The use of brand preference measures for market segmentation


Material Information

The use of brand preference measures for market segmentation
Physical Description:
vii, 210 leaves : ill. ; 28 cm.
O'Connor, Peter Joseph, 1945-
Publication Date:


Subjects / Keywords:
Market segmentation   ( lcsh )
Consumers' preferences   ( lcsh )
Marketing thesis Ph. D   ( lcsh )
Dissertations, Academic -- Marketing -- UF   ( lcsh )
bibliography   ( marcgt )
non-fiction   ( marcgt )


Thesis--University of Florida.
Bibliography: leaves 158-165.
General Note:
General Note:
Statement of Responsibility:
by Peter Joseph O'Connor.

Record Information

Source Institution:
University of Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
aleph - 026384973
oclc - 04165360
System ID:

Full Text








The undertaking of this study could not have been

accomplished without the guidance and helpfulness of a

number of individuals. Special thanks are due to my

chairman, Professor Gordon G. Bechtel, whose patience and

understanding have contributed largely to my intellectual

development. I would also like to thank other members of

my committee, Professors Albert R. Wildt and Marvin E.

Shaw, for their criticisms which enabled me to more

precisely understand the issues involved in this study.

A debt is also owed to Professors William L. Wilkie,

Joel B. Cohen, John R. Faricy, and Ralph B. Thompson for

helping me to conceptualize this research.

Also, a deep expression of gratitude goes out to my

wife, Karon, without whose help this work could never have

been completed.


ACKNOWLEDGEMENTS . . . . . . . . . . ii

ABSTRACT . . . . . . . . . . . . v

CHAPTER I INTRODUCTION . . . . . . . . 1

Purpose . . . . . . . . . . . 3
Organization . . . . . . . . . 9


Geography . . . . . . . . . .. 16
Demography . . . . . . . . . .. 16
Personality . . . . . . . . . .. 21
Psychographics . . . . . . . .. . 25
Benefit Segmentation . . . . . . .. 28

CHAPTER III METHODOLOGY . . . . . . . .. 37

Psychometric Scaling . . . . . . .. 38
Product Classes . . . . . . . . .. 44
Objectives . . . . . . . . . .. 45
Data Collection . . . . . . . . .. 48
Analysis . . . . . . . . . .. 50


Analgesics . . . . . . . . . .. 61
Floor Cleaners . . . . . . . . .. 108


Findings . . . . . . . . . .. 151
Limitations . . . . . . . . .. . 154
Future Research . . . . . . . . .. 156

REFERENCES . . . . . . . ... . . .. . 158




APPENDIX C EIGENVALUES . . . . . . . .. 185


BIOGRAPHICAL SKETCH . . . . . . . . .. 210

Abstract of Dissertation Presented to the
Graduate Council of the University of Florida
in Partial Fulfillment of the Requirements
for the Degree of Doctor of Philosophy



Peter Joseph O'Connor

March 1978

Chairman: Gordon G. Bechtel
Major Department: Marketing

The concept of market segmentation has a firm

theoretical base in microeconomic theory dealing with

imperfect competition and has become one of the most

pervasively discussed topics in the marketing litera-

ture of the past twenty years. Early research in

segmentation focused upon internalized characteristics

of the consumer and tended to ignore situation specific

aspects of the purchase sequence. As a result of this

deficiency much of this research was rather unproduc-

tive in developing predictive ability.

In response to this predictive weakness, an

orientation has developed more recently which focuses

on the interface between the consumer and the product

class. Known as the product stream, it considers

values, needs, and attitudes relevant to the product

under analysis to be the preferred segmentation bases.

Unfortunately, proponents of this approach have not

presented any formalized methodology, and other researchers

have limited their inquiry to the use of measures obtained

from various formulations of expectancy value models.

Such procedures are fraught with problems relating

to product attribute specification and data collection.

This study proposes that brand preference measures provide

an alternative segmentation base which is an unobtrusive

way of obtaining similar information as well as being more

closely related to overt behavior and serving as an input

for multidimensional scaling procedures. A geometric

space representative of the consumer's perception of the

product class which can serve as a guideline for strategy

development can be derived from multidimensional scaling


A survey of housewives from three cities in Florida

was conducted covering two product categories for the

purpose of comparing preference measures with attribute

importance measures for isolating market segments. The

comparison was undertaken to determine which measure lead

to segments which fulfilled stated criteria for meaningful

market segmentation. Specifically, the issues of concern

dealt with segment identification, brand purchasing, and

responsiveness to marketing mix elements.

Results indicated that both measures are equally

useful for identifying segment membership. However, the

preference measures lead to segments which are much more

highly associated with brand choice, although only

marginally more sensitive to marketing mix elements.

These results were more stable and significant for the

analgesic market than for the hard-surface floor cleaner


The geometric product spaces that were derived from

the preference measures were found to be less useful for

strategy development than had been thought. Still, their

validity seems well founded due to their relationship with

brand usage data. While some study objectives were not

achieved, taking the results as a whole indicates that

brand preference measures are useful for understanding the

constructs underlying market segmentation.


The propositions underlying the concept of market

segmentation are that consumers are different, that these

differences are related to differences in market demand,

and that segments of consumers can be isolated within the

overall market (Engel, Fiorillo, and Cayley, 1972, p.2).

The management strategy of targeting the firm's efforts

toward particular segments results from this market

segmentation concept. Thus, it is apparent that market

targeting is a two stage process. First, it requires the

identification of a number of consumer groups which are

distinct from each other in a meaningful way. Second,

elements of the firm's marketing mix need to be modified

in order to provide a marketing program specifically

tailored to one or more of the groups so as to provide for

maximum profitability.

The early published research in market segmentation

focused upon internalized characteristics of the consumer

and tended to ignore situation specific aspects of the

purchase process. As a result, much of this research proved

to be rather unproductive in developing predictive ability.

A more recent orientation has developed in marketing

research which focuses on the interface between the consumer

and the product class. The primary concern of this orienta-

tion is the identification of consumers of particular brands

rather than products (Boyd, Westfall, and Stasch, 1977).

Wilkie (1970) has named this person/product orienta-

tion the "product stream" and it consists of two main

avenues of investigation. First, the focus is on the

importance various product attributes have for brand choice

decisions. Labeled benefit segmentation by Haley (1968),

it considers needs and values relevant to the specific

product class as the preferred segmentation bases. Second,

the focus can be on the perceptions consumers have of the

various brands within the product class.

The product stream approach is managerial in nature

in that it is primarily concerned with brand level choice

behavior. The idea behind the orientation is to identify

target segments of consumers who either desire similar

product attribute configurations or who perceive similar

brand attribute relationships. Research findings can then

be used to create new brands or to position existing brands

so they will differentially appeal to specific segments.

Thus, under this orientation segmentation is associated with

both the person and the product and not merely inherent

characteristics of the consumer.

Although the applicability of brand perceptions as a

segmentation base is accepted by many theoreticians (Boyd,

Westfall, and Stasch, 1977), published reports using the


product stream orientation have all focused on product bene-

fits desired by consumers. These reports have primarily

consisted of post hoc explanations of market success which

have made very favorable claims for benefit segmentation's

usefulness as a segmentation tool.

Unfortunately, the proponents of benefit segmentation

have not presented any formalized procedures, and while

benefit segmentation methodologies of a proprietary nature

that work well in practice undoubtably exist, there has not

been any general agreement among researchers in the field

on the value of alternative procedures. Thus, development

of a product segmentation methodology as an alternative to

benefit segmentation that would provide a framework for the

practitioner as well as conceptual vigor would be beneficial

to the marketing discipline. This is especially true

considering the limitations of current knowledge dealing

with consumer responses.


The worth of a segmentation methodology can be aided

by the use of a theoretical framework for relating the

methodology to specific segmentation criteria. An

individual behavioral framework as viewed from a segmentation

perspective is presented in Figure 1. While not intended to

be a model of consumer behavior, the figure should provide a

simplified framework for relating alternative segmentation

bases to overt behavior.

Figure I



The first box, consumer characteristics, was the focus

of early segmentation research and used such descriptors as

demography and personality. These descriptors make up the

"empirical stream" (Wilkie, 1970) and are discussed in

Chapter II. The next two boxes are the domain of the product

stream. Desired product benefits have been used in a number

of cases to isolate benefit segments who have similar needs

with regard to the product class. As noted previously,

perceptions of brand benefits have been proposed as a segmen-

tation base but as yet no published reports have appeared

using these measures.

Within the attitude literature a good deal of theory

exists indicating that desired benefits and brand perceptions

are combined according to some logical process to form brand

preferences (Fishbein, 1967; Rosenberg, 1956; Cohen and

Ahtola, 1971). Such preferences serve as direct antecedents

of behavior. Therefore, the use of brand preference measures

as the segmentation base has the advantage over benefits

desired and perceptions of embracing information contained in

both of these measures as well as being more closely related

to overt behavior. Another advantage of employing brand

preference is that they can serve as input for certain

preference models, to be discussed below, which provide a

pictoral representation of the consumers' cognitive structure

regarding the product class.

Accordingly, this study proposes that brand preference

measures should be added to the product stream orientation


since there is theoretical justification for linking the

measures to specific behavioral criteria. The purpose of

the study, then, is to compare preference measures against

those used by researchers to create benefit segments as to

their efficacy for isolating market segments. It must be

kept clearly in mind that the goal is not to create unique

or interesting segments, but rather to create segments that

are meaningful in the sense that they provide for specific

targeting strategy. This can be accomplished by careful

delineation of market segmentation criteria that meet this



A number of researchers have advanced criteria for

fulfilling the goals of market segmentation. Kotler (1976,

p. 143) lists "measurability, accessibility, and

substantiality." The triumvirate of "identifiability,

variation in demand, and variation in response to marketing

variables" has been proposed by Frank (1968, p. 136). Wilkie

(1970, p. 13) has suggested "homogeneity within and hetergen-

eity between groups, usefulness as a correlate of behavior,

and efficiency as a target for marketing tools."

In general these criteria have been proposed by

researchers who are primarily interested in behavioral

consequences and thus have focused on such issues as

measurement and behavioral correlation. Using these and


attempting to approach the concept as related specifically

to the product stream, the following three criteria are

proposed as encompassing the requirements needed for

successful segmentation:

1. Identification of homogeneous consumer groups.

2. Brand usage.

3. Sensitivity to marketing mix elements.

It must be clearly understood that each criterion is

a necessary but not sufficient condition for segmentation.

