Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

WAYNE S. DESARBO, RAJDEEP GREWAL, and CRYSTAL J.

SCOTT*
The segmentationtargetingpositioning conceptual framework has
been the traditional foundation and genesis of marketing strategy
formulation. The authors propose a general clusterwise bilinear spatial
model that simultaneously estimates market segments, their composition, a brand space, and preference/utility vectors per market segment;
that is, the model performs segmentation and positioning simultaneously.
After a review of related methodological research in the marketing,
psychometrics, and classification literature streams, the authors present
the technical details of the proposed two-way clusterwise bilinear spatial
model. They develop an efficient alternating least squares procedure that
estimates conditional globally optimum estimates of the model parameters within each iteration through analytic closed-form expressions.
The authors present various model options. They provide a conceptual
and empirical comparison with latent-class multidimensional scaling.
They use an illustration of the new bilinear multidimensional scaling
methodology with an actual commercial study sponsored by a large U.S.
automotive manufacturer to examine buying/consideration intentions for
small sport-utility vehicles. The authors conclude by summarizing the
contributions of this research, discussing the marketing implications for
managers, and providing several directions for further research.

Keywords: market segmentation, positioning, clusterwise analyses,


bilinear spatial models, sport-utility vehicles

A Clusterwise Bilinear Multidimensional


Scaling Methodology for Simultaneous
Segmentation and Positioning Analyses
Competitive market structures are an essential ingredient
in the strategic market-planning process (Day 1984; Myers
1996; Wind and Robertson 1983). DeSarbo, Manrai, and
Manrai (1993) describe the primary task of competitive
market structure analyses as deriving a spatial configuration
of products/brands/services in a designated product/service
class on the basis of some competitive relationships
between these products/brands/services (see also Day,
Shocker, and Srivastava 1979; Fraser and Bradford 1984;
Lattin and McAlister 1985). As Reutterer (1998) notes, to

provide managers with more meaningful decision-oriented


information needed for the evaluation of actual brands
positions with respect to competitors, competitive market
structure analyses have been enhanced with positioning
models that incorporate actual consumers purchase
behavior/intentions, background characteristics, marketing
mix, and so forth (Cooper 1988; DeSarbo and Rao 1986;
Green and Krieger 1989; Moore and Winer 1987). Thus,
contemporary approaches for empirical modeling of
competitive market structure now take into account both
consumer heterogeneity (e.g., market segmentation) and the
competitive positioning of products/brands/services. Such
analyses are integrated into more encompassing frameworks, such as Lilien and Rangaswamys (2004)
segmentationtargetingpositioning framework. In such a
framework, in which a firm targets one or more groups or
market segments with its offerings, positioning becomes a
segment-specific concept.

*Wayne S. DeSarbo is Smeal Distinguished Research Professor of Marketing (e-mail: wsd6@psu.edu), and Rajdeep Grewal is Deans Faculty
Fellow and Professor of Marketing (e-mail: rug2@psu.edu), Smeal College of Business, Pennsylvania State University. Crystal J. Scott is Assistant Professor of Marketing, School of Management, University of
MichiganDearborn (e-mail: cjscott@umd.umich.edu). Michel Wedel
served as associate editor for this article.

2008, American Marketing Association


ISSN: 0022-2437 (print), 1547-7193 (electronic)

280

Journal of Marketing Research


Vol. XLV (June 2008), 280292

A Clusterwise Bilinear Multidimensional Scaling Methodology


Empirical modeling approaches in this area have taken
two forms: (1) sequential use of multidimensional scaling
(MDS) and cluster analysis and (2) parametric finite mixture or latent-class MDS models. The more traditional
approaches have employed the sequential use of MDS and
cluster analysis first to portray spatially the relationship
between competitive brands and consumers and then to segment the resultant consumer locations to form market segments. Several methodological problems are associated
with such a naive approach. First, each type of analysis
(MDS and cluster analysis) typically optimizes different
loss functions. Indeed, many forms of cluster analysis optimize nothing. As such, different aspects of the data are
often ignored by the disjointed application of such sequential procedures. Second, there are many types of MDS and
cluster analyses procedures, as documented in the vast psychometric and classification literature, and each procedure
can render different results (DeSarbo, Manrai, and Manrai
1994). In addition, there is little a priori theory to suggest
which methodological selections within these alternatives
are most appropriate for a given marketing application. The
various combinations of types of MDS and cluster analyses
typically render different results as well. Finally, as Holman
(1972) notes, the Euclidean distance metric used in many
forms of MDS is not congruent with the ultrametric distance formulation metric used in many forms of (hierarchical) cluster analysis.
Efforts to resolve such methodological problems associated with this sequential approach of administration MDS
and cluster analyses have resulted in the evolution of parametric finite mixture or latent-class MDS models (DeSarbo,
Manrai, and Manrai 1994; Wedel and DeSarbo 1996). In
particular, several authors over the past 15 years have
proposed latent-class MDS models for the analysis of
preference/dominance data, either employing scalar
products/vector (Slater 1960; Tucker 1960) or unfolding
(Coombs 1964) representations to two-way preference/
dominance data. In such latent-class MDS models, vectors
or ideal points of derived segments are estimated instead of
parameters for every individual consumer. Thus, the number of parameters is significantly reduced relative to
individual-level models. Latent-class MDS models are traditionally estimated with the maximum likelihood method
(E-M [expectationmaximization] algorithms [see Dempster, Laird, and Rubin 1977] are typically employed). For
example, DeSarbo, Howard, and Jedidi (1991) develop a
latent-class MDS vector model (MULTICLUS) for normally distributed data. DeSarbo and colleagues (1991)
extend this latent-class MDS model to a weighted idealpoint model. De Soete and Heiser (1993) and De Soete and
Winsberg (1993), respectively, extend these two latent-class
MDS models by accommodating linear restrictions on the
stimulus coordinates. Bckenholt and Bckenholt (1991)
develop simple and weighted ideal-point latent-class MDS
models for binary data. DeSarbo, Ramaswamy, and Lenk
(1993) develop a vector latent-class MDS model for
Dirichlet-distributed data, and Chintagunta (1994) develops
a latent-class MDS vector model for multinomial data.
Wedel and DeSarbo (1996) extend the entire family of
exponential distributions to such latent-class MDS models
for two-way preference/dominance data.
Note that such latent-class MDS models have several
limitations. First, they are parametric models that require

