AP / ADMS 3300

DECISION ANALYSIS

1
Week 11 & 12
Chapter 17

Conflicting Objectives II
Multiattribute Utility Models
with Interactions
2
Objective
In this session, we will learn the concept that two
attribute may be substitute for one another to
some extant. Such interaction cannot be captured
by the additive utility function. This model is
essentially an additive combination of preferences
for individual attributes.
To capture the kinds of interactions that we are
talking about here, as well as risk attitudes, we
must think more generally. We have to think in
terms of utility surface.
Although, it is possible to think about many
attributes at once, we will develop multi attribute
utility theory concepts using only two attributes.
3
Conflicting Objectives II

4
Presence of Attribute Interactions
Last
Chapter

5
Conflicting Objectives II

6
Multiattribute Utility Functions: Direct Assessment
Consider the reference gamble method. The
appropriate reference gamble has the worst pair
(x-, y-) and the best pair(x+, y+) as the two possible
outcomes:
Win (x+, y+) with probability p
Win (x-, y- ) with probability 1-p
Now find the probability p to use in the reference
gamble that will make you indifferent between ( x,y)
and the reference gamble.
As before, you can use p as your utility U(x,y) as
U(x+,y+) = 1, and U(x-, y-) = 0
7
 (p)
X+, y+.

(1-p)
X-, y-

X, y

This decision tree that represents the assessment situation. This is simply
the standard probability-equivalent utility assessment technique that we
have seen before.
21

8
 Multiattribute Utility Functions: Direct Assessment
Using this technique, for a reasonable number of
(x,y) pairs one can assess many utility numbers.
There may be several pairs with the same utility, and
you should be indifferent among such pairs.
 Thus, (x,y) pair with the same utilities must fall on
an indifferent curve.
One approach to understanding multiattribute
preferences is simply to plot the assessed points on a
graph, and sketch rough indifference curves.

9
 Multiattribute Utility Functions: Direct Assessment
DRAWBACK:
You must assess utilities for a substantial number of
points.
Even though it is straight forward to see how this
approach might be extended to three or more
attributes, the more being considered, the more points
you must assess, and graphical representation become
more complicated.

10
Multiattribute Utility Functions: Using
Individual Utility Functions
Another approach that would ease the assessment burden would be
to think about a multiattribute utility function that is made up of the
individual utility functions. Mathematically, we can represent the
most general case as
U(x,y) = f[Ux(x), Uy(y)]
The f[.,.] notation means that U(x,y) is a function of the individual
utility functions Ux(x) and Uy(y). In the last session, we used
U(x,y) = kxUx (x) + Ky Uy(y)
In this session we will consider
U(x,y) = c1+c2Ux(x) + c3Uy(y) + c4Ux(x)Uy(y)
Being able to break down the multiattribute utility function this way
sometimes is called separability; the overall utility function can be
“separated” into chunks that represent different attributes.

11
 Independence Conditions
Mutual Preferential Independence.
 An attribute Y is said to preferentially independent
of X if preference for specific outcome of Y do not
depend on the level of attribute X.
 As an example, let Y be the time to completion of a
project and X its cost. If we prefer a project time of
5 days to one of 10 days, assuming that the cost is
100 in each case, and if we also prefer a project
time of 5 days to one of 10 days if the cost is 200 in
both cases, then Y is preferentially independent of
X; it does not matter what the cost is---we still
prefer the shorter completion time.
12
It is easy to imagine situations in which preferential independence
might not hold.
Consider a decision with outcomes that affect both the place
where you live and the automobile that you drive. Let X be an
outcome variable that could denote either Los Angeles or an
African farm, and Y an outcome variable denoting either a
Cadillac or a Land Rover.
The value of X (whether you live in Los Angles or on an African
farm) may well affect your preference for a Cadillac or a Land
Rover. Therefore, Y would not be preferentially independent of X.
Consider the reverse: You may prefer Los Angles to an African
farm (or vice versa) regardless of the car you own.
Thus, one attribute would be preferentially independent of the
other, but two are not mutually preferentially independent.

