GIRISH PUNJ and DAVID W.

STEWART*

Applications of cluster analysis to marketing problems are reviewed. Alternative

methods of cluster analysis are presented and evaluated in terms of recent empirical
work on their performance characteristics. A two-stage cluster analysis methodology
is recommended: preliminary identification of clusters via Ward's minimum variance
method or simple average linkage, followed by cluster refinement by an iterative
partitioning procedure. Issues and problems related to' the use and validation of
cluster analytic methods are discussed.

Cluster Analysis in Marketing Research: Review

and Suggestions for Application

Cluster analysis has become a common tool for the of the confusing array of names and methods of cluster
marketing researcher. Both the academic researcher and analysis confronting the marketing researcher. As this
the marketing applications researcher rely on the tech- confusion is resolved and as additional information
nique for developing empirical groupings of persons, about the performance characteristics of various cluster-
products, or occasions which may serve as the basis for ing algorithms becomes available, such skepticism may
further analysis. Despite its frequent use, little is known disappear. Recent work on clustering algorithms affords
about the characteristics of available clustering methods a basis for establishing some general guidelines for the
or how clustering methods should be employed. One in- appropriate use of cluster analysis. It is useful to note
dication of this general lack of understanding of clus- that many of the problems associated with cluster anal-
tering methodology is the failure of numerous authors ysis also plague multivariate statistics in general: choice
in the marketing literature to specify what clustering of an appropriate metric, selection of variables, cross-
method is being used. Another such indicator is the ten- validation, and external validation.
dency of some authors to differentiate among methods Two general sets of issues confront the marketing re-
which actually differ only in name. searcher seeking to use cluster analysis. One set of issues
The use of cluster analysis has frequently been viewed involves theoretical properties of particular algorithms.
with skepticism. Green, Frank, and Robinson (1967) These issues are considered in the literature on cluster
and Frank and Green (1968) have discussed problems analysis (Anderberg 1973; Bailey 1974; Cormack 1971;
with determining the appropriate measure of similarity Hartigan 1975), and are not addressed here. The second
and the appropriate number of clusters. Inglis and John- set of issues are more practical and pertain to the actual
son (1970), Morrison (1967), Neidell (1970), and use of clustering procedures for data analysis. These is-
Shuchman (1967) have also expressed concern about the sues are the foci of our article, in which we review ap-
use of cluster analysis. More recently, Wells (1975) has plications of clustering methodology to marketing prob-
expressed reservations about the use of cluster analysis lems, provide a systematic treatment of the clustering
unless very different, homogeneous groups can be iden- options open to the marketing researcher, and use both
tified. Such skepticism is probably justified in the light theoretical and empirical findings to suggest which clus-
tering options may be most useful fer a particular re-
search problem.
*Girish Punj is Assistant Professor of Marketing, School of Busi-
ness Administration, University of Connecticut. David W. Stewart is
Cluster analysis has most frequently been employed
Assistant Professor of Management, Owen Graduate School of Man- as a classification tool. It has also been used by some
agement, Vanderbilt University. researchers as a means of representing the structure of
The authors acknowledge the support of the 1981 Dean's Fund of data via the construction of dendrograms (Bertin 1967;
the Owen Graduate School of Management, Vanderbilt University, Hartigan 1967) or overlapping clusters (Arabie et al.
which facilitated the completion of this article.
1981; Shepard and Arabie 1979). The latter applications

134

Journal of Marketing Research

Vol. XX (May 1983), 134-48
CLUSTER ANALYSIS IN MARKETING RESEARCH 135

are distinct from the use of cluster analysis for classifi- market segmentation research. In concluding his review
cation and represent an alternative to multidimensional of the segmentation literature, Wind suggests that one
scaling and factor analytic approaches to representing important area of future research should be the "eval-
similarity data. Whereas classification is concerned with uation of the conditions under which various data ana-
the identification of discrete categories (taxonomies), lytical techniques are most appropriate" (1978, p. 334).
structural representation is concerned with the develop- All segmentation research, regardless of the method
ment of a faithful representation of relationships. Both used, is designed to identify groups of entities (people,
uses of cluster analysis are legitimate, but the objectives markets, organizations) that share certain common char-
of these applications are very different. The best clus- acteristics (attitudes, purchase propensities, media hab-
tering algorithm for accomplishing one of these objec- its, etc.). Stripped of the specific data employed and the
tives is not necessarily the best for the other objective. details of the purposes of a particular study, segmenta-
We restrict our treatment of cluster analysis to the more tion research becomes a grouping task. Wind (1978)
common of the two applications, classification. notes that researchers tend to select grouping methods
Cluster analysis is a statistical method for classifica- largely on the basis of familiarity, availability, and cost
tion. Unlike other statistical methods for classification, rather than on the basis of the methods' characteristics
such as discriminant analysis and automatic interaction and appropriateness. Wind attributes this practice to the
detection, it makes no prior assumptions about important lack of research on similarity measures, grouping (clus-
differences within a population. Cluster analysis is a tering) algorithms, and effects of various data transfor-
purely empirical method of classification and as such is mations.
primarily an inductive technique (Gerard 1957). Though A second and equally important use of cluster analysis
some theorists have not been favorably disposed toward has been in seeking a better understanding of buyer be-
the use of cluster analysis, and criticism of the ad hoc haviors by identifying homogeneous groups of buyers.
nature of clustering solutions is common, classification Cluster analysis has been less frequently applied to this
is an important and frequently overlooked tool of sci- type of theory-building problem, possibly because of
ence. Wolf (1926) has suggested that classification is theorists' discomfort with a set of procedures which ap-
both the first and last method employed by science. The pear ad hoc. Nevertheless, there is clearly a need for
essence of classification is that certain things are thought better classification of relevant buyer characteristerics.
of as related in a certain way. Indeed, the final outcome Bettman (1979) has called for the development of tax-
of other methods of study may well be a new classifi- onomies of both consumer choice task and individual
cation. difference characteristics. Cluster analysis is one means
Kemeny (1959) and Kantor (1953), discussing the for developing such taxonomies. Examples of such use
philosophy of science, point to the fundamental impor- may be found in articles by Claxton, Fry, and Portis
tance of classification. Wolf (1926) holds that verifica- (1974), Kiel and Layton (1981), and Furse, Punj, and
tion of laws of science may occur only after classifica- Stewart (1982).
tion has been completed. Thus, whether the classification Cluster analysis has been employed in the develop-
exercise is completed explicitly or implicitly, it must ment of potential new product opportunities. By clus-
occur. Cluster analysis provides one, empirically based, tering brands/products, competitive sets within the
means for explicitly classifying objects. Such a tool is larger market structure can be determined. Thus, a firm
particularly relevant for the emerging discipline of mar- can examine its current offerings vis-a-vis those of its
keting which is still wrestling with the problems of how competitors. The firm can determine the extent to which
best to classify consumers, products, media types, and a current or potential product offering is uniquely posi-
usage occasions. tioned or is in a competitive set with other products
(Srivastava, Leone, and Shocker 1981; Srivastava,
USES OF CLUSTER ANALYSIS IN MARKETING Shocker, and Day 1978). Although cluster analysis has
The primary use of cluster analysis in marketing has not been used frequently in such applications, largely
been for market segmentation. Since the appearance of because of the availability of other techniques such as
Smith's now-classic article (1956), market segmentation multidimensional scaling, factor analysis, and discrimi-
has become an important tool for both academic research nant analysis, it is not uncommon to find cluster analysis
and applied marketing. In a review of market segmen- used as an adjunct to these other techniques. Cluster
tation research and methodology, Wind (1978) identifies analysis has also been suggested as an alternative to fac-
both the impact of this most fundamental of marketing tor analysis and discriminant analysis. In such applica-
tools and some rather significant problem areas. Not the tions it is important for the analyst to determine whether
least of these problems is the plethora of methods that discrete categories of products are desirable or whether
have been proposed for segmenting markets. This mul- a representation of market structure is desirable. The lat-
tiplication of techniques has served to confuse many ter may be more useful in many market structure appli-
marketers, shift discussions of researchers from more cations, in which case cluster analysis would not be used
substantive issues to issues of method, and impede the as a classification technique and the analyst would face
development of meta-research directed at integrating a different set of issues from those addressed here.
136 JOURNAL OF MARKETING RESEARCH, MAY 1983