281

the assumption of specific distributions. Often, continuous


support distributions are used in the finite mixture and are
applied to traditional discrete response scales (e.g., semantic differential, Likert), which cannot possibly obey such
distributional assumptions. As such, violations of such distributional assumptions may invalidate the use of the procedure. Second, such latent-class MDS procedures require the
underlying finite mixture to be identified (McLachlan and
Peel 2000), which is often problematic when the underlying
distributions deviate from the exponential family. When
multivariate normal distributions are employed, identification problems can arise in the estimation of separate full
covariance matrices by derived market segment. Third,
most of the latent-class MDS procedures are highly nonlinear in nature and require intensive computation. Such procedures typically use an E-M approach (see McLachlan and
Krishnan 1997) or gradient-based estimation procedures,
which may take hours of computation time for a complete
analysis to be performed. Fourth, locally optimum solutions
are typically reached, and the analyses must be repeated
several times for each value of the dimensionality and number of groups. Fifth, the available heuristics employing various information criteria typically result in different solutions being selected. For example, the Bayesian information
criterion and consistent Akaike information criterion are
considered more conservative measures that result in the
selection of fewer dimensions and groups than the Akaike
information criterion and the modified Akaike information
criterion, which are considered more liberal (see Wedel and
Kamakura 2000). As Wedel and Kamakura (2000) discuss,
other heuristics are also used for model selection for such
finite mixture models (e.g., ICOMP [information-theoretic
measure], NEC [normalized entropy criterion]) that are
equally plausible but may also result in different solutions
being selected. Finally, the underlying framework assumed
by these latent-class MDS procedures involves a partitioning of the sample space, though the estimated posterior
probabilities of membership often result in fuzzy probabilities of segment membership, which may be difficult to
interpret or justify in applications that require partitions.
Indeed, pronounced fuzziness of the resultant posterior
probabilities of segment membership may be indicative of
poor separation of the conditional support functions
centroids.
To overcome some of these limitations of current procedures, our objective is to devise a new deterministic, clusterwise procedure for the analysis of two-way preference/
dominance data, which uses a spatial bilinear scalar
product-based vector representation. The goal is to derive
simultaneously a single joint space in which segments are
represented by vectors, brands are represented through
coordinate points, and their interrelationship in the space
denotes some aspect of the structure in the data. The proposed approach does not require distributional assumptions,
such as latent-class MDS procedures, but provides a concise spatial representation for the analysis of preference or
dominance data, as illustrated in our application to buying
considerations for small sport-utility vehicles (SUVs). No
finite mixture distribution identification is required, unlike
latent-class MDS counterparts. The estimation procedure
developed is fast and efficient and converges in a matter of
minutes on a personal computer. Globally conditional optimum estimates of parameters are obtained within each

282

JOURNAL OF MARKETING RESEARCH, JUNE 2008

iteration cycle of the estimation (though overall locally


optimum solutions are also possible here). Finally, our proposed procedure accommodates both overlapping segments
(Arabie et al. 1981) and hard partitions.
We organize the remainder of the article as follows: In
the next section, we present the technical details of the clusterwise bilinear spatial regression model together with the
estimation procedure developed for the model. Subsequently, we present the application of the model to the
simultaneous estimation of market segments and a joint
multidimensional space for a commercial application investigating purchase intentions/considerations for small SUVs
in a segmentationtargetingpositioning application. We
compare the proposed procedure with the MULTICLUS
(DeSarbo, Howard, and Jedidi 1991) latent-class MDS vector model in terms of calibration fit and predictive validity.
We end by providing some concluding remarks on the utility of the proposed clusterwise model for various marketing
research problems, and we suggest some extensions of the
methodology and directions for further research.
THE PROPOSED CLUSTERWISE BILINEAR SPATIAL
MDS MODEL
The Model
Given that we illustrate the proposed methodology in the
context of simultaneous positioning and market segment
estimation, we use the following scenario to describe the
models structure and parameters:
Let i = 1, , N consumers; j = 1, , J brands; s = 1, ,
S market segments (unknown); r = 1, , R dimensions
(unknown); and ij = the preference/consideration/intention
to buy brand j by consumer i. Then, we model the observed
data as
S

ij =

(1)

P X
is

s =1

jr Ysr

+ b + ij ,

r =1

where
Xjr = the rth coordinate for brand j;
Ysr = the rth coordinate for segment s (vector);
0 if consumer i is not classified in segment s, and
Pis =
1 if otherwise;
such that
Pis {0, 1},
Ss = 1Pis = 1 for partitions, or
0 < Ss = 1Pis S for overlapping segments;
ij = error; and
b = an additive constant.
As in the ordinary restricted MDS of preference/choice
(see CANDELINC by Carroll, Pruzansky, and Kruskal
1980; GENFOLD2 by DeSarbo and Rao 1986), we allow
as an option the linear reparameterization of the brand coordinates (X) as functions of designated attributes, marketing
mix, features, as so forth, with the following:
X jr =

(2)

jk kr ,

Zjk = the value of attribute/feature k for brand j; and


kr = the impact coefficient for attribute/feature k on
dimension r.
Visually, we posit a scalar products or vector MDS display of the structure in the data (a joint space of brand
points and segment vectors) and simultaneously classify
consumers into segments, allowing for partitions or overlapping segment memberships. As with traditional vector
MDS models (e.g., MDPREF by Carroll 1980), the orientation of the estimated segment vector points in the direction
of higher utility, and the projection of a brand onto the segment vector indicates the magnitude of preference or utility
of that brand for that particular segment. In Figure 1, we
present a hypothetical solution with two underlying dimensions (R = 2), three market segments (S = 3), and ten brands
(J = 10, labeled AJ) to describe the spatial relationships
provided by our proposed methodology. In Figure 1, Panel
A, we present the estimated brand locations (X) for these
ten hypothetical brands. As we mentioned previously, these
are estimated coordinate points in this two-dimensional
space. Their particular locations render insight into their
relationships to these two dimensions. For example, Brand
I loads highest on Dimension 1 (Brand G loads the lowest),
and Brand J loads the highest on Dimension 2 (Brand E
loads the lowest). In Figure 1, Panel B, we add the segment
vectors indicated by the dashed lines. As we mentioned previously, the directions of these vectors give the direction of
highest utility for each of the derived segments. That is, a
brand has higher preference the farther out in the direction
of the vector it is positioned (i.e., the larger the orthogonal
projection of that brand onto that segments preference vector). In this hypothetical illustration, consumers classified
in Segment 1 prefer Brands D, F, and J; consumers classified in Segment 2 prefer Brands E and I; and consumers
classified in Segment 3 prefer Brands A, G, and J. Brands C
and H are not highly preferred by any of the derived market
segments in this illustration, because both brands project
near the origin of the space for all three segment vectors.
Thus, sample heterogeneity with respect to the sample consumers and derived segments is represented in relation to
different vector orientations in the derived space. Finally,
Figure 1, Panel C, adds the attribute vectors obtained by
typically normalizing and plotting kr or by property fitting
methods that regress each attribute onto the two dimensions
(with no intercept term). This feature is particularly useful
in marketing applications because the direction of these
attribute vectors indicate the predicted movement of a brand
as it increases/improves with respect to the attribute in
question. Occasionally, these attribute vectors are plotted
without normalization, so the length of these attribute vectors provides an indication of how related the attribute is to
the dimensions. Alternatively, a table of correlations
between the brand attributes and the brand coordinates on
these dimensions can be provided. Thus, Panel C provides a
concise summary of the structure in these data that can be
easily conveyed to marketing managers concerned with the
segmentationtargetingpositioning problem.

