THE ANALYTIC

HIERARCHY
PROCESS
EXTENSIONS

AHP VALIDATION EXERCISE

This exercise helps to validate the AHP.
You will make judgments on the relative sizes of
the areas of five shapes to find the percentage
each shape contributes to the total area.
The hierarchy has only two levels: the goal and the
five shapes.

AHP VALIDATION EXERCISE

For example, the results might indicate that one
shape is 30% of the total areas of the five
shapes.
We could use plane geometry to compute the
exact areas.
Using the AHP should provide estimates that are
close to the actual values.
The five shapes are on the next slide.

AHP VALIDATION EXERCISE

A
B

E
D
C

AHP VALIDATION EXERCISE

And now for the moment that you have all been
waiting for.........
The relative size of the five shapes are:
Circle A:
0.471
Triangle B:
0.050
Square C:
0.234
Diamond D:
0.149
Rectangle E:
0.096

MULTI-LEVEL HIERARCHIES
Tom Saaty suggests that hierarchies be limited to
six levels and nine items per level.
This is based on the psychological result that
people can consider 7 +/- 2 items
simultaneously (Miller, 1956).
Brainstorming can identify several dozen criteria.
In this case, related items are grouped into
categories, creating additional levels in the
hierarchy.

MULTI-LEVEL HIERARCHIES
The levels can be: goal, criteria, subcriteria, and
alternatives.
In Expert Choice, subcriteria are entered by
highlighting the desired criterion and selecting
the Edit and Insert Child of Current Node
commands.
Alternatively, if many subcriteria are entered at
one time, they can be dragged and dropped
under the desired criteria.

MULTI-LEVEL HIERARCHIES
Consider our car evaluation problem where ten
evaluation factors have been identified.
CARMULTI.AHP shows how these factors can be
grouped to form a four level hierarchy: goal,
criteria, subcriteria, and alternatives.
Notice that the Safety criterion has no subcriterion.
Also, pairwise comparisons are needed for each set
of subcriteria.

MULTI-LEVEL HIERARCHIES
Another important point is that all items on the
same level should be within one order of
magnitude of importance.
For example, NPV might be more than ten times
more important than initial market size and
appear one level above initial market size.
However, all market criteria taken together
might be comparable to NPV and appear on
the same level.

MULTI-LEVEL HIERARCHIES
We now display two additional examples of multilevels hierarchies using Expert Choice.
Both are based on student projects.
They appear in files VENDOR.AHP and SITE.AHP.
Others are found in the samples folder in Expert
Choice.

RATINGS: Background
Multilevel hierarchies are needed when there are
many criteria - but what happens if we have
many alternatives?
The ratings approach is used when there are a
large number of alternatives to be evaluated.
For example, if there are 50 employees to be
evaluated, then 1,225 (50(49)/2) pairwise
comparisons would be required for each
criterion!

RATINGS: Background
It is impractical to make that many alternative
pairwise comparisons.
The ratings approach requires setting up a
ratings scale under each criterion.
For example, in evaluating an employees
organizational skills, a manager could rate
the employee as either Excellent, Very Good,
Good, Fair, or Poor.

RATINGS: Background
It is crucial to define what Excellent means and
how it is attained.
Pairwise comparisons are needed to determine
the relative importance of each ratings scale
category (intensity).
For example, with respect to the organizational
skills criterion, how much better is an
Excellent rating compared to a Very Good
rating?

RATINGS: Background
The answer to this question might be different if
we changed the criterion from organizational
skills to implementation skills.
In fact, you may decide to use different
intensities for each criterion.
It is important to understand that alternatives are
not pairwise compared in a rating model,
rather alternatives are rated for each criterion.

RATINGS: Background
Ratings models are a part of everyday life.
Assigning grades to any course is a ratings
exercise.
Since an A is assigned a score of 4.00 and a C is
assigned a score of 2.00, it follows that an A is
twice as good as a C.
We never met a student who agreed with this!
Do you?

RATINGS: Background
Consider the following example.
Although a 91 is only two points higher than an
89, assigning an A to the 91 and a B to the 89
means that the 91 is really 1.33 (4.00/3.00)
times better than the 89.
These and other problems are discussed at the
Expert Choice web site
(www.expertchoice.com) under Annie Person.