Cluster analysis has also been employed by several clustering algorithms. Though they mention some em-
researchers in the problem of test market selection pirical work on the characteristics of these measures and
(Green, Frank, and Robinson 1967). Such applications algorithms, their report is primarily a catalog of tech-
are concerned with the identification of relatively ho- niques and some marketing applications. Relatively little
mogeneous sets of test markets which may become in- guidance is provided the researcher seeking to discover
terchangeable in test market studies. The identification the characteristics and limitations of various grouping
of such homogeneous sets of test markets allows gen- procedures. Indeed, the Sherman and Sheth report may
eralization of the results obtained in one test market to mislead some readers because its categorization of clus-
other test markets in the same cluster, thereby reducing tering algorithms suggests substantive differences among
the number of test markets required. identical algorithms which differ only in name. Frank
Finally, cluster analysis has been used as a general and Green (1968) also provide an introduction and re-
data reduction technique to develop aggregates of data view of clustering methodology but make no specific
which are more general and more easily managed than recommendations to guide the user of the methodology.
individual observations. For example, limits on the num- After reviewing the problems and issues facing the user
ber of observations that can be used in multidimensional of cluster analytic procedures, we offer clarification of
scaling programs often necessitate an initial clustering the similarities and differences among various clustering
of observations. Homogeneous clusters then become the algorithms and some suggestions about their use.
unit of analysis for the multidimensional scaling proce-
dure. Fisher (1969) discussed the use of cluster analysis PROBLEMS IN USING CLUSTER ANALYSIS
for data reduction from the perspective of econometrics Unlike other data analytic methods, cluster analysis
and argued that cluster analysis is most appropriate is a set of methodologies that has developed outside a
whenever the data are too numerous or too detailed to single dominant discipline. Factor analysis and various
be manageable. Such data simplification and aggrega- scaling methods were developed within the discipline of
tion are carried out for the convenience of the investi- psychology and one would look to that discipline for
gator rather than in the interest of theory building. guidance in the use of these methods. Regression,
Table 1 is a brief description of some recent appli- though used in a variety of disciplines, has tended to be
cations of cluster analysis to marketing problems. Al- the special province of econometricians, who have de-
though not a complete set of all applications of cluster veloped a large body of literature on the technique. In
analysis in marketing, it illustrates several points. First, contrast, no single discipline has developed and retained
the array of problems addressed by these studies is strik- clustering methodology. Rather, numerous disciplines
ing. Equally striking is the diversity of clustering meth- (econometrics, psychology, biology, and engineering)
ods employed. In constructing this table we had diffi- have independently approached the clustering problem.
culty discerning the specific clustering algorithm used Often working in parallel, researchers in these disci-
by the researchers. Cluster analysis methods were often plines have arrived at similar solutions but have given
identified by the name of the program used, e.g., them different names. For example, Blashfield (1978)
BMDP2M, BCTRY, or Howard and Harris, rather than reviewed the literature on hierarchical clustering meth-
by the specific clustering algorithm used. Only by con- ods and found as many as seven different names for the
sulting a particular program's manual could we identify same technique. This diversity of names for identical
the method actually employed. For one of the studies techniques has tended to prevent comparisons of algo-
cited we could not find any information on the clustering rithms across disciplines. It has also served to confuse
method used. the data analyst by implying a much greater number of
The lack of specificity about the method of clustering available clustering methods than actually exists.
employed in these studies is illustrative of the problems Also confronting the potential user of cluster analysis
associated with the use of cluster analysis. The lack of is the problem of cluster definition. There are currently
detailed reporting suggests either an ignorance of or lack no clear guidelines for determining the boundaries of
of concern for the important parameters of the clustering clusters or deciding when observations should be in-
method used. Failure to provide specific information cluded in one cluster or another. Cattell (1978) has sug-
about the method also tends to inhibit replication and gested that clusters are "fuzzy" constructs. The crite-
provides little guidance for other researchers who might rion for admission to a cluster is rather arbitrary. There
seek an appropriate method of cluster analysis. Use of are no well-established rules for the definition of a clus-
specific program names rather than the more general ter. The preferred definition of a cluster seems to vary
algorithm name impedes interstudy comparisons. with the discipline and purpose of the researcher.
This situation suggests a need for a sound review of Clusters have most frequently been defined by rela-
clustering methodology for the marketing researcher. tively contiguous points in space (Stewart 1981). Cor-
Previous reviews on this subject appeared prior to the mack (1971) suggested that clusters should exhibit two
publication of much of the research on the performance properties, external isolation and internal cohesion. Ex-
characteristics of clustering algorithms. Sherman and ternal isolation requires that objects in one cluster be
Sheth (1977) discuss selected similarity measures and separated from objects in another cluster by fairly empty
CLUSTER ANALYSIS IN MARKETING RESEARCH 137

Table 1
SOME RECENT APPLICATIONS OF CLUSTER ANALYSIS IN MARKETING

Application Purpose of research Nature of data Clustering method used

Anderson, Cox, and To identify the determinant Determinant attribute scores on Iterative partitioning-MIKCA
Fulcher (1976) attributes in bank selection several bank selection variables (McRae 1973)
decisions and use them for
segmenting commercial bank
customers
Bass, Pessemier, and To identify market segments with Attribute scores on several media Average linkage cluster analysis
Tigert (1969) respect to media exposure exposure variables (Sneath and Sokal 1973)
Calantone and Sawyer To examine the stability of Attribute scores on several bank K-means (Howard and Harris
(1978) market segments in the retail selection variables 1966)
banking market
Claxton, Fry, and Portis To classify furniture and Attribute scores on several Complete linkage cluster analysis
(1974) appliance buyers in terms of prepurchase activity measures (Johnson 1967; Lance and
their information search Williams 1967a)
behavior
Day and Heeler (1971) To classify stores into similar Factor scores on several store (1) Complete linkage analysis
strata attributes (2) Iterative partitioning (Rubin
1965)
Green, Frank, and To identify matched cities for test Factor scores on several city Average linkage cluster analysis
Robinson (1967) marketing characteristics (Sneath and Sokal 1973)
Greeno, Sommers, and To identify market segments with Q sorts on 38 product items Ward's minimum variance
Kernan (1973) respect of personality variables method (Ward 1963)
and implicit behavior patterns
Kernan (1968) To identify groups of people Scores on several personality and Ward's minimum variance
along several personality and decision traits method (Ward 1963)
decision behavior
characteristics
Kernan and Bruce (1972) To create relatively homogeneous Characteristics of census traits Ward's minimum variance
configuration of census traits method (Ward 1963)
Kiel and Layton (1981) To develop consumer taxonomies Factor scores derived from Average linkage cluster analysis
of search behavior by several search variables (Sneath and Sokal 1973)
Australian new car buyers
Landon (1974) To identify groups of people Scores on self-image and Iterative partitioning (BCTRY;
using purchase intention and purchase intention variables Tryon and Bailey 1970)
self-concept variables
Lessig and Tollefson To identify similar groups of Scores on several buyer behavior Ward's minimum variance
(1971) consumers along several buyer variables method (Ward 1963)
behavior variables
Montgomery and Silk To identify opinion leadership Scores on several interest and Complete linkage cluster analysis
(1971) and consumer interest segments opinion leadership variables (Johnson 1967)
Moriarty and Venkatesan To segment educational Importance scores on financial- K-means (Howard and Harris
(1978) institutions in terms of benefits aid management services 1966)
sought when purchasing
financial-aid MIS
Morrison and Sherman To determine how various Ratings of advertisements by Iterative Partitioning (Friedman
(1972) individuals interpret sex appeal respondents and Rubin 1967)
in advertising
Myers and Nicosia To develop a consumer typology Scores of supermarket image Iterative partitioning (BCTRY;
(1968) using attribute data variables Tryon and Bailey 1970)
Sethi (1971) To classify world markets Macrolevel data on countries Iterative partitioning (BCTRY;
Tryon and Bailey 1970)
Sexton (1974) To identify homogeneous groups Brand and product usage rate Type not specified
of families using product and data
brand usage data
Schaninger, Lessig, and To identify segments of Scores on several product usage K-means (Howard and Harris
Panton (1980) consumers on the basis of variables 1966)
product usage variables
Green and Carmone To identify similar computers Performance measures for K-means (Howard and Harris
(1968) (strata in the computer market) different computer models 1966)