where
k = 1, , K brand attributes or features (note that K <
J for use of this option);

The Estimation Procedure


Given and values of S and R, our goal is to estimate
P = ((Pis )), X = (( X jr )) or = ((kr)), b, and Y = ((Ysr )) to
minimize the following sums-of-squares error (SSE):

A Clusterwise Bilinear Multidimensional Scaling Methodology

283

Figure 1
ILLUSTRATIVE EXAMPLE: S = 3 MARKET SEGMENTS, R = 2 DIMENSIONS, J = 10 BRANDS (I.E., BRAND A = BRAND J), AND K = 7
ATTRIBUTES
A: Brand Space

B: Brand and Segment Space

C: Brand Segment and Attribute Space

(3) Min P, X, Y, b =
ij

s
i = 1 j = 1
J

X jr Ysr b

=1

P
is

=1

.
2
ij

(4)

where * = b. We now want to estimate b, P, Y, andX or


, given and values of S and R, so as to minimize the following SSE:

(5)

Letting I J = (( ij )), we can rewrite the model for this twoway data case as

* = PYX + ,

(6)

Min = Min tr ( ) tr = ( * PYX)( * PYX)

= tr * * * P YX XYP * + XY P PYX

284
(7)

JOURNAL OF MARKETING RESEARCH, JUNE 2008


= tr ( * * ) 2 tr ( * P YX) + tr ( XY P P Y X).

We now describe the alternating least squares algorithm


devised for complete clusterwise model estimation with
five cycled estimation steps.
1. Estimate X
Beginning with the first-order conditions, we calculate
the partial derivatives of the SSE expression (Equation 3)
with respect to X:
(8)

=
2 tr ( AX) + tr ( XBX) ,
X X

where
(9)

A = * PY and

(10)

B = Y P PY.

Thus,
(11)

(16)

= 2 A + X( B + B),
X

given the properties of the trace operation and its partial


derivatives. Because is symmetric, = and the firstorder conditions give the following:
(12)

2 A + 2 XB = 0.

Solving for X,
(13)

haps this selection problem can be somewhat ameliorated


when data reduction is used first and then the loadings of
the raw attributes on the derived factors are examined,
though some purists may argue that such a procedure
involves multiple use of the same set of data for an analysis. Third, each single attribute can be property fit onto the
space post hoc through regression methods (with no intercept). The problem that arises here is that the brand locations no longer become exact functions of the attributes,
and situations may arise in which the derived dimensions
may not be explained well by the attributes. As such, there
are potential difficulties associated with any of these
options for the case when K > J.
2. Estimate Y
Beginning with the first-order conditions, we calculate
the partial derivatives of the SSE expression (Equation 3)
with respect to Y:

= AB1
X
= * P Y(Y P PY)1 ,

which is estimable only for R S (one of the identification


restrictions for overlapping segments; R < S for partitions).
For use of the reparameterization option in Equation 2,
where X = , when we use the chain rule for derivatives
and the developments in Equations 1113, the stationary
equations reduce to the following:
(14)

2 2 = 0, and

(15)

a = ()11,


=
tr 2 * P YX + XY P PYX

Y Y

(17)


tr (2 * P YX) + tr ( XY P PYX)

(18)

2 tr (CYX) + tr ( XY QYX) ,
Y

where
(19)

C = * P and

(20)

Q = P P.

Then,
(21)

2 X * P + (P P)Y( X X) + (P P) Y( X X) = 0,

which follows from the properties of the trace operator and


its derivatives (Magnus and Neudecker 2002); thus,
= ( X X)1 X * P (P P)1.
Y

(22)

3. Estimate P
Note that Pis {0, 1}; this represents the segment membership indicator binary variables such that

(23)

which exists for the same-order conditions regarding R and


S that we mentioned previously, plus full column rank of Z
and K < J. For applications involving K > J, several options
are available. First, some sort of data reduction of the attributes (e.g., principal components analysis/factor analysis)
can be performed, and the resultant component/factor
scores can be used in Z. Unfortunately, the ability to easily
track the influence of changes in a single attribute on
brands in the space is typically lost. In addition, it may be
difficult to operationalize exact changes/manipulations in a
factor score when that factor score is a linear combination
of a host of different attributes. Given that each factor is a
linear combination (i.e., compensatory) of the various
attributes, there would be an infinite number of ways to
increase/decrease performance on a factor. Second, some
subset of the K > J attributes could be selected and used in
Z. This solution resolves the problems we alluded to previously with factor scores. However, other issues evolve
regarding which of the subset of attributes to select. Per-

1, i , and

is

(24)

is

> R , s.

Then, can be rewritten as


(25)

= tr ( ),

(26)

= tr (),

and
I

(27)

Hii ,

i =1

where H = ; thus,
I

(28)

i =1

*
i

P i YX)( i P i YX).

A Clusterwise Bilinear Multidimensional Scaling Methodology


Because *i and Pi only affect in the ith observation, the
optimization here is separable over i. That is, Equation 28
can be conditionally minimized by observation to obtain a
conditionally global optimum P, given X, Y, and . For
each i, we minimize i = i i with respect to Pi. Here, we
enumerate over all solution options (S options for the case
of partitions; 2S 1 solution options for overlapping segments) for each Pi (ignoring the 0 solution of no membership in any derived segment) to minimize i i.
4. Estimate b
We first define
S

(29)

ij =

X
Pis

s =1

jr Ysr ,

r =1

and let
L = vec( ij ),
K = (1, M),
1 = (1, 1, ..., 1), and
M = vec( ij ).
Then, we can formulate this estimation problem as a simple
least squares one and calculate the following:
(30)

b
= (K K )1 K L.
a

b i
= (K i K i )1 K i L i ,
a
i

which is applied for each consumer in the data. However,


this option often adds a substantial number of additional
parameters to the analysis (2 I = 600 in the SUV application to follow).
5. Test for Convergence
We can calculate an overall variance-accounted-for
(VAF) statistic (DeSarbo and Carroll 1985) akin to an R2 as
follows:
N

(32)

VAF = 1

ij

ij )2

i = 1j = 1
N
J

,
ij

.. )2

i = 1j = 1

where
(33)

.. =

1
NJ

.
ij

i = 1j = 1

If VAF(IT) VAF(IT 1) .0001, output all parameters estimated and stop; otherwise, increase IT = IT + 1 and return
to Step 1.
Note that Steps 14 of this alternating least squares algorithm provide a global optimum solution within iterate conditioned on holding fixed all the other parameter sets. In
addition, Steps 1, 2, and 4 are analytical closed-form
expressions that do not require much computational time.
However, these desirable properties do not guarantee a
global optimum solution after convergence. Similar to its
latent-class MDS counterparts, the proposed methodology
is also subject to locally optimum solutions, and thus the
procedure needs to be executed numerous times from different random starting points to check for globally optimum
solutions, as in the case of latent-class MDS. Model selection (selection of R and S) is determined by associated
scree plots of VAF in Equation 32 for sequential values of
R and S, as in traditional MDS and optimization-based
clustering. In addition, as with all multivariate procedures,
model selection is also guided by an inspection and interpretation of the results.
MARKETING COMMERCIAL APPLICATION: SMALL
SUVs
Study Background