That is, an attempt to maximize a subset of them at the

expense of the remainder will not lead to optimal

segmentation. Each must be considered in relation to the


The identification of homogeneous consumer groups is,

of course, the very essence of segmentation. The develop-

ment of differential marketing programs is dependent upon

the ability to identify distinct consumer groups.

Homogeneity refers to the fact that these groups share a

common property that can be used to discriminate among them.

In the case of benefit segmentation the common property

used to discriminate the groups is directly related to their

desire for similar product benefits, whether they are values

or needs. For brand preference segmentation the common

property deals with consumer's preference for brands within

the product class. The degree to which group members share
the property is a relative one which depends on the measure

used, the situation, and the time frame.


From a managerial perspective, market targeting is

primarily concerned with purchasing at the brand level.

Thus, the second criterion follows directly since segments

should exhibit differential brand usage. Brand usage has

three major dimensions: the brand consumers purchase, their

loyalty to the brand, and the amount they purchase.

Although researchers have tended to ignore this

aspect, sensitivity to marketing mix elements is absolutely

essential to the segmentation process. If segmentation is

to contribute to a profitable managerial strategy, the

crucial consideration must be whether consumers in different

segments respond in different degrees to changes in the

firm's price, promotion, media, copy, and distribution

policies (Frank, Massey, and Wind, 1972, pp. 133-34).

Regardless of how well defined a target segment might be,

if the segment members do not exhibit differential

responsiveness to the firm's marketing decision variables,

it is impossible to design a specific marketing program

tailored for them.

After segmentation has been accomplished, the

development of targeting strategy is of primary importance.

Consideration must then be given to the size and competitive

nature of the market segments identified. It is these two

factors which primarily determine the profitability that can

be achieved by cultivation of a particular market segment.

Profitability, of course, is the ultimate goal of any

strategy employed by the firm.


As previously noted, a firm base is used in this

study to relate segments to the above criteria. That is,

the concern will be with developing a segmentation scheme

for identifying homogeneous consumer groups, and for

relating these groups to their brand usage and sensitivity

to marketing mix elements. Thus, the use of preference

measures will be tested against the benefit scheme for

their efficacy in achieving the three criteria and analyzed

in terms of the implications for strategy they provide.


Since benefit segmentation and brand preference

segmentation are subsets of market segmentation, they

can best be understood as they relate to the entire

segmentation field. Therefore, Chapter II (Historical

Perspective on Segmentation) outlines the entire field

with emphasis given to the various bases used to segment

markets and discusses the major research findings associ-

ated with each.

Chapter III (Methodology) describes the preference

models underlying this study and discusses the specific

methods used to segment the consumers comprising the data

base, the issues to be investigated, and procedures used

to draw inferences.

An analysis of the results obtained and the

conclusions drawn, as well as the meaning these have for


marketing management, is contained in Chapter IV (Results

and Discussion).

Chapter V (Summary and Conclusions) presents an

overview of the study findings and their implications,

as well as some direction for future research.


With the criteria for segmentation outlined in

Chapter I firmly in mind, this chapter will now examine

the history of segmentation research from the perspective

of the various bases that have been advanced for its

implementation. Wilkie (1970) has shown that these bases

fall into two distinct approaches which he named the
"empirical stream" and the "product stream." The former

consists of the main stream of segmentation research and

has focused on enduring personal descriptors of the target

population attempting to relate segments defined in terms

of these descriptors to product class purchasing behavior.

The latter has consisted of both brand perception and

benefit segmentation and has been a more recent innovation.

Since this "stream" is more concerned with the interface

between the consumer and the product, it has dealt with

such variables as product and brand benefits, values,

perceptions, and use occasions. For the reasons outlined

in Chapter I it is proposed that brand perference measures

be added to the "product stream" and a comparison of these

two orientations is presented in Figure 2.

Throughout the history of consumer behavior a

controversy has raged over whether variations in behavior


Empirical Stream




Product Class Purchase

on Bases

Benefits Sought
Brand Perceptions
Brand Preferences

of Interest

Brand Choice


Comparison of the Major Approaches
to Market Segmentation

Product Stream


are due to stable traits residing within the individual or

to the difference in the environmental situations. In point

of fact, neither can be neglected, for even if situational

variations were able to explain behavior completely, the

question of how the individual transforms the situational

input into behavioral output would still be unresolved.

Thus, an understanding of the goals of the product

stream approach to segmentation and the state of knowledge

which brought about its emergence requires a complete

review of the segmentation literature. The various bases

used for segmentation have been advanced in a roughly

chronological order as each has attempted to improve upon

earlier ones. While surveying the field it should be

kept in mind that each base is not necessarily an alterna-

tive method but can be used in a complementary fashion.

Market segmentation received its primary impetus as

a topic of research interest from a pioneering article by

Wendell Smith (1956). Beginning from a firm theoretical

base in microeconomic theory dealing with imperfect

competition, he discussed the heterogeneity present among

the components of both market demand and supply. He

outlined two strategies available to the firm to deal with

this situation. One of the strategies:

. may consist of a program
designed to bring about the
convergence of individual
market demands for a variety
of products upon a single or
limited offering to the market.
This is often accomplished by

the achievement of product
differentiation through
advertising and promotion.
(Smith, 1956, p. 30)

Alternatively, under a strategy of market segmentation

it is:

S. better to accept diver-
gent demand as a market char-
acteristic and to adjust pro-
duct lines and marketing
strategy accordingly. This
implies ability to merchandise
to a heterogeneous market by
emphasizing the precision with
which a firm's products can
satisfy the requirements of
one or more distinguishable
market segments. (Smith, 1956,
p. 30)

It was his position that the latter strategy is

preferable since it evolves from variations on the demand

side of the market and is a more precise adjustment of

marketing effort to consumer needs. Due to this optimiza-

tion of consumer satisfaction, segmentation provides for a

stable market position with its attendant increased


On the other hand, Reynolds (1965, p. 111) argues

that "market segments are more apparent than real." It

is his contention that different products do not appeal

to different segments but rather succeed because of the

large number of switchers among the consumers comprising

the "uniformity of the general American culture."

Espousing the view that variety is the key element in

describing consumer purchases he states:


. One of the many charac-
teristics people have in common
is that they shift from one
brand to another more or less
frequently. Another is that
they tend to be attracted to
new brands and new products.
(Reynolds, 1965, p. 107)

In a similar vein, Evans urged markets to build

ambiguity into their products to avoid limiting their

markets. Given a product with an ambiguous image, he

believed that "there is a tendency for customers to read

into it what they want. The things they value highly

they attribute to their brand" (Evans, 1961, p. 366).

The above is an indication of the controversies

surrounding the segmentation issue. Other controversies

have arisen in that most of the early segmentation research

was rather unproductive in developing predictive ability.

In their zest to obtain correlations between purchase

behavior and inherent characteristics they have, for the

most part, ignored situation specific aspects of the pur-

chase sequence. The assumption has been that internalized

attributes of the individual are constant across time and

environment, and that they will lead to consistent behavior

responses regardless of any other external influences acting

upon the decision. In defense of this position, De Bruicker

(1973, p. 1) has pointed out that "measurement problems

prevented the use of any but the most basic of personal

descriptors in early segmentation analyses."



Historically, business enterprises first segmented

markets geographically (Haley, 1968). While geographical

areas are the bases of differentiated marketing effort

here, they are usually not due to any conscious effort to

appeal to diverse consumer groups. Rather it results from

either limited resources precluding national distribution

or in response to specific product usages. Obviously then,

while geography is an important concern for the firm it is

of little consequence for the researcher.


Demographics have been the most widely used procedure

for segmenting large markets into smaller groups. A

profile of the consumer type in each group can be enunciated

using primarily demographic variables. The usual focus is

on such characteristics as income, age, social class,

marital status, race, occupation, and the like. The usual

procedure is to use these characteristics as discriminators

of consumer choices regarding product usage and brand


Such procedures have been widely used for media

selection since it is assumed that there is a link between

target groups so identified and responsiveness to communica-

tion messages. Such an approach has value and provides an


efficient use of demographic profile building. However,

these variables have been shown to be incapable of

differentiating consumer choices among a product class.

Their only success seems to be in discriminating between

various product class.

During the decade of the 1960's a number of studies

were published searching for correlations between demogra-

phically defined segments and such behavioral variables as

average purchase rate and brand loyalty. To mention some

of the more noteworthy, Frank, Massey, and Boyd (1967)

used the Chicago Tribune panel data for 1960 and 1961.

The data consisted of purchase history for 491 households

for each of 57 grocery product categories. The average

proportion of the variance explained was .11 using 14

demographic variables as predictors. Rice cereals provided

the highest proportion at .29, while the proportion was

less than .20 for 46 of the 57 categories and less than

.10 for 25 of the 57 categories.

The J. Walter Thompson panel data were reported on

by Koponen (1960). He added personality measures to the

demographic variables for two unidentified products and

obtained .13 and .06 as the respected proportion of the

variance explained. Using the same panel Hildegaard and

Kruegar (1964) studied the toilet tissue purchasing

behavior of 3206 households for the Advertising Research

Foundation. They obtained a coefficient of determination

of .12 for one-ply tissue and .06 for two ply tissue.

More recently, Gensch and Ranganathan (1974) used the

Brand Rating Index data to study the purchasing behavior of

1550 consumers for six female grooming products and four

grocery products. The product usage data were regressed

against thirteen demographic variables and the proportion

of the variance explained ranged from .02 to .26 with a

mean of .09.

These studies were carried out within a regression
framework and the results, in terms of R were largely

disappointing. In a survey of consumer behavior research,

Ferber reported that cross-sectional analyses of household

expenditures on food, clothing, shelter, and services


. that these consumer purchases
are influenced by a wide variety
of socioeconomic characteristics;
but, nevertheless, the proportion
of variance in individual household
purchases explained by these
numerous factors is small, often on
the order of .3 or less. (Ferber,
1962, p. 62)
The consistently low levels of R achieved by using

demographic variables as predictors lead to the conclusion

that their usefulness as segmentation bases is severely

limited at best. In reviewing the segmentation literature

Frank concluded that:

. for the most part socio-
economic characteristics are
not particularly effective
bases for segmentation whether
in terms of their association
with household differences in
average purchase rate or in


response to promotion . In
spite of the reliability that
households exhibit with regard
to brand loyalty, research
efforts aimed at identifying
the brand-loyal customer have
been notably unsuccessful.
(Frank, 1968, p. 152)

In rebuttal Bass, Tigert, and Lonsdale (1968) have

supported the feasibility of defining segments with

demographic measures. While not disputing the evidence,

they argue that it has been misinterpreted. It is their

contention that the regression model is inconsistent with
the purposes of segmentation and the low R 's are

indicative of high within group variance while differences

in group means may still exist. They state:

The absense of a satisfactory
theory of individual behavior
does not necessarily imply the
absence of valid propositions
about the groups' behavior.
For marketing strategy, it is
the behavior of groups, not
persons, that is primarily
important. (Bass, Tigert, and
Lonsdale, 1968, p. 274)
Morrison (1973) agrees that R is an inappropriate

statistic to use in evaluating segmentation results.

