Faculty
Working Papers
HOW TO GET THE MOST
OUT OF MULTIVARIATE METHODS
Jagdish
N.
Sheth
#243
College of
Commerce and
Business Administration
University of Illinois at Urbana-Champaign
FACULTY WORKING PAPERS
College of Commerce and Business Administration
University of Illinois at Urbana-Champalgn
April 16, 1975
HOW TO GET THE MOST
OUT OF MULTIVARIATE METHODS
Jagdish
N.
Sheth
#243
HOW TO GET THE MOST OUT OF MULTIVARIATE METHODS
Jagdish N. Sheth
University of Illinois
Urbana, Illinois 61801
-3-
HOW TO GET THE MOST OUT OF MULTIVARIATE METHODS
Jagdish N. Sheth
University of Illinois
Today multivariate methods are widely used (and often misused) by many
marketing professionals and academic researchers.
Multivariate methods should
or actually have practically replaced the more traditional statistical
analyses such as frequency distributions and cross-tabulations in marketing
research.
It is,
therefore, almost impossible today to come across
a
research study
reported in academic journals such as Journal of Marketing Research which does
not utilize some type of multivariate analysis of the data.
If multivariate
analysis is somehow missing in the study, it is often recognized as
ness to be rectified in a follow-up research.
a
weak-
While this widespread use of
multivariate methods has certainly increased the respectability of marketing
as a discipline among the more "scientific" and traditional disciplines,
it
has also brought upon many the problems of communication and understanding.
Probably
a
majority of JMR readers have difficulties comprehending the
articles published in it.
This problem seems more vivid between the academic
and the professional researchers and especially between the researchers and
the managers of the marketing function in the organization.
Accordingly,
there are two objectives of this paper:
1.
Provide a nonstatistical description of the multivariate methods
and discuss their potential to solve marketing problems.
2.
Provide guidelines to the researcher for getting more out of the
multivariate methods.
Digitized by the Internet Archive
in
2012 with funding from
University of
Illinois
Urbana-Champaign
http://www.archive.org/details/howtogetmostouto243shet
-4-
DESCRIPTION AMD APPLICATIONS OF MULTIVARIATE METHODS
Multivariate methods refer to those statistical techniques which focus
upon, and bring out in bold relief, the structure of simultaneous relation-
ships among three or more phenomena.
It is important to remember that what
matters in multivariate analysis is the analysis of the simultaneous relationships among phenomena.
Many sequential or hierarchial statistical analyses
of large number of variables such as the Automatic Interaction Detection (AID)
are,
therefore, not truly multivariate in nature but simply repeated
applications of simpler statistical techniques.
Multivariate methods differ from simple univariate (single phenomenon)
statistical techniques in terms of a shift in focus away from the levels
(averages) and distributions (variances) of the phenomena, and instead
concentrating upon the degree of relationships (correlations or covariances)
among these phenomena.
They also differ from the bivariate (two phenomena)
statistical techniques by shifting focus away from pairwise relationships
to the more complex simultaneous relationships among phenomena
[5,6],
Multivariate methods can be broadly categorized into two types: Functional and structural multivariate techniques.
Functional Multivariate Methods
The functional multivariate methods are most appropriate for building
predictive models with which the researcher can forecast, or explain one or
more phenomena from the knowledge of other phenomena based on their relationships.
In order to satisfactorily utilize the functional multivariate methods
it is essential that the researcher has considerable knowledge or theory
about the market behavior with which to properly conceptualize a realistic
-5-
model.
The functional multivariate methods provide to the researcher
estimates of both the directionality and the magnitude of relationships among
phenomena.
As such, they border on being the most precise quantitative
models of market behavior.
a
Of course, how realistic these models may be is
direct function of the imagination and the experience of the model builder.
Depending on the nature and the number of phenomena the researcher wishes
to predict or explain,
there are several different types of functional
multivariate methods.
The first most commonly known and used multivariate
method is multiple regression, and its many variations, which enables the
researcher to predict the level or the magnitude of
sales volume or market share of a brand.
is to search for the best possible
a
phenomenon such as the
The objective in multiple regression
(optimum) simultaneous relationship between
the distribution of the predicted phenomenon and those of the many other
correlated or causal phenomena resulting in establishment of a functional
relationship between the criterion and the predictor variables.
market shares of grocery products may well be a function of
a
For example,
number of
marketing mix variables such as average unit price, customer loyalty, media
advertising, store coverage, store display and point-of-purchase promotion
[1].