Note that the multiplicative constant (a) is not identifiable,


because it can be (and is) directly embedded into X or Y
and then set equal to 1.00.
For consumer conditional analyses, identifiable additive
and multiplicative constants can be estimated by consumer
(segment-level constants are not identifiable) with the following development: Let Li = vec( i), Ki = (1, Mi), 1 = (1,
1, ..., 1), and Mi = vec( i). Then, we can formulate this estimation problem as a simple least squares one and calculate
the following:
(31)

285

A large U.S. automotive consumer research supplier


administered a tracking study in 2002 to gauge automotive
marketing awareness and shopping behavior among typical
consumers in the marketplace. The surveys used in these
tracking studies were conducted among new vehicle intenders (i.e., a prospective buyer that will be in-market or
has plans to purchase a new vehicle within the next 612
months) and were collected from an automotive consumer
panel of more than 600,000 nationally representative households. This image study had been conducted among new
vehicle intenders semiannually, in June and December, for
the past 20 years. Each survey respondent rated each brand/
model corresponding to the particular product segment in
which he or she intended to purchase. The ending completed sample resulted in approximately 200300 respondents per product segment. The information collected
included familiarity with each make/model, advertising
recall, overall opinion, purchase consideration, image
attribute ratings, and awareness of model redesign. The
image study was administered across 16 car, 10 light truck,
and 3 SUV product segments.
We selected the small-SUV segment for use in this study.
Data were collected in December 2002, with 360 consumers rating 18 different compact SUVs. On a four-point
response scale (4 = definitely would consider, 3 = probably would consider, 2 = probably would not consider,
and 1 = definitely would not consider), the respondents
rated how likely they would be to consider each brand
given that they would soon be in-market. Table 1 lists the
small-SUV brands used in this study with their 2002 sales
and market shares. In addition to rating their likelihood of
purchasing the vehicle, the respondents used a five-point
interval scale to rate each vehicle subjectively on some 23
image attributes (we discuss these subsequently).
In 2002, the year of data collection, market trends
revealed that SUV sales continued to increase as new vehi-

286

JOURNAL OF MARKETING RESEARCH, JUNE 2008


Table 1
SMALL SUV MARKET SHARE

Brand

2002 Sales

Jeep Liberty
Honda CR-V
Ford Escape
Toyota RAV4
Hyundai Santa Fe
Subaru Outback
Saturn VUE
Jeep Wrangler
Subaru Forrester
Mazda Tribute
Chevrolet Tracker
Suzuki XL-7
Land Rover Freelander
Suzuki Grand Vitara
Mitsubishi Outlander
Suzuki Vitara
Isuzu Rodeo Sport
Honda Element

171,212
146,266
145,471
86,601
78,279
77,917
75,477
64,351
53,992
44,989
42,212
27,295
15,021
11,529
11,346
6549
3745
957
1,063,209

Market Share (%)


16.1
13.8
13.7
8.1
7.4
7.3
7.1
6.1
5.1
4.2
4.0
2.6
1.4
1.1
1.1
.6
.4
.1
100.0

Outback, and Chevy Tracker. These 11 brands accounted


for 88% of 2002 compact SUV sales. Other compact SUVs
included in the data set, but not included among the top 11
market share brands, were the Isuzu Rodeo Sport, Suzuki
Vitara, Suzuki Grand Vitara, Suzuki XL-7, Mitsubishi Outlander, Honda Element, and Land Rover Freelander (we
used these 7 brands for predictive validation). The 2002
market share data are congruent with the mean consideration ratings in this image study as the Jeep Liberty and
Honda CR-V are the top vehicles in sales and also rank the
highest in mean consideration for these new vehicle intenders. Note that because the number of available brand
attributes (K = 23 + mean consideration and market share =
25 attributes in total) far exceeds the number of brands in
the calibration sample (J = 11), the reparameterization
option specified in Expression 2 is not identified, and thus
we estimate the brand coordinates X alone and property fit
post hoc the various brand attributes.
The Proposed Model Joint Space Solution

cle buyers abandoned traditional sedans and minivans. In


1997, midsize cars represented 27.1% of the total market,
and SUVs were 12.3%. In the first four months of 2002,
SUV sales were up to 20.2% share, and midsize cars had
fallen to 17.1% share (J.D. Power and Associates 2003a, b).
The small-SUV segment was the fastest-growing subsegment of the SUV arena. In 2002, light trucks outsold cars,
claiming a full 51% of the market. This growth was driven
by the small-SUV segment, in which there was a sales
increase of 9.2% from 2001 to 2002, compared with a .36%
increase for all light trucks and SUVs during the same
period; product segments experiencing decreasing sales
were full-size SUVs, vans, and pickup trucks (J.D. Power
and Associates 2003a, b).
We reduced the 18 vehicles included in this image study
to 11 calibration brands (used for estimation) on the basis
of 2002 market share information (i.e., we selected the top
11 brands). The resultant data set for this analysis included
the following brands: Jeep Liberty, Honda CR-V, Ford
Escape, Toyota RAV4, Hyundai Santa Fe, Saturn VUE,
Jeep Wrangler, Subaru Forester, Mazda Tribute, Subaru

Table 2 shows the various model selection heuristics values for determining R and S for the proposed methodology
without brand reparameterization since K > J. For sequential values of R and S, we performed five runs and selected
the best-fitting solution for values of S and R. We ran the
proposed procedure for S R (solutions for R > S cannot
be identified in either latent-class MDS or such clusterwise
procedures) for both overlapping and nonoverlapping cluster solutions on the row standardized input data, and Table
2 displays the results for the overlapping cluster solutions,
which dominated the corresponding nonoverlapping ones in
terms of VAF. On the basis of the goodness-of-fit values
(VAF) for the incremental search procedure across values
of R and S and subsequent interpretation, we selected the
R = 4 dimensions/S = 5 overlapping segments solution as
most parsimonious. Note that the overlapping solution rendered a 34.1% improvement in VAF compared with the
nonoverlapping solution for this same R and S. (The resultant VAF was .525 for the addition of individual-level additive and multiplicative constants/parameters versus .472 for
just the one additive constant parametera trivial improvement given the addition of 600 additional parameters.)

Table 2
CLUSTERWISE VECTOR MDS RESULTS
R = Number of Dimensions

S = Number of Clusters

Number of Iterations

SSE

VAF

1
2
3
4
5
2
3
4
5
3
4
5
4
5
5

3
9
13
20
23
16
15
9
22
11
12
16
16
10
17

2858.8
2719.8
2658.8
2643.8
2639.9
2590.9
2358.1
2291.4
2248.1
2214.8
2071.9
1988.4
1934.6
1740.8
1629.1

.133
.175
.194
.198
.200
.214
.285
.305
.318
.328
.372
.397
.413
.472
.506

1
1
1
1
1
2
2
2
2
3
3
3
4
4a
5
aSelected

as the most parsimonious solution.