Through the use of probability models he indicated that:

it is perfectly consistent
to have models which correctly
identify variables that are
useful for segmentation purposes
and are still virtually useless
in predicting individual
purchases. (M1orrison, 1973, p.

It would seem though, that inability to predict

purchase rates because of high within group variances


when there are differences in group means results from

considerable overlapping of the distributions of group

purchase rates. Under such conditions segmentation could

be improved by redefining the groups so as to minimize

such overlap.

Twedt (1964) has answered critics of the regression

approach by proposing that categorization of the dependent

variable into a dichotomous one representing light and

heavy users provides for more meaningful segmentation. He

suggests that direct measurement of product usage is

likely to be more efficient than regression analysis and

profile matching. This contention is based on the fact that

in most product classes 50% of the users account for in

excess of 80% of the purchases. While such an approach

has merit there are two major weaknesses in it. First, not

all heavy users are available as targets for a particular

brand since they may not differ in other respects. Second,

focusing on the heavy users eliminates an effort to turn

nonusers into users and light users into heavy users. It

may well be that such an effort would result in profitable


Thus, it is apparent that there is controversy over

the usefulness of demographic variables serving as bases for

segmentation. Regardless of the position espoused, the

results have been far from impressive, most likely due to

the fact that such an analysis investigates only direct

links between specific characteristics and overt action.


There is no attempt made to understand the mediating

responses and influences that are a part of the behavior.

Without such an understanding, it seems only reasonable

that predictive power is limited.


Partly in response to the limitations of demography

researchers turned to personality to provide richer

descriptions of consumer types. Personality is one of the

more absorbing concepts used in applied marketing research.

Virtually every component of consumer behavior has, at one

time or another, been linked to personality. Unfortunately

there has not been any general consensus among researchers

on the meaning of the term personality.

After considering the communalities of a multitude of

personality definitions, about all that can be said is that

personality is the sum total of all the internalized

attributes of the individual which result in consistent

responses to the surrounding stimulus environment. Of

primary importance in determining personality are

constitutional, group membership, role, and situational

influences (Lawrence and Seiler, 1965).

The typical approach here has been an attempt to find

a relationship between batteries of personality tests and a

myriad of product purchases. Such an approach has been

variously referred to as "shot-in-the-dark" (Wells, 1975),

or "shotgun" (Kassarjian, 1971). The problem with the

personality traits comprising the test batteries is that

most of them are theoretically irrelevant to the product

under consideration (Goldberg, 1976).

Perhaps the most widely debated study correlating

product usage has been Evans' (1969) use of the Edwards

Personal Preference Schedule to discriminate between Ford

and Chevrolet owners. Evans' failure to find significant

differences sparked a number of rejoinders and reanalyses

during the early 1960's (Steiner, 1961; Winick, 1961;

Evans, 1961; Kuehn, 1963; Evans and Roberts, 1963).

While some other investigations of these data found some

significant differences between Ford and Chevrolet owners,

the results were far from impressive.

The Edwards Personal Preference Schedule has been

used in a number of other studies including Koponen's

(1960) analysis of the J. Walter Thompson panel data.

He indicated that cigarette smoking was linked positively

to sex dominance, aggression, and achievement needs but

negatively to order and compliance needs. Claycamp (1965)

used this instrument to discriminate between customers of

banks versus savings and loan associations.

The Thurstone Temperament Schedule was used by

Westfall (1962) to compare personalities of automobile

owners and he was unable to find differences between

brands as well as between compact and standard size owners.

Noting the number of items left unanswered by subjects on


this instrument, Kamen (1964) concluded that opinion

proneness is not related to food preferences.

The California Personality Inventory has been used to

develop measures of innovativeness and opinion leadership

in the areas of food, clothing, and appliances (Robertson

and Myers, 1969, 1970; Bruce and Witt, 1970). Boone

(1970) tried to relate these variables to the adoption of

a community antenna television system.

The Gordon Personal Profile provides for measures of

ascendancy, responsibility, emotional stability, and

sociability. One or more of these variables were found

to be weakly correlated with the use of analgesics,

vitamins, mouthwash, chewing gum, and alcoholic beverages

by Tucker and Painter (1961). Using canonical correlation

Kernan (1968) found relationships between the personality

variables and sets of decision behaviors even though

individual correlations were insignificant.

Considering the weaknesses of personality measures

as predictors of consumer behavior, it must be remembered

that these instruments were originally developed to tap

major personality traits underlying racial prejudice,

marital incompatibility, proneness to commit suicide, and

the like. The prediction of an individual's brand choice

is a vastly different matter. Items used to measure

psychological stability or imbalance are obviously not

particularly germane to everyday marketing activities

(Dhalla and Mahatoo, 1976).


In reviewing the use to which personality measures

have been used, Kassarjian (1971) also notes that the

reason for the low amount of variability explained may

be due to the purpose for which these tests have been

validated. Pointing out the need to rectify this

situation he states:

Clearly, if unequivocal results
are to emerge, consumer behavior
researchers must develop their
own definitions and design
their own instruments to measure
the personality variables that
go into the purchase decision
rather than using tools designed
as part of a medical model to
measure schizophrenia? or mental
stability. (Kassarjian, 1971,
p. 415)

On the other hand, Wells (1966) calls into question

the fact that people have personality traits at all. It

is his contention that situational influences are so

important that most often the environment has greater

impact on behavior than anything which can be conceived

of as a personality trait. He suggests that the basic

idea asserting that personality traits exist needs to be

reexamined. In summary he states:

. that the environment is so
important in determining consumer
behavior and the problems of
measuring personality traits
are so difficult to handle in
consumer studies that personality
tests may never have much value
in consumer work if by value
you mean correlations high
enough to predict behavior
accurately. (Wells, 1966, p.


On the whole then, when used on their own as segmen-

tation bases, personality traits have proven to be very

disappointing. At best personality is more likely to be

associated with behavior towards a set of products,

representing a pattern of behavior, than to any one product

in particular (Goldberg, 1976). However, they may still be

a useful adjunct to demographic data in describing a priori

defined consumer groups. Specific to media selection, the

availability of personality profiles of existing audiences

would seem to offer an opportunity for message segmentation.


The need for psychological measures which have been

developed with consumption in mind as well as the integra-

tion of personality theories with the statistical

sophistication of computer technology has led to an approach

called psychographics, or lifestyle segmentation. It has

its roots in the empirical reality that for some products,

purchase is more dependent on the individual's stage in the

life cycle than on personality.

The concept of lifestyle and its relationship to

marketing was introduced by Lazer, who referred to it as:

. a systems concept. It
refers to a distinctive mode
of living in its aggregate and
broadest sense . It embodies
the patterns that develop and
emerge from the dynamics of
living in a society. (Lazer,
1963, p. 141)


Since then, the methods for measuring lifestyle

patterns and the determining of psychographic segments has

been refined. The usual procedure is to divide the market

into segments by means of paper and pencil inventories of

activities, interests, opinions, values, attitudes and

the like using techniques of numerical taxonomy and factor

analysis. Unfortunately though, the segments uncovered by

these methods, when correlated with actual purchase

behavior, account for only 10% or less of the variance

(Kassarjian, 1971).

It must be pointed out however, that these techniques,

while extremely appealing, lack a conceptual framework and

are marked by the absence of theoretical foundations.

Thus, the possibility of obtaining a spurious structure is

so high and results so inconsistent that it suggests

statements made about psychographic segments are on very

weak ground (Kinnear and Taylor, 1976). Yet they seem to

show some value and thus, their integration into a consis-

tent theoretical frame should be of immediate importance.

Closely related to this lack of a conceptual framework

are the issues of validity and reliability. These are

particularly salient due to the widespread use of "homemade"

psychographic items and scales in a number of studies

(Wells, 1975). Commenting on that issue he says:

In view of the reservations
already expressed as to the
reliability and validity of
segmentation procedures it
might seem that the wisest
course would be to ignore


psychographics until the pro-
cedures have been thoroughly
validated. But again, one
must consider the alternatives.
Marketers know the customers
for a product or a service are
frequently not much alike.
They know that empirical
segmentation procedures hold
out the possibility of new
insights into how consumers
may be divided into groups.
And they know that the
reliability and validity of
segmentation procedures have
not been established beyond
all doubt. Given that dilemma,
many marketers have elected to
conduct and to use segmentation
studies even when fully aware
of the art's imperfections.
(Wells, 1975, p. 208)

The degree of specificity inherent in psychographic

measures has also received a caveat from Wilkie and Cohen

(1976). They point out that general psychographic measures

meant to be applicable across product categories can not

also discriminate among product characteristics. Thus,

there can be hardly any basis for between brand analysis.

This problem has stimulated Dhalla and Mahatoo (1976) to

propose that psychographic measures are meaningful only

when they are situation specific and not of a generalized

nature. The danger here is that when carried to the

extreme, the search for predictive validity can lead to

measures which are redundant of the behavior under consid-

eration. What is needed is to find a middle ground between

the extremes of measuring a basic underlying orientation,

unlikely to have predictive validity, and measuring a

situation specific orientation unlikely to have construct


validity (Goldberg, 1976). Obviously, this is a formidable

task for researchers to handle.

Benefit Segmentation

All of the above segmentation bases fall under the

rubric of the "empirical stream" (Wilkie, 1970) in that

they have been concerned with personal descriptors of the

consumer. In recent years as the need for more situation

specific measures has become apparent, an orientation has

developed which focuses more on the interface between the

consumer and the product. Wilkie (1970) has dubbed this

approach the "product stream" since its orientation is more

managerial in nature in that it is primarily concerned with

brand differences within the product class. Thus, this

orientation tends to ignore product class issues and

concentrates instead on brand positioning and image.