space. Internal cohesion requires that objects within the provide for overlapping clusters. Although a few algo-
same cluster be similar to each other. Everitt (1974) of- rithms have been developed for identifying overlapping
fered a similar concept which he defines as a natural clusters (Jardine and Sibson 1971; Peay 1975; Shepard
cluster. The requirement of external isolation does not 1974), these methods are primarily concerned with the
138 JOURNAL OF MARKETING RESEARCH, MAY 1983

representation of structure rather than classification. Ap- chology literature have forcefully argued that factor
plications of these methods have been few and are not analysis is inappropriate as a method for identifying
reviewed here. clusters. Skinner (1979) discusses some relationships
In the absence of a generally accepted or definitive between factor analysis and cluster analysis. AID is not
definition of a cluster, various algorithms have been de- included because it operates on a rather different prin-
veloped which offer particular operational definitions. ciple than the clustering procedures. AID requires the
Differences among clustering algorithms are frequently prior specification of independent and dependent vari-
related to how the concept of a cluster is operationalized. ables and seeks to identify sets of nominal independent
Thus, to develop a set of recommendations for the ap- variables which group observations in a manner that
plication of cluster analysis, we must first develop a rec- minimizes the variance of the dependent variable within
ognition of the clustering algorithms available to the each group. Cluster analysis procedures require no such
marketing researcher and an understanding of the per- a priori specification of independent and dependent vari-
formance of these methods in relation to one another. ables.
These clustering algorithms exist in various forms but
CLUSTERING ALGORITHMS most have been programmed. Several software programs
Table 2 provides a description of the more common are currently available for cluster analysis. They differ
clustering algorithms in use, the various alternative in their comprehensiveness and ease of use. Table 3
names by which the algorithms are known, and a brief briefly describes several of the more common clustering
discussion of how clusters are formed by each of these software programs; identifies the types of clustering
methods. Table 2 shows clearly that there are relatively methods available within each program, and cites the
few clustering methods from which to choose, far fewer original source of the program. Selecting an appropriate
than one might suspect from a reading of the literature cluster analytic method or software package requires
on cluster analysis. Four primary hierarchical methods some knowledge of the performance characteristics of
are available, single linkage, complete linkage, average the various methods.
linkage, and Ward's minimum variance method. Al-
though there are several variations of the average linkage EMPIRICAL COMPARISONS OF CLUSTERING
method, only one, simple average linkage, is widely METHODS
used. In addition, two variants of the average method, One method for evaluating clustering methods that has
the centroid and median methods, have very undesirable been used with increasing frequency involves comparing
properties (Aldenderfer 1977; Sneath and Sokal 1973) the results of different clustering methods applied to the
which recommend against their use. The weighted av- same data sets. If the underlying characteristics of these
erage linkage method has been shown to produce results data sets are known, one can assess the degree to which
very similar to those produced by the simple average each clustering method produces results consistent with
method (Blashfield 1977). these known characteristics. For example, if a data set
There is more variety among the nonhierarchical consists of a known mixture of groups, or subpopula-
methods, though all work on similar principles. These tions, the efficacy of a cluster solution can be evaluated
iterative partitioning methods begin by dividing obser- by its success in discriminating among these subpopu-
vations into Some predetermined number of clusters. lations. This mixture model approach to the evaluation
Observations are then reassigned to clusters until some of clustering algorithms has recently been employed by
decision rule terminates the process. These methods may several researchers. Table 4 summarizes the findings of
differ with respect to the starting partition, the type of 12 such studies.
reassignment process, the decision rule used for termi- The number of clustering algorithms, distance mea-
nating clustering, and the frequency with which cluster sures, and types of data that might be incorporated in
centroids are updated during the reassignment process. a mixture model study is so large as to preclude anyone
The initial partition may be random or based on some comprehensive study of the relative efficacy of cluster-
prior information or intuition. One method (MIKCA) ing methods. We can look across the studies in Table
uses several different random starting partitions to en- 4, however, and begin to draw some conclusions about
sure an efficient solution. Two types of reassignment are clustering methods. Three procedures seem to warrant
generally employed, K-means and hill-climbing. These special consideration. Ward's minimum variance method,
methods are briefly discussed in Table 2 as are the ter- average linkage, and several variants of the iterative par-
mination decision rules used with each method. Cluster titioning method appear to outperform all other methods.
centroids may be updated after each membership move Ward's method appears to outperform the average link-
or only after a complete pass through the entire data set. age method except in the presence of outliers. K-means
Not included in Table 2 are two methods frequently appears to outperform both Ward's method and the av-
used for cluster analysis: Q factor analysis and automatic erage linkage method if a nonrandom starting point is
interaction detection (AID) (Morgan and Sonquist 1963). specified. If a random starting point is used K-means
Q factor analysis is not included because Stewart (1981) may be markedly inferior to other methods, but results
in the marketing literature and Cattell (1978) in the psy- on this issue are not consistent. Nevertheless, the K-
CLUSTER ANALYSIS IN MARKETING RESEARCH 139

Table 2
CLUSTERING METHODS

Primary name Alternative names Method offorming clusters

Hierarchical methods
Single linkage cluster analysis Minimum method (Johnson 1967); link- An observation is joined to a cluster if it has a certain
age analysis (McQuitty 1967); nearest level of similarity with at least one of the members of
neighbor cluster analysis (Lance and that cluster. Connections between clusters are based on
Williams 1967a); connectiveness links between single entities.
method (Johnson 1967)
Complete linkage cluster analysis Maximum method (Johnson 1967); rank An observation is joined to a cluster if it has a certain
order typal analysis (McQuitty 1967); level of similarity with all current members of the
furthest neighbor cluster analysis cluster.
(Lance and Williams 1967a); diameter
method (Johnson 1967)
Average linkage cluster analysis Simple average linkage analysis (Sneath These are actually four similar methods. In all four meth-
and Sokal 1973); weighted average ods an observation is joined to a cluster if it has a cer-
method (McQuitty 1967); centroid tain average level of similarity with all current mem-
method (Gower 1967); median bers of the clusters. These methods differ in the
method (Lance and Williams 1967a) manner in which the average level of similarity is de-
fined. The weighted average method and median
method provide for an a priori weighting of the aver-
ages based on the number of entities desired in each
cluster. The centroid method provides for an initial
computation of the centroid of each cluster. Average
similarity is based on this centroid. Only the simple
average linkage procedure has been widely used.
Minimum variance cluster analy- Minimum variance method; Ward's The minimum variance method is designed to generate
sis method; error sum of squares method clusters in such a way as to minimize the within-clus-
(Ward 1963); HGROUP (Veldman ter variance. Unlike other hierarchical clustering meth-
1967) ods, Ward's method optimizes an objective statistic: it
seeks to minimize tr W, where W is the pooled
within-clusters sum of squares and cross-products ma-
trix. Ward's method is somewhat similar to the aver-
age method in that variance is a function of deviations
from the mean. Some authors have included Ward's
method as a special case of the average method (Bai-
ley 1974). It is an average linkage method because it
does not seek to minimize distance between one mem-
ber of the cluster and the entity, or all members of the
cluster and the entity as in single linkage and complete
linkage, respectively, but minimizes the average dis-
tance within the cluster.
Iterative partitioning methods These methods begin with the partition of observations
(nonhierarchical methods) into a specified number of clusters. This partition may
be on a random or nonrandom basis. Observations are
then reassigned to clusters until some stopping crite-
rion is reached. Methods differ in the nature of the
reassignment and stopping rules.
K-means Cases are reassigned by moving them to the cluster
whose centroid is closest to that case. Reassignment
continues until every case is assigned to the cluster
with the nearest centroid. Such a procedure implicitly
minimizes the variance within each cluster, tr W.
Hill-climbing methods Cases are not reassigned to the cluster with the nearest
centroid but are moved from one cluster to another if a
particular statistical criterion is obtained. Reassignment
continues until optimization occurs. The objective
function to be optimized may be selected from one of
four options, tr W, tr [(W-1BJ, Iwl, and the largest ei-
genvalue of [(W-1BJ, where W is the pooled within-
cluster covariance matrix and B is the between cluster
covariance matrix.
Combined K-means and hill- Uses a combination of K-means and hill-climbing meth-
climbing methods ods.
140 JOURNAL OF MARKETING RESEARCH, MAY 1983

