Professional Documents
Culture Documents
Chiranjeev S. Kohli
&
Lance Leuthesser
Chiranjeev S. Kohli and Lance Lancaster are Associate Professors of Marketing at California State
University, Fullerton, California.
Chiranjeev S. Kohli would like to thank Dr. Franklin Acito for his valuable suggestions on an earlier
draft of the article.
COMPARISON OF ALTERNATIVE PERCEPTUAL MAPPING TECHNIQUES
Abstract
mapping techniques are frequently used to aid managers in making product positioning decisions.
This paper presents an overview of perceptual mapping, explains the conceptual foundation, and
compares three widely used techniques -- factor analysis, discriminant analysis, and
positioning has been defined as the act of designing the image of the firm's offering so that target
customers understand and appreciate what the product stands for in relation to its competitors.
Each brand within a set of competitive offerings is thought of as occupying a certain position in a
positioning studies, an understanding of the basics of product positioning techniques will enable
these managers to make better evaluations of proposed research undertakings, and ultimately better
decisions. Accordingly, in the first section of the paper we review three frequently used perceptual
mapping techniques: (1) factor analysis, (2) discriminant analysis, and (3) multidimensional scaling.
Keeping with our objective of explaining the concepts, we discuss the techniques in their basic form.
In the second part of the paper, we highlight the relative strengths and limitations of each of these
approaches to perceptual mapping and suggest guidelines for their use which should assist
managers facing product positioning decisions. Focusing on the differences in these techniques, we
In general, marketers have two broad objectives in mind when undertaking perceptual
mapping. One objective is to determine where a target brand is positioned versus the competition.
The other objective is to help identify product attributes which are determinant in influencing
customer choice for the product class. Determinant attributes are those which are important to
customers and also exhibit differences across brands. No matter how important a product attribute
is, if brands are not perceived to differ on that attribute, then the attribute will not be influential in
customers' choice decisions. Often, the determining aspects of a product are latent, unobservable
1
constructs which subsume a number of manifest, observable attributes. Perceptual mapping
Factor Analysis
The Concept. Factor analysis is essentially a data reduction technique in which the objective
is to represent the original pool of attributes in terms of a smaller number of underlying dimensions
or factors. After the factors have been identified, the brands' ratings on these factors are used to
Input Data. An example will clarify the type of data which is used as input to factor analysis.
Consider m brands of automobiles, which we wish to map. Each of the m brands is rated on p
attributes by n respondents. The resulting input data form a matrix containing mn rows and p
columns.
Generation of the Perceptual Map. Table 1 shows hypothetical rating data for the automobile
example. In this example, three brands (m = 3) are rated on four attributes (p = 4). Three
individuals (n = 3) each rates all brands on the four attributes (typically, likert-like scales containing
from five to nine rating points are employed for such measures). The next step is generation of a
correlation matrix for the p attributes. For this example, the correlation matrix appears in Table 2.
An examination of the correlation matrix reveals that the attributes "reliability" and "resale value"
are highly correlated. Similarly, "exterior design" and "aerodynamic body" are highly correlated.
However, the first pair of attributes shows very low correlation with the second pair. Thus we may
conclude that there are really two underlying dimensions represented in our respondents' ratings of
these automobiles, not four, as the number of observed attribute measures might suggest. On
statistical grounds, at least, the data suggest that "reliability" and "resale value" might be two
indicators of some underlying concept. For example, we might hypothesize that these two measures
both tap the concept "intrinsic value." Likewise "exterior design" and "aerodynamic body" might be
indicators of, say, "styling."† We cannot, however, rule out the possibility that the correlation
2
between these measures is due to another, unobserved (and unmeasured), variable. Thus any
conclusions about the existence of meaningful factors must ultimately be supported by theoretical
considerations.
Having identified the factors, the brands are then evaluated on these factors by computing
an overall "factor score." The factor score is simply the average of all the attribute evaluations
contributing to that factor, which may be averaged over n individuals for the brand's overall
position on the map.‡ Brands are then positioned by factor scores. (Refer to Figure 1 for the
automobile example.) The positioning of the brands for each individual is shown in Figure 1 as J1,
D1, S1...J3, D3, S3. For example, John's positioning of Brand 1 (point J1) appears in the lower left
portion of Figure 1 as J1(3, 1.5). (Disregard the ellipses at this point.) The overall brand position is
shown by the respective number of the brand and the mark "."
Discriminant Analysis
The Concept. Discriminant analysis also requires that respondents provide attribute ratings.
