You are on page 1of 58

From Verbal Scales to

Numerical Scales in
Decision Making

The Fundamental Scale about Absolute


Numbers
1 Equal importance two activities contribute equally to the objective
2 Weak or slight3 Moderate importance-experience and judgment slightly favor one
activity

over another

4 Moderate plus
5 Strong importance-experience and judgment strongly favor one
activity over

another

6 Strong plus
7 Very strong or demonstrated importance -an activity is favored
very

practice

strongly over another; its dominance demonstrated in

8 Very very strong


9 Extreme importance-the evidence favoring one activity over another is
of the

highest possible order of affirmation

Saatys argument for comparative


questions
In addition, when we think of an alternative we
are already conscious of other alternatives and
rate it silently as if we are thinking of it alone,
which is rarely ever the case. In fact, once we
have rated the first alternative, we cannot forget it
when we rate the next one and the one after. All
subsequent ratings depend on what passes
through our minds before
The conclusion is that assuming independence
and rating things one at a time is a convenient
and unscientific way to assign numers subjectively
to alternatives. The numbers are pulled out of a
hat even when an expert is involved

Saatys argument for 1-9 scale


Research in psychology that relates the use of the
fundamental scale of the absolute numbers 1-9 to
represent judgments in making pairwise comparisons of
homogeneous elements with respect to a common
property:
Stanislas Dehaene : The number sense, How the mind
creates mathematics
Introspection suggests that we can mentally represent
the meaning of numbers from 1 to 9 with actual acuity.
Indeed, these symbols seem equivalent to us.In
summary, the invention of numerical symbols should
have freed us from the fuzzines of the quantitative
representation of numbers
Also, integers from 1-9 used in pairwise comparison
judgments can be derived from stimulus response theory.

The first argument against the 1-9


scale:
irregular
changings
in
the
corresponding priority vector

Previous formulas gave rise to an


interesting observation
(see Salo and Hamalainen(1997)).

So, if Saatys numerical scale is used,


with integers from 1 to 9 and their
reciprocals,
to a change of one unit on the scale function
correspond
irregular changes
in the corresponding priority
vectors.

The second argument against the 1-9 scale:


the distribution of the corresponding priority
vector

In the end of this discussion regarding a


(2,2) decision matrix one perhaps would
want to notice the particular convenient
form of the priority vector which
allowed for all this considerations.

When one upgrades to a (3,3) decision


matrix or it is considering instead a
hierarchy, all the previous scheme of
reasoning becomes inapplicable.

The consistency index and the random


index

The random index (RI) is a measure of the range


of the consistency index for matrices of the
same dimensions.
The upper diagonal of a matrix is filled with
randomly extracted elements from the Saatys
scale function {1/9,..1,9},
the elements below the main diagonal are
completed by taking the reciprocals,
on the main diagonal are ones
and
the correspondent CI is calculated, for 50,000
trials.

The correspondent average is the


random index RI.
The consistency ratio (CR) is the ratio
among consistency index and the
random index.
Saaty recommends that the
consistency ratio should always be
smaller than 0.1.

Numerical scales in the literature


This section is revising the existent literature
so
far alternative numerical scales mainly
regarding

based (yet not exclusively) on the work done


by Dong&all (2008), Elliot (2010),Salo and
Hamalainen (1997).

Matching a certain numerical scale to a


broadly accepted linguistic scale raises issues
which are not currently sorted out.

The variety of criterions taken into


consideration when a certain numerical scale
is adopted can shed a light to the implications
of a certain choice over the other alternatives.

This broad range of criterions assessing the quality of a certain


numerical scale led to a more formalized way of looking at the
linguistic variables, in the attempt to impose on these a minimal set of
axioms.
Thus, following Dong & all (2008), if we accept that

linguistic variables
are judgments vaguely quantifications in
respect to certain questions
(like for instance
extremely poor,
very poor ,
poor,
slightly poor,
fair,
slightly good,
good,
very good,
extremely good),

a label
represent a possible way in which
the linguistic variables are ordered.

