Академический Документы
Профессиональный Документы
Культура Документы
DeSARBC
The authors present a new measurement methodology of perceived
value, based on latent structure multidimensional scaling, that derives
simultaneously the underlying dimensions of the perceived value of various brands and market segment heterogeneity in terms of how such
value evaluations are made. This latent structure, ordered probit, multidimensional scaling (MDS) based methodology improves on existing industry techniques of illustrating perceived customer value because it enables
researchers to infer the underlying dimensions of perceived value from
the data without specifying these a priori, as is common in existing methods. The authors compare the proposed model against more traditional
MDS approaches in an empirical illustration involving the perceived value
of compact cars. Finally, the authors discuss managerial implications of
this technique and provide directions for further research.
Customer value management (CVM) recently has become a major focus in contemporary marketing, as value
marketing has become a watchword annmg marketing practitioners (cf. BusinessWeek 1991; Gale 1994; Treacy and
Wiersema 1995). Perceived value has been called the "new
marketing mania" and "the way to sell in the 9()s" (BusinessWeek 1991). In the marketplace, value often is defined
as "quality at the right price" iProi^ressive Grocer 1984)
and is seen as more important to consumers than quality, because value is quality that the consumers can afford. Industry gurus such as Gale (1994) and Treacy and Wiersema
(199.')) have called tor better management of the perceived
customer value of a firm's consumers as the uppermost priority of its executives. To that extent, customer value analysis has been accorded particular importance, and a host of
value mapping approaches have been advanced in the industry (see, e.g., Brayman 1996; Gale 1994). It has been reported that the AT&T board now is supplied regularly with
three metrics; customer value, overall quality, and price
competitiveness (Gale 1994). Ongoing research at AT&T
has provided evidence of a significant relaiionship between
improved customer value ratings by consumers and overall
market share gains (Kordupteski and Vogel 1989).
236
237
Sources of Heterogeneity
ii
A
Bi'tlvr<'usluiii.'r
Valuv
.Super
Various sources of heterogeneity can be identified regarding the perception of value that arises from differences
among consumers, product classes, and consumptive situations (Holbrook 1994; Zeithaml 1988). Empirical evidence
ha.s demonstrated that intersegment differences inOuence
the weighting ofthe perceived value dimensions. For example, Zeithaml (1988) reports that one segment judged value
from only quality, another from only priee. a third from both
quality and price, and, finally, a fourth from all "get" and
"give" eomponents. Segment characteristics (i.e.. background variables) such as age and income al.so have been
shown to affect the perception of value (Boiton and Drew
1991). Customer value analysis is also dependent on the
product class insofar that tbe salient value dimensions for a
238
different brands are perceived by different consumer segments in tenTis of value, that is, if the brand locations are
found to be positioned above, in, or below the thresholds.
METHODOLOGY
The VALUBMAP mode! is implemented as a latent .structure, bilinear, MDS vector model with reference thresholds
that use consumers* empirical categorizations of perceived
value (relative to expectations) as input data. Therefore,
subjects are asked in the empirical study only if the perceived value of a given brand is better than expected, same
as expected, or worse than expected^just as a shopper
browsing at a slore can determine quickly which are the
"good deals,"" the "so-so value" brands, and the "rip-offs."
Varian (1987) notes that few people have been known to
rate the value of individual brands in a numerically reliable
fashion. (Economists circumvent this issue by assuming
ordinal utility.) Measurement researchers find that if a stimulus is not highly discriminable, a scale with several categories does not improve accuracy and can be confusing (cf.
Komorita and Graham 1965). There is also some empirical
evidence showing tbat a scale with fewer response categories is considered easier to rate by subjects (see Diefenbach, Weinstein. and O'Reilly 1993). Although the
trichotomous category scale used here makes minimal
demands on the subjects" reporting a judgment as complex
as perceived value, the predicted perceived value of brands
from the model in the brand-segment space is still interpretable on a metric scale.
Let us define the notations as i = I, ..., I consumers; j = 1,
..,, J brands; k = I, ..., K attributes; s - I, ..., S segments or
latent classes; t - 1,..., T dimensions; and r = 1, ..., R replications (e.g., time, consumptive situations, experimental
treatments).