Table 3
COMMON CLUSTERING PACKAGES/PROGRAMS

Name of
package/pro- Where
gram available? / authors Clustering methods' Comments
ANDERBERG In the appendices of book entitled Cluster S, C, A, W, K, H, KH 1. No missing value treatment
Analysis for Applications by M. R. An- 2. Only binary data type with octal coding
derberg (1973). scheme
3. User manual not available
BCTRY D. Bailey and R. C. Tryon, Tryon-Bailey K 1. No MANOVA statistics are optimized
Associates, Inc., c/o Mr. Peter Lenz, 2. Initial partition has to be user specified
2222 S.E. Nehalem St., Portland, OR 3. Factor analysis of variables may be per-
formed as well
BMDP W. J. Dixon (ed.), Health Sciences Com- S,C,A,K 1. Single and complete linkage available for
puting Facility, School of Medicine clustering
UCLA, Los Angeles, CA 2. Binary data not permissible
3. Continuous type similarity measure
4. Method for clustering cases and variables
simultaneously is available
5. User cannot supply only similarity matrix
for cases
CLUS H. Friedman and J. Rubin (1967 JASA ar- K,H I. Fixed number of clusters
ticle), IBM SHARE system 2. Expensive to use
CLUSTAN D. Wishart, Computer Centre, University S, C, A, W, K, H 1. High versatility (38 s/dis measures)
College of London, 19 Gordon St., 2. Initial partition for iterative partitioning
London, WCIH OAH, Great Britain methods has to be user specified
3. Binary and continuous data types
4. Binary data in 3 coding schemes
5. Variable transformations not available
6. Permits overlapping clusters
HARTIGAN In the appendices of book entitled Clus- S,A,K 1. Fixed number of clusters for iterative par-
tering Algorithms by J. J. Hartigan titioning methods
(1975). 2. No variable transformations available
3. No user manual available
HGROUP D. J. Veldman (1967), FORTRAN Pro- W 1. Part of the University of Texas EDSTAT
gramming for the Behavioral Sciences. statistics package
HICLUS S. C. Johnson (based on 1967 Psychome- SC I. No user manual available
trika article), Bell Telephone Labs, 2. No missing value treatment
Murray Hill, NJ 3. No standardization of variables
4. No transformation of variables
5. User must supply similarity matrix (hence
is versatile in some sense)
HOWD Britton Harris, F. J. Carmone, Jr., Uni- K 1. No user manual available
(Howard- versity of Pennsylvania, Philadelphia, 2. No MANOVA statistics
Harris) PA 3. Number of clusters fixed
ISODATA Daviel Wolf, SRI, 333 Ravenswood Ave- K 1. No user manual available
nue, Menlo Park, CA 2. No MANOVA statistics optimized
MIKCA D. J. McRae, Coordinator, Testing & K,H,KH I. No user manual available
Computer Applications, Jackson Public 2. 4 MANOVA statistics optimized
Schools, Jackson, MI 3. 3 different distance measures
NT-SYS F. James Rohlf, John Kishapugh, David S, C, A I. Permits overlapping clusters
Kirk, Dept. of Ecology and Evolution, 2. Alphanumeric coding scheme for binary
SUNY at Stony Brook, Stony Brook, data
NY 3. Moderately versatile
OSIRIS Institute of Survey Research, University C
of Michigan, Ann Arbor, MI
SAS James H. Goodnight, SAS Inst. Inc., P. C 1. Continuous similarity measure
O. Box 10066, Raleigh, NC
'S = single linkage.
C = complete linkage.
A = average linkage.
W = Ward's minimum variance method.
K = K-means.
H = hill climbing.
KH = joint K-means, hill climbing.
CLUSTER ANALYSIS IN MARKETING RESEARCH 141

Table 4
EMPIRICAL COMPARISONS OF THE PERFORMANCE OF CLUSTERING ALGORITHMS

Data sets
Reference Methods examined employed Coverage' Criteria Summary of results
Cunningham and Single, complete, aver- Normal mixtures Complete Measures of Average linkage outperformed
Ogilvie age linkage with Eu- "stress" to com- other methods
(1972) clidean distances and pare input simi-
Ward's minimum var- larity/ dissimilar-
iance technique ity matrix with
similarity rela-
tionship among
entities portrayed
by the clustering
method
Kuiper and Single, complete, aver- Bivariate normal mix- Complete Rand's statistic Ward's technique consistently
Fisher (1975) age, centroid, median tures (Rand 1971) outperformed other methods
linkage, all using Eu-
clidean distances and
Ward's minimum var-
iance technique
Blashfield Single, complete, aver- Multinormal mixtures Complete Kappa (Cohen Ward's technique demon-
(1976) age linkage, all using 1960) strated highest median ac-
Euclidean distance curacy
and Ward's minimum
variance technique
Mojena (1977) Simple average, Multivariate gamma Complete Rand's statistic Ward's method outperformed
weighted average, distribution mix- other methods
median, centroid, tures
complete linkage, all
using Euclidean dis-
tances and Ward's
minimum variance
technique
Blashfield Eight iterative partition- Multinormal mixtures Complete Kappa For 15 of the 20 data sets ex-
(1977) ing methods: Ander- amined, a hill-climbing
berg and CLUST AN technique which optimized
K-means methods, W performed best, i.e.,
each with cluster sta- MIKCA or CLUS. In two
tistics updated after other cases a hill-climbing
each reassignment method which optimized
and only after a com- tr W performed best,
plete pass through the CLUS.
data; CLUS and
MIKCA (both hill-
climbing algorithms),
each with optimiza-
tion of tr Wand W
Milligan and Single, complete aver- Data sets differing in Complete Rand's statistic and Average linkage and Ward's
Isaac (1978) age linkage, and degree of error per- kappa technique superior to single
Ward's minimum var- turbation and complete linkage
iance technique, all
using Euclidean dis-
tances
Mezzich (1978) Single, complete link- Psychiatric ratings Complete Replicability; agree- K-means procedure with Eu-
age, and K-means, ment with "ex- clidean distances performed
each with city-block pert" judges; best followed by K-means
and Euclidean dis- goodness of fit procedure with the city-
tances and correlation between raw in- block metric; average link-
coefficient, ISO- put dissimilarity age also performed well as
DATA, Friedman and matrix and matrix did complete linkage with a
Rubin method, Q fac- of O's and I's in- correlation coefficient &
tor analysis, multidi- dicating entities city-block metric & ISO-
mensional scaling clustered together OAT A; the type of metric
with city-block and used (r, city-block, or Eu-
Euclidean metrics and clidean distance) had little
correlation coeffi- impact on results.
cients, NORMAP /
NORMIX, average
142 JOURNAL OF MARKETING RESEARCH, MAY 1983

Table 4 (Continued)