A Clusterwise Bilinear Multidimensional Scaling Methodology


Figures 2 and 3 present the resultant joint space (varimax
rotated) for this R = 4/S = 5 solution. Figure 2 presents
Dimension 1 (horizontal) versus Dimension 2 (vertical),
and Figure 3 displays Dimension 3 (horizontal) versus
Dimension 4 (vertical). The brands are represented as
points in black, and the segments are represented as red
dashed vectors. For interpretation purposes, we property fit
each of the 25 attributes (23 image attributes + mean consideration and market share) onto this space (plotted as
solid blue vectors whose length reflects their relationship to
the dimensions plotted). In addition, Table 3 presents the
correlations between the four rotated dimensions and the 25
attributes. From an inspection of these figures and the correlations, we can easily ascribed interpretations to these
four derived dimensions. Dimension 1 describes a rugged
off-road latent construct, which has high correlations with
ride handling off-road (.965), built rugged tough (.967),
and fun to drive (.942). Dimension 2 can be described as
economical considerations, which displays high correlations with gas mileage (.712), value for the money

287

(.815), lasts a long time (.849), and high trade-in value


(.851). Dimension 3 can be labeled as a size or practicality dimension, which has high correlations with the attributes good vehicle for family use (.969), interior passenger room (.956), large cargo space (.970), and easy to
enter/exit (.931). Finally, Dimension 4 can be labeled as
popularity, given that its largest correlations are with
mean consideration (.702) and market share (.424).
Note that some of the directions of the attribute vectors
were reversed, and negative labels were used for ease of
plot interpretation, given the spacing in these figures. For
example, we used difficult to enter/exit instead of easy
to enter/exit in Figure 3. However, the fit to the dimension
remains the same, as illustrated by the length of the attribute arrow.
The five red dashed vectors representing the five segments span all four quadrants in Figure 2 with respect to the
first two dimensions, with substantial dispersion among the
various vector orientations, indicating heterogeneity of
preference among these five derived market segments.

Figure 2
CLUSTERWISE SOLUTION: DIMENSION 1 VERSUS DIMENSION 2

288

JOURNAL OF MARKETING RESEARCH, JUNE 2008


Figure 3
CLUSTERWISE SOLUTION: DIMENSION 3 VERSUS DIMENSION 4

Mean consideration
Market
Not prestigious
Does not last a long time
Poor gas mileage
Poor value for the money
Poor cargo space
Difficult to enter/exit
Difficult to load/unload

looking
Fun GoodReasonably
priced
drive
Passenger room
Excellent acceleration
Good vehicle for family use

Not
sporty
Workmanship

However, in Figure 3, there are similarities with respect to


Segments 1 and 2 and between Segments 3 and 4, given
how close (in terms of angles) these two pairs of vectors
are. Conversely, Segment 5 is well distinguished from the
other four derived segments in this figure with respect to
Dimensions 3 and 4. As an alternative indication of relationships (overlap) between the derived market segments,
we present the probabilities of classification in Table 4,
computed as (PP)/300, which provides some information
about overlap and cluster sizes. The main diagonal of this
table gives an indication of the number of consumers classified into each derived market segment (148, 130, 146, 164,
and 180, respectively, for each of the five derived market
segments). Thus, each derived segment is indeed sizable,
containing between 43.3% and 60.0% of the sample. The
off-diagonal elements of this table indicate the average
probabilities that membership in segments i and j jointly
occur in the sample. This two-way overlap ranges from
.210 to .350. These sizable off-diagonal elements reflect the
nature of the overlapping segment solution in this particular
application (the nonoverlapping solution would render

Luxurious
Good looking
Technically advanced
High trade-in value

zeros in such off-diagonal elements). Note that the overlapping solution rendered a 34.1% improvement in VAF compared with the nonoverlapping solution for this same R and
S. Other calculations can also be explored to quantify
higher-order overlap (i.e., memberships in 3, 4, or all 5
segments).
Table 5 displays the calculated scalar product utility predictions for each of the five derived market segments for
each of the 11 calibration brands. Segment 1 prefers the
Chevy Tracker, Jeep Wrangler, and Hyundai Sante Fe (offroad); Segment 2 prefers the Toyota RAV4, Honda CR-V,
and Mazda Tribute (economical considerations); Segment 3
has overwhelming preference for the Jeep Liberty (popularity); Segment 4 prefers the Suburu Forrester and Saturn
VUE (size); and Segment 5 prefers the Hyundai Sante Fe,
Ford Escape, and Honda CR-V (practicality). Thus, there
are rather distinct preference sets among these five derived
segments. Note that predicted utilities for overlapping segments can also be formed by adding the appropriate
columns of this table to reflect the particular combination
of market segments desired.

A Clusterwise Bilinear Multidimensional Scaling Methodology


Traditional Latent-Class MDS Solution
We also ran the MULTICLUS latent-class vector MDS
procedure on this same data set as a comparison with or
alternative to the proposed model. We present an overview
of the MULTICLUS results in this section, and details are
Table 3
CORRELATIONS BETWEEN DIMENSIONS AND ATTRIBUTES
Dimension
Attribute
Gas mileage
Value for the money
Workmanship
Good ride/handling
Easy to load/unload
Luxurious
Good safety
Ride/handling off-road
Built rugged and tough
Reasonably priced
Sporty
Good looking
Good vehicle for family use
Fun to drive
Interior passenger room
Easy to enter/exit
Dependable
Excellent acceleration
Cargo space
Lasts a long time
Prestigious
Technically advanced
High trade-in value
Mean consideration
Market share

.454
.125
.467
.290
.488
.280
.332
.965
.967
.382
.887
.856
.111
.942
.124
.486
.545
.825
.239
.562
.698
.397
.519
.340
.349

.712
.815
.881
.685
.460
.460
.568
.182
.214
.070
.339
.244
.283
.208
.238
.303
.834
.436
.323
.849
.615
.749
.851
.058
.176

.287
.535
.428
.779
.821
.911
.865
.184
.184
.535
.115
.514
.969
.204
.956
.931
.420
.608
.970
.351
.654
.679
.432
.295
.409

.106
.011
.215
.096
.112
.021
.135
.040
.066
.240
.132
.244
.029
.201
.196
.146
.146
.048
.069
.187
.115
.140
.197
.702
.424