A second functional multivariate method is multiple discriminant analysis
and many of its variations, which are extremely useful if the researcher is
interested in predicting the likelihood of an event happening sometime in the
future.
For example, what is the likelihood that a tax payer's return will be
audited by the IRS?
Or what is the likelihood that
a
telephone customer will
convert to TouchTone service when it is promoted by the company?
The objec-
tive in multiple discriminant analysis is to identify those key discriptors
-6-
on which various predefined events have statistically significant differences,
and to build a functional model out of them which will enable the researcher
to predict likelihoods of events happening as best as possible.
Thus, a tax
payer with more than $50,000 taxable income or one who belongs to certain
occupations such as medical doctors may have significantly higher likelihood
of being audited than the average tax payer.
Similarly,
household in upper
a
socioeconomic class, younger life cycle or with high mobility may have
significantly greater likelihood becoming a TouchTone subscriber than the
average customer.
A third functional multivariate method is multivariate analysis of
variance (MANOVA) which is more useful for testing the impact of various
For
levels of one or more experimental factors on a variety of phenomena.
example, what is the impact of doubling the advertising budget (weight) on
market awareness, attitudes, and purchase behavior toward
a
product?
The
objective in MANOVA is to test for significant differences on a set or pro file of variables due to some changes in one or more causal factors.
Thus,
for example, doubling the advertising budget may be highly effective in
significantly increasing the awareness and attitude levels among customers
but may have little impact on their immediate purchase behavior.
A fourth major functional multivariate method is canonical correlation
analysis which enables the researcher to build
a
predictive model with which
he can simultaneously forecast or explain several phenomena based on his
knowledge of their correlates.
For example, the researcher may be interested
in the nature and magnitude of price competition among various brands of a
product class.
The objective in canonical correlation analysis is to
simultaneously regress
a set of
criterion variables on
a
set of predictor
-7-
variables in Che hopes of bringing out the functional relationships both within and between the two sets of variables.
Thus it is quite possible that while
price competition may be prevalent within the national brands and within the
store brands but not between the two types of brands.
Structural Multivariate Methods
The structural multivariate methods, on the other hand, are more des-
criptive and less predictive in nature.
They are essentially data reduction
techniques which simplify complex and diverse relationships among phenomena
in a manner which enables the researcher to gain insights into the underlying
and nonintuitive structure of relationships.
The structural multivariate
methods are thus analogous to the search for the needle in the hay stack.
The most popular of the structural multivariate methods is factor analysis
and many of its variations.
Factor analysis enables the researcher to gain
insights into the common underlying bonds or dimensions by which otherwise
highly divergent phenomena tend to correlate among themselves.
For example,
what is the common bond between income, education and occupation of a household?
Or is there any systematic pattern of preferences in the viewership
of vast variety of television programs?
The objective in factor analysis is
to decompose into meaningful components or dimensions the extent of relation-
ships empirically observed among a set of divergent phenomena.
Thus, the
common underlying dimension of social class may be responsible for the strong
positive correlations found between income, education and occupation.
Similarly, interest in situation comedy, quiz shows, soap operas, westerns,
police or detective stories, etc. may be the common bondages among the vast
variety of television programs.
-8-
A second structural multivariate method is cluster analysis which enables
the researcher to classify,
segment or disaggregate entities into homogeneous
subgroups based on their similarities on a profile of information.
For example,
what are the different psychographic segments of self-medicated drug users;
or what are the different benefit segments among bank customers?
The objective
in cluster analysis is to meaningfully classify a group of entities into
There are many differ-
clusters based on some judgmental or statistical rule.
ent algorithms proposed for cluster analysis and very few have any statistical
inferential properties so that cluster analysis is more
statistical technique.
a
heuristic than
a
However, it does provide insights into the typology
or segments present in the data.
Thus, it is possible to find psychographic
segments such as hypochondriacs, skeptics, realists and authority-seekers in
the self-medicated drug case [8]
,
and segments such as social interaction-
oriented, banking tasks-oriented and money borrowing-oriented customers in the
bank services case.