Furthermore, we define
5||p = an empirical categorization given by consumer i of
brand j in replication r if value is less tban expected, same as expected, or more than expected (coded
arbitrarily as - I , 0, I, respectively);
a^, = vector coordinate for segment s on t'*" dimension;
bj, = brand j location on t"' dimension;
U^ = upper threshold for segment s; and
L^ = lower threshold for segment s.
In the usual latent-class assumption, we assume there are
S (unknown) segments or latent classes, and each consumer
belongs to only one of these. Let the probability of consumer i belonging to a segment s ( - I
S) be given by A.^,
where 0 < X., < 1 and l.X = '
Now, let us denote the perceived value of a brand j for
consumer i (conditional on being in segment s) in replication r to be denoted by "^i\f\^~ and we represent perceived value in a scalar products or vector formulation (cf. Tucker
I960)as
{1}
a b
I
^'
Jl
ijr s
239
(2)
Ijrs
ijt\.
J -Va b
a b.
>U -
=P
^^
M Jt
=0
- P L < V
< U^
= PI
< e . I <
P5
=-1 \=P\ V
<L U
L - > ;i h
where:
a.ir - 1, if 5ijr = - I , 0 otherwise;
p = 1, i f 6,.r - 0, 0 otherwise; and
Y ,. - 1, if'S,.f- 1 , 0 otberwise.
Tbe unconditional likelihood is then L; = S^X^L^j^, where
the X^s are the mixing proportions thai satisfy tbe usual
probability constraints. The complete likelihood function
across all consumers can be written as
(6) ^ =
, hence In L = /
In
(7)
where Pj,, represents the posterior probability of the i''' consumer of belonging to segment s. A^ is the estimated segment weight, and L||^ is the estimated likelihood function for
the individual subject at any given iterate.
Given the A - ||(6|j|)|} and Z matrices, the objective is to
obtain maximum-likelihood estimates of a,b,(or T),U,L,X, to
maximize the log-likelihood in Equation 6. As such, the proposed methodology can be viewed roughly as a multidimensional and latent class extension of Thurstone's Law of
Categorical Judgment (Thurstone 1927, 1959; Torgerson
1958) that involves replication over persons (Class II) and
successive intervals (Condition D). This latent structure
MDS lormulaiion extends the methodological developments of Carroll (1972, 1980), DeSarbo and Cho (1989),
Bockenholt and Gaul (1989). and DeSarbo, Howard, and Jedidi (1991). Carroll (1972, 1980) developed MDPREI-a
metric MDS vector model hased on singular value decomposition for the analysis of metric preference or dominance
data. Here, eacb consumer is represented by a vector pointing in the direction of increasing preference, and brands are
represented by points in this T-dimensional joint space. As
such, our proposed methodology is not restricted solely to
the analysis of perceived value, but also can be used to represent preference or cboice judgments obtained from a trinary ordered scale. DeSarbo and Cho (1989) present a stochastic MDS threshold vector model for pick any/N binary
choice data that enables reparameterization of Ihe stimulus
space, but the vectors also are estimated at the individual
level. Bockenholt and Gaul (1989) present vector and unfolding MDS models for a latent structure analysis of binary pick any/N data, as well as a general framework for ordered category data (without opcrationali/ing il in tcmis of
a demonstrated st>ftware program). DeSarbo, Howard, and
Jedidi (1991) generalize this latent structure approach
(MULTICLUS) to continuous data (De Soete and Winsberg
(19931 generalized the DeSarbo, Howard, and Jedidi |I991|
approach two years later to accommodate linear restrictions). The present lalent structure MDS approach operates
on trinary ordinal scale data (not binary choice or continuous dala), estimales segment level vectors, and allows for a
reparameterization of the stimulus/brand space, with reference value thresholds. We chose to represent perceived value by a vector model (versus an ideal point model) because
the monotonicity assumption ("the more the belter") makes
explicit sense (cf. Dodds, Monroe, and Grewal 1991) iii this
240
241
Maittifacturer
Ri'liahilitv
(tnilf\ jH'r
i<alli>n)
Safety
Cosl
Factor
A iTt-a\-i'
Dcprcciatidti
Pi'rftirmance
Price
87%
Betier than
average
Average
% 12.(X)0
Saltirn
GM
Average