Table 4
CLASSIFICATION PROBABILITIES FOR THE CLUSTERWISE
OVERLAPPING SOLUTION

Segment
Segment
1
2
3
4
5

.493
.210
.280
.323
.277

.210
.433
.250
.227
.273

.280
.250
.487
.267
.263

.323
.227
.267
.547
.350

.277
.273
.263
.350
.600

289

available in the Web Appendix (see http://www.marketing


power.com/jmrjune08). Given the row standardization preprocessing we performed, it is not possible to estimate full
segment-level covariance matrices, because they are not
identifiable with the rank reduction implied by this preprocessing. Thus, for the MULTICLUS procedure, the estimation relied on segment-distinct diagonal covariance matrices performed in R S = 1, , 5, as was done with the
clusterwise procedure we presented previously. Here, too,
we performed five analyses per run and selected the largest
fit values given the possibility of locally optimum solutions.
An immediately evident problem was that the two model
selection heuristics that Wedel and Kamakura (2000) discussminimum Bayesian information criterion and consistent Akaike information criterion (the criteria that Rust and
colleagues [1995] recommend)point to the R = 3/S = 4
solution, whereas the Akaike information criterion heuristic, notorious for selecting overparameterized models, suggests that the R = S = 5 solution is the most parsimonious
(also indicating that further fitting beyond five dimensions
and five segments may be needed). According to an inspection of the VAF statistic, the R = 4/S = 5 solution is the
most parsimonious. Thus, as we mentioned previously, different solutions are obtained depending on the criterion
used for model selection in such latent-class MDS models.
The later solution (R = 4/S = 5) is what we use for MULTICLUS to compare with our proposed methodologys
results.
The correlation coefficients between dimensions and
attributes for the first two dimensions were not very large in
absolute magnitudes, and interpretation is somewhat
ambiguous for at least one of them. For Dimension 1,
attributes with somewhat higher correlations, such as high
trade-in value (.527), workmanship (.510),
dependable (.489), and lasts a long time (.468),
seemingly indicate a quality dimension. Dimension 2 is
difficult to interpret because only one attribute (mean consideration) has a correlation coefficient larger than .400 in
absolute value. Dimension 3 resembles Dimension 4 from
the proposed clusterwise procedure popularity, given the
highest correlations with mean consideration (.769) and
market share (.576). Dimension 4 seemingly reflects the
economic considerations construct, with somewhat high
correlations with reasonably priced (.714), sporty
(.527), and gas mileage (.492). Furthermore, the canonical correlation comparison between the two brand configu-

Table 5
PREDICTED UTILITY VALUES OF 11 SUVS BY DERIVED SEGMENT
Segment
Brand
Chevy Tracker
Jeep Wrangler
Hyundai Santa Fe
Subaru Outback
Toyota RAV4
Honda CR-V
Ford Escape
Subaru Forrester
Mazda Tribute
Saturn VUE
Jeep Liberty

1.058
.946
.746
.786
.102
.368
.219
.645
.418
.503
.147

.152
.490
.549
.055
.815
.613
.823
.348
.613
.066
.648

.695
.466
.614
.439
.287
.371
.151
.417
.346
.155
.994

.040
.871
.343
.306
.183
.267
.067
.691
.162
.650
.206

.798
.556
.835
.052
.266
.596
.726
.400
.432
.400
.221

290

JOURNAL OF MARKETING RESEARCH, JUNE 2008

rations derived for MULTICLUS and the proposed clusterwise procedure suggests that, at most, two of the four
dimensions from both analyses can be deemed to be somewhat similar because only the first two canonical correlations are statistically significant. Thus, the two procedures
render different results.
The mixing proportions denoting segment sizes for
MULTICLUS were .230, .262, .094, .300, and .114 for the
five segments, respectively. The MULTICLUS predicted
utility scores per derived segment suggest that Segment 1
prefers the Honda CR-V and Ford Escape, Segment 2
prefers the Mazda Tribute and Ford Escape, Segment 3
prefers the Hyundai Sante Fe, Segment 4 prefers the Jeep
Liberty and Jeep Wrangler, and Segment 5 prefers the
Hyundai Sante Fe and Jeep Liberty.

cles. We performed similar stepwise regression analyses to


obtain predictions for the 7 brands in our clusterwise
derived space. Using the calibration data for the 11 brands,
we formed segment classifications on the basis of nearestneighbor decision rules to the centroids of the computed
possible segments and then resultant predictions of intentions for the proposed clusterwise procedure. Table 7 shows
the correlations between actual and predicted intentions by
brand and overall for both methodologies. Again, the proposed clusterwise procedure performs better than MULTICLUS with out-of-sample predictions. In particular, the
proposed clusterwise methodology out-predicts MULTICLUS with respect to all 7 validation brands. Overall, there
is a (.596 .433)/.433 = 37.6% improvement in prediction
validation for the proposed clusterwise spatial model over
MULTICLUS.

Comparative Predictive Validation


To compare these two procedures (MULTICLUS and the
proposed clusterwise procedure) properly with respect to
their performance on this particular data set, we first computed calibration correlations between the actual and the
predicted buying intentions for each of these two methodologies with respect to the 11 brands used in the analysis
for calibration (in-sample fit). Table 6 depicts the overall
and brand-specific comparative results for the two procedures. It appears that the clusterwise procedure produces a
much higher fit to the actual data than the MULTICLUS
solution (.687 versus .492 overall, an improvement of
[.617 .516]/.516 = 25.4%). In examining the comparative
fits by brands, we observe that the proposed clusterwise
procedure dominates MULTICLUS in the fitting of 9 of the
11 calibration brands.
Finally, we return to the 7 brands not used in the calibration sample and the 60 consumers held out of the estimation analysis. Recall that the 7 brands were Rodeo Sport,
Vitara, Grand Vitara, Freelander, Suzuki, Outlander, and the
Element. For MULTICLUS, we assigned the 60 consumers
not used in model calibration to the segment that provided
the highest likelihood for each consumer. On the basis of
stepwise regression analyses by dimension performed for
the brand configurations derived in the MULTICLUS
analysis, we formed predicted brand locations in this four
dimensional MULTICLUS space for these 7 brands not
included in the original MULTICLUS solution. Using the
classification matrix calculated, we formed intention-to-buy
predictions for these 60 consumers for each of the 7 vehiTable 6

The Reparemeterized Solution


We present a summary of our analysis performed with
our proposed methodology using the brand reparameterization option. Recall that for K > J, all the attributes cannot
be explicitly used in reparameterizing the brand locations
as exact linear functions of the attributes, and thus we performed principal components analysis on the entire set of
25 attributes. Using the eigenvalue > 1 rule, we extracted
four components. Notably, the four rotated components
were similar in interpretation to the four dimensions
reported in the nonreparameterized solution. We performed
a canonical correlation between these rotated principal
components analysis scores for the brands and the four
dimensions obtained from the proposed clusterwise procedure as an approximate configuration matching procedure
and obtained canonical correlations of .999, .999, .998, and
.893, which showed strong consistency between the two
configurations. We then ran the reparameterized model for
S R = 1, , 5, as we had done for the nonreparameterized model. The S = 5/R = 4 solution rendered SSE =
1774.54 and VAF = .463 (versus SSE = 1740.8 and VAF =
.472 for the nonreparameterized model). Thus, the structure
in the attribute data appears to overlap substantially with
the structure in the preference data, suggesting that consumers seemingly use these particular attributes to evoke
their preferences in this application. Table 3 somewhat confirms this, given the high correlations.
DISCUSSION
We introduced a new clusterwise bilinear spatial MDS
model for use in marketing that performs data reduction
and classification simultaneously, which is useful for simul-