13
It is fair to say that mutual preferential independence holds
for many people and many situations, or that at least it is a
reasonable approximation.
In fact, mutual preferential independence is exactly the
condition that is needed for the additive utility function to be
appropriate when making a decision with no uncertainty.
That is, when a decision maker’s preferences display mutual
preferential independence, then the additive utility function,
assessed using any of the technique described in the last
session, is appropriate for decisions under certainty.
Once we move to decision under uncertainty, however,
mutual preferential independence is not quite strong enough.
Thus, we must look at stronger conditions.
Namely, Mutual utility Independence.
14
 Utility Independence
 Utility independence is slightly stronger than
preferential independence. An attribute Y is considered
utility independent of attribute X if preferences for
uncertain choices involving different levels of Y are
independent of the value of X.
 Utility independence clearly is analogous to preferential
independence, except that the assessments are made under
conditions of uncertainty.
For the project evaluation example above, suppose we
assess that the certainty equivalent for an option giving,
say, a 50% chance of Y = 5 and a 50% chance of Y = 10 does
not depend on the level at which cost X is fixed. As long as
our preference for lotteries in the completion-time
attribute are the same(say, measured by their certainty
equivalent) regardless of the fixed level of cost, then the 15
Determining whether Independence
Exists
How can you determine whether your
preferences are preferentially independent?
The simplest approach is to imagine a series of
paired comparisons that involve one of the
attributes. With the other attribute fixed at its
lowest level, decide which outcome in each pair
is preferred. Once it is done, imagine changing
the level of the fixed attribute. Would your
comparisons be the same? Would the
comparisons be the same regardless of the fixed
level of the other attribute? If so, then
preferential independence holds.
See the sample dialogue on page 649 in the text. 16
Using Independence
If a decision maker’s preference show mutual
utility independence, then a two attribute utility
function can be written as a composition of the
individual utility functions. As usual, the least
preferred outcome (x-,y-) is assigned the utility
value 0, and the most preferred pair(x+,y+) is
assigned the utility value 1.
Under mutual utility independent preference, the
two-attributes utility function can be written as
U(x,y) = kxUx(x) + kyUy(y) + (1-kx-kY)Ux(x)Uy(y)

17
Where
Ux(x) = utility function on X scaled so that
Ux(x-)=0 and Ux(x+) = 1
Uy(y) = utility function on Y scaled so that
Uy(y-)=0 and Uy(y+) = 1
Kx = U(x+,y-)
Ky = U(x-,y+)
The product term Ux(x)Uy(y) in this utility
function is what permits the modeling of
interactions among attributes. To understand
the last two conditions, we must plug the
individual utilities into the equation. For
example
18
 U(x+,y-) = kxUx(x+) + kyUy(y-) +
(1-kx-ky)Ux(x+)Uy(y-)
=kx(1) +ky(0) + (1-kx-ky)(1)(0)
= kx
This multiattribute utility function is called a multi-
linear expression.

19
Additive Independence
Look again at the multiattribute utility function
U(x,y) = kxUx(x) + kyU(y) + (1-kx-kY)Ux(x)Uy(y)
If kx+ky=1, then the utility function turns out to be
simply additive:
U(x,y) = kxUx(x) + (1-kx)Uy(y)
To be able to model preferences accurately with this
additive utility function, we need additive
independence, an even stronger condition than
utility independence.

20
The statement of additive independence is the
following: suppose X and Y are mutually utility
independent, and you are indifferent between
Lotteries A and B:
A (x-,y-) with probability 0.5
(x+,y+) with probability 0.5
B (x-,y+) with probability 0.5
(x+,y-) with probability 0.5
If this is the case, then the utility function can be
written as the weighted combination of the two
utility functions, U(x,y) = kxUx(x) + (1-kx)Uy(y)

21
You can see by writing out the expected utilities of the lotteries that
they are equivalent:
EU(A) = 0.5[kxUx(x-) + (1-kx) Uy(y-)] + 0.5[kxUx(x+) + (1-kx)Uy(y+)]
= 0.5[kxUx(x-) + (1-kx) Uy(y-) + kxUx(x+) + (1-kx)Uy(y+)]
EU(B) = 0.5[kxUx(x-) + (1-kx) Uy(y+)] + 0.5[kxUx(x+) + (1-kx)Uy(y-)]
= 0.5[kxUx(x-) + (1-kx) Uy(y+) + kxUx(x+) + (1-kx)Uy(y-)]
= EU (A)
The intuition behind additive independence is that, in assessing
uncertain outcomes of both attributes, we only have to look at one
attribute at a time, and it does not matter what the other attribute’s
value are in the uncertain outcomes.

