THE ANALYTIC HIERARCHY PROCESS: DECISION MAKING IN COMPLEX

ENVIRONMENTS

Thomas L. Saaty

University of Pittsburgh
Pittsburgh, PA 15260

INTRODUCTION

The Analytic Hierarchy Process (ARP) is a problem solving frame-

work. It is a systematic procedure for representing the elements
of any problem. It organizes the basic rationality by breaking down
a problem into its smaller constituent parts and then calls for only
simple pairwise comparison judgments, to develop priorities in each
hierarchy.

The AHP provides a comprehensive framework to cope with the

intuitive, the rational, and the irrational in us at the same time.
It is a method we can use to integrate our perceptions and purposes
into an overall synthesis. The AHP does not require that judgments
be consistent or transitive. The degree of consistency (or incon-
sistency) of the judgment is revealed at the end of the ARP process.

When dealing with scientists, corporate managers, the academic

community, lay people and others in solving problems or planning,
we have observed repeatedly that people provide subjective judgments
based on feelings and intuition rather than on well worked out logical
reasoning. Also when they reason together people tend to influence
each other's thinking. Individual judgments are altered slightly
to accomodate the group's logic and the group's interests. However,
people have very short memories and if asked afterwards to support
the group judgments, they instinctively go back to their individual
judgments.

One also observes that people find it difficult to justify

their judgments logically and to explain how strong these judgments
are. As a result people make great compromise in their thinking
to accomodate ideas and judgments.

285
R. Avenhaus et al. (eds.), Quantitative Assessment in Arms Control
© Plenum Press, New York 1984
 Designing an analytic hierarchy - like the structuring of a
problem by any other method necessitates substantial knowledge of
the system in question. A very strong aspect of the AHP is that the
knowledgeable individuals who supply judgments for the pairwise
comparisons usually also play a prominent role in specifying the
hierarchy.

THE ANALYTIC HIERARCHY PROCESS: A BRIEF DESCRIPTION

When people make a decision, probably they would not list all
the factors that are essential to this decision and explicitly com-
pare the significance of each. Nevertheless people constantly make
comparisons and implicitly indicate preferences among different
choices. In making decisions, we have observed repeatedly that peo-
ple provide subjective judgments based on feelings and intuition,
as well as their "logical" understanding.

The Analytic Hierarchy Process (AHP) is a multiobjective,

multicriterion decision-making approach which employs a pairwise
comparison procedure to arrive at a scale of preferences among
sets of alternatives. To apply this technique, it is necessary to
break down a complex unstructured problem into its component parts;
arraying these parts, or variables, into a hierarchic order; assign-
ing numerical values to subjective judgments on the relative impor-
tance of each variable and synthesizing the judgments to determine
which variables have the highest priority and should be acted upon
to influence the outcome of the situation.

In Figure 1, we show the structure of a hierarchy with several

levels. The first level indicates the objective or focus concerned
with the best choice of a mode of transportation from home to school.
In the second level of the hierarchy are the criteria considered
important by the person concerned to choose from several modes. In
the third level are the modes themselves. The structure can be
revised to include the amount of detail desired. Usually it is
possible to capture the richness and complexity of most problems
within hierarchies using a few levels.

The application of the pairwise comparison and the prioritization

prodedure proceeds as follows: the factors at level II in Figure 1
are now compared with one another in relation to their relative im-
portance in making the best mode selection. We ask, for instance,
how much stronger an impact does "time" make than "exercise" on
determining the mode?

The pairwise comparisons are entered in a matrix, sometimes

referred to as a reciprocal matrix, since the a .. element of this
1.J
matrix is l/a .. or the reciprocal of the a .. element. A reciprocal
J1. J1.

286
matrix also has the propoety that its principal diagonal elements
are unity reflecting the fact that a factor when compared with itself
should obviously produce a judgment of "equal importance." Thus
if a level includes n factors, a total of n(n-l)/2 comparisons are
needed since the diagonal elements and the reciprocals are entered
automatically. For instance, with six factors at level II, we
need to make fifteen pairwise comparisons.

The scale used to indicate the relative judgmental preference

of one factor over another is shown in Table 1.