Under this approach, demographic and personality

descriptors are demoted to the role of ancillary measures.

Indeed, Yankelovich (1964) feels that the old assumption

that demography is the best way of partitioning markets

needs to be discarded. Reynolds (1965) places more emphasis

on the product than on the consumer and asserts that needs

for variety and specific uses result in the same individual

purchasing different brands at different times.

As previously noted, reports of research using consumer

perceptions of brands have been virtually non existent. In


the main this stream has been concerned with the measurement

of consumer's desire for various product benefits.

The bases used for benefit segmentation have been the

values, needs, and attributes relevant to the product under

analysis. The concept was first discussed in the marketing

literature by Haley (1968), who introduced it as a panacea

for the shortcomings of earlier segmentation variables. In

regard to geography, demography, and purchase volume he


However, each of these three
systems of segmentation is
handicapped by an underlying
disadvantage inherent in its
nature. All are based on an
ex-post facto analysis of
the kinds of people who make
up various segments of a
market. They rely on -
descriptive factors rather
than casual factors. For
this reason they are not
efficient predictors of future
buying behavior that is of
central interest to marketers.
(Haley, 1968, p. 198)

Unfortunately, proponents of this approach have not

presented any formalized methodology or procedures. As

a matter of fact, Haley (1968) has even stated that

successful segmentation can result from some form of

esoteric intuition. This has been the problem with this

entire avenue of investigation. It consists primarily of

interesting findings and post hoc presentations of

marketing successes attributed to the process. While this

state of affairs is understandable due to the proprietary

nature of the studies undertaken, researchers interested

in validating the usefulness of benefit segmentation have

been forced to devise their own methods for partitioning

of the market and testing their efficacy (Wilkie, 1970;

Mitchell, 1974). Thus, there is great difficulty in

adequately determining the value of the approach vis-a-vis

more traditional approaches.

Market structure analysis is a related technique

developed by Volney Stefflre from a background in

psycholinquistics and anthropology (Barnett, 1969). This

technique deals with ratings of consumer preferences for

descriptions of proposed new products. When such a

description is preferred over current brands, it is

proposed that a market potential exists which can be

capitalized on by developing the product so consumers

perceive it as matching its description. Barnett (1969)

indicates that this approach has been successful in

introducing new brands but does not specifically present

quantitative results.

Reynolds (1965) disagrees with this notion believing

that segments tend to be somewhat artifactual, especially

those defined a posteriori in terms of the product

purchased. Given merely that two products seem to be

different, he contends that it is an error to conclude

that the consumers who purchase them must also be different

and comprise separate market segments. It may well be that

they appeal to the same general market's need for variety

(Reynolds, 1965).


The fact that consumers may not differ on variables

other than their purchase behavior does not diminish the

usefulness of the segmentation concept according to Haley

(1971). The real payoff of segmenting markets, he asserts,

has to do with copy efficiencies which can increase the

impact of advertising by as much as three-fold. Explaining

his position, he states:

In short, benefit segmentation
is a tool for improving your
communications with the group
or groups of consumers selected
as the market target by
selecting themes which improve
your chance of capturing the
attention of your prospects and
of involving them in your
advertising. (Haley, 1971, p. 3)

While benefit segmentation has been primarily under-

taken in a proprietary setting, the concept has interested

a number of academic researchers who have used similar

formulations to form segments. In general, benefit segmen-

tation studies have been carried out by measuring consumers'

beliefs about the attributes of a product and then cluster-

ing them on the basis of the relative importance of each in

contributing toward satisfaction of their needs and desires.

Some studies have used determinant attribute scores

(Anderson, Cox, and Fulcher, 1976), or attribute affect

scores (Mitchell, 1974). The list of attributes employed

may include redundancies which are either real or perceived

as real by consumers. The typical procedure is to use a

large set of attributes (Calantone, 1976).

The belief underlying benefit segmentation is that

consumers who desire similar attribute, or benefit,

bundles will purchase the same brand. That brand is the

one that they perceive as most adequately fulfilling their

needs due to the congruence of its attributes. The

benefits sought by consumers are presumed to be stable

over time in a manner similar to the way demographics and

personality exhibit stability (Wilkie, 1970). Also, there

is some evidence that differences in benefits sought by

consumers exist across product classes (Haley, 1968).

While most benefit segmentation results have not been

subject to critical evaluation, they have sparked a trend

toward the use of product specific attitude measures

employed in a more rigorously demanding framework. These

have not yet provided the hoped for degree of success.

De Bruicker (1973) was unable to discriminate between

consumers' brand preferences using perceived product

benefits, attitude, and activity measures as alternative

bases for segmentation. Wilkie (1970) found that brand

purchasing behavior was predicted much better by behavioral

intention scores than by segmenting consumers on the basis

of the importance they attach to product attributes.

These results may in part be due to the fact that

relying on natural clusters runs the risk of achieving

spurious segments, since most clustering algorithms are

data reducing tools devoid of theoretical grounds for

specifying what segments should comprise (Dhalla and


Mahatoo, 1976). However, it still seems obvious that

market segmentation's focus upon attitude domain measures

will provide keen insight into the mediating processes

involved in purchase behavior.

Given that benefit segmentation studies have used

consumers' importance ratings for various product attri-

butes, there are two broad issues involved in determine

the attributes to be measured. The first, initial spec-

ification, concerns the manner in which a proposed list

of product attributes, encompassing perceptual as well as

objective characteristics, is generated by the investiga-

tor. The usual procedure has been to use some type of

unstructured pretest to generate attribute lists.

Given this attribute list, inclusion in the model

deals with the way the researcher uses this data. The

question is, are all of the generated attributes

applicable to all respondents? The typical study has

ignored the possibility of idiosyncratic attributes and

used them all for every respondent, assuming that all

consumers have the same perceptual structure.

By ignoring differences in the perceptions of

individuals, these studies seem to imply that product

attributes are product specific and large in number.

Most all would agree, however, that only salient attributes

should be included. In most papers an a priori assumption

is made that this is the case.

This issue has been investigated by Wilkie and

Weinreich (1973), who hypothesize that improvements can be

achieved when individual differences are allowed in the

number and type of attributes used. The main concern then

was with the criteria used for attribute inclusion. These

researchers focused on attribute determinism as proposed

by Myers and Alpert (1968).

Determinant attributes are those which are directly

involved in brand preference or actual purchase behavior.

When important product characteristics are approximately

equally represented in the set of brands under considera-

tion, then such an attribute is not particularly salient

to the choice of any one brand. Thus, determinism incor-

porates both the idea of importance and variation in the

evaluation of brands on an attribute.

The results of Wilkie and Weinreich's (1973)

investigation indicate that attitudes can be efficiently

described in predictive terms with fewer attributes than

are typically gathered in marketing research. They also

conclude that incorporation of only those attributes

salient to the individual leads to significantly better

results than inclusion of all available ones, and that

the practice of using all attributes for every respondent

is likely to understate the predictive power of an atti-

tude model.

Resolving the two problems alluded to above,

attribute specification and inclusion, for each individual


would require the researcher to gather a larger quantity

of as well as more accurate information from each subject.

The difficulty is that as data gathering instruments

increase in complexity the validity of the subjects'

responses goes down.

One reason for this is that many individuals in

general are not aware of their beliefs and attitudes and,

thus, cannot accurately report them. From a marketing

perspective, Greenberg and Green have commented:

In the investigation of a
product category, it is
frequently misleading to depend
upon consumers to "truthfully"
tell us what are the most
salient dimensions or which
product attributes are the
primary determinants in their
purchase decisions. (Greenberg
and Green, 1969, p. 52)

Also, for the purpose of responding to simple written

questionnaires, it has been estimated that at least 10% of

the adult population in this country is illiterate. The

percentage is much higher for complex questionnaires

(Selltiz, Wrightsman, and Cook, 1976, p. 296). While this

percentage may seem high at first glance, it may well be an

underestimate in light of a recent report released by the

Comptroller General of the United States (1975). According

to the report in excess of 23 million American adults are

functionally illiterate. Specifically, 39 million adults

cannot compute their paychecks if they include overtime;

48 million adults cannot determine the correct amount of

change from a purchase; 52 million adults cannot read


help wanted advertisements; 86 million adults cannot

compute the gasoline consumption rate of a car; and so on.

In light of the above statistics, it is no wonder

that research employing self report techniques is fraught

with error. In regard to benefit segmentation specifically,

the problem is that attribute elicitation techniques are not

very useful for tapping a consumer's cognitive domain.

Considering all of the research conducted on segmenta-

tion, there is still a considerable need for improvement in

isolating segments. It is believed that many of the

existing problems can be overcome by the use of brand

preference measures. Preference measures are more closely

related to over behavior and thus more relevant to the pur-

chase process. Also, they are more readily obtainable and

not subject to the same type of errors as are attribute

elicitation techniques.


In light of the discussion presented in Chapter III

concerning segmentation, it is apparent that previous

research efforts aimed at isolating market segments have

been less than successful. This is especially true for

the earlier segmentation bases since the constructs they

measure are so far removed from actual purchasing behavior.

The product stream, on the other hand, is concerned

directly with consumers' motivations in making purchase

decisions and is more closely related to overt behavior.

However, problems in measurement have prevented researchers

(Wilkie, 1970; Mitchell, 1974; Anderson, Cox, and Fulcher,

1976) from achieving the promise of the product stream


Some of these measurement problems can be eliminated

by the use of certain kinds of psychometric procedures,

which present a more unobtrusive method for achieving

similar results. By focusing upon brand preferences, these

psychometric procedures lead to segmentation results from

which the attributes of the product class can be inferred

with greater confidence.

Psychometric Scaling

A number of different techniques and algorithms

have been developed by psychometricians for locating a

complex stimulus along a number of dimensions simultan-

eously. These various techniques have been grouped under

the name multidimensional scaling (MDS) and are concerned

with two basic types of input data, either proximity

measures or dominance measures (Coombs, 1950). Both of

these approaches consist of a body of techniques for

locating stimulus objects in a geometric space of minimum

dimensionality which is representative of the respondent's

cognitive structure.

A brand or product can be considered as composed

of both objective and perceived attributes or dimensions.

