work that 10% of total sales will be in year 1, 20% will be
can be done on an accelerated basis in a 6-month in each of the next 2 years, 30% will be in year 4,
period, but the cost of sectioning will then increase and 20% will be in year 5. They also estimate that
from $1.8 million to $2.4 million. the probability distribution of total sales is discrete,
Frankly, the Director of Engineering, Dr. with possible values $25 million, $30 million, $50
Smith, wants to try developing the improved million, and $70 million, and corresponding
extrusion process. This is not only cheaper (if probabilities 0.10, 0.45, 0.30, and 0.15.
successful) for the current project, but its expected To guard its interests, B&C has imposed the
side benefits for future projects could be sizable. following restriction to any royalty agreement. It
Although these side benefits are difficult to gauge, places a cap on the amount Ligature can earn
Dr. Smith’s best guess is an additional $2 million. through the royalty scheme. Specifically, the
(Of course, these side benefits are obtained only if royalties, discounted back to the beginning of year 1
both steps of the modified extrusion process are at a 10% discount rate, cannot exceed 33% of
completed successfully.) Ligature’s $4.15 million costs. Obviously, this limits
a. Develop a decision tree to maximize Ventron’s B&C’s downside exposure, regardless of the
EMV. This includes the revenue from this negotiated royalty rate or how well the books sell.
project, the side benefits (if applicable) from an Ligature is interested in maximizing the NPV
improved extrusion process, and relevant costs. of its profit from this project (discounted back to the
You don’t need to worry about the time value of beginning of year 1), using a 10% discount rate. The
money; that is, no discounting or NPVs are following steps lead you through the required
required. Summarize your findings in words in calculations to “solve” the problem. No decision
the spreadsheet. tree is required for this problem.
b. What value of side benefits would make Ventron a. The file P10_65.XLS supplies the inputs in an
indifferent between the two alternatives? input section (blue border), and it has a
c. How much would Ventron be willing to pay, calculation section (red border). First, calculate
right now, for perfect information about both the upper part of the calculation section. To do

so, enter any trial value for total sales in cell G8 assessed to be 0.0025. If a fire occurs, then the
and do the necessary calculations to eventually cleanup cost could be high ($80 million) or low ($20
find (in cell G17) the NPV to Ligature from the million). The probability of a high cleanup cost,
royalty agreement. At this point, you can use any given that a fire occurs, is assessed at 0.2.
royalty rate in the RoyRate cell (C28). a. If the company uses EMV as its decision
b. Using the calculations from part a, complete the criterion, should it replace the transformer?
data table in the middle part of the calculation b. Perform a sensitivity analysis on the key
section. It should show the NPV to Ligature for parameters of the problem that are difficult to
any potential value of total sales. Then use these assess, namely, the probability of a fire, the
NPVs to calculate the expected NPV to Ligature probability of a high cleanup cost, and the high
in the ExpNPV cell (G27). and low cleanup costs. Does the optimal decision
c. Suppose the current “offer on the table” is a 3% from part a remain optimal for a “wide” range of
royalty rate. In the bottom part of the calculation these parameters?
section, use IF comparisons to see which c. Do you believe EMV is the correct criterion
arrangement, fixed fee or royalty, each party to use in this type of problem involving
would favor. environmental accidents?
d. Continuing part c (with the 3% offer on the
table), what do you think the two parties will 68. Based on Mellichamp et al. (1993). Construction
eventually agree upon? That is, will they equipment managers typically have many large
stick with the 3% royalty rate, move to a pools of engines, transmissions, and other equip-
different royalty rate, or settle on the fixed fee ment units to maintain. One approach to this
arrangement? Answer below cell B36. maintenance is to use oil analysis, where the oil
66. The American chess master Jonathan Meller is from any of these is subjected periodically to an
playing the Soviet expert Yuri Gasparov in a inspection. These inspections can sometimes signal
two-game exhibition match. Each win earns a player an impending failure (for example, too much iron in
one point, and each draw earns a half point. The the oil), and preventive maintenance is then
player who has the most points after two games performed (at a relatively low cost), eliminating the
wins the match. If the players are tied after two risk of failure (failure would result in a relatively
games, they play until one wins a game; then the high cost). However, oil analysis costs money, and it
first player to win a game wins the match. During is not perfect. That is, it can indicate that a unit is
each game, Meller has two possible strategies: to defective when in fact it is not about to fail, and it
play a daring strategy or to play a conservative can indicate that a unit is nondefective when in fact
strategy. His probabilities of winning, losing, and it is about to fail. As a possible substitute for oil
drawing when he follows each strategy are shown in analysis, the company could simply change the oil
Table 10.30. To maximize his probability of winning periodically, thereby reducing the probability of a
the match, what should the American do? failure.
Suppose the company has four alternatives:
(1) do nothing, (2) use oil analysis only, (3) replace
TABLE 10.30 Probabilities for Chess oil only, or (4) replace oil and do oil analysis. For
Problem option (1) the probability of a failure is p1 , and the
Strategy Win Loss Draw cost of a failure is C1 . For option (2), the probability
of a failure remains at p1 . If the unit is about to fail,
Daring 0.45 0.55 0.00 the oil analysis will indicate this with probability
Conservative 0.00 0.10 0.90 1 − α; if the unit is not about to fail, the oil analysis
will indicate this with probability 1 − β. (Therefore,
α and β are the error probabilities of the oil
67. Based on Balson et al. (1992). An electric utility analysis.) The oil analysis itself costs C2 , and if it
company is trying to decide whether to replace its indicates that a failure is about to occur, the oil will
PCB transformer in a generating station with a new be changed, at cost C3 , and preventive maintenance
and safer transformer. To evaluate this decision, the will be performed. The cost of maintenance to
utility needs information about the likelihood of an restore a unit that is about to fail is C4 , whereas the
incident, such as a fire, the cost of such an incident, cost of maintenance for a unit that is not about to
and the cost of replacing the unit. Suppose that the fail is C5 . The only difference between options (3)
total cost of replacement as a present value is and (4) is that the probability of a failure decreases
$75,000. If the transformer is replaced, there is to p2 after changing the oil. The values of these
virtually no chance of a fire. However, if the current parameters for a particular class of units (engines in
transformer is retained, the probability of a fire is light trucks, say) appear in Table 10.31 (page 556).
10.7 Conclusion 555