Table 1. Scale of Relative Importance

Intensity of
Relative
Importance Definition Explanation

1 Equal importance Two activities contri-

bute equally to the
objective.
3 Moderate importance Experience and judgment
of one over another slightly favor one ac-
tivity over another.
5 Essential or strong Experience and judgment
importance strongly favor one
activity over another.
7 Demonstrated impor- An activity is strongly
tance favored and its dominance
is demonstrated in prac-
tice.
9 Absolute importance The highest possible order
of affirmation.
2,4,6,8 Intermediate values When compromise is needed
between the two ad-
jacent judgments

Reciprocals of If an activity has

above non-zero one of the above num-
numbers bers assigned to it
when compared with a
second activity, then
the second activity
has the reciprocal
value when compared
with the first

287
 The use of this scale is contingent on the satisfaction of the
assumption that the factors being compared are of similar orders of
magnitude. When a factor does not appear strictly comparable with
another, clustering techniques can be applied so the comparisons
are made possible at the level of the clusters.

The AHP does not require that judgments be consistent or even

transitive. The consistency (or inconsistency) of the judgments
is revealed at the end. One might ask. if the judgments were tot-
ally random in nature. what kind of consistency would the AHP interpret
them to have? The consistency of such random judgments should be
much worse than the consistency of informed judgments and can be
used to evaluate how good those judgments are.

If a o represents importance of criterion i over criterion j

1J
0

and a jk represents importance of criterion j over criterion k. then

the importance of criterion i over criterion k. must equal a o
1J
0

a ok , for the judgments to be consistent; i.e., we must have the

oJ
1mportance 0
f the l th over t h e J th t1mes t h e 1mportance 0 f t h e Joth
0 0 0 0

th
over the k . However, all the comparison matrices need not be
consistent.

It can be shown that the eigenvector corresponding to the lar-

gest eigenvalue of the matrix provides the relative "local" prior-
ities of the factors. Thus, we obtain from the matrix of paired
comparisons a vector of weights which reflects the relative impor-
tance of the various factors. Local priorities become global
priorities when they are weighted by multiplying each of them by
the global priority of the corresponding criterion and then adding
over all the criteria with respect to which the comparison is made.
See section 3.

It can be proved that for a consistent matrix. Amax = n. where

Amax is the largest eigenvalue of a reciprocal matrix of order n;
hence, as a measure ~f deviation from consistency we use the consis-
/I. -n
max
tency index, C.I.
n-l
By simulating random reciprocal matrices of different orders
(sample size: 500) the average consistency indices (known as the
random index (R.I.)) have been established as follows:

n 2 3 4 5 6 7 8

R.1. 0.00 0.58 0.90 1.12 1.24 1.32 1.41

288
 The ratio of the C.I. to the R.I. for the same order matrix is
called the inconsistency ratio. Consistency is a central concern of
all analysis. Inconsistency allows one to readjust a system of judg-
ments to accomodate new data. Both are important for learning but
consistency is more important by one order of magnitude. Incon-
sistency must be admissible without dominating or confusing consis-
tency. An overall measure of consistency for a hierarchy can be
obtained by weighting each consistency index by the priority of
its criterion, adding for the entire hierarchy and dividing by a
similar sum with corresponding random indices. An inconsistency
ratio of the order of 0.10 is considered tolerable (Saaty, 1980).
When this level is exceeded significantly revisions of the judgments
are recommended.

THE WEIGHTING PROCESS

Assume that we are given n stones, AI' ... , An whose weights

wI' ... , wn ' respectively, are known to us. Let us form the matrix
of pairwise ratios whose rows give the ratios of the weights of each
stone with respect to all others. Thus we have the matrix;

Al A2 A
n
I
wI /w 2 w1/w 2
I Al
I
I
wi/wn wI wI
w2,w
I
I w2/w 2 w2/wn w2 w2

A ~ n

A wn/wr wn/w2 w /w w w
n n n n n

we have multiplied A on the right by the vector of weights w. The

result of this multiplication is nw. Thus, to recover the scale
from the matrix of ratios, we must solve the problem Aw = nw or
(A - nI) w = O. This is a system of homogeneous linear equations
It has a nontrivial solution if and only if the determinant of A -
nI vanishes; i.e., n is an eigenvalue of A. Now A has unit rank
since every row is a constant multiple of the first row. Thus,
all the eigenvalues except one are zero. The sum of the eigenvalues
of a matrix is equal to its trace, and in this case the trace of A

289
is equal to n. Thus n is an eigenvalue of A and we have a nontrivial
solution. The solution consists of positive entries and is unique
to within a multiplicative constant, by the Perron-Frobenius theorem
since A is irreducible; i.e., is not decomposable into blocks of the
form

To make w unique, we normalize its entries by dividing by their sum.