COMPARISONS OF PREDICTIVE VALIDITY: WITHIN SAMPLE


MULTICLUS
Chevy Tracker
Jeep Wrangler
Hyundai Santa Fe
Subaru Outback
Toyota RAV4
Honda CR-V
Ford Escape
Subaru Forrester
Mazda Tribute
Saturn VUE
Jeep Liberty
Overall

.556
.351
.563
.239
.601
.468
.566
.276
.633
.294
.256
.492

Table 7

Clusterwise
.458
.589
.736
.426
.780
.654
.708
.817
.814
.263
.824
.687

COMPARISONS OF PREDICTIVE VALIDITY: OUT OF SAMPLE

Rodeo Sport
Vitara
Grand Vitara
Freelander
Suzuki
Outlander
Element
Overall

MULTICLUS

Clusterwise

.447
.485
.423
.455
.395
.394
.280
.433

.507
.643
.646
.857
.413
.399
.472
.596

A Clusterwise Bilinear Multidimensional Scaling Methodology


taneous segmentation and positioning. Unlike its latentclass MDS counterparts, the procedure does not require distributional assumptions for estimation purposes. In addition, the alternating least squares estimation algorithm
developed for parameter estimation renders conditionally
globally optimum estimates at each stage of the estimation
per iteration relatively quickly, instead of more computationally intensive latent-class MDS methods, which may
require hours for convergence. The procedure can accommodate internal or external analyses (i.e., it can be used
when X, Y, and/or P is fixed in the analysis) to test for
specified model structures. More important, unlike its
latent-class MDS counterparts, it accommodates either partitions or overlapping segments and does not present difficulties related to identification problems associated with
such mixture distributions (McLachlan and Krishnan 1997;
McLachlan and Peel 2000). We also developed consumer
conditional analyses and brand reparameterization options.
We demonstrated the proposed methodology in the context of segmentation and positioning for an actual commercial study sponsored by a large U.S. automotive manufacturer involving the class of small SUVs. We derived a
five-segment, four-dimensional solution that carries substantial face validity, given what is known about that marketplace at the time of the study (one of the authors is an
industry expert). Comparison of in-sample and out-ofsample fit values between the proposed methodology and
that of the MULTICLUS confirms the superiority of the
proposed procedure, with 25.4% improvement in in-sample
fit and 37.6% improvement in out-of-sample fit.
The spatial maps that the proposed methodology estimates are of immense use in several segmentation
targetingpositioning marketing decisions, including new
brand introductions, brand repositioning, targeting, and
management of brand portfolios, among others. As with
GENFOLD2 (DeSarbo and Rao 1986), the reparameterization option in Equation 2 makes it possible to perform
reverse mapping from the attribute domain into the brand
space. For example, policy simulations can be performed
under what-if scenarios for new brand introductions with
known attributes, features, marketing mix, and so forth, to
obtain predicted positions in the joint space, just as we did
with the predictive validation exercise discussed previously
with the holdout set of seven brands. Thus, predictions of
market share and segment demand can be made at the brand
and segment level for new brands. Similarly, such simulations can also be performed in repositioning exercises with
existing brands and various redesign strategies. Insights
into targeting can be obtained when target market segments
are identified and when the goal is to maximize the appeal
of new and/or existing brands in terms of alterations in the
attributes, features, marketing mix, and so forth, to these
targeted market segments. When a particular manufacturer
possesses multiple brands in the product class under investigation, it is possible to examine the effects of new brands
and/or repositioned brands on an entire portfolio of brands
under the same manufacturer and to examine cannibalization under such policy simulations.
Indeed, such maps hold immense potential for decisions
about marketing strategy. In addition to problems involving
segmentation and positioning, the proposed bilinear MDS

291

methodology can be gainfully employed to examine other


various strategic marketing problems. For example, this
proposed methodology could be applied to profile data
(firms or strategic business units in an industry measured
on various strategic and performance variables) to estimate
strategic groups and the dimensions underlying the strategic
groups (e.g., Frazier and Howell 1983). Here, groups of
firm vectors (strategic groups) would be identified in a
dimensional space of strategic and performance variables in
which the orientation of the estimated vectors indicated
directions of increasing ability or performance for the
derived strategic groups. Alternatively, the proposed clusterwise bilinear spatial methodology could be applied for
positioning analyses when a matrix of brands (rows) and
attributes (columns) is obtained for a designated product/
service class. Here, groupings of brands (vectors) might be
indicative of competitiveness in the derived joint space in
which the attributes would be represented by coordinate
points. The proposed methodology suffers from the same
type of limitations as all MDS models in terms of dimension reduction and extending predictions outside the range
of the brands used for calibration (e.g., attempting to make
market share predictions for new brands that are introducing new technology to the product/service class).
In addition to applications to other marketing research
problems, further research could extend our methodology
in several directions. First, extensions of this procedure to
three-way data and dynamic analyses would be insightful.
Second, a fully nonmetric version would be useful for
applications involving ordinal scales of measurement.
Third, extending the utility model specification to an ideal
point or unfolding model would prove advantageous, especially for optimal positioning efforts, given that the underlying utility assumptions for the vector model (the more
the better) are not conducive to such analyses without
severe constraints. Fourth, more extensive testing of the
procedure and comparisons with alternative traditional and
latent-class MDS models would be desirable through more
rigorous Monte Carlo testing with synthetic data structures.
Fifth, research on developing better heuristics for model
selection is desirable, especially penalized heuristics that
adjust for the number of free parameters estimated in such
clusterwise MDS and classification procedures. Sixth,
relaxing the assumption of segment homogeneity in terms
of a single brand space applicable across all segments (an
assumption in both latent-class MDS and our proposed
clusterwise approach) could be achievable by estimating
Xjrs in this alternating least squares procedure. Seventh,
extending the methodology to accommodate fuzzy clusters
might prove beneficial in certain applications in which
there is reason to suspect fractional membership in multiple
segments. Eighth, research that develops the normative
aspects of the methodology would prove beneficial to managers. Here, mathematical solutions to problems associated
with optimal positioning, cannibalism, and portfolio management, given the estimated space, could be explored.
Finally, as we mentioned previously, applying the procedure across a multitude of different marketing applications
and data would better reveal the utility of such a spatial
methodology. We hope that our research will spur additional studies on these important issues.