TABLE 10.31 Parameters for Oil Analysis TABLE 10.32 Probabilities for Water
Problem Pollution Problem
Event Probability
Parameter Value
Significant market 0.6 ± 0.15
p1 0.10
Technically feasible 0.6 ± 0.15
p2 0.04
Board sanctions plant expenditures 0.8 ± 0.2
α 0.30
Commercial success 0.8 ± 0.2
β 0.20
C1 $1200.00
C2 $20.00 indicates the company’s uncertainty about the true
C3 $14.80 probabilities.
C4 $500.00 The primary economic factors are the
C5 $250.00 following:
■ the research expenses to identify a new
production process for the product

a. For these parameters, develop a decision tree to ■ the marketing development cost to determine
find the company’s best decision and the whether there is a significant market
corresponding expected cost. ■ the process development costs, including
b. If the company has 500 units, what should it do? presanction engineering
What is the expected cost for the entire fleet? ■ the commercial development costs, both before
c. Suppose that the company has different types of and after the board’s sanction
units. For example, the cost of an oil change ■ the venture value (net present value) if successful
might be higher for some, or the cost of a failure The estimates of these values are shown in Table
might be higher or lower. Run a sensitivity 10.33. Again, the plus-or-minus values indicate the
analysis on any of the parameters you believe company’s considerable uncertainty about the
might be “key” parameters and see whether the values. All values are in millions of dollars.
optimal decision changes in ways you would The timing of events is as follows:
anticipate. ■ Decide whether to abandon product now. (This is
69. Based on Hess (1993). A company that is heavily really the only nontrivial decision the company
involved in R&D projects believes it might have the will make.) If not, then:
■ Spend on research and marketing development. If
potential to develop a very lucrative commercial
product that would (if successful) reduce pulp mill marketing development indicates an insignificant
water pollution. At the current stage, however, market for the product or research indicates that
everything is quite uncertain, and the company is the process is technically infeasible, cut expenses
trying to decide whether to go ahead with its R&D and quit. Otherwise:
■ Spend on process and commercial development.
or abandon the product. The following are the
primary risks: If company board then declines to sanction
money for plant, cut expenses and quit.
■ Would market tests confirm that there is a
Otherwise:
significant market for the product? ■ Spend on further commercial development. By
■ Could the company develop a new process for
this time, the company has made all of its
making this product—that is, is it technically decisions. If the venture turns out to be a
feasible? commercial success, then it gains the venture
■ Even if there is a significant market and the
value for a success (less expenses so far).
process is technically feasible, would the Otherwise, the company has lost the money spent
company’s board sanction the new plant so far, but that is all.
capital necessary to produce the product on a Analyze the company’s problem. Obviously,
commercial scale? with the high degree of uncertainty, sensitivity
■ Assuming the answers to the above questions are
analysis is the key. Note that there are many
all yes and the plant is built, would the venture uncertainties about the input parameters in
turn out to be successful? Tables 10.32 and 10.33. In fact, there are far too
We assume that each of these questions has a many to allow you to try every combination.
yes or no answer. The probabilities of yes answers Therefore, just try a few combinations that you
are shown in Table 10.32. The plus-or-minus value believe might be the most important.