Thus, given the comparison matrix we can recover the scale. In this
case the solution is any column of A normalized. Note that in A we
have a .. = l/a .. the reciprocal property. Thus, also, a .. = 1. Also,
J1 1J 11
A is consistent; i.e., its entries satisfy the condition

a .. = a.k/a ..
1J 1 1J
Thus, the entire matrix can be constructed from a set of n elements
that form a chain across the rows and columns.

In the general case we cannot give the precise values of w./w.

1 J
but only estimates of them. For the moment, let us consider an
estimate of these values by an expert who we assume makes small er-
rors in judgment. From matrix theory we know that small perturbation
of the coefficients implies small perturbation of the eigenvalues.
Our problem now becomes A'w' = A w' where A is the largest
max max
eigenvalue of A'. To simplify the notation, we shall continue to
write Aw = A w where A is the matrix of pairwise comparisons.
- max
The problem is now, how good is the estimate w? Note that if we
obtain w by solving this problem, the matrix whose entries are w./
1
w. is a consistent matrix which is our consistent estimate of the
1
matrix A. A itself need not be consistent. In fact, the entries
of A need not even be ordinally consistent; i.e., Al may be pre-
ferred to A2 all A2 to A3 but A3 is preferred to AI' What we would
like is a measure of the error due to inconsistency. It turns out
that A is consistent if and only if A n and that we always have
max
A > n. This suggests using A - n as an index of departure from
max - max
consistency. But
n
A - n = - LA., A = AI'
max i=11 max
where A., i = 1, .•• , n are the eigenvalues of A. We adopt the av-
1
erage value(A -n)/(n-l), which is the (negative) average of A.,
max 1

290
i = 2, ••• , n (some of which may be complex conjugates). On cal-
culating this value we compare the result with those of the same
index obtained as an average over a large number of matrices of the
same order whose entries are random. However, we preserve the rela-
tions aji = l/aji' aii = 1 in these matrices to improve consistency.
The reason for this is that if one stone is estimated to be k
times heavier than another, it does not seem unreasonable to re-
quire that the second stone be estimated to be l/k times the
weight of the first. If the ratio of our index to that from random
matrices is significantly small, we accept the estimates. We shall
not go into the details of this procedure.

Note that in making the estimates and to keep the comparisons

relevant, an individual has to keep in mind all the elements being
compared. It is known that an individual cannot simultaneously com-
pare more than 7 ± 2 elements. If this is so, then how are we able
to have measurement across wide classes of objects? The answer is
by hierarchical decomposition. The elements are grouped ordinally
(as a first estimate) into comparability classes of about seven el-
ements each. The element with the highest weight in the class of
lighter weight elements is also included in the next heavier class
and serves as a pivot to uniformize the scale between the two clas-
ses. The procedure is repeated from a class to an adjacent one
until we have all the elements appropriately scaled.

Another way which has proved very useful in practice is to div-

ide the criteria into subcriteria and rank each of these into desir-
ability levels: high, medium, low for which priorities are deter-
mined and global priorities computed. Suppose then below this level
we have a level ~omposed of a large number of alternatives. We take
each one and simply check the high, medium or low value as sub-
criteria according to which it satisfies best. We then add the glo-
bal priorities for each and these checked values overall value for
the alternative. It is then accepted or rejected depending on how
it compares with the others.

In passing, we note that the eigenvector approach to measurement

(as one might expect) preserves ordinal preferences among the alter-
natives; i.e., if an alternative is preferred to another, its eigen-
vector component is larger than that of the other.

The AHP has already been successfully applied to a variety of

fields; these include: a plan to allocate energy to industries;
designing a transport system for the Sudan; planning the future of
a corporation and measuring the impact of environmental factors on
its development; design of future of scenarios for higher education
in the United States; the candidacy and election selections; setting
priorities for the top scientific institute in a developing country;
a faculty promotion and tenure p~oblem and oil price predictions.