RATINGS: Background
Many organizations use ratings or scoring
models for evaluation.
For example, in evaluating carpet suppliers, a
company might assign the values 3, 1, 2 for
cost, support, and quality, respectively.
Typically, they assign 5, 4, 3, 2, and 1 to ratings
of excellent, very good, good, fair, and poor,
respectively.

RATINGS: Background
Suppose supplier A is judged to be good in cost,
excellent is support, and good in quality.
Supplier As score would be 3*3+1*5+2*3=20.
Assume that supplier B is judged to be excellent
in cost, fair in support, and very good in
quality.
Supplier Bs score would be: 3*5+1*2+2*4=25.

RATINGS: Background
Can we say that supplier B is 25% better than
supplier A?
Absolutely not! The numbers assigned as
criteria weights and as intensity weights are
not necessarily ratio-scaled.
Ratio-scaled comparisons, such as dividing
supplier total scores are meaningless in such
cases.

RATINGS: Background
Ratio-scaled measurement assumes, for
example, that cost is 3 times (3/1) more
important than support, and that an excellent
rating is 1.25 times (5/4) better than a very
good rating for each criterion.
This is rarely, if ever, the case for such scoring
systems!
The AHP is preferred because it applies ratioscale measurement throughout the evaluation
process.

EXPERT CHOICE: Ratings

Goal and criteria (and possibly subcriteria) are
entered in a ratings model in the same fashion
they were entered in standard AHP.
Criteria (and possible subcriteria) pairwise
comparisons are next performed.
Next, select the Data Grid button (looks like a
spreadsheet).
Highlight a cell in the first criteria column and
select the Formula Type and Ratings commands.

EXPERT CHOICE: Ratings

Enter each rating scale intensity (for example,
excellent, very good, good, fair, and poor) in
the Intensity Name column.
When finished select the Assess command.
You can now enter the pairwise comparisons
for the rating scale intensities.
After recording judgments, select the Close
command.

EXPERT CHOICE: Ratings

If the rating scale intensities and their pairwise
comparisons are not the same for all criteria,
highlight a cell in the second criteria column
and repeat the process.
If the intensities and pairwise comparisons are
the same for all criteria, then select the
Formulas Grid button (looks like Y=f(x)).
(If this button does not appear, select the Model
View button and then the Data Grid button.)

EXPERT CHOICE: Ratings

To copy the intensities and pairwise comparisons
(from criterion 1) to other criteria (criteria 2 and
3), highlight the Ratings cell in the Type column
of criterion 1 and select the Edit and Copy
Formula commands.
Next, highlight the Ratings cells for criteria 2 and 3
and select the Edit and Paste Formula commands.
You have now copied all of the ratings intensities
and their pairwise comparisons from criterion 1 to
criteria 2 and 3.

EXPERT CHOICE: Ratings

Select the Data Grid button and you are ready to
enter the alternatives.
Remember that alternatives are NOT entered in
the hierarchy.
Highlight the first cell in the Alternative column
and enter each alternative in turn.
When finished, highlight the cell corresponding
to the first alternative (row 1) and the first
criterion (column 1).

EXPERT CHOICE: Ratings

Select the desired rating scale intensity and repeat for
all criteria for all alternatives.
For a given alternative (row), as the user highlights
each criterion (column), the appropriate intensities
appear and the user selects the desired one.
The final step is to select the View and Totals column
commands to see the final scores for each alternative.
To sort, highlight any final weight and select the Edit
and Sort, Descending commands.

EXPERT CHOICE: Ratings

Criterion intensity scores are computed similarly to
ideal synthesis without the normalization step.
First, all intensity weights are divided by the
largest intensity weight.
Second, the adjusted intensity weight selected by
the user is multiplied by the criteria weight and
the results added to the total score.

EXPERT CHOICE: Ratings

An AHP ratings model for our carpet supplier
problem is in a file called CARPET.AHP.
The local weights for each rating scale intensity
are: 0.419, 0.263, 0.160, 0.097, and 0.062.
Dividing by 0.419 yields adjusted weights of:
1.000, 0.627, 0.382, 0.232, and 0.148.
For example, if we select a good rating for cost,
then 0.382 times the cost weight of 0.558 or
0.213 is added to the total score.