Like factor analysis, an objective of this method is to reduce the number of attributes to a smaller
number of underlying dimensions. However, unlike factor analysis, discriminant analysis focuses
on attributes which show differences between brands. This approach will tend to ignore attribute
ratings which show a large variation within brands; these latter dimensions will not be "significant"
In product positioning studies, the set of ratings for each brand constitutes a group. As its name
suggests, discriminant analysis will identify those underlying dimensions which are most useful in
Input Data. Input data requirements for discriminant analysis are similar to those for factor
analysis.
Generation of the Perceptual Map. Refer, again, to Figure 1. Using the same data employed for
factor analysis (see Tables 1 and 2), discriminant analysis identifies the first dimension of "intrinsic
3
value" (represented by the X axis in Figure 1) as the first discriminating dimension. The ellipses in
the figure indicate the variation in the ratings of each brand by different respondents. Note the
relatively short axes of the ellipses along the X axis (intrinsic value) and the relatively long axes in
the Y direction (brands 1 and 2 in particular). It is clear from Figure 1 that respondents show
substantially greater overall agreement when rating the two attributes representing "intrinsic value"
than when rating the two attributes representing "styling." Stated another way, knowing a value for
styling conveys little information about a particular brand's identity. A styling rating of 4, for
example, could pertain to any of the three brands. On the other hand, the intrinsic value dimension
conveys complete information as to brand; an intrinsic value of 4 uniquely identifies brand 2. In this
case, therefore, discriminant analysis will identify only the "intrinsic value" dimension as significant.
4
Table 1
Attributes
Factor 1 Factor 2
Reliability Resale Exterior Aerodynamic
Value Design Body
Respondents
5
Table 2
Reliability 1.00
6
Multidimensional Scaling (MDS)
The Concept. MDS enables us to map objects (brands) spatially, such that the relative
positions in the mapped space reflect the degree of perceived similarity between the objects (the
closer in space, the more similar the brands). When the map has been generated, the relative
positioning of the brands, together with knowledge of the general characteristics of the brands,
Input Data. Respondents evaluate brands in pairs, judging the overall similarity between the
paired brands. Respondents may either rate (metric scaling) or rank-order (non-metric scaling) the
similarity for each pairing of brands. This is a major distinction from both factor and discriminant
analysis. Unlike those methods, MDS asks respondents to rate brands on overall similarity, not
individual attributes.
Table 3
SIMILARITY MATRIX
A B C D
A X
B 3 X
C 2 1 X
D 1 5 6 X
Entries in the matrix show the rank-order of similarity of the two brands in a pair. Thus, brands "B"
and "C" are most similar, and brands "C" and "D" are least similar.
Generation of the Perceptual Map. For illustration sake, we take a hypothetical "similarity
matrix" shown in Table 3. First, consider three brands only: A, B, and C. Note that the matrix has
7
dimensions to map the brands, where m is the number of brands. Thus, in our example we can
represent the three brands on a one dimensional map with perfect fit. This is illustrated in Figure 3.
Now, assume we introduce a fourth brand, D. Its similarity rankings when paired with the
other brands are also shown in Table 3. If we now try to reconstruct the map for the four brands,
three additional conditions need to be satisfied (these conditions arise from the data appearing in
the bottom row of Table 3). First, AD (i.e., the distance between A and D) should be greater than AB,
AC, and BC. Since AC and BC are already less than AB, the new condition dictates that AD should
be made greater than AB, without disturbing the earlier set up. Thus, referring to Figure 3, D should
lie somewhere outside a circle centered at A, with a radius AB. Placing D at the position shown in
the map satisfies this condition, and preserves our one-dimensional solution. The next condition is
that BD must be greater than all the above distances, and this is already satisfied with the existing
positions. Finally, however, the condition that CD be the largest of all inter-brand distances creates
a problem if we wish to maintain a one-dimensional map. With a bit of trial and error it will soon
become apparent that mapping the four brands cannot be accommodated with a one-dimensional
solution. However, a possible outcome in two dimensions is shown in Figure 3, with positions for
the four brands depicted as A, B, C', and D'. In this case the two-dimensional (m-2, where m = 4)
solution allows perfect fit. Similar to the simple trial and error process suggested here, algorithms
used for MDS generally entail an iterative estimating routine wherein each step seeks to reduce the
8
COMPARISON OF THE PERCEPTUAL MAPPING TECHNIQUES
The following issues are important when comparing MDS with factor analysis and
discriminant analysis:
1. Under MDS, it is desirable to have as many stimuli (brands) as possible because the number
of stimuli puts a limit on the number of dimensions which can be extracted (m - 1 for metric
scaling and m - 2 for non-metric scaling). Too few stimuli may lead to unstable solutions or
the obscuring of subtle dimensions which may differentiate the brands. As a rule of thumb,
at least nine stimuli are desirable for a two-dimensional solution, thirteen for three
2. Despite its desirability, a large number of stimuli significantly increases the complexity of
9
respondents' tasks. For m stimuli, Cm2 similarity judgments are required, which implies C92
= 36 paired comparisons if the Kruskal-Wish criteria are followed. This may result in
One way to alleviate this problem is to increase the number of respondents q times, thereby
requiring each respondent to perform only one qth of the total number of judgments
selection of one qth of all possible pairs. MDS has been found to be robust to either
modification.6
3. Because MDS requires only similarity judgments for brand pairings, it is not necessary a
priori to specify all product attributes which are relevant to consumers' choice decisions.