291
EXAMPLES

We now give three different kinds of examples to illustrate the

method.

Transportation

In this section we give a systematic analysis of a decision

faced everyday -- how should I transport myself to school on a day
when I have appointments and I have to be dressed up; how should
I do it on an ordinary, no appointment day? This choice involves
the consideration of many factors in the environment which change
from day to day such as weather, or during a day such as during
slack or rush hours. The transportation options considered in this
analysis include walking, driving an automobile, taking the bus,
or riding a bicycle. Each of these alternatives has advantages and
disadvantages unique to that alternative.

Alternatives

Walking. Walking has several advantages. It is inexpensive,

healthful, enjoyable and reliable. However, some circumstances make
the disadvantages of walking more apparent. It provides little
protection from the elements, it takes considerable time (depending
on how far one lives). It can also be dangerous at night (depend-
ing on the neighborhood in which one lives). It can be an undesir-
able experience after a long day. Finally it can be quite uncom-
fortable due to perspiration, and can cause one's shoes to wear out
faster.

Driving. Driving an automobile to school is very convenient

for covering the distance. When running late, it is more reliable
than taking the bus, since one has an unpredictable waiting time in-
volved when catching the bus. Alternatively, the unknown aspect
of driving a car is finding a parking place. There are 2 trade-offs
when seeking a parking spot -- time and money -- both of which are
influenced by the time of day one arrives at school. Early in the
morning or in the evening one has a reasonably good chance of find-
ing an unmetered spot within a 5 to 10-minute walk of the school.
Between 9:00-5:00, however, one must choose between paying at a meter
or walking much farther.

Other aspects of driving include the risks of accidents, cost of

gasoline and maintenance, the risk of theft and vandalism, and the
risk of some mechanical break down.

Busing. The nearest bus stop in this case is about 2 blocks

away. It is quite dependable, but requires a certain amount of
slack-time in the travel plans. Once one becomes accustomed to the
bus system, one finds that it is reliable, convenient, induces a low

292
risk to one's safety, is nonstressful (compared to driving), easy
to use, and cheaper than driving and paying to park. The major
disadvantages are that it is more expensive than walking or biking,
it has inflexible schedules (e.g., when seconds are important), it
can be quite uncomfortable in hot weather and crowded conditions,
it may require additional walking or transfers to reach destination,
and it is slower than driving one's car.

Bicycling. The bicycle has many advantages. It is faster

than walking, busing, or driving because one has flexibility to leave
when one is ready, travelling the same route as a car, one can ride
directly to a bike rack on a campus which simpliefies the parking
situation. Bicycling provides good exercise, is usually refreshing,
and over a long period is cheaper than maintaining a car or riding
a bus.

Bicycling is inappropriate when the weather is poor, when one

needs to wear formal attire, or when the distance is extremely long
(greater than 3 miles). It becomes more dangerous if one rides
at night or in fast traffic. It also poses greater accident and
theft risks.

Assumptions

This decision problem was studied under certain assumptions

about the environmental conditions. The first required that the
schedule for the day include an interview and that neatness, con-
venience, and timeliness were most important. The second assumes
that during the "typical" day, weather, time/cost considerations and
exercise would be more dominant in the decision.

The individual lives about 2 miles from school. It takes about

35 minutes to walk this distance, 10 minutes to ride a bicycle,
from 15 to 35 by bus depending on the waiting time, about 15 to 20
minutes to drive and park (5 to 10 minutes) and walk (about 10 min-
utes after parking).

What follows is a detailed rationale for the hierarchical struc-

ture of the problem and the judgments used to formulate the hier-
archy.

Structuring the Hierarchy

Goal. The goal or objective of this exercise is to arrive at

the "best" Hode of transportation for travelling from home to school
on any particular occassion.

Criteria. The criteria considered in the selection of a trans-

portation mode to school are grouped into six major classes; SCHEDULE,
WEATHER, TIME OF DAY, TIME/COST, EXERCISE, AND RISKS. The schedule

293
for the day will determine the importance of appearance, timeliness,
and flexibility required, all of which influence the best mode of
transport. This is the most important factor, since certain things
about the schedule would determine the transportation used, no mat-
ter what the weather, time of day, travel time and costs, exercise
value, or risks involved.