Thus, for the linguistic variable fair,


one could associate the label 0.

Similarly,
the linguistic variable slightly good
could be associated with the label 1,

the linguistic variable good


could be associated with the label 2
and the linguistic variable poor
could be associated with the label -2.

Therefore, a linguistic label set


with odd number of
elements
can be represented as
S={ s /=-t,-t+1,-1,0,1,..t-1,t}

where
with the label s can be, in principle,

attached any possible real number .

Following the previous example, the linguistic variable


fair corresponds to the linguistic label s 0,

the linguistic variable slightly good corresponds to


the linguistic label s1 and so on,
and so,
instead of considering a set of linguistic variables like

S={extremely poor.. extremely good}

we admit that there must be some order in these


quantification

and consider instead a set of linguistic labels

S={ s /=-t,-t+1,-1,0,1,..t-1,t}
where

t=4,
s-4=extremely poor,,
s0=fair,.,
s4=extremely good.

In the AHP framework for instance, Saaty preestablished the next linguistic label set:

S={s-8,s-7,s-1,s0,s1,s8}

(or, shorter,
S={ s /=-8,-7,-1,0,1,..7,8})
where
s-8=extremely less important,
s-7=very, very strongly less important,..
s-1=weakly less important,
s0=equally important,
s1=weakly more important,
s7=very, very strongly more important,
s8=extremely more important.

In general (see Dong &all), it is required


that the labels s,

as elements in the linguistic label set S,

should satisfy the following characteristics:

> if and only if (iff)


s >s
there is a negation operator:
neg(s)= s-

Given a certain discrete linguistic label set


S={s/=-t,..0,t},

the scale gradation of a certain element


s in S
is its correspondent lower index,

noted with I(s).

More specific, if s= s, then I(s)= .

A scale function associates to every


linguistic label a real number.

Thus, if S is a linguistic label set


(discrete or continuous),

a scale function is a monotonically


increasing function f:SR+ with f(s)f(s)=1 .

As one can notice, the way in which to the


linguistic labels are associated numerical
quantifications is not uniquely determined.

Consider the next example:

S={ s /=-t,-t+1,-1,0,1,..t-1,t}
where t=2,
s-2=extremely poor,
s-1=poor,
s0=fair,
s1=good,
s2=extremely good.

The next alternatives of associating


numerical values to the linguistic label
set could be considered:

f1: SR ,
f1(s-2)=-2,
f1(s-1)=-1,
f1(s0)=0,
f1(s1)=1,
f1(s2)=2

f2: SR ,
f2(s-2)=1,
f2(s-1)=2,
f2(s0)=3,
f2(s1)=4
f2(s2)=5

f3: SR ,
f3(s-2)=1/3,
f3(s-1)=1/2,
f3(s0)=1,
f3(s1)=2,
f3(s2)=3

f4: SR ,
f4(s-2)=1/ c3/2,
f4(s-1)=1/c,
f4(s0)=c1/2,
f4(s1)= c,
f4(s2)= c3/2, c>0.

Function f1 is not a scale function in the sense of


the previous definition since it does not take
positive values.

Function f2 is neither a scale function since the


condition f(s)f(s-)=1 is not checked.

One would perhaps recognize functions f 1 and f2 as


being alternatives of the so called Likert scale, so
in the developing field of linguistic variables and
linguistic preferences relations, this scale is at
least inexistent.

An alternative support for the previous remark can


be found in Saaty (2010):

In a rather haphazard way people have the


habit of using ordinal numbers to create
scales used to assign numbers to anything. In
addition they perform meaningless arithmetic
operation on such numbers that make no
sense. A frequently used example is the
psychometric scale commonly used in
questionnaires the Likert scale named after
Rensis Likert (1932) []It is disturbing to see
such scales used in unjustifiable
mathematical ways. It is not clear what a
person relates such judgments to because
peoples likes and dislikes are so variable
even about the same thing but at different
times.