Data sets
Reference Methods examined employed Coverage' Criteria Summary of results
linkage with correla-
tion coefficient
Edelbrock Single, complete, aver- Multivariate normal 70,80,90, Kappa Ward's method and simple
(1979) age, and centroid, mixtures, standard- 95, 100% average were most accu-
each with correlation ized & unstandard- rate; performance of all al-
coefficients, Euclid- ized gorithms deteriorated as
ean distances, and coverage increased but this
Ward's minimum var- was less pronounced when
iance technique the data were standardized
or correlation coefficients
were used. The latter find-
ing is suggested to result
from the decreased extrem-
ity of outliers associated
with standardization or use
of the correlation coeffi-
cient.
Edelbrock and Single, complete, aver- Multivariate normal 40,50,60, Kappa and Rand's Ward's method and the aver-
McLaughlin age, each with corre- mixtures & multi- 70,80,90, statistic age method using one-way
(1980) lation coefficients, variate gamma 95, 100% intraclass correlations were
Euclidean distances, mixtures most accurate; performance
one-way and two-way of all algorithms deterio-
intraclass correlations, rated as coverage increased.
and Ward's minimum
variance technique
Blashfield and Ward's minimum vari- Multivariate normal Varying levels Kappa Group average method best at
Morey (1980) ance technique, group mixtures higher levels of coverage;
average linkage, Q at lower levels of coverage
factor analysis, Lorr's Ward's method and group
nonhierarchical proce- average performed simi-
dure, all using Pear- larly.
son product moment
correlations as the
similarity measure
Milligan (1980) Single, complete, group Multivariate normal Complete Rand's statistic; the K-means procedure with a de-
average, weighted av- mixtures, standard- point biserial cor- rived point generally per-
erage, centroid & me- ized and varying in relation between formed better than other
dian linkage, Ward's the number of un- the raw input dis- methods across all condi-
minimum variance derlying clusters similarity matrix tions I. Distance measure
technique, minimum and the pattern of and a matrix of selection did not appear
average sum of distribution of O's and I's indi- critical; methods generally
squares, minimum to- points of the clus- eating entities robust across distance mea-
tal sum of squares, ters. Data sets clustering to- sures. 2. Presence of ran-
beta-flexible (Lance ranged from error gether dom dimensions produced
& Williams 1970a b), free to two levels decrements in cluster recov-
average link in the of error perturba- ery. 3. Single linkage
new cluster, Mac- tions of the dis- method strongly affected by
Queen's method, Jan- tance measures, error-perturbations; other
cey's method, K- from containing no hierarchical methods mod-
means with random outliers to two lev- erately so; nonhierarchical
starting point, K- els of outlier con- methods only slightly af-
means with derived ditions, and from fected by perturbations. 4.
starting point, all no variables unre- Complete linkage and
with Euclidean dis- lated to the clusters Ward's method exhibited
tances, Cattell's to one or two ran- noticeable decrements in
(1949) rp , and Pear- domly assigned di- performance in the outlier
son r mensions unrelated conditions; single, group
to the underlying average, & centroid meth-
clusters. ods only slightly affected
by presence of outliers;
nonhierarchical methods
generally unaffected by
presence of outliers. 5.
Group average method best
among hierarchical methods
CLUSTER ANALYSIS IN MARKETING RESEARCH 143

Table 4 (Continued)

Data sets
Reference Methods examined employed Coverage' Criteria Summary of results
used to derive starting point
for K-means procedure. 6.
Nonhierarchical methods
using random starting
points performed poorly
across all conditions.
Bayne, Beau- Single, complete, cen- Six parameterizations Complete Rand's statistic K-means, trace W, and Iwi
champ, Be- troid, simple average, of two bivariate provided the best recovery
govich, and weighted average, normal populations of cluster structure. NOR-
Kane (1980) median linkage and MIX performed most
Ward's minimum var- poorly. Among hierarchical
iance technique, and methods, Ward's technique,
two new hierarchical complete linkage, variance
methods, the variance & rank score methods per-
and rank score meth- formed best. Variants of
ods; four hierarchical average linkage method
methods: Wolfe's also performed well but not
NORMIX, K-means, as well as other methods.
two variants of the Single linkage performed
Friedman-Rubin pro- poorly.
cedure (trace W &
IWI). Euclidean dis-
tances served as-simi-
larity measure.
'The percentage of observations included in the cluster solution. With complete coverage, clustering continues until all observations have been
assigned to a cluster. Ninety percent coverage could imply that the most extreme 10% of the observations were not included in any cluster.

means procedure appears to be more robust than any of solution. A similar effect is obtained if data are stan-
the hierarchical methods with respect to the presence of dardized prior to clustering.
outliers, error perturbations of the distance measures, One characteristic of data appears to have a marked
and the choice of a distance metric. It appears to be least decremental effect on the performance of all clustering
affected by the presence of irrelevant attributes or di- methods-the presence of one or more spurious attri-
mensions in the data. butes or dimensions. A variable that is not related to the
One conclusion in several of the studies is that the final clustering solution, i.e., does not differentiate
choice of a similarity/dissimilarity measure, or distance among clusters in some manner, causes a serious dete-
measure, does not appear to be critical. Despite the con- rioration of the performance of all clustering methods,
siderable attention given such measures (Green and Rao though this problem is least severe with the K-means
1969; Morrison 1967; Sherman and Sheth 1977), the procedure and is probably less serious for other iterative
selection of a similarity measure appears to be less im- partitioning methods as well. This finding indicates the
portant for determining the outcome of a clustering so- need for careful selection of variables for use in clus-
lution than the selection of a clustering algorithm. Two tering and the need to avoid "shotgun" approaches
cautions should be observed in taking this conclusion at where everything known about the observations is used
face value, however. First, the number of studies of the as the basis for clustering. Clearly one cannot know in
relative import of distance measures for determining advance what variables may differentiate among a set of
clustering solutions is small and many types of data have as yet unidentified clusters. Nevertheless, it is not un-
yet to be examined. There may be types of data for reasonable for a researcher to have some rational or the-
which the selection of a distance measure is critical to oretical basis for selecting the variables used in a cluster
the clustering solution. Second, clustering algorithms analysis.
which are sensitive to the presence of outliers (e.g., A final conclusion can be drawn from the empirical
complete linkage specifically, and more generally all of findings on the performance of clustering algorithms: as
the hierarchical methods of clustering) seem to produce a clustering algorithm includes more and more obser-
better solutions when Pearson product moment or intra- vations, its performance tends to deteriorate, particularly
class correlation coefficients are used. Such similarity at high levels of coverage, 90% and above. This effect
measures tend to reduce the extremity of outliers in re- is probably the result of outliers beginning to come into
lation to Euclidean distance measures. This, in turn, re- the solution. Clustering all observations may not be a
duces the influence of outliers on the final clustering good practice. Rather the identification and elimination
144 JOURNAL OF MARKETING RESEARCH, MAY 1983

of outliers or the use of a decision rule to stop clustering DATA TRANSFORMATION ISSUES
short of the inclusion of all observations is probably ad- Although issues related to the choice of a similarity /
vantageous. Suggestions for identifying outliers are pro- dissimilarity measure have received considerable atten-
vided hereafter. The K-means procedure has shown less tion (Green and Rao 1969; Morrison 1967) the results
decrement in performance as coverage increases than of the empirical studies cited above suggest that the
have the hierarchical methods. choice is not crucial to the final clustering solution. The
Though a reasonable amount of evidence suggests that same appears to be true of the standardization issue. To
iterative partitioning methods are superior to hierarchical the extent that a particular measure of similarity or stan-
methods, particularly if nonrandom starting points are dardization reduces the extremity of outliers, the per-
used, it is not yet clear which of the iterative partitioning formance of some algorithms which are sensitive to out-
methods are superior. K-means procedures and tr Wand liers may be improved. Otherwise the selection of a
Iwi hill-climbing procedures all appear to perform well. similarity measure or the standardization of data prior
Some evidence (Blashfield 1977) suggests that hill- to clustering appears to have minimal effect. We do not
climbing methods which minimize IWI have an advan- suggest that the choice of a similarity measure should
tage over other iterative partitioning methods. be indiscriminant; the measure should be appropriate for
RECOMMENDATIONS FOR USING CLUSTER the type of data being considered. Rather, the choice of
ANALYSIS a correlation coefficient, a Euclidean distance, or a city-
block metric does not seem to produce much difference
It should be clear from the preceding discussion that in the final outcome of a clustering exercise involving
the research analyst must make several decisions which data for which each of the similarity measures is appro-
affect the structure of a cluster solution. These decisions priate.
can be grouped in the following broad categories: Some measures of similarity/dissimilarity explicitly
1. Data transformation issues correct for interdependencies. Other measures do not
A. What measure of similarity/dissimilarity should consider interdependencies. Interdependencies among
be used? variables may exist by design or, more often, are the
B. Should the data be standardized? How should unexpected result of the research design. Careful selec-
nonequivalence of metrics among variables be tion of variables may reduce unwanted interdependen-
addressed? cies but the problem is likely to remain even in the best
C. How should interdependencies in the data be of circumstances. Bailey (1974) provides an illustration
addressed? of the problem, the effect of which is to weight more
2. Solution issues heavily certain dimensions along which clustering will
A. How many clusters should be obtained? be carried out. When this is desirable for some theoret-
B. What clustering algorithm should be used? ical or practical purpose, correcting for interdependen-
C. Should all cases be included in a cluster anal- cies is inappropriate. When the researcher desires that
ysis or should some subset be ignored? all dimensions or attributes be given equal weight in the
3. Validity issues clustering process, it is necessary to correct for inter-
A. Is the cluster solution different from what dependencies. This can be achieved by selecting a sim-
might be expected by chance? ilarity measure which corrects for interdependencies,
B. Is the cluster solution reliable or stable across Mahalanobis D 2 or partial correlations. Correction may
samples? also be achieved by completing a preliminary principal
C. Are the clusters related to variables other than components analysis with orthogonal rotation. Compo-
those used to derive them? Are the clusters use- nent scores may then be used as input for the compu-
ful? tation of a similarity or distance measure. Skinner (1979)
4. Variable selection issues gives an example of this latter approach.
A. What is the best set of variables for generating SOLUTION ISSUES
a cluster analytic solution?
The selection of the clustering algorithm and solution
Often these decisions are not independent of one another characteristics appears to be critical to the successful use
because the choice of a means for addressing one of of cluster analysis. Empirical studies of the performance
these issues may constrain the options available for ad- of clustering algorithms suggest that one of the iterative
dressing other issues. For example, choosing to use a partitioning methods is preferable to the hierarchical
Pearson product moment correlation coefficient also de- methods. This holds, however, only when a nonrandom
termines that the data will be standardized because starting point can be specified. In addition, iterative par-
standardization is implicit in the computation of the cor- titioning methods require prior specification of the num-
relation coefficient. Thus, it is not possible to offer rec- ber of clusters desired. Hierarchical methods require no
ommendations for the resolution of anyone of these is- such specification. Thus, the user is confronted with
sues without an explicit understanding of the interactions determining both an initial starting point and the number
among these decisions. of clusters in order to use the methods that have dem-
CLUSTER ANALYSIS IN MARKETING RESEARCH 145