A more recent structural multivariate method is multidimensional scaling
which enables the researcher to explore and infer underlying criteria or
dimensions that people utilize to form
perceptions about similarities between,
and preferences among, various products or services.
For example, how do
people judge similarities among automobiles or toothpastes?
in multidimensional scaling is to map the alternatives in
a
The objective
multidimensional
space in such a way that their relative positions in the space reflect the
degree of perceived similarity between alternatives.
In the process, it
provides the researcher insights into the complexity or the number of salient
criteria which underlie a person's judgment.
Thus, prestige and styling may
be the most salient criteria a person uses when he compares various automobiles.
-9-
Similarly, decay prevention and brightening of teeth may be the two criteria
underlying his judgment about various brands of toothpaste.
Research Needs of Management
The above nonstatistical description of multivariate methods and their
applications to marketing research problems clearly suggests that they are
highly useful and relevant to marketing.
Now let us also look at the potential
usefulness of multivariate methods from the perspective of the major types of
research needs or inputs for managerial planning.
There are four types of
research inputs one generally encounters in an organization.
1.
They are:
Diagonostic research which provides a snapshot representation of the
present realities related to products or customers.
2.
Prognostic research which trends or forecasts the position in which
organization's products or customers are likely to be in sometime in the
future.
3.
Strategy research by which the organization can assess possible
impact of changes in actionable programs on market behavior.
This often
takes the form of either field experiments or laboratory type simulations
A.
Statistical research to ensure that the quality and quantity of
information to be analyzed is least biased and most satisfactorily
calibrated.
The statistical research essentially concerns itself with
questions of sampling and nonsampling errors in data, and how to adjust
for them by way of analytical strategies.
Table
I
summarizes the linkage between these four types of research needs
and specific multivariate methods relevant to each of those needs.
The
diagnostic market research is generally exemplified by three areas of research.
-IO-
TA BLE
I
Linking Research Needs With Multivariate Methods
Research Needs
A.
Diagnostic Research
Structural Methods
1.
Market Segmentation
Cluster or Factor Analysis
2.
Product or Corporate Typology
Factor or Cluster Analysis
3.
Customer Perceptions &
Multidimensional Scaling or
Preferences
B.
Multivariate Methods
Prognostic Research
1.
Sales Forecasting
Conjoint Measurement
Functional Methods
Multiple Regression or
Canonical Correlation
2.
Market Potentials
Multiple Discriminant
Analysis
C.
Strategy Research
1.
Field Experiments
Functional Methods
MANOVA or Discriminant
Analysis
2.
Laboratory Simulation
MANOVA or Discriminant
Analysis
D.
Statistical Research
1.
Heterogeneity Reduction
Structural Methods
Cluster Analysis or Factor
Analysis
2.
Measurement Errors
Factor Analysis or Multi-
dimensional Scaling
3.
Indexing or Data Consistency
Factor Analysis
4.
Normal Distributions
Factor Analysis
-11-
The first is market segmentation based on some relevant information such as
the psychographics, the demographics or the consumption patterns for which both
factor analysis and cluster analysis are most appropriate techniques.
The
second area consists of product, brand or company typology or imagery for which
also factor analysis or cluster analysis are useful techniques.
The third
type of diagnostic market research deals with the why aspect of customer
perceptions and preferences about products for which multidimensional scaling
techniques are quite relevant.
The prognostic market research is exemplified by at least two types of
predictive activities.
The first is the forecasting research related to
company, industry or product sales either as a time series analysis or as
complex function of environmental and organizational factors.
a
As we discussed
earlier, multiple regression and canonical correlation are directly relevant
for this area of prognostic research.
The second area of research entails
estimates of market potentials for new products as well as customer segmentation for existing products.
In short, this type of research is directly
related to various aspects of the product life cycle.
The techniques of
multiple discriminant analysis are directly relevant for this type of research.
The strategy market research often entails field experimentation or test
marketing.
It involves systematic manipulation of marketing mix in selected
markets in order to assess their impact on market behavior such as awareness,
attitudes and purchase behavior.
It is obvious that multivariate analysis of
variance is directly relevant here.
The final category is statistical research.
different aspects of data error and consistency.
There are at least four
The first is the question
-12-
of heterogeneity.