Functions f3and f4 are both of them scale


functions according to the previous definition,
provided c>1.

If c<1, function f4 is no longer a scale function


since it is not a monotonically increasing
function.

In the happy event that c>1, the next question


arises: what are criterions for deciding which of
the possible scale functions is most appropriate?

If a scale function f: SR+ is considered,


then the condition dijdji=s0 moves into f(dij)f(dji)=1

which is exactly the reciprocity condition for the decision


matrices as introduced by Saaty.

On the other hand,


the consistency condition,
also referred as transitivity condition (dij djk=dik)
is not automatically fulfilled,
for any scale function

(for instance, the usage of scale function f3 does not


assure consistency, while the usage of scale f4 does it).

Considering once again the AHP


linguistic scale as introduced by Saaty,
in the next table will be explicitly
described the most important four
numerical scales.

Next, the correspondent scale


functions, using the scale gradation
will be formally presented.

Linguistic
variable

extremely
less
important
very,very
strongly
less
important
demonstratedely
less important
strongly plus less
important
strongly
less
important
moderately plus
less important
moderately less
important
weakly
less
important
equally
important
weakly
more
important
moderately more
important
moderately plus
more important
strongly
more
important
strongly
plus
more important
demonststratedl
y more important

Linguis Scale
tic
gradation
label
I(s)=

Saaty
scale

Geometrical Ma-Zheng
scale,
scale

Salo-Hamalainen
scale

c>0

e=1/20 or

s-8

-8

1/9

c-4

1/9

e=1/17
(0.5-8e)/( 0.5+8e)

s-7

-7

1/8

c-7/2

2/9

(0.5-7e)/( 0.5+7e)

s-6

-6

1/7

c-3

3/9

(0.5-6e)/( 0.5+6e)

s-5

-5

1/6

c-5/2

4/9

(0.5-5e)/( 0.5+5e)

s-4

-4

1/5

c-2

5/9

(0.5-4e)/( 0.5+4e)

s-3

-3

c-3/2

6/9

(0.5-3e)/( 0.5+3e)

s-2

-2

1/3

c-1

7/9

(0.5-2e)/( 0.5+2e)

s-1

-1

c-1/2

8/9

(0.5-e)/( 0.5+e)

s0

s1

c1/2

9/8

(0.5+e)/( 0.5-e)

s2

9/7

(0.5+2e)/( 0.5-2e)

s3

c3/2

9/6

(0.5+3e)/( 0.5-3e)

s4

c2

9/5

(0.5+4e)/( 0.5-4e)

s5

c5/2

9/4

(0.5+5e)/( 0.5-5e)

s6

c3

9/3

(0.5+6e)/( 0.5-6e)

Recalling that the scale gradation of a


linguistic label s is I(s)= ,

Saaty scale function can be formalized


as follows:
f Saaty (s)=I(s)+1 if ss0
and
f Saaty (s)=1/(1-I(s)) if s< s0

Geometrical scale function can be formalized as follows:


fgeometrical(s)=cI(s)/2 ,c>0

The Ma-Zheng scale function can be formalized as


follows:

f Ma-Zheng (s)=9/(9-I(s)) if ss0 and f Ma-Zheng (s)=(9+I(s))/9


if s< s0

The Salo-Hamalainen scale function can be formalized


as follows:

f(s)=(0.5+I(s)e)/( 0.5-I(s)e), where e=1/20 or e=1/17.

Perhaps the most important thing to know is what the arguments are,
based on which, one decides the best numerical scale to be applied.
In literature, so far, there are three types of arguments. In the next,
these will be presented in general lines.
First argument refers to the so called priority vectors spatial
distribution.