Figure 1 Similar two-stage clustering approaches have been sug-

TWO-STAGE CLUSTERING gested by Hartigan (1975) and Milligan (1980). Figure
1 is a schematic representation of the procedure. Only
four cluster analytic software packages provide average
AVERAGE LINKAGE OR WARD'S linkage or Ward's method and an iterative partitioning
MINIMUM VARIANCE METHOD algorithm: BMDP, CLUSTAN, and the Anderberg and
Hartigan series.
VALIDITY ISSUES
Even after careful analysis of a data set and the de-
termination of a final cluster solution, the researcher has
PRELIMINARY CLUSTER SOLUTION
I no assurance of having arrived at a meaningful and use-
ful set of clusters. A cluster solution will be reached
even when there are no natural groupings in the data.
This problem is similar to that encountered with a variety
of other procedures ranging from factor analysis to
regression analysis. Some test or set of tests must be
1. SELECT CANDIDATE NUMBER OF CLUSTERS applied to determine whether the solution differs signif-
2- OBTAIN CENTROIDS OF CLUSTERS icantly from a random solution. Milligan and Mahajan
3. ELIMINATE OUTLIERS (1980) and Milligan (1981) reviewed several such meth-
ods for testing the quality of a clustering solution and
found them wanting on a number of dimensions. A
method suggested by Arnold (1979) appears to over-
come the problems of other methods. Arnold (1979) pro-
posed using a statistic first suggested by Friedman and
ITERATIVE PARTITIONING ALGORITHM Rubin (1967) as a test of the statistical significance of
USING CLUSTER CENTROIDS OF
PRELIMINARY ANALYSIS AS STARTING
a cluster solution. The statistic is given by
POINTS (OUTLIERS NOT INCLUDED)
C = log (max /ITI/IWI»
where ITI is the determinant of the total variance-covari-
ance matrix and [w is the determinant of the pooled
within-groups variance-covariance matrix. A number of
iterative partitioning methods seek maximization of the
ratio of ITI to Iwl: MIKCA, CLUSTAN, and CLUS. For
I FINAL CLUSTER SOLUTION
algorithms which do not optimize ITl/lWI the test be-
comes even more conservative. Arnold generated distri-
butions of the C statistic for 2, 4, and 8 group solutions,
onstrated superior performance. Information for deter- 10, 20, 50, 100, 200, 500, and 1000 entities, and 5, 10,
mining starting points in the form of a priori descrip- 20 attributes and indicated values of C which allow re-
tions of expected clusters may be available. In the jection of the null hypotheses that the data arise from
absence of such information a means for obtaining start- unimodal or uniform distributions. He presented data to
ing points and an estimate of the number of clusters is support the use of the statistics and provided formulas
required. A two-stage procedure may be employed to for the derivation of other values.
cope with this problem. As with other multivariate statistics, one must dem-
In the first step one of the hierarchical methods which onstrate the reliability and the external validity of a clus-
has demonstrated superior performance, average linkage ter solution as well as its statistical significance. Reli-
or Ward's minimum variance method, may be used to ability may be established by cross-validation. External
obtain a first approximation of a solution. By examining validation requires a demonstration that the clusters are
the results of this preliminary analysis, one can deter- useful in some larger sense. Numerous authors have rec-
mine both a candidate number of clusters and a starting ommended cross-validating cluster solutions (see, e.g.,
point for the iterative partitioning analysis. In addition, Sherman and Sheth 1977) and several methods of cross-
this preliminary analysis can be used for examining the validation have been proposed. One of the more fre-
order of clustering of various observations and the dis- quently used methods involves dividing the sample in
tances between individual observations and clusters. half and carrying out clustering on each half. Descriptive
This provides an opportunity for the identification of statistics of the two sets of clusters are compared to de-
outliers which may be eliminated from further analysis. termine the degree to which similar clusters have been
The remaining cases may then be submitted to an iter- identified. The problem with such an approach is that
ative partitioning analysis for refinement of the clusters. no objective measure of reliability is obtained.
146 JOURNAL OF MARKETING RESEARCH, MAY 1983