2.1
A verage
Subaru Iniprc/a
OJM'
Better than
average
29
Beller ihan
average
91%
Average
Betier than
average
14,(KH>
Chevy Cavalier
GM
Average
26
Average
94%
Average
Beller Ihan
average
13.000
Eagle Sumiiiii
Chrysler
Worse than
average
34
Worse ihati
average
92%
Worse than
average
Worse Ihan
average
12.000
Dodge Neon
Chrysler
Average
31
Better than
average
93%
Average
Better than
average
11.000
Ford Escort
Ford
Average
33
Average
92%
Worse than
average
Worse than
average
!I.(HH)
Geo Metro
GM
Average
33
Average
93%
Average
Average
9.000
Honda Civic
Honda
Belter than
average
29
Average
90%
Woi>c than
average
Better than
average
I5.(K)O
MazJa Protege
OJM
Average
33
Average
90%
Worse than
average
Better than
average
15.(KM)
Mitsuhi.shi Mirage
OJM
Worse than
average
30
Worse than
average
90%
Average
Worse than
average
12.000
Toyoia Corolla
Toyoia
Betier than
average
30
Average
88%
Average
Beller than
average
16.(MK)
Nissan Sentra
OJM
Belter than
average
30
Average
88%
Average
Better than
average
13,(KM)
10
1 1
12
Rcpi/rts (1995).
cluster analyzing such MDS results. To illustrate these difficulties, we pcrtbnncd cluster analyses of the untiormalizcd
MDPREF vectors, includitig LI K-meatis cluster analysis atid
Ward hierarchical cluster analysis. The Ward dendrogram
pointed to two or three clusters, whereas the scree plot of the
within sum-ol'-squares by numbers of clusters in K-means
denoted three clusters. We focus on two clusters, given the
analyses to follow. A 2 x 2 cotitingeney table was constructed after optimally permuting cluster labels. Twentytwo subjects were classified differently between these two
schemes, and the resulting chi-square = 36.46.'S was significant beyond p = .001. which indicates that the two segmentation schemes appear utirelatcd lo each other, even though
they were computed on the basis of the same dataa typical problem encountered with naive, two-stage approaches.
To attempt to uncover market segments (cf. Zeithaml
1988), we applied the MULTICLUS latent structure vector
MDS model to this - I , 0. I vaiue data. As with MDPREF,
preprocessing will alfeet the results, so we sitiiply analyzed
the raw data in T, S = I
4. According to all intbmiation
heuristics, the T - S - 1 solution is the most parsinu)nit)us
solution, though it is rather uninteresting to exhibit.
Nonetheless, we still can convey the unidimensionality and
one segtnetit vector by examitiing the T = S = 2 solution in
Figure 3. This figure aptly describes a one-dimensional subplane on which the 12 automobiles order themselves. In addition, there is little angular separation between the two segments" vectors, which indicates virtually no heterogeneity
between these value perceptions. All that cati be gleaned
242
Price
Ruliabilily
Safety
/////'
/ill ' !
Cosl
The VALUEMAP two-segment, two-dimensit>nal unc(Mistrained solution is displayed In Figure 4. The estimated segment weights (X^s) are .64 and .36. Two distinct segments in
the data are indicated by the clear separation of the segment
vectors in the two-dimensional space, as well as by an entropy-based measure (bounded between 0 and 1) that examines the centroid separation of the conditional parametric
distributions (.see DeSarbo et al. 1992). The entropy value in
this case was .86 (a value close to I indicates that the centroids are well separated for the ntimber of segments specified). The (rotated) location for the first segment is 1.01 and
.43 in two dimensions. Therefore, Segment I weights the
horizontal dimension (Di[nension I) much more than the
vertical one. In contrast. Segment 2 weights the vertical dimension (Dimension 2) more than the horizontal, as its (rotated) location is .43 and -.75. We note in Figure 4 that, for
Segment I. the Subaru Impreza, Nissan Sentra, and Dodge
Neon are considered good values, whereas the Mazda Protege, Honda Civie, Toyota Corolla, Saturn. Geo Metro,
Chevrolet Cavalier, and Ford Hscort are projected near reference values. The Civic and Protege are poor values for
Segment 2, whereas the Impreza. Sentra, Neon, and Metro
are good values. Note that for both segments Mitsubishi Mirage and Fagle Summit are perceived as poor values.