Such perceived attribute spaces consist of brand positions

in relation to dimensions reflecting the consumer's

perception of the product class. This perceptual space

may vary over individual perceivers as well as over time

and context within the same individual. Also, the

dimensionality of the space may vary over stimulus object


Nonetheless, for the purpose of consumer perception,

the derived dimensions are the relevant ones. That is,

the dimensions that are recovered are ones upon which the

brands can be discriminated, and other dimensions on which

the brands do not differ are not part of the solution.


Thus, MDS fulfills the requirements of deterministic

attributes (Myers and Alpert, 1968) far more efficiently

than attribute elicitation techniques. Such solution

efficiencies are another advantage of MDS procedures

over the attribute elicitation approach besides the

decreased complexity of data gathering.

This study will deal with dominance data which

provide a spatial configuration of the stimulus objects

as well as a position for the individual generating the

data. The axes' projections of this position are the

individual's weights for the attributes underlying the

product class. Thus the attributes as well as the

importance an individual attaches to each for a product

category are recovered.

Dominance Data

Under the domain of psychophysical measurement,

dominance refers to the fact that one object possesses

more of an attribute than does another. Thus, it can

be said that one object is larger, louder, brighter,

heavier, and so on; in short, it dominates the other on

the characteristic of interest. In psychological

processes though, dominance has primarily come to mean

preference. Thus, there has been a development within

psychometrics wherein the interest has been in uncovering

the dimensions underlying preference.

It seems entirely reasonable that there should be a

relationship between perception and preference. When

asked to judge which object is preferred among a set, that

judgment should be made in light of the dimensions which

are used to discriminate the objects. The dimensions may

be differentially important in judging preference, but the

same basic dimensions should be used. Some support for

this assumption has indeed been found (Green and Carmone,

1969; Carroll, 1972).

The dominance data concern pairs of objects which

provide more evidence for constructing scales than when

judgments are made on sets of more than two objects

(Slater, 1960) Specifically, the element Pi(jk) is the

degree to which stimulus j is judged dominant over, or

preferred to, stimulus k by individual i. A linear

transformation of these preference strengths, the Pi(jk)'

into descriptive scale values has been presented by Bechtel

(1976). These scale values are least squares estimates of

the parameters contained in the following orthogonal anal-

ysis of variance model:

P. 5.. .- Si + U- + e-
i(jk) = Sij ik + Ujk + eijk


Pi(jk) = individual i's preference for object j over k

Sij = individual i's scale value for object j
Sik = individual i's scale value for object j
Sik = individual i's scale value for object k
Ujk = unscalability associated with the pair j, k

eijk = random error

The individual scale values are tantamount to

Thurstonean scale values (cf. Torgerson, 1958, pp.

169-173). The similarity between this model and the more

familiar analysis of variance model is seen more clearly

when one considers that Sij and Sik are analogous to main

effects and that Uijk is analogous to an interaction term.

Detailed procedures are outlined by Bechtel (1976) not

only for estimating the parameters but also for determin-

ing their significance. That is, the sum of squares

associated with each effect can be obtained and used to

derive independent F tests of hypotheses of equality among

various sets of parameters.

Besides the analysis of variance tests of signifi-

cance, Bechtel (1976) also shows how indications of the

reliability and validity of the scale values can be

determined. In the statistical sense reliability refers

to freedom from random error (Wells, 1975), and therefore

the correlation between the P i(jk)'s and the first three

parameters in the above model are indicative of the

reliability of the data. In a similar fashion, since the

unscalability stems from systematic error the correlation

between the P i(jk)'s and the differences in the scale

values indicate the degree of validity inherent in the data.

Obviously, the validity coefficient has an upper bound at

the reliability coefficient, this correspondence occurs

when all of the unscalabilities are equal to zero.

The multidimensional model that has been advanced to

deal with individual preference that is directly analogous

to some attitude models (Rosenberg, 1956; Fishbein, 1967)

is the vector model (Tucker, 1960; Slater, 1960). The

vector model proposes that an individual's scale value for

an object is the scalar product of the object's coordinates

in the space and a vector of weights describing the

differential importance of each dimension (plus some

random error). That is:

S.. = 1 0. I. = e..
i3 r=i 3r ir 13


S.-. = individual i's scale value for object j
0. = the projection of object j on dimension r
ir = the importance of dimension r for person i

e. = random error

t = the number of dimensions

Inspection of the vector model indicates that it

assumes preferences change monotonically with increments

in the objects' position on the dimensions. That is, the

more of an attribute any object possesses the higher its

scale value. For those dimensions weighted negatively

by an individual, the less of an attribute the object

possesses the higher the scale value. The analogy with

attitudes is apparent by noting that in reviewing the

literature on multiattribute attitude models, Wilkie and

Pessemier (1973) found that fully 21 of the 42 studies

used preference as a surrogate for attitude.

The present study then is concerned with the

development of a segmentation scheme using preference

data obtained in a paired comparison format and comparing

this procedure with benefit segmentation. The paired

comparison format provides the best evidence for the

internal consistency of a scale (Slater, 1960). The

use of preference data adds a motivational flavor to the

representation of an individuals cognitive structure

in line with the goals of the product stream and is

more relevant for product choice processes (Wilkie and

Pessemier, 1973).

In one sense, the brand preference segmentation

scheme used here and the benefit segmentation procedure

usually employed by researchers are related in that

they are both concerned with identifying market segments

that have similar behavioral patterns with respect to

brand purchasing. Also, both deal with the importance

individuals have for the attributes of a product class.

In essence, the difference between the two procedures

is really an issue of measurement. Researchers inter-

ested in benefit segmentation have focused on importance

weights as the applicable measure. Specifically,

Mitchell (1974) found that benefit segments based upon

importance scores were more useful and interpretable

than segments based upon affective scores.


Product Classes

Granted that segmentation is a useful management tool,

it is not meant to be implied that segmentation is equally

applicable to all products and markets. Indeed, many

industrial and government markets which rely heavily on

personal selling deal with such specifically defined markets

that segmentation as a research strategy is superfluous. It

is in the consumer package goods area that segmentation

would seem to have its most advantageous use. In fact, this

area has accounted for almost all of the segmentation

literature since there is considerable uncertainty as to the

optimum strategy to follow in marketing consumer package


This uncertainty results from the characteristics of

hetergeneous mass markets, frequent purchase, low price, and

aggressive competition (Wilkie, 1970). Therefore, this

study will be concerned with delineating segments for

consumer package goods, although the techniques may well

apply in other areas as well. Specifically, the study will

deal with two product classes, analgesics and hard-surface

floor cleaners. The analgesic class includes the following

brands: Bayer, Bufferin, Anacin, Excedrin, Empirin, B.C.,

Tylenol, and Datril. The hard-surface floor cleaning class

includes: Aero Wax, Glo Coat, Klear, Future, Step Saver,

Mop & Glo, and Floor Shine.

These two product categories were selected for analy-

sis for the following reasons. First, it is believed that

housewives enjoy a certain amount of autonomy in choosing

these products and are not limited by the influence of

other family members. Thus, by employing psychometric

procedures the resulting configuration should directly

tap their cognitive domain. Secondly, the limited number

of brands named above account for virtually the entire

product category whereas many others meeting the first

reason consist of a myriad number of brands. For instance,

recent sales figures for the analgesic market show Tylenol

to be the sales leader with 15.9% of the market; followed

by Anacin with 13.7%, Bayer with 10.7%, Bufferin with 9.9%,

Excedrin with 7.3%, and Datril with 1.8% (Giges, 1977).

Thus, just these six brands alone account for 60% of the

analgesic sales in this country. Therefore, the resulting

product space configuration should include all of the

respondents cognitive structure concerning the product

class while requiring only a limited amount of input.


The objectives of the study can be subsumed under

three main areas of inquiry, which have been derived from

the criteria for meaningful segmentation proposed in

Chapter I,


In review, these criteria are:

1. Identification of homogeneous consumer groups.

2. Brand usage.

3. Sensitivity to marketing mix elements.

In practice, segmentation researchers have begun by

isolating groups on the basis of one of the criteria and

then testing to see if these groups also fulfill the other

criteria. Early in the history of segmentation research

groups were isolated with respect to the first criterion

so as to be homogeneous in their demographic or personality

characteristics and then examined for their fulfillment of

the latter two. More recently there has been a tendency to

start by isolating groups that have similar purchasing

proclivities thereby fulfilling the second criterion, and

then focusing on the two remaining ones. The point being,

that efficiency is best served by beginning with segments

that fulfill one of the criteria by definition and then

examining them for their performance on the remaining ones.

Thus study will begin with the first criteria and

identify homogeneous consumer groups. The benefit

segmentation procedure will start with groups that are

homogeneous with respect to their profile of importance

ratings and the product segmentation will begin with

groups that have similar scale value profiles. These two

alternative methods will then be compared for their efficacy

in achieving the other two criteria.


First, it is believed that the segments derived by the

use of preference measures will exhibit more differential

behavior with respect to brand usage than will the segments

derived from the importance measures. This result is

expected since, as discussed in Chapter I, preferences are

more closely related to behavior.

Secondly, it is also believed that the preference

segments will be more differentially responsive to elements

of the marketing mix. This should result since preferences

are less grounded to immutable intrapersonal characteristics

and are partly determined by external sources. It is felt

those individuals who have developed similar preferences

have similar information gathering proclivities. Thus,

these segments should provide greater stability in the sense

that the segments will be more significantly different in

their cognitive structure.

Also, the stimulus configurations developed by the

psychometric procedures for each segment will be marked by

solutions of lower dimensionality than would result from

scaling all respondents into a common stimulus space. Thus,

such a procedure will provide for not only differences in

preference among the respondents but also differences in

their perceptions of the product class. This procedure will

be an improvement over the usual way of using MDS and will

result in solutions which are of lower dimensionality, and

which are idiosyncratic to the segment. By focusing on

fewer dimensions for each segment, clearer guidelines for

strategy development will be available.

Lastly, while the segments are homogeneous with

respect to their scale values and importance ratings, it

is desired to see if they are identifiable in terms of

some more readily available ancillary descriptors. Thus

they will be tested to see if they can be identified by

background demographic variables.

It should be kept clearly in mind that all of these

objectives are united by the common thread of furnishing

marketing management with a methodology for identifying

market segments that will lead to the development and

implementation of meaningful strategy. Obviously, a

procedure which can do this efficiently and reliably is

of great value.