Several authors have recommended the use of discrim- VARIABLE SELECTION ISSUES
inant analysis for cross-validation (Field and Schoe~feldt
The findings of empirical studies of cluster methods
1975' Nerviano and Gross 1973; Rogers and LInden
suggest that attention to initial yariable selec~ion is cru-
1973). The approach involves using cluster membership
cial because even one or two Irrelevant vanables may
as the group membership variable in a discriminant anal-
distort an otherwise useful cluster solution. The basis for
ysis. After a cluster solution has been d~veloped .on one
classification must be carefully identified to ensure that
sample, discriminant functions are denved wh~ch are
extraneous characteristics do not distort an otherwise
applied to a second sample. The degree to ~hICh the
useful cluster analysis. There should be some rationale
assignments made with the discriminant fu?ctIons agree
for the selection of variables for cluster analysis. That
with assignments made by a cluster analysis ?! the sec-
rationale may grow out of an explicit theory or be based
ond sample serves as an estimate of the. s~abIhty of the
on a hypothesis. Clearly more attention needs to be paid
cluster solution across samples. A coefficient of agree-
to this critical issue. As a science develops, researchers
ment, such as kappa, may be used to provide an objec-
must agree on those dimensions which are most relevant
tive measure of such stability. Using discriminant anal-
to classification for a particular purpose. Much debate
ysis for validating cluster analysis has seve:al drawbacks.
in the science of marketing involves the issue of variable
Discriminant coefficients may be poor estimates of pop-
selection. Thus, it is not surprising that a variety of dif-
ulation values and need to be cross-validated them-
ferent classification systems have been developed for
selves. This procedure is not cost-effective and th~ s~m­
similar phenomena. Indeed, it is probably unrealistic ~o
ple size available may be insuffic~ent.fo: cross-vahd.atIng
expect that a single classification system will emerge In
both the cluster analysis and a discriminant analysis,
any area of marketing in the foreseeable future. Rather,
McIntyre and Blashfield (1980) discussed an alterna-
there are likely to be numerous competing systems. The
tive approach to cross-validation which is recommen?ed
development of diverse systems is healthy and has been
here. The procedure is relatively simple and easy to Im-
observed in other sciences. Experience with rival sys-
plement on a computer. Cluster anal~sis is first carri~d
tems and a comparison of their usefulness ultimately
out on one half of the observations available for analysis.
provide a basis for selection of one system ~ver another.
Once a statistically significant clustering solution has
Cluster analysis has much to offer as an aid for devel-
been identified, centroids describing the clusters are ob-
oping classification systems. To the .exte~t t~~t ~lassi­
tained. Objects in the holdout data set are then assigned
fication is both the first and last step In scientific Inves-
to one of the identified clusters on the basis of the small-
tigation, cluster analysis should have increasing application
est Euclidean distance toa cluster centroid vector. The
in marketing.
degree of agreement between the nearest-centroid as-
signments of the holdout sample and ~he re~ult.s o! a
cluster analysis of the holdout sample IS an indication REFERENCES
of the stability of the solution. A coefficient of agree- Aldenderfer, M. S. (1977), "A Consumer Report on Cluster
ment, kappa, may be used as an objective measure of AnalysisSoftware: (2) Hierarchical Methods," working pa-
stability. If an acceptable level of stability is obtained per, Department of Anthropology, Pennsylvania State Uni-
the data sets may be combined to obtain a final solution. versity.
The demonstration of the statistical significance and Anderberg, M. R. (1973), Cluster Analysis for Applications.
stability of a cluster solution is necessary before one can New York: Academic Press.
accept and use the classification developed by the meth- Anderson, W. T., Jr., E. P. Cox, III, and D. G. Fulcher
odology. The acceptance of a particular classification (1976), "Bank Selection Decisions and Market Segmenta-
tion," Journal of Marketing, 40, 40-45. .
system, whether developed through cluster analysis or
Arabie, P., J. D. Carroll, W. DeSarbo, and J. Wmd (1981),
some other method, requires a further demonstration of "Overlapping Clustering: A New Method for Product Po-
utility, however. Classification is only useful if it assists sitioning," Journal of Marketing Research, 18 (August),
in furthering an understanding of the phenomena of in- 310-17.
terest. Clusters, or classes, must have demonstrable im- Arnold, S. J. (1979), "A Test for Clusters," Journal of Mar-
plications for hypothesis generation, theory building, keting Research, 16 (November), 545-51.
prediction, or management. The ultimate test of a set of Bailey, K. D. (1974), "Clusters Analysis," in Sociological
clusters is its usefulness. Thus, the user of cluster anal- Methodology, D. Heise, ed. San Francisco: Jossey-Bass.
ysis should provide a demonstration that clusters are re- Ball, G. H. (1970), Classification Analysis. Menlo Park, CA:
lated to variables other than those used to generate the Stanford Research Institute.
solution. Ideally, only a small number of variables Bass, F. M., E. A. Pessemier, and D. J. Tigert (1969), "A
Taxonomy of Magazine Readership Applied to Problems in
should be required to classify individuals. This classi- Marketing Strategy and Media Selection," Journal ofBusi-
fication should then have implications beyond the nar- ness, 42, 337-63.
row set of classification variables. The task of classifi- Bayne, C. K., J. J. Beauchamp, C. L. Begovich, and V. E.
cations is not finished until these broader implications Kane (1980), "Monte Carlo Comparisons of Selected Clus-
have been demonstrated. tering Procedures," Pattern Recognition, 12,51-62.
CLUSTER ANALYSIS IN MARKETING RESEARCH 147

Bertin, J. (1967), Semiology Graphique. Paris: Gauthier-Vil- vances in Consumer Research, Vol. 9, Andrew Mitchell,
lars. ed., 379-84.
Bettman, J. R. (1979), An Information Processing Theory of Gerard, R. W. (1957), "Units and Concepts of Biology,"
Consumer Choice. Reading, MA: Addison-Wesley. Science, 125, 429-33.
Blashfield, R. K. (1976), "Mixture Model Tests of Cluster Gower, J. C. (1967), "A Comparison of Some Methods of
Analysis: Accuracy of Four Agglomerative Hierarchical Cluster Analysis," Biometrics, 23, 623-37.
Methods," Psychological Bulletin, 83, 377-88. Green, P. E. and F. J. Carmone (1968), "The Performance
___ (1977), "A Consumer Report on Cluster Analysis Structure of the Computer Market: A Multivariate Ap-
Software: (3) Iterative Partitioning Methods," working pa- proach," Economic and Business Bulletin, 20, 1-11.
per, Department of Psychology, Pennsylvania State Uni- - - , R. E. Frank, and P. J. Robinson (1967), "Cluster
versity. Analysis in Test Market Selection," Management Science,
___ (1978), "The Literature on Cluster Analysis," Mul- 13, B-387-4oo.
tivariate Behavioral Research, 13, 271-95. - - - and V. R. Rao (1969), "A Note on Proximity Mea-
___ and L. C. Morey (1980), "A Comparison of Four sures and Cluster Analysis," Journal of Marketing Re-
Clustering Methods Using MMPI Monte Carlo Data," Ap- search, 6 (August), 359-64.
plied Psychology Measurement, in press. Greeno, D. W., M. S. Sommers, and J. B. Kernan (1973),
Calantone, R. J. and A. G. Sawyer (1978), "Stability of Ben- "Personality and Implicit Behavior Patterns," Journal of
efit Segments," Journal of Marketing Research, 15 (Au- Marketing Research, 10 (February), 63-9.
gust), 395-404. Hartigan, J. A. (1967), "Representation of Similarity Matrices
Cattell, R. B. (1949), "Rp and Other Coefficients of Pattern by Trees," Journal of the American Statistical Association,
Similarity," Psychometrika, 14, 279-98. 62, 1140-58.
- - - (1978), The Scientific Use of Factor Analysis in the - - - (1975), Clustering Algorithms. New York: John Wiley
Behavioral and Life Sciences. New York: Plenum Press. & Sons, Inc.
Claxton, J. D., J. N. Fry, and B. Portis (1974), "A Taxonomy Howard, H. and B. Harris (1966), "A Hierarchical Grouping
of Pre-purchase Information Gathering Patterns," Journal Routine, IBM 360/65 FORTRAN IV Program," University
of Consumer Research, 1, 35-42. of Pennsylvania Computer Center.
Cohen, J. (1960), "A Coefficient of Agreement for Nominal Inglis, J. and D. Johnson (1970), "Some Observations On,
Scales," Educational and Psychological Measurement, 20, and Developments In, the Analysis of Multivariate Survey
37-46. Data," Journal of the Market Research Society, 12, 75-8.
Cormack, R. M. (1971), "A Review of Classification," Jour- Jardine, H. and R. Sibson (1971), Mathematical Taxonomy.
nal of the Royal Statistical Society (Series A), 134, 321-67. New York: John Wiley & Sons, Inc.
Cunningham, K. M. and J. C. Ogilvie (1972), "Evaluation Johnson, S. C. (1967), "Hierarchical Clustering Schemes,"
of Hierarchical Grouping Techniques: A Preliminary Study," Psychometrika, 32, 241-54.
Computer Journal, 15, 209-13. Kantor, J. R. (1953), The Logic of Modern Science. Bloom-
Day, G. S. and R. M. Heeler (1971), "Using Cluster Analysis ington, IN: Principle Press.
to Improve Marketing Experiments," Journal of Marketing Kemeny, J. G. (1959), A Philosopher Looks at Science. New
Research, 8 (August), 340-7. York: Van Nostrand.
Dixon, W. J. and M. B. Brown (1979), BMDP: Biomedical Kernan, J. B. (1968), "Choice Criteria, Decision Behavior,
Computer Programs, P Series, 1979. Los Angeles: Univer- and Personality," Journal of Marketing Research, 5 (May),
sity of California Press. 155-69.
Edelbrock, C. (1979), "Comparing the Accuracy of Hierar- - - - and G. D. Bruce (1972), "The Socioeconomic Struc-
chical Clustering Algorithms: The Problem of Classifying ture of an Urban Area," Journal of Marketing Research,
Everybody," Multivariate Behavioral Research, 14, 367- 9 (February), 15-18.
84. Kiel, G. C. and R. A. Layton (1981), "Dimensions of Con-
- - - and B. McLaughlin (1980), "Hierarchical Cluster sumer Information Seeking Behavior," Journal of Market-
Analysis Using Intraclass Correlations: A Mixture Model ing Research, 18 (May), 233-9.
Study," Multivariate Behavioral Research, 15,299-318. Kuiper, F. K. and L. A. Fisher (1975), "A Monte Carlo Com-
Everitt, B. S. (1974), Cluster Analysis. London: Halsted parison of Six Clustering Procedures," Biometrics, 31,
Press. 777-83.
Field, H. S. and L. F. Schoenfeldt (1975), "Ward and Hook Lance, G. N. and W. T. Williams (1967a), "A General The-
Revisited: A Two-Part Procedure for Overcoming a Defi- ory of Classificatory Sorting Strategies. I. Hierarchical Sys-
ciency in the Grouping of Two Persons," Educational and terns," The Computer Journal, 9, 373-80.
Psychological Measurement, 35, 171-3. - - - (1967b), "A General Theory of Classificatory Sorting
Fisher, W. D. (1969), Clustering and Aggregation in Eco- Strategies. II. Clustering Systems," The Computer Journal,
nomics. Baltimore, MD: Johns Hopkins Press. 10,271-7.
Frank, R. E. and P. E. Green (1968), "Numerical Taxonomy Landon, E. L. (1974), "Self Concept, Ideal Self Concept, and
in Marketing Analysis: A Review Article," Journal of Mar- Consumer Purchase Intentions," Journal of Consumer Re-
keting Research, 5 (February), 83-98. search, 1, 44-51.
Friedman, H. D. and J. Rubin (1967), "On Some Invariant Lessig, V. P. and J. D. Tollefson (1971), "Market Segmen-
Criteria for Grouping Data," Journal of the American Sta- tation Through Numerical Taxonomy," Journal of Market-
tistical Association, 62, 1159-78. ing Research, 8 (November), 480-7.
Furse, D. H., G. Punj, and D. W. Stewart (1982), "Individual McIntyre, R. M. and R. K. Blashfield (1980), "A Nearest-
Search Strategies in New Automobile Purchase," in Ad- Centroid Technique for Evaluating the Minimum-Variance
148 JOURNAL OF MARKETING RESEARCH, MAY 1983