While the present sampling theory
in obtaining a representative sample,
a
homogeneous sample.
very useful to assist
is
there is nothing comparable to ensure
On the other hand,
a
heterogeneous sample has
a
direct
adverse effect on the correlation coefficient which is often reduced to
statistical artifact of aggregating apples and oranges so to speak.
a
The
techniques of clustering and factor analysis are, therefore, often used as
intermediate stages of analysis to provide insights into the heterogeneity
problem.
Another area of statistical research is concerned with the question
of nonsampling measurement errors inevitable in marketing data.
Often it
becomes essential to eliminate this error from the data by making appropriate
transformations of the data.
Once again, factor analysis and multidimensional
scaling become very useful intermediate procedures to remove the measurement
error from the data.
The third area is concerned with the question of con-
Often it is impossible to represent
sistency of data.
such as attitudes or brand loyalty by a single scale.
essential to use
a
a
complex phenomenon
It becomes, therefore,
variety of indicators which then must be properly indexed
to produce a composite score.
relevant for indexing purposes.
Once again, factor analysis becomes highly
Finally, often the raw data does not meet
certain statistical assumptions of functional models.
true with respect to the normality assumption.
This is especially
With the use of multivariate
methods, it is possible to transform the data so that they are more normally
distributed.
HOW TO GET THE MOST OUT OF MULTIVARIATE METHODS
While multivariate methods have direct relevance to marketing problems
as we discussed above,
it is not easy to successfully implement them in the
-13-
research program of Che organization due to their novelty, complexity and
variety.
Therefore, a number of practical guidelines are described below
which should be followed by the researcher if he is committed to the idea of
integrating multivariate methods in his research program.
First, try not to be technique-oriented.
It is not uncommon to find
researchers who are comfortable with, and experienced in,
multivariate method
a
particular
such as multidimensional scaling factor analysis or
multiple regression and try to use that technique across all research problems.
They seem to be literally in search for problems which will fit the technique
rather than the other way around.
Often this leads to redefinition of the
problem just so it meets the specifications of the technique.
No single
technique can solve all research problems, however, and this "Tom Swift and
his electronic machine" attitude has resulted in many misapplications of multi«
variate methods.
While it is easy to explain this attitude as due to narrow
specializations and discipline biases, it is highly hazardous to the longIn fact, this technique-
term survival of multivariate methods in marketing.
oriented myopic attitude of the researcher may well become the cause for the
downfall of multivariate methods just as it did for operations research models
several years ago
[6].
Second, consider multivariate models as information inputs to managerial
decisions rather than as their substitutes.
Often
a
researcher gets carried
away in building models and attempts to replace managerial judgment with the
model.
Unfortunately this is suicidal in view of the fact that marketing
research is only
a
staff function whose legitimate role is to provide the
necessary inputs for managerial decisions.
Most managers tend to be
satisficers rather than optimizers given the complexity of decisions and
-14-
being continuously pressed for time.
They regard research as useful input to
their judgmental process but do not wish their judgment skills to be replaced
by models and computers.
In short, it is in the best interest of the research-
er to be customer-oriented where his customers are the managers.
Third, multivariate methods or any other technique are not substitutes
for researcher's skills and imagination in the proper design of the study.
Statistics has nothing to do with causality and can never replace prior theory
or experimental design.
Unless the problem is adequately conceptualized, it
is very easy in today's world of fast, efficient and inexpensive computerized
calculations to evoke the GIGO principle (garbage in
-
gospel out)!
Fourth, half the battle in market research is proper communication of
techniques and display of results.
It is not at all uncommon to find a
brilliant researcher totally competent in multivariate analysis whom the
management or even others in the research department simply cannot understand.
His communication about the beta weights, heteroscedasticity eigenvalues,
varimax rotations, vectors, configurations and Kruskal's Stress are Chinese
and Sanskrit to the management.
Consequently, the most carefully designed
study with highly relevant results for managerial planning go wasted because
the management simply cannot understand let alone utilize them as inputs to
its decision-making process.
It is indeed a sad state of affairs in marketing
research that too little emphasis is placed on the art and science of display
and communication and too much emphasis is placed on the marginal elegancies
of techniques and computer programs.
Fifth, avoid making statistical inferences about the parameters of multi-
variate models.
It is simply impossible in social sciences due to the sub-
stantial existence of nonsampling or measurement errors in the data.
No
-15-
sampling theory can as yet offset this nonsampling error even if one has the
resources to sample the total population.