This argument has initially presented in Salo and Hamalainen (1997) at


it basically looks at the dispersion of local priorities as a function of
ratio estimate.

Elliot (2010) is revising this argument in his paper (devoted, as the


previous one, to the introduction of another scale function) by
computing all possible priority vectors for a 22 positive, reciprocal
decision matrix and respectively for a 33 one, when Saatys numerical
scale, Salo-Hamalainen and geometrical scales are considered.

The spatial representation in each of these three numerical


scales reveals the fact that no matter the choices for a 22 or
33 decision matrix, when Saaty scale is applied, there is a
narrow range of priority vectors which can be deduced, while
the most uniform distribution for the priority vectors is
deduced when a geometrical scale is applied.

This can be interpreted as having (or not) sensitivity of the


output with respect to changes in the input, or a uniform range
of possible outputs (as priority vectors) over the 2 or 3dimensional hyper planes.
Not going more into specific details of these two cited paper, it
has to be said that only 22 or 33positive, reciprocal
decision matrices were taken into consideration, while perhaps
considering higher order such matrices could reveal more
information.

Secondly, one can notice that there is an inverse relationship


among the constant increases in the numerical scale function and
the correspondent ones for the priority vectors.

For instance, the Salo-Hamalainen scale belongs to the so-called


balanced scales where it is first imposed for the priority vectors to
increase at a constant rate and depending on this, the
correspondent numerical scale is deduced.
For example, for a 22 decision matrix, the priorities 0.1, 0.15, 0.2
0.9 correspond to the scale 1, 1.22, 1.5, 1.86, 2.33, 3...9.

Again, such a scale function is derived for two-dimensional


decision matrices (where the relation among the matrix and the
priority vector is easy to control) and then extended to decision
matrices of higher dimensions.

Also, under the consideration of a certain hierarchy with criterions


and alternatives in variable amounts, the distribution of the
possible priority vectors can change substantially.

Third observation refers to the consistency condition.

This is imposed by the usage of geometrical, Ma-Zheng or SaloHamalainen scales

and indeed the result will be

priority vectors with the property that the elements in the


decision matrices are precisely the ratio of the correspondent
elements in the priority vector.

Thus, by simply adjusting a numerical scale function,

all the (in) consistency in judgments is ruled out,

making any comparisons among the decision makers,

based on consistency index for example, useless.

The second argument refers to the problem of


minimizing the distance between
the decision matrix reflecting ones judgments

and

the matrix determined so that its every


element is the inverse of the numerical scale
considered,
applied to the ratio of correspondent elements
in the priority vector previously computed.

To be more specific, consider D=(d ij)i,j=1..n a certain positive,


reciprocal LPCM matrix in which every element is d ij= s.
Suppose that to every element d ij a scale function f is
thereafter applied, f(dij),
and as a result a matrix whose elements are positive real
numbers is deduced.
Next, consider w=(wi)i=1,..n the corresponding priority
vector.
Afterwards, the matrix D f=(d-1ij) i,j=1..n is computed so that
its elements are d-1ij=f-1(wi/ wi).
Obviously, if D is not only positive and reciprocal,
but also a consistent matrix, then f(d ij)= wi/ wj,
so that f(s)= wi/ wj , s= f-1(wi/ wj) and in the end dij= s= d1 ,
ij
meaning that
the two matrices D and Df are identical
and therefore the distance between them is zero.

In Saaty(2010) there are presented several examples


(the Distance Problem, the Optics Problem, the Nation
Wealth Problem and the Weight Estimation Problem)
in which the true priority vector is known.

Based on these examples, one could for instance


decide which of the existent, alternative scale
functions is the most appropriate to be used.

Of course, such a decision has nothing in itself to


guarantee that similar choice, on different examples,
will be the best one.

Even if some of these examples were presented in the


previous section, we will briefly recall them,
expressing numerical decision matrices using
linguistic labels. The presentation follows the
Dong&all (2008) paper.

You might also like