243
Figure 3
MULTICLUS T = S - 2 SOLUTION
* Saturn
* Sentra
*lmpreza
Segment 2
*Cavalier
*Neon
*Metro
C orolla
Segment 1
^Civic
*Protege
*Escort
'Summit
*l\/lirage
Table 2
MODEL CALIBRATION RESULTS: KEY STATISTICS
Crilcrui
Free parameters (#)
t,og-likelihood
Modified Ate
BiC
CAIC
S^
t T= I
14
-84L77
17''5.54
1747.^9
1761 29
S^2T^t
18
-829.57
1713.14
1741.11
1759-11
.V
^2T=2
29
-795.68
I678.-K1
1723.42*
1752.42*
S = 3T^ 1
22
-819.25
1704.50
17.18 69
1760.69
S = _i T = 2
34
-788.43
1678.86
1731.69
1765.69
.V = 3 T ^ J
44
-768-29
1668.58*
1736.95
1780.95
244
Figure 4
VALUEMAP OF 12 CARS
Ma; da Protege
Honda (I!!ivic
Segment 1
Upl
.2
Tpyola Corolla
Mitsubishi Mirage
L()2
. . . . . .
^y^
Chevy
Cavalier
Lol \
Eagle
Summit
Saturn
\,
/ ^ :., Ford
\ Escort \
Subaru
Impreza
Nissan
Sentra
Scgmctit 2
Geo
Metro
Dodge Neon
Dimension 1
Table 3
MODEL CALIBRATION RESULTS INTERPRETING THE
DIMENSIONS
Attribute
Amencan/impori
Reliahiliiy
COM factor
Pcrlormaiico
Price
Mileage
Safety"
Depreciation
Dimension t
Dtniensioti 2
-.1129
.466
-.072
.569
-.09
,093
.536
.077
-.653
.261
-.677
..^61
.825
.095
-288
.026
"Worse than Average" ratings hy Consumer Reports on performance, safety, and reliability atirihutes. Japanese cars,
such as the Ma/_da Protege, Honda Civic, and Toyota Corolla, are located high in Dimension 2, which is dominated by
the import origin, price, and cost factor. Low priced American cars, such as the Ford Escort, Geo Metro, and Dodge
Neon, are found to be low on this dimension. We note that, in
general, Japanese cars are more expensive than American
cars in this class and offer less bargaining leverage, which
explains why these cars are high on this dimension.
To compare these VALUEMAP results with those obtained from MDPREF and MULTICLUS more formally, we
used canonical correlation as an approximate configuration
matching procedure. In matehing the stimulus/brand space
between MDPRBF and VALUBMAP, we obtained canonical correlations of X\ = .998 and X2 - .957. which indicates
that hoth solutions can be rotated orthogonally tt) high congruence. Unfortunately, we have no easy way of comparing
the consumer versus segment vectors hecause of their different orders. We al.so compared the full MULTICLUS solution with that obtained frt)m VALUEMAP by stacking the
segment vectors (after similar normalizations with adjustment to the brand space) atop the brand coordinates. Here,
X^ = .995 and ^2 - -*^^ indicate one dimension in common
between these two solutions. (We also performed Ward and
K-means cluster analyses to the raw data in two clusters and
then pertbrmed a two dimensional MDPRFF analysis on
these averaged scores, as per the suggestion ot a reviewer.
The two sets of canonical correlations between these solutions and VALUEMAP were X, = ,745 and Xi ^ .351 for Kmeans, and k[ = .812 and ^2 = -313 for WARD. These results also indieate some, but not full, congruence with the
VALUEMAP solutions.) At the insistence of one reviewer,
we compared the K-means and Ward segmentation schemes
245
Table 4
MODEL CALIBRATION RESULTS CHARACTERIZING SEGMENT V
CoetfuieiUs
Value
.Stttiiilarcl Error
Intercept
Sex
Maniifucturcr
Oeprcciaiion
Siilety
Perlormance
I'ricc scnsitiviiy
Qualiiy seasitivily
Preference tor Japanese curs
-.512
.733
-.046
-.104
.03
.224
-.379
.157
.017
.191
.053
.013
.014
.015
.018
.024
,03
.007
t-vithii
-2.689
13.752**
-.3.521**
-7.002**^
1.988**
12.72**
16.086**
5.147**
2-445**
liitetprcttttitin
Mostly leniiile
Brand in.sensitive
Depreciation unimporlanl
Salety conscious
Performance consciotis
Price insensitive
Quality conscious
Prefers Japanese cars
246
Figure 5
VALUEMAP OF ^2 CARS (REPARAMETERIZED MODEL)
Honda Civic
Upl
^._
Mazda Protege
Toyota Corolla
X.