Thus in cases where it is not clear that the relevant attributes can be specified, or where only
a subset of such attributes are thought to be available, MDS is recommended over factor and
discriminant analysis.
4. Programs used for MDS, such as INDSCAL, do not allow changes in rank order among
subjects. At most they allow a complete reversal. No such constraints are imposed for the
other two techniques. Thus, MDS should not be used if there is a large variation in the way
5. MDS has been shown to be robust to embedding an existing stimuli set with new stimuli, as
concern? If not, it is unlikely that they can give meaningful responses to such
questions in a research setting. For example, MDS performs very well in mapping
10
relatively complex marketing application such as automobiles, where there is a large
actual internal cognitive space. However, it is possible that a relatively simple two
constraints; that is, the respondent's ability to make similarity judgments during the
task may be limited to using two or three dimensions only. If the actual
representation is much more complex, MDS will generate unstable dimensions that
are highly dependent on the task. The extent of this problem may be assessed by
repeating the task at different points in time. For example, mapping of health care
overall similarity judgments, others may use Gestalt impressions about the stimuli to
make that judgment. Thus, while some individuals may perceive the Mazda Miata
to be similar to the Chevrolet Corvette, both being in the overall category of "sports"
cars, others may find the Corvette to be nearer to the Ford Mustang based on specific
attributes such as engine size, horsepower, gas consumption, and domestic origin.
The validity of MDS is diminished in cases where there is large variation in the way
knowledgeable about the product category are usually able to make attribute level
judgments. Variation in the perception of similarity should be assessed after the task
11
by asking respondents to report the way in which they made similarity judgments.
7. MDS is fairly robust to the metric used; both metric and non-metric scaling techniques yield
similar dimensions.1 Accordingly, it is advisable to base the choice of scale on factors such
as the information load of respondents, or the ease with which the task may be completed.
Thus as discussed earlier in this section, when the number of stimuli becomes large, ranking
(non-metric) the similarity judgments becomes difficult. In this case, the problem may be
8. The "ease of interpretation" criterion has been advocated over statistical significance for
In summary, MDS is likely to be preferred when a large number of stimuli are available, and
only a subset of determinant attributes have been identified. It is also preferred when a large
variation in respondents' similarity perceptions is not expected, and the stimuli are not very
complex. Finally, it is preferred in situations where consumers naturally make overall similarity
judgments for the product class, and extract underlying dimensions as a part of their internal
cognitive process.
1. Recall that discriminant analysis defines dimensions which show maximum differences
between groups, tending to ignore those dimensions of brands which show large variation
across all consumers. For this reason, discriminant analysis is likely to generate fewer
dimensions than factor analysis, because factor analysis will include those dimensions that
account for a significant amount of the total variation even when such dimensions do not
show variation across groups.2 This may or may not be desirable, depending on the
objective at hand. To clarify this point, refer to the two-dimensional map shown in Figure 4,
which depicts three brands. Discriminant analysis will identify the dimension defined by
12
line aa as its first discriminating dimension. In contrast, factor analysis will identify the
dimension bb. The ellipses indicate the variation in the evaluation of the two brands on the
two dimensions mapped. As the diagram indicates, discriminant analysis gives more
weight to the Y dimension, and factor analysis gives more weight to the X dimension.
This distinction is very important. Typically, respondents are more likely to agree on
objective dimensions (length, weight, color, and so on) than on evaluative dimensions
(good/bad, like/dislike, and so on). Accordingly, discriminant analysis generally places less
emphasis on evaluative dimensions than does factor analysis.3 This is an important point to
keep in mind when designing positioning studies that include both types of measures. This
13
leads us to the next issue.