The next most important factor is the weather. If it is rain-

ing, walking and bicycling are less likely selections than when it
is a clear day. Also, driving is a less likely choice than busing
because driving usucally results in a 10-minute walk from the park-
ing place to school. The bus minimizes the time exposed to the
elements. In this stiuation the money spent on the bus is worth
more than the consequence of a 10-minute walk in the rain. If it
snows, driving is eliminated as an alternative since there were no
snow tires on the car. Snow has a different influence than rain
as the individual enjoys walking in the snow and values it more
than riding the bus.

On a clear day, bicycling and walking are the most desirable

alternatives since the time and money saved by these alternatives
dominate the unimportant benefits of the cover offered by the
car or bus.

The time of day influences the choice of vehicle. For example,

parking places for an auto are very scarce between 9 and 5. A night
class would probably lead to driving a car because parking is not
a major problem at that time and bicycling is more risky after dark.

Next to schedule and weather, time and cost are the most im-
portant factor. Each option has a unique set of time/cost char-
acteristics. Walking is the slowest alternative, but also the
cheapest. Biking is cheaper than busing and driving and is us-
ually the fastest means of transportation. However, if certain en-
vironmental conditions prevent biking, time/cost considerations are
used to help choose between driving and busing. While busing is
probably cheaper than driving and paying to park, it is not cheaper
than driving and walking 10 minutes from an unmetered parking area.
As a result, one would usually drive his car rather than take the
bus.

Exercise has been a significant motivating factor in selecting

the mode of transport. With a hectic schedule, there is little free
time for organizing athletic activity. Consequently, a 35 minute
walk twice a day would serve a good purpose in providing some ex-
ercise that would be lost if the bus or car were used. A bike also
provides exercise. The bike ride takes about 10 minutes, but the
exercise is of a more intense quality so the total benefit may be
slightly above that derived from walking.

294
 There is risk involved in using each alternative. Risk includes
such factors as theft of property, mugging, health risks such as catch-
ing a cold from walking in cold weather, and mechanical breakdowns.
The bus is probably the least risky alternative, followed by walking,
biking and driving. Driving seems to have the greatest number of
potentially expensive mishaps. Exhibit 1 summarizes the judgments de-
scribed here about the importance of these 6 criteria. The judgments
are first given qualitatively as explained in Table 1 and the cor-
responding number is entered in the matrix. The reciprocal is then
entered in the transpose position. These judgments are the response
given to the question: how much more important is a criterion shown
on the left than a criterion shown above in making a best choice
of a transportation mode? Our hierarchy has a structure shown in
Figure 1. Inconsistency ratios for each criterion with respect to
the alternatives are shown in Exhibit 2, given the situation that
I have an interview schedules. Exhibit 3 shows the results of the
program and the choice of the bus as the best alternative under this
condition.

Exhibits 4 and 5 show the same data and results for a decision
on a "typical" day when not involved in interviews. The answer sug-
gests riding the bike first, walking second, busing third and driv-
ing last under this situation. The results of the program tended to
confi.rm one's overall intuition about the outcome.

Even though the figure shows the entire problem under one focus,
this parti.cular example was worked out to obtain possibly a different
choice for each of the constraints: appointment, no appointment.
In the latter we only changed the priorities of the alternatives
with respect to schedule everything else remained the same. Note
that weather has two subcriteria: good weather and bad weather
and the problem is solved for each of these along with each of the
two subcriteria: slack time of day or rush hour time of day. Thus
in all there are four sets of outcomes or priorities for the alter-
natives under appointment and four under no appointment. The bus
alternative is by far the best whether for all four under appoint-
ment and the bicycle alternative under no appointment. The matrices
of judgments are as follows. The first gives the general weighting
of the criteria. Under appointment we have eight matrices for the
alternatives in terms of each criterion except that there are two
subcriteria for weather and two for time of day. Under no appoint-
ment we have a new matrix for comparing the alternatives with res-
pect to schedule. The other matrices are identical with the other
seven under appointment.