// c Segment. ,1
-.,..
Subaru
,
Inipreza
*
Nissan
/
Sentra
/
/
Cbevy
Dodge Neon
/
Satttrti
Cavalier
Lo2
. /
/
\,^^
X^
Mitsubishi Mirage
/
Lol
^^^-^
Geo
Metro
\ .
/Pord
SutTimii rbsciMl
Segment 2
Dimension 1
S=2T=2
classified entries. 44 (23^^) were off by one position (i.e.. 0
instead of 1. 0 instead of -1). and only two etitries were off
by two positions (I instead o f - 1 and vice versa). Figure 6
depicts the predicted locations for the two cars held out of
the analysis. Note how close they are to their actual positions in Figure 5. Consequently, the results of this validation
study using the etnpirical data enable us to establish further
the utility of the VALUEMAP methodology atid the estimated results describe!.! previously.
DISCUSSION
The VALUEMAP tnethodology allows for the sitnultaneous estimation of brand location paratneters in terms of
value perceptions and the segmentatioti of consumers in a
joint-dimensiotial space. The results help detertnine the
dimensions of perceived value for a given category for a
giveti pool of suhjects. The location ofthe brands also indicates how different brands fare on the identified diiiiensions
oi" value. For example. Figure 4 reveals that the Subaru
Itnpreza does well on the reliability, safety, and performance dimension, whereas tbe Dodge Neon fares well on
both dimensions, Tbe importance weights assigned by the
segments to the ditnensit)ns also are indicated by their estimated locations. The compositions of the dimensions were
247
Figure 6
VALUEMAP VALIDATION
Honda Civic
Upl
Mazda Protege
\ ^
Lo2
Mitsubishi Mirage
Segment 1
ToyotUyCorolla
Nissan
/
^Chevy Sentra Dodge Neon
/
Cavalier
r- i
/
Subaru
/
Saturn
Impreza
/
(predicted
(predicted
/
location)
location)
Lol
Geo Metro
/
'/..,
E^g'^ F o r /
Summit rhscdtt
Segment 2
Dimension 1
S = 2 T = 2 (Reparameterized)
LIEMAP provides insight as to how these suhjects integrate
this infonnation in deriving their value judgments. This
study could he performed with real consumers without such
prelitninary infonnation. which might bias value perceptiotis. Second, we bave incorpotated itidividual differences
as the only source of heterogeneityit is also important to
consider both product-class and situational beterogeneity.
Third, we used an obiective attribute tnatrix frotn Consutner
Reports to characterize the dimensions, hut it is reasonahle
to assume that a subjective perceptual attribute matrix might
capture reality better. Finally, though we purposely restricted consumer judgtnents to trinary ordered categorizations.
there i.s evidetice in the literature (e.g.. Cox !9S()) to suggest
that use of response scales with five to nine categories might
represent such perceptual evaluations more accurately. We
discuss this issue further iti the following section.
FUTURE RESEARCH AREAS
There are several ways in whieb this researeh on perceived value migbt be extended. Oti the substantive side, the
focus should be on investigating the process of how the
notion of value is fortned in the minds of consumers. Here,
the role of prior expectations as well as eontextual inlluences should be salient. Also, how value is perceived from
bundles tnight be an interesting research topic, because people generally regard bundles as better values Ibr the tnoney.
Finally, the recent surge in the popularity of store-lahel
brands over manufacturer labels might he of substantive
interest to the tnarketing researcher because this has potential implicatit)ns in the areas of eustomer value and brand
equity.