TABLE 10.33 Monetary Estimates for Water Pollution Problem
Expense or Gain Net Present Value
Research expense $0.8 ± 25%
Market development expense $0.2 ± 25%
Process development expense (presanction) $3.0 ± 25%
Commercial development expense (presanction) $0.5 ± 25%
Commercial development expense (postsanction) $1.0 ± 25%
Value if successful $25.0 ± 50%
10.7 Conclusion 557

C A S E 10.1
GMC Motor Company II

his case is a continuation of GMC I (from on the observed demand, GMC plans its produc-
T Chapter 6). Management at GMC is gener-

ally pleased with the modeling effort that
has been done for capacity planning in the coming
year. However, some managers have asked about
tion accordingly. For example, in the second year,
GMC must decide on its plant configurations before
the demand scenario is revealed, but can determine
its production plan after the demand scenario is re-
the effect demand forecasts for the second year out vealed. This sequence of events is consistent with
could have on the recommended strategy. the relative time periods involved. Reconfiguring a
Although demand forecasts for the coming year plant is a major undertaking that must be planned in
are considered to be quite reliable, forecasts two or advance, so this decision must be made before the
more years in the future have been less accurate. demand scenario is revealed. Production during a
Accordingly, analysts at GMC formulate several de- year can be altered to best meet the demand as it de-
mand scenarios in the future, and assign probabili- velops during the year. For modeling purposes, the
ties to each scenario. The situation for the coming production decision can be made after the demand
two years is summarized in Table 10.34. scenario is revealed. Also, no inventory is carried
Three demand scenarios are possible in the sec- from one year to the next.
ond year. Scenario A corresponds to a robust eco- The demand diversion matrix is assumed to be
nomic expansion and increasing market share for constant for both years. For convenience, it is re-
GMC cars. Scenario B represents little change from peated in Table 10.36.
the first year, although there is a relative shift away
from the smaller Lyra to the larger Libra and Hy-
dra models. Scenario C represents an economic re-
Questions
cession and decreased demand for all car lines. In GMC wants to decide whether to retool the Lyra and
scenario C, the decrease in demand for Libras and Libra plants in each of the coming two years. In addi-
Hydras is larger than for the economical Lyras. An- tion, GMC wants to determine its production plan at
alysts give scenario A a slightly higher probability each plant for each year. Based on the previous data,
of occurring than scenarios B and C. formulate a mixed integer programming model for
Management at GMC wants to consider all pos- solving GMC’s production planning–capacity ex-
sible configurations of capacity in the next two years. pansion problem for the coming two years. Assume
As before, the Lyra and/or Libra plants can be re- that GMC’s objective is to maximize total average
tooled, but retooling can be done in either the first profit for the two years. For simplicity, assume that
or second year. Because of the enormous costs of no discounting of profits is done for the second year.
changing a plant configuration, a plant that is re- In the past, GMC had solved problems sepa-
tooled in the first year cannot be returned to its orig- rately for each scenario. The three optimal solutions
inal configuration in the second year. The costs and were compared and then a final decision was made.
characteristics of the original and retooled plants are What are the three optimal solutions corresponding
the same in either year. For convenience, these are to each scenario? (For example, assuming that sce-
repeated in Table 10.35. nario A occurs with probability 1.0, what is the op-
In addition to selecting the plant configurations timal solution? Then repeat for scenarios B and C.)
for each year, GMC needs to determine the produc- How do the three separate optimal solutions com-
tion plan at each plant in each year. The sequence of pare to the overall optimal solution found before?8
events and decisions is as follows. At the beginning
of a year, GMC must decide on the plant configura-
8
tions. Demand occurs during the year, and based Acknowledgment: The idea for GMC I and II came from
Eppen et al. (1989).