Data Collection

Obviously, an undertaking of this magnitude will

require a large amount of data. Accordingly, a two phase

survey covering the two product categories will be

conducted in order to investigate the above discussed

objectives. The survey will consists of a non probability

sample of housewives residing in the state of Florida.

Housewives will be used since they will provide for more

realism plus the fact that some other convenience sample

(e.g. students) may not be heterogeneous enough to exhibit

distinct segments.

Accordingly, phase one will consist of a pretest to

elicit attributes salient to the purchase of analgesics and

hard-surface floor cleaners (see Appendix A). Such a

procedure is standard in the social sciences for obtaining

the attributes of an object. It is assumed that beliefs

about an object can be elicited in a free-response format

by asking respondents to list the characteristics, qualities,

and attributes of the object. It has been argued that

salient beliefs are elicited first, and, thus, those

elicited beyond the first nine or ten are probably not

salient for the individual. As a rule of thumb then, it has

been recommended to use the first five to nine attributes

(Fishbein, 1967).

The main emphasis is in the second phase where five

different types of data will be collected for each product

category. Measures will be obtained of each respondent's

(1) brand preferences in the paired comparison format, (2)

importance ratings for the product attributes, (3) product

usage, (4) sensitivity to marketing mix elements, and (5)

background demographic data (see Appendix B).

The paired comparison brand preferences are, of course,

the usual input for MDS procedures as outlined above, these

are presented in randomized order according to the procedure

outlined by David (1963) in order to prevent response bias.

The importance ratings are for the attributes obtained in

phase one. The product usage data refer to the brand

usually purchased, the degree of brand loyalty, and the

amount of the product used. The background demographic

data consist of the following measures: the respondent's

age, marital status, race, religion, number of children

living at home, number of children under six, education,

employment, occupation of the household head, household

income, and dwelling status.

The key area of sensitivity to marketing mix elements

will be patterned after the price, distribution, and promo-

tion variables. The price variables include the increase

required to induce brand switching, an index of dealing,

and coupon usage. The type of retail outlet purchases are

made at and an index of store loyalty are the distribution

variables. The promotional variables include the news-

papers and magazines read plus the time spent reading them

as well as the type of radio and television programing

attended to and the amount of time spent in these pursuits.

Also included is an index of personal communication regard-

ing product categories.


In order to define segments and compare procedures

for their efficacy in fulfilling the criteria of meaningful

segmentation there need to be several waves of data

analysis. Figure 3 presents the plan of analysis in flow

Preference Data
Vector Model

Figure 3



chart form and should prove helpful in conceptualizing

the research. All of the steps outlined below will,

of course, be carried out twice, once for each product


The first task will be to identify the benefit and

preference segments. This will be accomplished by submit-

ting the importance ratings and the unidimensional scales

derived from the paired comparison data to inverse factor

analysis, or Q analysis, for clustering purposes. Since

the emphasis is on the profile of the scales, the

covariance between individuals as employed in Q analysis

should provide good results.

Specifically, the technique used is principal

component analysis followed by varimax rotation. Rather

than use the raw factor loadings, each respondent's

loading on a factor will be indexed to the average loading

for that factor. Due to the fact that some loadings are

negative, ordinary arithmetic averages are not used, but

rather the square root of the average squared value for

each factor. Each respondent is then allocated to the

group where she has the highest index, even though she

may have a higher loading with some other factor. These
"normalized Q groups" tend to produce results that are

more reliable when the analysis is repeated with new data

(Johnson, 1974).

After obtaining the segments formed by the two

approaches, they will be compared by relating them to the

product usage data and the marketing mix measures. First,

crosstabulation of the groups with the nominally scaled

data and analysis of variance with the ration scaled data

will highlight on which particular variables the groups

are significantly different. Secondly, the statistically

significant variables will be used as independent variables

in a discriminant function which takes into account the

covariation between variables to predict segment membership

in order to assess which of the segmentation approaches is

more highly associated with the other variables.

A discriminant function will also be run using the

demographic measures as independent variables. This will

be done in order to determine if the groups are identifiable

in terms of these ancillary variables.

Next, the vector model will be used for deriving a

spatial representation containing both brand positions

and individual importance weight vectors for the entire

sample. Then, each of the preference segments will be

scaled by the vector model to provide separate spaces for

each. This will enable a comparison of these spaces with

the common space and it should turn out that the segment

spaces are of lower dimensionality and are idiosyncratic

to each segment.

By focusing on the spatial configurations, the

univariate results, and the discriminant function


concurrently the issue of segment interpretability can

be addressed. By noting the determinant attributes and

the significant independent variables, meaningful

segmentation guidelines can be formulated for each

segment. These guidelines will consist of both strategy

plans and the tactics for implementing them.


In the first phase of the analysis a sample of 50

housewives was asked to elicit in an unstructured format

(see Appendix A) specific attributes that they take into

consideration when purchasing analgesics and hard-surface

floor cleaners. The 28 attributes which were mentioned

for analgesics along with the number of respondents

mentioning them are presented in Table 1. It was decided

to use those attributes which were mentioned by 10% or

more of the sample in the phase two questionnaire. Thus,

the attributes rated for importance in the main portion

of the study included:

-effectiveness in eliminating pain
*doesn't upset the stomach
*fast acting
*doesn't cause side effects
-recommendation of your doctor
*works on different types of pain
-package size
*chemical ingredients
*reputation of the brand name

Table 2 lists the 25 attributes mentioned for hard-

surface floor cleaners and the number of respondents

mentioning each. Using the same 10% or more criterion

it was decided to include the following 11 attributes in

the phase two questionnaire:

Table 1

Frequency of Elicited Attributes
for Analgesics

Attribute Number

Effectiveness in eliminating pain


Doesn't upset the stomach

Fast acting

Doesn't cause side effects

Recommendation of doctor

Works on different types of pain

Package size

Chemical ingredients

Reputation of brand name


Time between doses

Recommendation of friends

Past experience

Suitability for all ages

Ease of swallowing


Dosage amount

Advertising image

Compatible with other drugs

Non-habit forming

Subjects Percentage























Table 1 continued


Effects with alcohol

Daytime or nightime use

Family preference

Child proof container

Consumer Reports evaluation

Prior medical tests

Shelf life

Number of Subjects Percentage

1 2

1 2

1 2

1 2

1 2

1 2

Table 2

Frequency of Elicited Attributes
for Hard-Surface Floor Cleaners



Ease of application

Cleaning effectiveness

Removes old wax

Long lasting shine

Non-yellowing shine

Quick drying

Resists scuffing

Can be used for many floor type


Shines after mopping

Package size

Safe to use around children anc


Reputation of brand name

Recommendation of friends


Application amount

Past experience


Advertising claims

Consumer Reports evaluation

Number of Subjects Percentage

34 68

28 56

21 42

15 30

15 30

15 30

9 18

8 16

Is 5 10

5 10

5 10

4 8

Ipets 4 8

3 6

3 6

3 6

2 4

2 4

2 4

1 2

1 2

1 2

Table 2 continued


Impervious to water

Other equipment needed


Number of Subjects Percentage

1 2

1 2

1 2


*ease of application
*cleaning effectiveness
*removes old wax
*long lasting shine
*non-yellowing shine
*quick drying
*resists scuffing
*can be used for many floor types
*shines after mopping
A random number table was used to determine the order

that both of these attribute lists appear in the phase two

questionnaire in order to prevent biasedness in responding.

The above task having been completed, the phase two

questionnaire was administered to 35 individuals for

protesting. Analysis of these results enabled the question-

naire to be rephrased in a number of places so as to provide

for greater clarity. Once this had been Pccomplished, the

final questionnaire (see Appendix B) was administered to 218

housewives throughout the state of Florida. These women

were contacted through various religious, social, and civic

organizations which received $2 for each participant

competing a questionnaire. Since this is a convenience

sample, it is not representative of the nation's housewives.

For example, there are more whites and catholics with higher

incomes, more education, and fewer children among the respon-

dents than found in the general population.

Preliminary inspection of the questionnaires indicated

the 19 of them were so full of errors and ambiguous as to be

totally useless for any purpose. These were immediately

eliminated and the remaining 199 respondents comprised the

sample for the study. While many of these questionnaires


were incomplete and thus there is much missing data, it was

felt that enough information was present to provide for

meaningful analyses. The data are concerned with two

product categories and each will be analyzed in turn.


The analgesic market will be analyzed with the

three criteria for meaningful segmentation developed in

Chapter I firmly in mind. Thus the concern is with the

issues of identification of segments, brand usage, and

sensitivity to marketing mix elements. Since these derive

the main areas of investigation outlined in Chapter III,

the analysis should provide an indication of the more

efficient segmentation procedure for isolating segments

which provide for more meaningful strategy development.

Identification and Usage

The first task to be undertaken in order to address

the first objective is to determine the various segments

of the analgesic market. Accordingly, each subject's

importance ratings for the 10 analgesic attributes were

submitted to inverse factor analysis, or Q analysis,

since the emphasis is on identifying consumers who have

similar profiles, or utility, for the various attributes.

Three of the respondents had given the same importance

weight for all 10 attributes and thus their profile was a
constant with no variance and they were eliminated from the

analysis. Inspection of the eigenvalues (see Appendix C)

associated with the remaining subject's correlation

matrix indicated that 6 factors should be retained which

explain 89.6% of the variance in the original data. The

loadings on the 6 factors were submitted to varimax

rotation and each respondent's loadings was indexed to

the average loading for that factor according to the pro-

cedure developed by Johnson (1974) outlined in Chapter III.

Assignment of each subject to the factor where she has the

highest index resulted in 38 individuals being allocated

to group 1, 37 to group 2, 37 to group 3, 25 to group 4,

28 to group 5, and 31 to group 6.

Analysis of each group's mean rating on the 10

attributes presented in Table 3 indicates that the brand's

effectiveness in relieving pain and its lack of unwanted

side effects are primarily important to the entire sample.

However, the relative importance they place on other

attributes differs. Group 1 seems to be more interested

in price, ingredients, a doctor's recommendation, reputa-

tion, and causation of stomach upset while being less

concerned about a generalized pain reliever and speed of

action. The speed of relief provided is relatively more

important to group 2 whereas a generalized pain reliever

and a doctor's recommendation are of very little importance.