Furthermore, it is not easy to
apply sampling procedures in social sciences where often we don't know the
population itself.
Unfortunately, too often multivariate methods have been
criticized, chastized and even discarded as irrelevant tools and techniques
because it is impossible to make statistical inferences.
While it is true
that multivariate methods require far more stringent requirements of multi-
variate normal distributions, it should be pointed out that distribution
assumptions underlying statistical techniques even in the univariate and
bivariate analysis are also impossible to meet in marketing research.
A better strategy, therefore, is not to discard the techniques as
irrelevant but to put them to use for other purposes such as for making substantive inferences or as descriptive statistical techniques by which large
data sets can be reduced to meaningful and concise summaries for managerial
inputs.
In other words, multivariate methods are more useful as data trans-
formation, data reduction and as data display techniques than as mathematical
models.
This is not the fault of the techniques but the limitations of exist-
ing methods of data collection.
Sixth, guard yourself against the danger of making substantive inferences
about market realities which may be an artifact solely due to the peculiarities
of a particular multivariate method.
Since multivariate methods are more
complex statistical procedures, there are many more underlying assumptions
required for the optimization (minimization or maximization) of statistical
decision rules.
Consequently, it is easier to inject substantive meanings in
the data even if the data are essentially random relationships.
This has been
especially true of those multivariate methods such as cluster analysis,
-16-
mul tridimensional scaling and conjoint measurement which possess no underlying
sampling theory, and therefore, are essentially heuristics often no better
than naive judgmental rules.
In order to guard against this danger, it is recommended that the same
data be subjected to at least two different techniques.
Often, this may be
limited to two or more variations of the same basic multivariate method.
The
replication principle underlying this recommendation will at least bring to the
researcher's attention the presence of
a
technique artifact in his data analysis
Finally, exploit the complimentary relationship inherent between the
structural and the functional multivariate methods.
For example, it is
extremely advantageous to subject the original predictor variables to a
factor analysis and utilizing the transformed factor scores as derived pre-
dictor variables in a multiple regression because it makes the data more
matching the requirements of lack of multicollinearity and nonsampling error
and the presence of normality of the distribution.
Similarly, it is best to
utilize cluster analysis first to define the number of mutually exclusive
groups or segments before attempting a multiple discriminant analysis.
In
short, this guideline urges the researcher to replace or at least substantiate
a
number of judgments he has to make in order to build functional multivariate
methods, with a structural multivariate analysis of the data.
For often the
researcher's judgment is highly tenuous and sometimes patently wrong which
increases the probability of building less useful multivariate models.
In conclusion, multivariate methods are highly relevant to marketing
problems.
However, due to lack of familiarity with them, their innate com-
plexity and large variety, it is easy to misapply these techniques.
Several
-17-
practical suggestions have been made in the paper to increase the likelihood
of getting more out of the multivariate methods.
Perhaps the single most
guideline to recommend is: don't be enamoured by them.
-13-
REFERENCES
1.
"Some Correlates of Coffee and Cleanser Brand Shares,"
Banks, Seymour.
Journal of Advertising Research
2.
,
1
(June 1961), 22-28.
Handbook of Marketing Research
Ferber, Robert.
.
New York: McGraw-Hill
Book Company, 1974.
3.
Gatty, Ronald.
"Multivariate Analysis for Marketing Research: An
Evaluation," Applied Statistics
4.
Sheth, Jagdish N.
,
Marketing
6.
,
35
,
15
(November 1966), 146-58.
"Multivariate Analysis in Marketing," Journal of
Advertising Research
5.
,
,
10 (February 1970),
29-39.
"Multivariate Revolution in Marketing Research," Journal of
(January 1971),
13-19.
"Some Thoughts on the Future of Marketing Models,"
unpublished Faculty Working Paper No. 232, February 1975, University of
Illinois.
7.
,
Multivariate Methods For Marketing Research
,
Chicago:
American Marketing Association (in press).
8.
Ziff, Ruth, "Psychographics for Market Segmentation," J ournal of
Advertising Research
,
11 (April 1971),
3-10.
UNIVERSITY OF ILLINOIS-URBANA
3 0112 060296701
EH
Irani
wBBBu&
ns
HP
BUMPS
H
B
Sffitt
u
HI
HIV