TABLE 10.34 Demand Forecasts and Probabilities for
GMC Case Study
Second Year
Model First Year Scenario A Scenario B Scenario C

Lyra 1400 1700 1300 1300
Libra 1100 1500 1200 800
Hydra 800 1100 850 600
Probability 1 0.4 0.3 0.3
TABLE 10.35 Plant Characteristics for GMC Case Study

Lyra Libra Hydra New Lyra New Libra
Capacity (in 1000s) 1000 800 900 1600 1800
Fixed cost (in $millions) 2000 2000 2600 3400 3700
Profit Margin by Car Line (in 1000s)
Lyra 2 – – 2.5 2.3
Libra – 3 – 3 3.5
Hydra – – 5 – 4.8
TABLE 10.36 Demand Diversion Matrix for

GMC Case Study
Lyra Libra Hydra
Lyra – 0.3 0.05
Libra 0 – 0.10
Hydra 0 0.0 –
Case 10.1: GMC Motor Company II 559

C A S E 10.2
Jogger Shoe Company

he Jogger Shoe Company is trying to de- will continue for the next 3 years if the current style
T cide whether to make a change in its most

popular brand of running shoes. The new
style would cost the same to produce, and it would
be priced the same, but it would incorporate a new
is retained. However, there is uncertainty about de-
mand for the new style, if it is introduced. The com-
pany models this uncertainty by assuming a normal
distribution in year 1, with mean 220,000 and stan-
kind of lacing system that (according to its market- dard deviation 20,000. The company also assumes
ing research people) would make it more popular. that this demand, whatever it is, will remain constant
There is a fixed cost of $300,000 of changing over for the next 3 years. However, if demand in year 1
to the new style. The unit contribution to before-tax for the new style is sufficiently low, the company can
profit for either style is $8. The tax rate is 35%. Also, always switch back to the current style and realize
because the fixed cost can be depreciated and will an annual demand of 190,000. The company wants a
therefore affect the after-tax cash flow, we need a strategy that will maximize the expected net present
depreciation method. We assume it is straight-line value (NPV) of total cash flow for the next 3 years,
depreciation. where a 15% interest rate is used for the purpose of
The current demand for these shoes is 190,000 calculating NPV.
pairs annually. The company assumes this demand

C A S E 10.3
Westhouser Paper Company

he Westhouser Paper Company in the state Distribution of Price
T of Washington currently has an option to
purchase a piece of land with good timber
forest on it. It is now May 1, and the current price
of the land is $2.2 million. Westhouser does not
TABLE 10.37
Price Decrease
$0
Decrease in May
Probability
0.5
actually need the timber from this land until the $60,000 0.3
beginning of July, but its top executives fear that $120,000 0.2
another company might buy the land between now
and the beginning of July. They assess that there is 1
chance out of 20 that a competitor will buy the land the probabilities of the possible price decreases dur-
during May. If this does not occur, they assess that ing May. Table 10.38 shows the conditional proba-
there is 1 chance out of 10 that the competitor will bilities of the possible price decreases in June, given
buy the land during June. If Westhouser does not take the price decrease in May. For example, if the price
advantage of its current option, it can attempt to buy decrease in May is $60,000, then the possible price
the land at the beginning of June or the beginning of decreases in June are $0, $30,000, and $60,000 with
July, provided that it is still available. respective probabilities 0.6, 0.2, and 0.2.
Westhouser’s incentive for delaying the pur- If Westhouser purchases the land, it believes
chase is that its financial experts believe there is a that it can gross $3 million. (This does not count the
good chance that the price of the land will fall signif- cost of purchasing the land.) But if it does not pur-
icantly in one or both of the next two months. They chase the land, it believes that it can make $650,000
assess the possible price decreases and their proba- from alternative investments. What should the com-
bilities in Tables 10.37 and 10.38. Table 10.37 shows pany do?
TABLE 10.38 Distribution of Price Decrease in June