Group 3 has more interest in a generalized pain reliever,

ingredients, and causation of upset stomach but unconcerned

about the brand's reputation. The package size and stomach

Table 3

Mean Attribute Rating for
Analgesic Importance Groups


Attribute 1 2 3 4 5 6

Effectiveness 6.61 6.97 6.86 6.72 6.93 6.68

Price 4.79 3.81 4.22 4.16 3.68 2.39

Different types of pain 3.66 4.73 5.78 5.12 5.50 5.74

Chemical ingredients 6.32 3.11 6.19 4.44 3.39 5.84

Doctor's recommendation 6.39 3.54 5.84 5.92 6.54 6.13

Brand reputation 4.97 4.22 2.43 4.48 4.75 5.87

Side effects 6.58 6.24 6.59 6.52 6.14 6.39

Fast acting 5.50 6.24 5.73 5.84 5.71 5.77

Package size 2.71 2.54 3.24 4.04 1.82 2.61

Doesn't upset stomach 6.16 5.67 6.11 6.12 4.14 5.93


upset effects are of more interest to group 4 while price

is not. Group 5 is more concerned about a doctor's

recommendation and less concerned about the package size

and causation of side effects. Group 6 is more interested

in a doctor's recommendation and reputation of a generalized

pain reliever but doesn't care about price.

The assumption underlying this segmentation procedure

is that consumers who desire similar attributes in a product

will buy the same brand. Therefore, since the above groups

have been identified in this manner it would be expected

that each group purchase the same brand and that different

groups purchase different brands. However, comparison of

the individual members of each group with the brand they

most frequently purchase indicates that this is not true. A

crosstabulation of this data is presented in Table 4 which

indicates that the groups are not significantly different in

their brand purchasing. Inspection of Table 4 discloses

that there are a number of cells with very low expected

frequencies and that a chi-square statistic may be

inappropriate in this case. In order to correct for this

deficiency a number of the rows were combined. Specifically,

the Anacin, B.C., Bufferin, Empirin, and Excedrin rows were

combined as well as the Datril and Tylenol rows. Thus,

there are now 6 groups times only 4 rows: Bayer, aspirin

containing brands, non aspirin brands, and other brands. As

shown, crosstabulation produces a highly non significant chi-

square and thus the hypothesis of independence can not be


Table 4

Crosstabulation of Analgesic Importance
Groups with Brand Purchased


Brand 1 2 3 4 5 6

Anacin 5 2 3 3 0 4

Bayer 5 9 9 3 8 9

B.C. 0 0 1 0 0 0

Bufferin 6 2 1 3 0 2

Datril 1 0 1 2 2 0

Empirin 1 1 1 0 1 1

Excedrin 2 5 2 1 5 3

Tylenol 12 11 12 10 9 8

Other 6 7 7 3 3 4

Chi-Square = 9.901*

DF = 15*

Probability = 0.8088*

* based upon a contingency table having combined rows
(see text)

The identification of segments based upon their paired

comparison preferences is slightly more involved. Initially

the variance contained in the paired comparisons needs to be

partitioned in the analysis of variance format discussed in

Chapter III to determine the significance of the individual's

scale values. This analysis of variance is presented in

Table 5 which shows that the scale values are highly

significant. While the F test for the unscalability is also

significant due to the large degrees of freedom, analysis of

the sum of squares column shows that 82.2% of the variance is

accounted for by the scale values whereas only 0.3% of the

variance is accounted for by the unscalability. Futhermore,

the overall validity correlation coefficient between the

original preferences and the scale values is .906 while the

reliability correlation coefficient between the original

preferences and the scale values plus the unscalability is

.907, also indicating that the data is relatively free of

systematic and random error. Thus, the data is highly

reliable and possesses a high degree of internal validity

which bodes usefulness for multidimensional scaling.

However, partitioning of the sum of squares attributable to

the scale values into orthogonal components for each subject

obtains 5 non significant F's at an alpha level of .05.

This indicates that their scale values are not significantly

different from zero and thus they are indifferent to the

analgesic brands. Further support is provided in that the

average validity correlation coefficient between the original

Table 5

ANOVA for Analgesic Preferences

Source SS

















*probability > .001




preferences and the scale values for these 5 subjects is a

lowered .812. These indifferent subjects were, therefore,

deleted from further analysis.

The remaining 194 respondents were submitted to the

same factor analytic procedure as outlined above. In this

case the correlation between individual profiles was based

upon the scale values for eight brands which sum to zero

and thus the correlation matrix has only 7 non zero

eigenvalues (see Appendix C). Thus, retaining all seven

factors will explain 100% of the variance in the scale

values. It was deemed appropriate to do this for two

reasons. First, all of the variance can be explained by

using a relatively small number of groups roughly equal to

the number of importance rating groups. Second, it is felt

that each brand has some cadre of loyal buyers who should

be considered as a segment.

Assignment of each subject to the factor where she has

the highest index resulted in 38 individuals being allocated

to group 1, 34 to group 2, 22 to group 3, 25 to group 4, 24

to group 5, 24 to group 6, and 27 to group 7. The average

scale value for each group is shown in Table 6.

While it was shown above that the groups based upon

importance ratings did not purchase significantly different

brands, a similar analysis of the preference groups shows

that these groups do have significantly different purchasing

proclivities. A crosstabulation of the preference groups

with the brand they most frequently purchase is presented in


Table 6

Mean Scale Values for
Analgesic Preference Groups


Brand 1 2 3 4 5 6 7

Anancin -1.00 -0.16 1.72 -0.74 -0.51 2.13 -0.94

Bayer -0.00 3.74 0.45 1.34 0.27 1.32 0.34

B.C. -2.37 -1.53 -1.77 -1.68 -2.48 -1.66 -1.33

Bufferin -0.40 0.42 0.41 3.08 -0.55 0.26 0.52

Datril 2.48 -1.43 -1.70 -1.22 -1.43 -1.54 -1.70

Empirin -1.27 -1.44 -1.23 -0.62 2.36 -0.99 -1.49

Excedrin -0.61 0.07 2.69 -0.81 0.53 -0.21 0.75

Tylenol 3.18 0.33 -0.57 0.65 1.82 0.70 3.85


Table 7. Inspection of this table reveals that a number of

the cells have very low expected frequencies and thus the

chi-squares statistic may not be appropriate. In order to

correct for this deficiency a number of the rows were

combined. Specifically, the Anacin, B.C., Bufferin,

Empirin, and Excedrin rows were again combined as well as

the Datril and Tylenol rows. Thus there are now 7 groups

times only 4 rows: Bayer, aspirin containing brands, non

aspirin brands, and other brands. Crosstabulation now

produces a highly significant chi-square indicating that

the procedure using preferences as the segmentation input

produces more meaningful segments with regard to brand

purchasing behavior as would be expected.

Marketing Mix Elements

Before further inquiry into the differential behavior

of the groups formed using the two types of data it is

appropriate to insure that the two procedures do in fact

result in different groupings of the subjects. Towards

this end the crosstabulation presented in Table 8 indicates

that while the two grouping methods are not independent the

relationship between them is rather weak as evidenced by

the low Cramer's V statistic. Thus, it is clear that these

two methods result in very different segments.

Having satisfied this point, the next step is to

ascertain the association between the groups and the various

Table 7

Crosstabulation of Analgesic Preference
Groups with Brand Purchased


Brand 1 2 3 4 5 6 7

Anacin 0 0 7 0 1 9 0

Bayer 4 22 1 8 2 4 3

B.C. 0 0 0 0 1 0 0

Bufferin 0 0 0 13 0 0 0

Datril 6 0 0 0 0 0 0

Empirin 0 0 1 0 4 0 0

Excedrin 1 0 7 0 3 4 3

Tylenol 20 3 3 2 12 5 18

Other 7 9 3 2 1 2 3

Chi-Square = 127.258*

DF = 18*

Probability = 0.001*

* based on a contingency table having combined rows
(see text)


Table 8

Crosstabulation of Analgesic Importance
Groups with Preference Groups

:nce Importance Group

Cramer's V = 0.238










marketing mix elements. There are 35 variables dealing with

price, distribution, and promotional issues as outlined in

Chapter III. Of these, two, the type of store where

subjects usually purchase analgesics and the place they

usually listen to the radio, were eliminated due to the

fact that such a large number of the sample, 57% and 21%

respectively, failed to answer these questions.

Importance groups. Of the remaining 33 variables, 8

consist of ratio scaled data and 25 are binary zero/one

indicating whether or not a subject reads a particular

type of magazine, listens to a particular type of radio

program, or watches a particular type of television program.

The importance groups were then tested to see if there

were differences between group means on these variables.

Crosstabulations were run on the discrete variables while

analysis of variance was performed on the ten continuous

variables. The significance levels for the resulting chi-

square and F statistics are presented in Table 9. It can

be seen that at an alpha level of .10 the importance group

means are significantly different on only six of the


-percent of time the brand is used
-listening to popular radio programs
*listening to news/weather radio programs
*watching action television programs
.watching talk show television programs
*time watching television programs

These results, being univariate in nature, do not take

into consideration the covariance structure in characterizing

Table 9

Significance of the Analgesic Importance
Group Means on the Marketing Mix Variables


Probability Level

Percent of time brand used

Amount used

Buy at the same store

Price increase for switching

Check ads for sales

Clip and use coupons

Number of newspapers read

Time reading newspapers

Magazine type readership







General Interest

Time reading magazines

Radio program listening


Easy listening


. 0288**



















Table 9 continued


Probability Level


Talk show

Time listening to radio

Radio frequency

Television program watching


Sporting events





Game shows

Talk shows

Time watching television

Consulting with others















* p>.10

** p>.05

*** p>.01

the groups and for this reason multiple discriminant

analysis was used to see what effect the covariation

would have on the results. Linear discriminant analysis

if a natural technique to use when the data consist of a

categorical criterion variable and a set of possible

predictor variables. This technique involves certain

assumptions about the form of the data which are not

true in this case and will be discussed below. Still,

application of this technique can provide useful

insights into the relationship between group membership

and marketing mix elements.

The variables were submitted to the SPSS stepwise

discriminant procedure to determine which variables

are most able to discriminate among the importance

groups. The criterion used for inclusion into the

model is the largest increase in the generalized

distance as measured by Rao's V. The variable which

contributes the largest increase in V when added to

the previous variables is the one selected for

inclusion into the model. This amounts to the largest

overall separation between the groups. The change in

V is distributed as a chi-square with g-l degrees of

freedom, where g equals the number of groups, and so

can be tested for significance. An alpha level of .10

was chosen as the cutoff point and 6 variables were

found to contribute to the classification model.