Clustering Procedure," Multivariate Behavior Research, Investigate Nonlinear and Interactive Relationships Between
15, 225-38. Personality and Product Usage," Journal of Marketing Re-
McQuitty, L. L. (1967), "A Mutual Development of Some search, 17 (February), 119-24.
Typological Theories and Pattern-Analytic Methods," Ed- Sethi, S. P. (1971), "Comparative Cluster Analysis for World
ucational and Psychological Measurement, 17,21-46. Markets," Journal of Marketing Research, 8 (August),
McRae, D. J. (1973), "Clustering Multivariate Observa- 348-54.
tions," unpublished doctoral dissertation, University of Sexton, D. E., Jr. (1974), "A Cluster Analytic Approach to
North Carolina. Market Response Functions," Journal of Marketing Re-
Mezzich, J. E. (1978), "Evaluating Clustering Methods for search, 11 (February), 109-14.
Psychiatric Diagnosis," Biological Psychiatry, 13,265-81. Shepard, R. N. (1974), "Representation of Structure in Sim-
Milligan, G. W. (1980), "An Examination of the Effect of ilarity Data: Problems and Prospects," Psychometrika, 39,
Six Types of Error Perturbation on Fifteen Clustering Al- 373-421.
gorithms," Psychometrika, 45, 325-42. - - - and P. Arabie (1979), "Additive Clustering: Repre-
- - (1981), "A Monte Carlo Study of Thirty Internal Cri- sentation of Similarities as Combinations of Discrete Over-
terion Measures for Cluster Analysis," Psychometrika, 46, lapping Properties," Psychological Review, 86, 87-123.
(2), 187-99. Sherman, L. and J. N. Sheth (1977), "Cluster Analysis and
- - - and P. D. Isaac (1980), "The Validation of Four Ul- Its Applications in Marketing Research," in Multivariate
trametric Clustering Algorithms," Pattern Recognition, 12, Methods for Market and Survey Research, J. N. Sheth, ed.
41-50. Chicago: American Marketing Association.
- - - and V. Mahajan (1980), "A Note on Procedures for Shuchman, A. (1967), "Letter to the Editor," Management
Testing the Quality of a Clustering of a Set of Objects," Science, 13, B688-96.
Decision Sciences, 11, 669-77. Skinner; H. A. (1979), "Dimensions and Clusters: A Hybrid
Mojena, R. (1977), "Hierarchical Grouping Methods and Approach to Classification," Applied Psychological Mea-
Stopping Rules: An Evaluation," Computer Journal, 20, surement, 3, 327-41.
359-63. Smith, W. (1956), "Product Differentation and Market Seg-
Montgomery, D. B. and A. J. Silk (1971), "Clusters of Con- mentation as Alternative Marketing Strategies," Journal of
sumer Interests and Opinion Leaders' Sphere ofInfluence," Marketing, 21, 3-8.
Journal of Marketing Research, 8 (August), 317-21. Sneath, P. H. A. (1957), "The Application of Computer to
Morgan, J. N. and J. A. Sonquist (1963), "Problems in the Taxonomy," Journal of General Microbiology, 17, 201-
Analysis of Survey Data, and a Proposal," Journal of the 26.
American Statistical Association, 58, 87-93. - - - and R. R. Sokal (1973), Numerical Taxonomy. San
Moriarty, M. and M. Venkatesan (1978), "Concept Evalua- Francisco: W. H. Freeman.
tion and Market Segmentation," Journal of Marketing, 42, Stewart, D. W. (1981), "The Application and Misapplication
82-6. of Factor Analysis in Marketing Research," Journal of
Morrison, B. J. and R. C. Sherman (1972), "Who Responds Marketing Research, 18 (February), 51-62.
to Sex in Advertising?" Journal of Advertising Research, Srivastava, R. K., R. P. Leone, and A. D. Shocker (1981),
12, 15-19. "Market Structure Analysis: Hierarchical Clustering of
Morrison, D. G. (1967), "Measurement Problems in Cluster Products Based on Substitution-in-Use," Journal of Mar-
Analysis," Management Science, 13, B-775-80. keting, 45 (Summer), 38-48.
Myers, J. G. and F. M. Nicosia (1968), "On the Study of - - , A. D. Shocker, and G. S. Day (1978), "An Explor-
Consumer Typologies," Journal of Marketing Research, 5 atory Study of Usage-Situational Influences on the Com-
(May), 182-83. position of Product-Markets," in Advances in Consumer
Neidell, L. A. (1970), "Procedures and Pitfalls in Cluster Research, H. K. Hunt, ed. Ann Arbor: Association for Con-
Analysis," Proceedings, Fall Conference, American Mar- sumer Research, 32-38.
keting Association, 107. Tyron, R. C. and D. E. Bailey (1970), Cluster Analysis. New
Nerviano, V. J. and W. F. Gross (1973), "A Multivariate York: McGraw-Hill Book Company.
Delineation of Two Alcoholic Profile Types on the 16PF," Veldman, D. J. (1967), FORTRAN Programming for the Be-
Journal of Clinical Psychology, 29, 370-4. havioral Sciences. New York: Holt, Rinehart, and Winston.
Peay, E. R. (1975), "Nonmetric Grouping: Clusters and Ward, J. (1963), "Hierarchical Grouping to Optimize an Ob-
Cliques," Psychometrika, 40, 297-313. jective Function," Journal of the American Statistical As-
Rand, W. M. (1971), "Objective Criteria for the Evaluation sociation, 58, 236-44.
of Clustering Methods, " Journal of the American Statistical Wells, W. D. (1975), "Psychographies: A Critical Review,"
Association, 66, 846-50. Journal of Marketing Research, 12 (May), 196-213.
Rogers, G. and J. D. Linden (1973), "Use of MUltiple Dis- Wind, Y. (1978), "Issues and Advances in Segmentation Re-
criminant Function Analysis in the Evaluation of Three search," Journal ofMarketing Research, 15 (August), 317-
Multivariate Grouping Techniques," Educational and Psy- 37.
chological Measurement, 33, 787-802. Wishart, D. (1969), "CLUSTAN IA: A FORTRAN Program
Rubin, 1. (1965), "Optimal Taxonomy Program (7090-IBM- for Numerical Classification," Computing Laboratory, St.
0026)," IBM Corporation. Andrew's University, Scotland.
Schaninger, C. M., V. P. Lessig, and D. B. Panton (1980), Wolf, A. (1926), Essentials of Scientific Method. New York:
"The Complementary Use of Multivariate Procedures to Macmillan Company.