Comparison of the criteria is the same with appointment as it

is with no appointment.

295
N
CD
0)
FOCUS: BEST CHOICE OF TRANSPORTATION MODE

CONSTRAINTS:

CRITERIA:

SUBCRITERIA:

ALTERNATIVES: 0GB8 0088

Criteria Alternatives
S Schedule Tic Time and Cost of B Bicycle
H Weather Transportation A Automobile
GW Good Weather E Exercise Bs Bus
BW Bad Weather R Low Risk Wa Walking
T Time of Day
ST Slack Time
RT Rush Hour
Which criterion is most important in deciding on a best mode of
transportation?

S W T T/C E R Priorities

S 1 6 7 6 5 7 .519

W 1/6 1 5 3 3 7 .204

T 1/7 1/5 1 1/3 1/2 3 .051

T/C 1/6 1/3 3 1 2 5 .112

E 1/5 1/3 2 1/2 1 5 .085

R 1/7 1/7 1/3 1/5 1/5 1 .029

Judgments to compare criteria for a typical day. Inconsistency

Ratio = 0.088.

Matrices of comparison under appointment.

Schedule B A Bs Wa Priorities

B 1 1/5 1/6 1/3 .060

A 5 1 1/2 3 .294

Bs 6 2 1 6 .529

Wa 3 1/3 1/6 1 .117

Inconsistency Ratio .043

Good Weather B A Bs Wa Priorities

B 1 1/6 1/4 1/3 .067

A 6 1 3 2 .493

Bs 4 1/3 1 2 .256

Wa 3 1/2 1/2 1 .184

Inconsistency Ratio .044

297
Bad Weather B A Bs Wa Priorities

B 1 1/5 1/7 1 .066

A 5 1 1/4 4 .245

Bs 7 4 1 7 .619

Wa 1 1/4 1/7 1 .069

Inconsistency Ratio .043

Slack Hour
Time of Da B A Bs Wa Priorities

B 1 1/5 1/6 1/2 .068

A 5 1 1/2 3 .313

Bs 6 2 1 4 .499
Wa 2 1/3 1/4 1. .120

Inconsistency Ratio = .013

Rush Hour
Time of Da B A Bs Wa Priorities

B 1 1/2 1/6 1/2 .085

A 2 1 1/4 1/2 .134

Bs 6 4 1 4 .590

Wa 2 2 1/4 1 .191

Inconsistency Ratio .026

Time/Cost B A Bs Wa Priorities

B 1 5 4 4 .575

A 1/5 1 1/2 1/3 .081

Bs 1/4 2 1 1/2 .133

Wa 1/4 3 2 1 .211
Inconsistency Ratio 0.36

298
Exercise B A Bs Wa Priorities

B 1 5 6 3 .552

A 1/6 1 2 1/4 .099

Bs 1/5 1/2 1 1/5 .063

Wa 1/3 4 5 1 .286

Inconsistency Ratio =.049

Low Risk B A Bs Wa Priorities

B 1 1/3 1/6 1/2 .077

A 3 1 1/3 1 .189

Bs 6 3 1 5 .584

Wa 2 1 1/5 1 .151

Inconsistency Ratio .020

Under no appointment

Schedule B A Bs Wa Priorities

B 1 7 8 5 .652

A 1/7 1 2 1/4 .081

Bs 1/8 1/2 1 1/4 .055

Wa 1/5 4 4 1 .212

Inconsistency Ratio .064

Synthesizing

If we multiply each column of priorities of the alternatives

by the priority of its criterion and add over the alternatives we
have results first under appointment but conditioned above it by the
weather condition and time of day. The second set is similarly gen-
erated under no appointment. For example under appointment we get:

299
 S GW ST Tic E R Composite
(.519) (.204) (.051) ( .112) ( . 085) ( . 029)

B .060 .067 .068 .575 .552 .077 .162

A .294 .493 .313 .081 .099 .189 .292

Bs .529 .256 .499 .133 .063 .583 .389

Wa .117 .184 .120 .211 .286 .151 .157

Here is a sample calculation

.060x.519+.067x.204+.068x.051+.112x.575+.552x.085+.077x.029 .162

The results for appointment are

GW+ST GW+ST BW+RT BW+RT

B .162 .163 .162 .163

A .292 .283 .241 .232

Bs .389 .394 .463 .468

Wa .157 .161 .l34 .l38

The corresponding results for no appointment in which only the com-

parisons with respect to schedule are different are:

B .467 :468 .467 .468

A .182 .173 .l3l .122

Bs .144 .149 .218 .223

Wa .207 .211 .184 .188

One should use the bus on days when he has appointments and ride the
bicycle otherwise.