These variables are:


*listening to popular radio programs
-watching action television programs
-percent of the time brand is used
*watching talk show television programs
*time reading magazines
*time watching television programs

Five of these variables are the same as univariate

results with the time reading magazines as the new addition.

These variables were then used to discriminate the impor-

tance group membership. Since a few of these variables

contain missing data, these observations are deleted from

the analysis and the size of the sample shrunk to 193

from the original 196. Of these, the model correctly

classified 66 of the observations as shown in Table 10, a

score of 34.2%

This statistic by itself does not indicate the

classificatory power of the model in uncovering structure.

To address this issue, Montgomery (1975) has devised a

procedure to assess the expected spurious percent of

correct classifications. His procedure involves random

assignment of each actual observation to the segments in

proportion to the observed sample incidence of these

groups. By running a discriminant analysis with the actual

observations in these randomized segments a base line for

spurious performance is obtained.

Accordingly, three of these randomized discriminant

runs were undertaken and they classified 54, 55, and 64

observations correctly; scores of 28.0%, 28.5%, and 33.2%

respectively. Thus, on the average predictability is

Table 10

Classification Matrix for Analgesic
Importance Groups Using Marketing Variables









Total number =

Number correct

into Group


= 66

Percent correct = 34.2%

increased only 4.3% by placing the observations in

the proper segments.

Preference groups. A similar process was under-

taken using the same variables to discriminate among

the preference groups. First, univariate tests were

performed to determine on which variables the prefer-

ence groups differed. The significance levels for the

variables are presented in Table 11 and it can be seen

that the preference group means differ on only 7

variables. These are:

-amount used
-time reading newspapers
-reading general interest magazines
-listening to popular radio programs
-listening to talk show radio programs
-watching variety television programs
-watching game show programs

Again, in order to take into consideration the

effect of the covariance between variables, multiple

discriminant analysis was employed. While the problem

of violation of the assumptions of the model still

exists (see below), it is not believed to have serious

impact on the implications of the results.

The stepwise procedure using the change in Rao's

V as the criterion was again used. At an alpha level

of .10, it was found that 5 variables were significant.

These are:

Table 11

Significance of the Analgesic Preference
Group Means on the Marketing Mix Variables


Probability Level

Percent of the time brand used

Amount used

Buy at the same store

Price increase for switching

Check ads for sales

Clip and use coupons

Number of newspapers read

Time reading newspapers

Magazine type readership







General Interest

Time reading magazines

Radio program listening


Easy listening





















Table 11 continued

Variable Probability Level

News/Weather .4821

Talk show .0198**

Time listening to radio .5308

Radio frequency .5589

Television program watching

News/Weather .4406

Sporting events .3484

Movies .5615

Action .6877

Comedy .8047

Variety .0894*

Game shows .0578*

Talk shows .2278

Time watching television .5055

Consultation with others .3153

* p>.lO

** p>.05

*** p>.Ol

*amount used
-listening to talk show radio programs
*reading beauty/fashion magazines
-listening to popular radio programs
-time reading newspapers

Four of these variables are the same as were found in

the univariate analysis with readership of beauty/fashion

magazines being the new addition. These variables were

then used to discriminate the preference groups. Again,

of course, a few of the variables contain missing data and

the number of observations retained drops to 187 from the

original 194. This model though classified 56, or 29.9%,

of the observations correctly as shown in Table 12.

While the percent classified correctly is lower for

the preference groups than for the importance groups, it

must be remembered that there is an extra group to be

classified with one less variable. For this reason, the

Montgomery (1975) procedure was employed to assess the

improvement this model provides over spuriousness.

Accordingly, three randomized discriminant runs were under-

taken which classified 42, 44, and 44 observations

correctly; 22.5%, 23.5%, and 23.5% respectively. Thus,

placing the observations in the proper segments improves

predictability by 6.7% on the average.

Since the significant variables tend to be a small

subset of the original 33, it was decided to see if the

addition of more variables would increase the performance

of the discriminant function. The stepwize procedure was

used specifying a multivariate partial F test on the mean

Table 12

Classification Matrix for Analgesic
Preference Groups Using Marketing Variables


into Group

21 11

10 20

Total number =







Number correct = 57

Percent correct = 29.9%


difference between groups as the criterion. An F of 1.0

was chosen as the cutoff point and 13 variables were found

to contribute to the importance group classifications and

19 variables contributed to the preference group classifi-


For the importance groups the 13 variables classified

42.9% correctly. Three randomized runs produced correct

classifications of 36.5%, 38.5%, and 39.8%. Thus, on the

average the real importance groups increase predictability

over spuriousness by 4.6%.

The 19 variables classified 54.3% of the preference

groups correctly. Correct classifications produced by the

randomized process were 40.7%, 42.6%, and 43.2%. Here the

increase in predictability amounts to 12.1% by using the

real preference groups.

Demographics. In order to determine if the groups

can be characterized by more readily available ancillary

data, an attempt was made to discriminate the groups on the

basis of 11 demographic variables. Five of the variables,

age, number of children, number of children under 6 years

of age, education completed in years, and income, are

ratio scaled. Two others, employment and dwelling owner-

ship, are binary zero/one. The remaining four, marital

status, race, religion, and occupation of the household

head, are nominally scaled and hence crosstabulations were

run to determine if any of the groups differed on these

variables. Crosstabulation of the importance groups with

the above four variables produced chi-square values of

20.569, 12.653, 14.986, and 24.373 respectively. At the

associated degrees of freedom none of these statistics is

significant above the .24 alpha level and thus all of

these were deleted from the analysis.

Crosstabulating the same four variables with the

preference groups produced chi-square values of 29.214,

13.151, 24.546, and 39.059 respectively. At the

associated degrees of freedom none of these statistics

is significant beyond the .21 alpha level so again these

variables were deleted from the analysis. The remaining

seven variables were then used attempting to discriminate

the two types of groups.

The results for the importance groups appear in

Table 13 which shows that use of these variables correctly

classifies 28.2% of the observations. This model,

considering the above results, is not very useful for

characterizing the membership of the various groups. It

should also be pointed out that while all of the variables

were used, only one, age, was significant at the .10 alpha

level according to the change in Rao's V criterion.

Table 14 contains the results for the preference

groups which shows that use of these demographic variables

classifies 29.4% of the observations correctly. Again,

these results are not very useful for identifying segment

members, especially since, according to the change in Rao's

V criterion, none of the variables are significant at the

Table 13

Classification Matrix for Analgesic
Importance Groups Using Demographic Variables


Classified into Group



Total number =

Number correct = 51

Percent correct = 28.2%

Table 14

Classification Matrix for Analgesic
Preference Groups Using Demographic Variables


Total number =

into Group


Number correct = 52

Percent correct = 29.4%









.10 alpha level. It seems then that the groups are not

significantly different and can not be predicted at more

than spurious levels by demographic variables.

Before leaving this section two caveats are in order.

First, linear functions provide the best discrimination

when the group covariance matrices of the independent

variables are equal. Testing indicates that this is not

true for this sample when using the small subsets of

variables. For the results of an analysis using a quadratic

function see Appendix D.

Second, inferences employing F statistics involve an

assumption of multivariate normality among the independent

variables. The use of binary variables in this study clearly

violate this. Therefore, the results of these discriminant

analysis should not be considered as absolutely conclusive in

their own right. Still, violation of the assumptions or not,

the results of the preceding section provide useful insights

that are somewhat greater than spurious performance.

Marketing Implications

Addressing the final issue requires initially

determining the dimensionality of the common product space

which is representative of all the subjects' perceptual

structure. This can be accomplished by inspection of the

eigenvalues derived from the matrix obtained by premulti-

plying the subjects' scale value matrix by its transpose.

The eigenvalues are listed in Table 15 and plotted in

Figure 4. There are 7 non zero eigenvalues since the

analgesic scale values sum to zero. Inspection of the

eigenvalues suggests retaining 4 dimensions.

This conclusion is confirmed by referring to Table

5 in conjunction with Table 15. It can be seen that the

first 4 eigenvalues account for 78.2% of the variance in

the scale value; that is, the sum of the squares attri-

butable to the 4 dimensions is 46167.312. Thus, the sum

of squares associated with the unused eigenvalues equals

12810.688 or 21.8% of the variance in the scales.

Comparison of this figure with the error sum of squares

indicates that the error increment in fitting a 4 dimen-

sional submodel to the data is not excessive.

Obviously it is impossible to graphically display 4

dimensions so the coordinates of the analgesic brands are

presented in Table 16. Inspection of this table can pro-

vide one with an appreciation of the difficulty involved

in interpreting solutions of this dimensionality. It

would appear from the fact that most subjects have a

negative weight for this dimension coupled with the prepon-

derance of Tylenol users in the sample (see Table 6) that

dimension 1 is an evaluative dimension concerned with the

analgesic's effectiveness in relieving pain. Inspection of

the brand order on dimension 2 indicates that this dimen-

sion represents aspirin content. The last two dimensions,

however, are much more arcane. They most probably represent

Table 15

Eigenvalues Obtained from
Analgesic Scale Values






































1 2 3 4 5 6 7 8

Figure 4

Eigenvalues Derived from
Analgesic Scale Values

Table 16

Coordinates of the Analgesic Brands
in the Common Product Space
















0 .271



Dimens ion




























some subjective aspects of the product class which can not

be inferred without ancillary information.

While it requires 4 dimensions to represent the

perceptual structure of all the subjects, this product

space may not in fact be representative of any one subject

but rather be generalizeable to some non existent average

group member. This situation might well be improved by

developing a product space for various segments of the

market. This has usually been accomplished by segmenting

on some a priori demographic variable and the results have

been somewhat less than easily interpretable. It would

seem to make much more sense to isolate segments on the

basis of their similarity in preferences. In this way, the

product space derived, since it contains the perceptual

structure of subjects with similar preference, will

result in solutions of lower dimensionality which are more

easily interpretable.

Accordingly, each preference group identified above

was submitted to a separate vector submodel analysis.

While there is no longer a relevant error sum of squares

for comparison, inspection of the eigenvalues for group 1

(see Appendix C) shows that 73.1% of the variance in the

scale values is explained by the first dimension and that

the remaining dimensions contribute a negligible amount in

explaining the variance. Thus, the common product space

for the individuals in group 1 is unidimensional in nature

and is shown in Figure 5. It appears from inspection of