2. Factor analysis and discriminant analysis should not be viewed as conflicting techniques. If
factor analysis gives more emphasis to a dimension than discriminant analysis, it is likely
that the response frequency distribution for that attribute is as shown in Figure 5.
Figure 5 suggests that though the reliability rating of B is superior to that for A, there
is a large variation in the way it is perceived. This information, by itself, is useful to the
manager responsible for Brand B, suggesting that he/she may wish to explore possible
reasons for lack of image clarity or focus concerning this attribute. As Figure 5 further
illustrates, narrowing the frequency distribution for Brand B to curve B' will strengthen the
perceived difference between Brands A and B, even though the mean score on Brand B
remains unchanged. In sum, then, factor analysis and discriminant analysis can be used
together to highlight attributes which differ significantly in the degree to which consumers
perceptions vary.
3. The maximum possible number of dimensions in factor analysis is the number of attributes
14
on which ratings have been obtained. In discriminant analysis, the maximum possible
brands and m is the number of attributes on which these brands are evaluated. Thus, it is
important that m and p be sufficiently large. As previously mentioned, too few brands leads
circumstances with a limited number of brands and/or attributes factor analysis is preferred.
In summary, (1) discriminant analysis should be generally preferred over factor analysis
when objective dimensions, as opposed to evaluative dimensions, are of interest, (2) factor analysis
and discriminant analysis may be used as complementary techniques to highlight those dimensions
which differ substantially in the level of perceptual agreement among consumers, and (3) factor
analysis is generally preferred when attribute ratings are available on very few brands.
CONCLUSION
This article has presented a brief overview of three methods of analysis frequently used in
product positioning research: (1) factor analysis, (2) discriminant analysis, and (3) MDS. MDS
differs from both factor and discriminant analysis in that, under the first method, consumers make
overall similarity judgments of brands, whereas under the latter two methods interval scaled ratings
must be provided for each product attribute. Accordingly, MDS is particularly useful in instances
where consumers naturally tend to make overall similarity judgments, or in circumstances where
attribute data are sparse. Conditions favoring the use of factor and discriminant analysis are
Factor analysis and discriminant analysis are both data reduction techniques, and both are
useful in condensing a large number of product attribute measures into a smaller set of meaningful
underlying dimensions. A key difference between factor analysis and discriminant analysis is that
the former method identifies product dimensions which account for overall variation in consumers'
responses, whereas the latter identifies only those dimensions that account for significant variation
15
between brands. Because the overall variance in consumers' responses for evaluative dimensions is
generally greater than for objective dimensions, an implication of the above difference is that factor
analysis tends to place greater emphasis on evaluative dimensions than does discriminant analysis.
The differences in these two methods make it useful to view them as complementary techniques.
16
References
2. Hauser, John R. and Frank S. Koppelman, 1979, "Alternative Perseptual Mapping Techniques:
3. Huber, Joel, and Morris B. Holbrook, 1979, "Using Attribute Ratings for Product Positioning,"
4. Kruskal, Joseph B., and Myron Wish (1978), Multidimensional Scaling. Beverly Hills, CA: Sage
Publications, Inc.
5. Malhotra, Naresh, 1987, "Validity and Structural Reliability of Multidimensional Scaling, Journal
6. Malhotra, Naresh, Arun Jain, and Christopher Pinson, 1988, "Robustness of Multidimensional
Scaling in the Case of Incomplete Data," Journal of Marketing Research, 25 (February), 95-102.
7. Shephard, Roger N. (1972), "A Taxonomy of Some Principal Types of Data and Multidimensional
Methods for their Analysis," in Multidimensional Scaling: Theory and Applications in the
Behavioral Sciences, Vols. 1 and 2, Roger N. Shephard and Sara B. Nerlove, eds. New York: Seminar
Press, 23-51.
8. Summers, John O., and David MaKay, 1986, "On the Validity and Reliability of Direct Similarity
17
Footnotes
† The first dimension identified by factor analysis is the one that shows maximum variance.
The second dimension has the next highest variance subject to the constraint that it is orthogonal to
the first one. In the example (using rough estimates), the first dimension identified will be the one
described by "styling," which is likely to explain greater variance than the one defined by "intrinsic
value." See Table 1 for approximate estimates of variance explained by each factor. Actual variance
‡ The precise method is to evaluate each respondent's brand attribute on the first and second
factor (i.e., the extracted orthogonal linear equations containing p elements), which may then be
averaged over n individuals for the brand's overall position on the perceptual map.
18