EXPERT CHOICE: Ratings

Another example of a ratings model with
subcriteria appears in EMPEVAL.AHP.
This model is based on a student project which
utilized the actual factors in an employee
evaluation system.
Others are found in the samples folder in
Expert Choice.

GROUP DECISION MAKING

How did the couple arrive at their combined
judgments in the original car evaluation
problem?
There are many ways of applying AHP to support
a group decision-making process.
For example, all of the parties discuss, debate, and
eventually agree on each pairwise comparison
entry.

GROUP DECISION MAKING

Alternatively, each individual provides their own
judgments in separate copies of the model.
These results could be summarized and used as a
basis to reach consensus.
Another approach is to create a hierarchy with
goal, participants, criteria, and alternatives.
Pairwise comparisons can determine each
participants weight in the process.

GROUP DECISION MAKING

One last approach is to achieve consensus
mathematically.
Each participant provides their own judgments for
each pairwise comparison and the results must
be averaged.
For example, suppose two individuals compared
cost to safety and provide judgments of 9 and
1/9.

GROUP DECISION MAKING

The arithmetic mean is 4.56 ((9+(1/9))/2). Do
you think this is the best estimate?
Probably not! Since both judgments are at
opposite ends of scale, we would expect the
combined judgment to be 1.00.
The geometric mean produces this result.
In general, if there are n individuals that provide
judgments, the geometric mean is defined as
the nth root of the product of the n judgments.

GROUP DECISION MAKING

As another example, in comparing cost to safety
suppose the judgments of three individuals are
2, 4, and 8.
The geometric mean is the cube root of their
product (64) which is 4.
Expert Choice manages the entire group decision
making process and achieves consensus
mathematically by computing the geometric
mean.

GROUP DECISION MAKING

First, create a hierarchy as described earlier.
Tell Expert Choice that this is a group model by
selecting the Go and Participant Table commands.
Next, select Edit and Group Enable, followed by
Edit, Add N Participants, and enter the number of
participants.
Click on a participant to change the name, enter any
demographic data, and select File and Close.

GROUP DECISION MAKING

At this point, there are N participants and a
facilitator.
The facilitator acts as the leader and may also
enter judgments, if desired.
When a group model is opened, you must
respond with either the facilitators name (you
have access to all information) or the name of
one of the participants (you only have access
to that participants information).

GROUP DECISION MAKING

The facilitator can enter pairwise comparisons for
all participants.
Select a participant from the Participants drop-down
list on the toolbar (under the Go command).
Choose a pairwise comparison mode and enter the
judgments for the participant.
Record the judgments when finished and repeat for
all parts of the hierarchy and for all participants.

GROUP DECISION MAKING

After all pairwise comparisons have been entered
for all participants, the judgments are combined.
This is accomplished by selecting Combined from
the Participants drop-down list.
Next, select Edit, Combine Participants
Judgments/Data, Entire Hierarchy, and Both.
This will combine judgments by computing all
necessary geometric means.

BUILDING LARGER MODELS

Useful Expert Choice features
File, Print Preview, File, Save as Word Document
commands creates a Word file of the entire
hierarchy. Use Options and Printing commands
to select desired output.
Drop and drag features are useful when
developing the hierarchy.
To get information from Word to Expert Choice
use the Edit, Paste Children from Clipboard
commands. This is useful if developing the
hierarchy while brainstorming in Word.

BUILDING LARGER MODELS

Lessons that we have learned about AHP.
Have experts develop their part of the hierarchy.
Develop hierarchy iteratively over several
sessions.
An alternate approach is to only develop a
benefits hierarchy. The benefits alternative
weights could be used in a cost/benefit analysis.
You could also have a benefits hierarchy and a
cost hierarchy.

BUILDING LARGER MODELS

Lessons that we have learned about AHP.
Rank alternatives or criteria before performing
pairwise comparisons. This helps consistency.
Many people are comfortable with graphical
mode of pairwise comparison.
After entering pairwise comparisons, Expert
Choice displays a graphical representation of the
weights. The user can move these bars if
necessary. Expert Choice computes the
corresponding pairwise comparisons that yield
these weights.

SUMMARY
In this module:
we provided an overview of classical decision
analysis; and
offered the AHP as an alternative decisionmaking process.

SUMMARY
AHP benefits include:
natural way to elicit judgments;
measure degree of inconsistency;
easy to use;
allows broad participation; and
fully supported by Expert Choice.