School Admission Hierarchy (A useful way to prioritize a very

large number of al~ernatives).

In this example we give the hierarchy without setting priorities

The example is a useful illustration of the case where there are
many alternatives. In admitting students to the undergraduate pro-

300
gram of study after they have completed their first t,~o years at
some college, the admissions office usually looks at their past
performance. If it is good, they are admitted, if it is bad they are
rejected and if it is marginal, then they are asked to provide let-
ters of recommendation and come for an interview. The hierarchy to
evaluate each student is shown in Figure 2.

Here the prioritization proceeds downward as follows: First

assume that "courses completed" is all there is and prioritize under
it the importance of the three factors in level 3 and then prioritize
under each of the~e factors the factors in level 4 and so on accum-
ulating the global priorities. In level 9 one would be assigning
relative values to intensities such as high, medium, low for the type
of school. As one examines a student's record, one checks the rel-
evant intensity for that item, e.g., if from Yale, under "place" one
checks H for high, so on. Finally one adds the priorities of the
checked items. If the total is, for example, 80% the student is
admitted since the absolute minimum required for admission may be
75%. Otherwise we go back and prioritize the relative importance
of courses completed versus other factors. Assume that the priority
of courses completed is 75% and other factors 25%. Then a student
whose overall priority falls below 75% in the first brush may qual-
ify better by considering other factors. The total for qualifying
in this case should also be not less than 75%.

301
w Level 1
o
N

Level 2

Level 3

Level 4

Level 5

Level 6

Level 7

Level 8

Figure 2. A Hierarchical Representation of the Student Selection Problem

Presidential Election 1980

As a final example we give without explanation the results of

a prediction prioritization of which of the two candidates, James
Earl Carter and Ronald Reagan was more likely to win by projecting
what the public (a cross section) thought of them on each of the
issues.

The judgments were provided by about 30 business people. When

there is no consensus, the geometric mean of the judgments is used.

The hierarchy is shown in Figure 3.

The matrices of pairwise comparison and the r.esu1ting priorities

are given below.

Priorities of Criteria

Election 1980 P1 P2 P3 P4 Priorities

Personality P1 1 1/5 1/7 3 .08

Politics P2 5 1 5 9 .61

Professional 7 1/5 1 7 .27

Aptitude P3
Physical P4 1/3 1/9 1/7 1 .04

Inconsistency Ratio = .17

Priorities of the Factors which Define the Criteria

Personality R/O CHR M/R APR Priorities

Region of Origin
R/O 1 1/7 1/5 1/2 .06
Charisma
CHR 7 1 1/2 3 .32
Media Relations
M/R 5 2 1 7 .52
Appearance
APR 2 1/3 1/7 1 .10

Inconsistency Ratio .05

303
w
o FOCUS:
~

CRITERIA:

REGION OF INTERNATIONAL I
SUBCRITERIA: H H EXPERIENCE AGE
ORIGIN STANDING

CHARISMA PARTY COMPETENCE APPEARANCE

MEDIA RELIGION CREDIBILITY

APPEARANCE RUNNING MATE LEADERSHIP

MONEY

FOREIGN SOCIAL NATIONAL

ISSUES: ENERGY ECONOMY
AFFAIRS II ORDER II DEFENSE
I - II If
CANDIDATES: [ CARTER REAGAN

Figure 3.
Politics I/S PTY REL R/M MON Priorities

International
Standing I/S 1 1/4 7 1/4 1/3 .11

Party PTY 4 1 9 2 1 .34

Religion REL 1/7 1/9 1 1/7 1/6 .03

Running Mate R/M 4 1/2 7 1 1/2 .22

Money MON 3 1 6 2 1 .30

Inconsistency Ratio .07

Professional
Aptitude EXP COM CRD LDS Priorities

EXPERIENCE EXP 1 1/3 1/5 1/3 .06

COMPETENCE COM 3 1 1/5 1/3 .12

CREDIBILITY CRD 5 5 1 3 .54

LEADERSHIP LDS 6 3 1/3 1 .28

Inconsistency Ratio .08

Physical Age Appearence Priorities

Age 1 1/5 .17

Appearance 5 1 .83

Inconsistency Ratio = 0

Overall Priorities of the Factors

(1) Region of Origin .005 (10) Experience .016

(2) Charisma .025 (11) Competence .032
(3) Media Relations .041 (12) Credibility .148
(4 ) Appearance .040 (13) Leadership .076
(5) International Standing .065 (14) Age .006
(6) Party .206
(7) Religion .019
(8) Running Mate .136
(9 ) Money .183

305
We group these under the following. four headings as follows:

Cluster Party Credibility Media International

Standing
Members (6)(8)(9) (12)(13) (3)(4) (5)(11)

Total .57 .25 .09 .10

Priorities
Normalized
for factors
Included

Priorities of the Issues with Respect to the Four Clusters of Factors

Party A B C D E Priorities

A 1 1/5 6 3 3 .21
B 5 1 6 9 5 .54

C 1/6 1/6 1 2 1/5 .08

D 1/3 1/9 1/2 1 2 .08

E 1/3 1/5 5 1/2 1 .11

Consistency .20

Credibility A B C D E Priorities

A 1 1 1 1 1 .20

B 1 1 1 1 1 .20

C 1 1 1 1 1 .20

D 1 1 1 1 1 .20

E 1 1 1 1 1 .20

Consistency 0

A = Energy
B Economy
C Foreign Affiars
D Social Order
E = National Defense

306
Media A B C D E Priorities

A 1 1/2 1/4 1/3 3 .12

B 2 1 5 3 5 .39

C 4 1/5 1 1/4 1/4 .11

D 3 1/3 4 1 5 .27

E 1/2 1/5 4 1/5 1 .11

Consistency .35

International
Standing A B C D E Priorities

A 1 6 1/4 5 1/7 .17

B 1/6 1 1/3 5 1/3 .09

C 4 3 1 5 1 .31

D 1/5 1/5 1/5 1 1/5 .04

E 7 3 1 5 1 .39

Consistency .23

Priorities of Carter and Reagan with Respect to the Issues

Ener Carter Rea an Priorities Econom Carter Rea an Priorities

Carter 1 1/3 .25 Carter 1 1/3 .25

Reagan 3 1 .75 Reagan 3 1 .75

Foreign
Affairs Carter Rea an Priorities Social Carter Rea an Priorities

Carter 1 2 .67 Carter 1 3 .75

Reagan 1/2 1 .33 Reagan 1/3 1 .25

Defense Carter Rea an Priorities Composite Weights:

Carter 1 1/5 .17 Carter .35

Reagan .65
Reagan 5 1 .83

307
 The overall or hierarchic consistency for this example is not
very good but iteration with another group produced the same ranking
of the candidates but with different composite weights.

CONCLUSION

The Analytic Hierarchy Process is a useful model for dealing

with complex decision problems. The process has been made avail-
able for the IBM Personal Computer by Decision Support Software of
Arlington, Virginia together with a very readable manual explain-
ing the process to beginners. The AHP is currently being used ex-
tensively in businessm and government and in personal decision
making. For details about the process see the references.

REFERENCES

Saaty, T.L., 1977, A scaling method for priorities in hierarchical

structures, Journal of Math. Psychology, 15: 234-281.
Saaty, T.L., 1980, "The Analytic Hierarchy Process," McGraw Hill,
New York.
Saaty, T.L. and Vargas, L.G., 1981, "The Logic of Priorities,"
Kluwer-Nijhoff, Boston.
Saaty, T.L. 1982, "Decision Making for Leaders," Lifetime Learning
Publications, Belmont, California.
Saaty, T.L., and Aczel, J., 1983, On synthesizing judgments, Journal
of Math. Psychology, (forthcoming).
Saaty, T.L., and L.G. Vargas, 1983, Comparison of eigenvalue, logar-
ithmic least squares and least squares methods in estimating
ratios, Psychometrika, (forthcoming).

308