Price Decrease in May
$0 $60,000 $120,000
June Decrease Probability June Decrease Probability June Decrease Probability
$0 0.3 $0 0.6 $0 0.7
$60,000 0.6 $30,000 0.2 $20,000 0.2
$120,000 0.1 $60,000 0.2 $40,000 0.1
Case 10.3: Westhouser Paper Company 561

Decision Making Under Uncertainty: Decision and Risk Analysis at Du Pont

Uploaded by

Copyright:

Available Formats

Decision Making Under Uncertainty: Decision and Risk Analysis at Du Pont

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Decision Making Under Uncertainty: Decision and Risk Analysis at Du Pont

Uploaded by

Copyright:

Available Formats

C H A P T E R 10

DECISION AND RISK ANALYSIS AT DU PONT

F at the most strategic levels of a company’s planning process and

494 Chapter 10 Decision Making Under Uncertainty

10.1 Introduction 495

10.2 ELEMENTS OF A DECISION ANALYSIS

TABLE 10.1 Data for Bidding Example

496 Chapter 10 Decision Making Under Uncertainty

TABLE 10.2 Payoff Table for SciTools Bidding Example

No Bid <115 >115, <120 >120, <125 >125

10.2 Elements of a Decision Analysis 497

TABLE 10.3 Alternative Payoff Table for SciTools Bidding Example

498 Chapter 10 Decision Making Under Uncertainty

10.2 Elements of a Decision Analysis 499

500 Chapter 10 Decision Making Under Uncertainty

FIGURE 10.2 Result of Folding Back to Obtain Optimal Decision

10.2 Elements of a Decision Analysis 501

502 Chapter 10 Decision Making Under Uncertainty

10.2 Elements of a Decision Analysis 503

504 Chapter 10 Decision Making Under Uncertainty

10.2 Elements of a Decision Analysis 505

FIGURE 10.12 Structure of Completed Tree

506 Chapter 10 Decision Making Under Uncertainty

10.2 Elements of a Decision Analysis 507

Sensitivity Analysis We have already stressed the importance of a follow-up sensi-

508 Chapter 10 Decision Making Under Uncertainty

10.2 Elements of a Decision Analysis 509

510 Chapter 10 Decision Making Under Uncertainty

Skill-Building Problems TABLE 10.5 Distribution of Possible

TABLE 10.6 Payoff Table for Carlisle’s Decision Problem

10.2 Elements of a Decision Analysis 511

TABLE 10.7 Revenue Table for Techware’s Decision Problem

512 Chapter 10 Decision Making Under Uncertainty

10.2 Elements of a Decision Analysis 513

each, the purchasing agent is intrigued to learn that

10.3 MORE SINGLE-STAGE EXAMPLES

A discussed so far. We first identify the possible decision alternatives, assess

514 Chapter 10 Decision Making Under Uncertainty

TABLE 10.12 Distribution of Medical Expenses

10.3 More Single-Stage Examples 515

TABLE 10.13 Employee Cost Table for

516 Chapter 10 Decision Making Under Uncertainty

10.3 More Single-Stage Examples 517

518 Chapter 10 Decision Making Under Uncertainty

10.3 More Single-Stage Examples 519

520 Chapter 10 Decision Making Under Uncertainty

Skill-Building Problems TABLE 10.15 Distribution of Possible

10.3 More Single-Stage Examples 521

TABLE 10.17 Distribution of Annual TABLE 10.19 Distribution of Unit

TABLE 10.18 Distribution of Annual TABLE 10.20 Distribution of Unit

522 Chapter 10 Decision Making Under Uncertainty

10.3 More Single-Stage Examples 523

10.4 MULTISTAGE DECISION PROBLEMS

524 Chapter 10 Decision Making Under Uncertainty

10.4 Multistage Decision Problems 525

526 Chapter 10 Decision Making Under Uncertainty

10.4 Multistage Decision Problems 527

528 Chapter 10 Decision Making Under Uncertainty

10.4 Multistage Decision Problems 529