22
When we are considering a choice among risky
prospects involving multiple attributes, if
additive independence holds, then we can
compare the alternatives one attribute at a time.
If you have a clear preference for one or the
other, then additive independence cannot hold.

23
Substitutes

and Complements
In the multiattribute utility function, the
interaction between the attributes is captured by
the term (1-Kx –ky)Ux(x)Uy(y).
The sign of (1-kx –ky) can be interpreted in terms
of whether x and y are complements or substitutes
for each other.
Suppose (1-kx –ky) is positive. Now examine the
multiaribute utility function:
U(x,y) = kxUx(x) + kyUy(y) +
(1-kx-ky)Ux(x)Uy(y)

24
Preferred values of X and Y will give high values to the
conditional utility functions, and the positive (1-kx-ky)
will drive up the overall utility for the pair even higher.
Thus, if (1-kx-ky) is positive, the two attributes
complement each other.
On the other hand, if (1-kx-ky) is negative, high values
on each scale will result in a high product term, which
must be subtracted in the multiattribute preference
value. In this sense, preferred values of each attribute
work against each other. But if one attribute is high and
the other low, the subtraction effect is not as strong.
Thus,if (1-kx-ky) is negative, the two attributes are
substitutes.

25
Three or More Attributes
When the decision problem involves three or more
objectives, modeling preferences is more difficult.
Building a utility function that will permit
interactions across many attributes can become
complex. Under certain conditions, however, the
multiplicative utility function can be used.

26
Interaction between Attributes
Attributes interact when preferences for outcomes on one
attribute depend in some way on the level of another
attribute.
In particular, because we often are concerned about
decision making under uncertainty, interaction among
attributes in a utility sense means that preferences for
lotteries over outcomes of one attribute may differ when
lotteries over other attributes change.
Another way to say this is that your certainty equivalent for
a specific lottery in attribute A may depend on exactly what
uncertainty you face regarding attribute B.

27
 Interaction between Attributes
For example, consider a job choice. The risks you are willing to take
regarding salary may depend on the chances associated with
advancement.
Suppose that two jobs offer the same prospects in terms of future
raises in salary.
One job, though, is with a smaller firm, and gives potential for both
more control in the firm as well as possibly losing the job
altogether.
The other job is a relatively stable one with a much more gradual
promotion path.
An individual might find the salary uncertainty more acceptable
(accord to it a higher CE) when it is accompanied with the riskier
promotion path, because the combination of a higher salary and
a lot of control in the company is a much preferred alternative as
compared to either a lot of control alone or a high salary alone. In
this sense, the two attributes interact in a complementary way.
The additive utility function discussed in last session does not
permit interaction.

28
Preferential Independence
Preferential independence means that preferences over sure
consequences in one attribute do not depend on sure
consequences in another. For most of us, preferences over cost
and quality of goods that we purchase are preferentially
independent.
Imagine a group of products that all have the same price but
differ in specific and known ways on quality. Your rankings of
preferences in terms of quality probably would not depend on
the common price; better quality is preferred. Likewise, if
quality were fixed and price were varied, your rankings of
preference over prices would not vary; lower prices would be
better.
Examples of attributes that are not preferentially independent
are more difficult to find.
The attributes in question would be beverages and weather. On
a cold day you might prefer a hot drink, but on a warm day you
might prefer something cold. Regardless of the beverage on
hand, though, you might prefer a warm day to a cold one.
29
 Difference between Preferential ,Utility and
additive Independence
Consider two attributes, A and B. Preferential
independence means that preferences over sure
consequences in A do not depend on sure consequences of
B; no matter where B is set, the preferences (rankings) over
sure consequences in A are the same.
In contrast, utility independence means that we are
looking at lotteries over A. If A and B are utility
independent, then preferences over lotteries (or CEs) in A
do not change with different sure consequences in B.
With additive independence, preferences for lotteries in
Attribute A do not depend on lotteries in attribute B.

30