In our study, we itnpose sparse requiretnents about the input data-subjects were asked merely to categorize the given hrand as a better-than, same-as, or less-than-expected
value. The idea was that the response task should not be too
demanding, such as requiring them to provide quantitative
ratitigs on something as complex as v;tlue. One direction lor
furtber research would he to generalize this VALUEMAP
tnethodology to tnultivalued ordered category scales, as are
found in more getieralized studies involving successive category scaling. Also, the measurctncnt properties of the perceived value concept must be investigated more thoroughly.
To do this formally, hoth internal-consistency reliability and
tesi-retest reliability might need to he exatnined. as well as
the discriminant validity of perceived value established
from other related constructs such as utility, hrand equity,
customer satisfaction, and purchase intentioti. The multitrait, niultimethod approach and structural equation model-
248
ing provide ways of empirically establishing the measurement properties of theoretical constructs such as perceived
value.
On the tnethodological side, the VALUEMAP model can
be extended in sevetal directions. A feature of this model is
that it enables the segtnent tnembership probabilities of the
subjects (i.e,, the ^^s) to be reparameterized a priori, if desired, as an explicit logit function ofthe concomitant demographic variables, such us Gupta and Chintaguttta (1994)
etiiploy in their article, in this special instance, all the parameters can he estitnated simultaneously: however, the estimation procedure might become somewhat onerous, hi addition, such repataineterization otily affects the sizes o( these
derived segments. Still another methodological area for further research of the VALUEMAP approach involves an ideal point/unfolding tnodel specification. Here, we have assumed frotn the available empirical evidence (sec Dodds,
Monroe, and Grewal 1991) that perceived value is monotonic with respect to its underlying dimensions (e.g.. perceived
quality and price). However, in practice, some of these ditnensions migbt not be tnonotonic with the fundatnental
eompotient attributes of the stitnuli/brands. In such etnptrical applications, a "tnixed model" might be assutned in
which some dimensions are vector-like and otbers ideal
point-like.
Another potential extension might be to devise a spatial
conjoint tnodel that will allow the measuretnent ofthe partworths of the attributes of perceived value and the identification of tbe bypt)tbetiL-;tl "best value"" hrand. Such a fratiiework is thus conducive to determining the way to position a
certain brand optimally in terms of perceived value. In addition, extension of VALUEMAP to an unfolding model
seems appropriate, though there does seetn to be more operational difficulties with that type ol model, giveti the prevalence of degenerate solutions. All in all, much work remains
to be done toward explicating the perceived value ct)nstruct.
As articulated by Woodruff (1997), given the pervasive
global importance of value marketing and growitig international cotnpetition in trade and eommerce. this area should
continue to be a fertile area for research for quite sotne time.
REFERENCES
Akaike, H. (I974J. 'A New Look ;it Slatistical Mode! IdcntiMcation." IEEE Transactions on Automatic Control. 6. 716-2.V
BockenhoU, I. antJ W. Gaul (1989), "GetieralJzcd Lateiil Class
Analysis: An New Methodology for Market Structure Atialysis," in Concepttuil and Nutnerical An(il\.\i.s oj Data. O. Opit/,.
ed. Berlin: Sprinyer Verlag, 367-76,
Bolton, Ruth N. atid James H. Drew (1991), "A Multi.stage Model
of Customers" Assessments of Service Quality and Value.'"
Journal of Consumer Hesearch. 17 (March). 375-84.
Bozdogan, hi. (1987). ""Model Selection and Akaike's Inloimation
Criterion (AIC): The General Theoty and Its Analytical Kxicnsions," Psvchometrika. 52 (September), 345-70.
Brayman, Joann M, Davis (1996), "Building a Bcllei Business
Tool^Custotner Value Analysis," paper presented at the AMA
Conference on Cusioiner Satisfaction. (January).
Busine.vsWeek (1991). -Value-Marketing." (November I I).
132^0.
CarToll, J. Douglas (1972). ""Individual Differences anJ MuUiJimensional Scaling." in Multidimensional Scaling: Theotr and
Applications in the Behavioral Sciences, R. N. Shepard, A. K.
Romney, and S. B. Nerlove, eds. New York and London: Semi-
Marketing Utters.
3 (3). 273-88.
249
Of .Scaling.
Science.
25 {2). 1 3 9 - 5 3 .