292

JOURNAL OF MARKETING RESEARCH, JUNE 2008


REFERENCES

Arabie, Phipps, J. Douglas Carroll, Wayne S. DeSarbo, and Yorum


Wind (1981), Overlapping Clustering: A New Method for
Product Positioning, Journal of Marketing Research, 18
(August), 31017.
Bckenholt, Ulf and Ingo Bckenholt (1991), Constrained Latent
Class Analysis: Simultaneous Classification and Scaling of Discrete Choice Data, Psychometrika, 56 (4), 699716.
Carroll, J. Douglas (1980), Models and Methods for Multidimensional Analysis of Preferential Choice Data, in Similarity and
Choice, E.D. Lantermann and H. Feger, eds. Bern, Switzerland:
Hans Huber, 23489.
, Sandra Pruzansky, and Joseph B. Kruskal (1980), CANDELINC: A General Approach to Multidimensional Analysis of
Many-Way Arrays with Linear Constraints on Parameters, Psychometrika, 45 (1), 324.
Chintagunta, Pradeep K. (1994), Heterogeneous Logit Model
Implications for Brand Positioning, Journal of Marketing
Research, 31 (May), 304311.
Coombs, C.H. (1964), A Theory of Data. New York: John Wiley &
Sons.
Cooper, Lee G. (1988), Competitive Maps: The Structure Underlying Asymmetric Cross Elasticities, Management Science, 34
(June), 707723.
Day, George S. (1984), Strategic Market Planning: The Pursuit of
Competitive Advantage. St. Paul, MN: West Publishing Co.
, Allan D. Shocker, and Rajendra K. Srivastava (1979),
Customer-Oriented Approaches to Identifying Product Markets, Journal of Marketing, 43 (October), 819.
Dempster, A.P., M.N. Laird, and D.B. Rubin (1977), Maximum
Likelihood Estimation from Incomplete Data via the E-M Algorithm, Journal of the Royal Statistical Society, Series B, 39 (1),
138.
DeSarbo, Wayne S. and J. Douglas Carroll (1985), Three-Way
Metric Unfolding via Alternating Weighted Least Squares,
Psychometrika, 50 (3), 275300.
, Daniel J. Howard, and Kamel Jedidi (1991), MULTICLUS: A New Method for Simultaneously Performing Multidimensional Scaling and Cluster Analysis, Psychometrika, 56
(1), 12136.
, Kamel Jedidi, Karel Cool, and Dan Schendel (1991),
Simultaneous Multidimensional Unfolding and Cluster Analysis: An Investigation of Strategic Groups, Marketing Letters, 2
(2), 12946.
, Ajay K. Manrai, and L.A. Manrai (1993), Non-Spatial
Tree Models for the Assessment of Competitive Market Structures, in Marketing, Jehoshua Eliashberg and Gary L. Lilien,
eds. North Holland: Elsevier Science, 193258.
, , and (1994), Latent Class Multidimensional Scaling: A Review of Recent Developments in Marketing
and Psychometric Literature, in Advanced Methods of Marketing Research, Richard P. Bagozzi, ed. Cambridge, MA: Blackwell, 190222.
, Venkatram Ramaswamy, and Peter Lenk (1993), A
Latent Class Procedure for the Structural Analysis of Two-Way
Compositional Data, Journal of Classification, 10 (2), 15993.
and Vithala R. Rao (1986), GENFOLD2: A Constrained
Unfolding Model for Product Positioning, Marketing Science,
5 (1), 119.
De Soete, Geert and Willem J. Heiser (1993), A Latent Class
Unfolding Model for Analyzing Single Stimulus Preference
Ratings, Psychometrika, 58 (4), 54565.

and Suzanne Winsberg (1993), A Latent Class Vector


Model for Preference Ratings, Journal of Classification, 10
(2), 192218.
Fraser, Cynthia and John W. Bradford (1984), Competitive Market Structure Analysis: A Reply, Journal of Consumer
Research, 11 (3), 84247.
Frazier, Gary L. and Roy D. Howell (1983), Business Definition
and Performance, Journal of Marketing, 47 (April), 5967.
Green, Paul E. and Abba M. Krieger (1989), Recent Contributions to Optimal Product Positioning and Buyer Segmentation,
European Journal of Operational Research, 41 (2), 12741.
Holman, Edward W. (1972), The Relation Between Hierarchical
and Euclidean Models for Psychological Distances, Psychometrika, 37 (4), 41723.
J.D. Power and Associates (2003a), Sales Report: Highlights,
(August), (accessed February 21, 2008), [available at http://
www.jdpa.com/library/publications/srSample/index.htm].
(2003b), Sales Report: Model Summary, (August),
(accessed February 21, 2008), [available at http://www.jdpa.
com/library/publications/srSample/sr0803f.pdf].
Lattin, James M. and Leigh McAlister (1985), Using a VarietySeeking Model to Identify Substitute and Complementary Relationships Among Competing Products, Journal of Marketing
Research, 22 (August), 33039.
Lilien, Gary L. and Arvind Rangaswamy (2004), Marketing Engineering: Computer Assisted Marketing Analysis and Planning.
Reading, MA: Addison-Wesley.
Magnus, Jan R. and Heinz Neudecker (2002), Matrix Differential
Calculus with Applications in Statistics and Economentrics, rev.
ed. New York: John Wiley & Sons.
McLachlan, Geoffrey and T. Krishnan (1997), The EM Algorithm
and Extensions. New York: John Wiley & Sons.
and David Peel (2000), Finite Mixture Models. New York:
John Wiley & Sons.
Moore, William L. and Russell S. Winer (1987), A Panel-Data
Based Method for Merging Joint Space and Market Response
Function Estimation, Marketing Science, 6 (1), 2542.
Myers, John H. (1996), Segmentation and Positioning for Strategic Marketing Decisions. Chicago: American Marketing
Association.
Reutterer, Thomas (1998), Competitive Market Structure and
Segmentation Analysis with Self-Organizing Feature Maps, in
Proceeding of the 27th EMAC Conference, Vol. 5, P. Andersson,
ed. Stockholm: EMAC, 85115.
Rust, Roland T., Duncan Simester, Roderick J. Brody, and V.
Nilikant (1995), Model Selection Criteria: An Investigation of
Relative Accuracy, Posterior Probabilities, and Combinations of
Criteria, Management Science, 41 (2), 32233.
Slater, Patrick (1960), The Analysis of Personal Preferences,
British Journal of Statistical Psychology, 13, 11935.
Tucker, L.R. (1960), Intra-Individual and Inter-Individual Multidimensionality, in Psychological Scaling, H. Gulliksen and S.
Messick, eds. New York: John Wiley & Sons, 15567.
Wedel, Michel and Wayne S. DeSarbo (1996), An ExponentialFamily Multidimensional Scaling Mixture Methodology, Journal of Business & Economic Statistics, 14 (4), 44759.
and Wagner A. Kamakura (2000), Market Segmentation:
Conceptual and Methodological Foundations. Boston: Kluwer.
Wind, Yoram and Thomas S. Robertson (1983), Marketing Strategy: New Directions for Theory and Research, Journal of Marketing, 47 (April), 1225.

You might also like