Chapter 10 MDS and Correspondence Analysis

Chapter 10 MDS and
Correspondence Analysis
Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-1

Chapter 10 MDS and
LEARNING OBJECTIVES
Upon completing this chapter, you should be able to do the
following:
• Define multidimensional scaling and describe how it is
performed.
• Understand the differences between similarity data and
preference data.
• Select between a decompositional or compositional
approach.
• Determine the comparability and number of objects.
• Understand how to create a perceptual map.

Multidimensional Scaling (MDS)
Overview
• What is it?
• Why use it?

Multidimensional Scaling Defined
Multidimensional scaling (MDS) . . . also

known as perceptual mapping, is a procedure
that allows a researcher to determine the
perceived relative image of a set of objects (firms,
products, ideas, or other items associated with
commonly held perceptions).
The purpose of MDS is to transform consumer
judgments of overall similarity or preference (e.g.,
preference for stores or brands) into distances
represented in multidimensional space.

Multidimensional Scaling Analysis (MDS)
To perform a multidimensional scaling analysis,

there are three basic steps . . .
1. Gather measures of similarity or preference

across the entire set of objects to be analyzed.
2. Use MDS techniques to estimate the relative
position of each object in multidimensional
space.
3. Identify and interpret the axes of the
dimensional space in terms of perceptual
and/or objective attributes.

Multidimensional Scaling Mapping
 Perceptual maps represent the positions of products on a
set of latent dimensions:
• Preferences can be incorporated into the maps – so-called joint
space.
• Segment-specific preferences can also be represented.
 Analytical methods for producing perceptual maps:
• Factor analysis of customer ratings.
• Use focus groups and/or prior research to identify important
attributes for product class.
• Ask 100 to 300 customers to rate the products with which they are
familiar on a set of metric measurement scales for these attributes.
• Determine number of dimensions
• Use factor loadings to name underlying dimensions.
• Plot average factor scores (across customers) for each brand in
map.
Multidimensional Scaling Mapping
• Using a metric measure of stated preferences, regress preference on the
factor scores to get the slope of the preference vector.
• If multiple segments are desired, a finite mixture regression model can be
used to produce multiple preference vectors.
• Factor analysis approach depends on the identification of a well-specified
set of attributes.
• Not model-based – involves a two-stage process.
 Maps can also be formed by fitting models to preference/choice data:
• A persons-by-stimuli two-way data matrix.
• Attributes may or may not be utilized.
• These models allow for statistical inferences to be made, unlike the factor
analysis approach and MDS.
• Preference vectors represent segments rather than individuals, easing
interpretation.
 Maps can be formed from cross-tabulated data using
correspondence analysis.

Multidimensional Scaling
 Using Multidimensional Scaling (MDS)

o Ask customers to rate the similarity of pairs of brands on a
metric scale – no attributes are involved!
o These measures are averaged across all customers (or
segments of customers) to produce a proximity matrix whose
entries represent the similarity/dissimilarity among the products.
o Given the number of dimensions, MDS program finds locations
of brands that minimize “stress,” a measure of lack of fit.
o Interpreting dimensions will be more difficult without attributes:
 Include some attribute labels (“Prestige”) along with brands
and plot the labels alongside the brands.
 Measure attribute ratings as in factor analysis, regress
attribute ratings on brand locations, and use regression
coefficients (like loadings) to name dimensions.

• Comparison of techniques:
o Ratings on attributes can be linked to potential product
improvements.
o When factor analysis is used, we cannot identify any
perceptual dimensions that are not represented by the
attribute measures.
o Similarity measures indicate which products are likely to be
considered as substitutes.
o Similarity scaling is useful if some attributes of a product
are difficult to scale (allure of a perfume) or difficult to
articulate (ambiance of a restaurant).
o Difficult to name and interpret similarity-based dimensions.

 Why perceptions?
• Perceptions are not equal to reality (e.g.,
perceived vs. actual quality of German
cars, particularly Mercedes, VW).
• Customers buy based on perceptions.

• Aggregate vs. Disaggregate analysis:
o MDS can be performed at individual level, segment level, or market level.
o Combining information across individual-level maps is a challenge.
o For a more aggregated map, average similarity judgments prior to
analysis.
o Clustering the coordinates from individual-level maps is not
recommended.
o Mixture of both (INDSCAL):
 Form perceptual map on basis of all consumers (a shared space);
brands are represented as points.
 Represent respondents’ preference weights for the dimensions
separately; each respondent is represented by a vector whose slope
is determined by regression coefficients.
 Length of vector represents % variance explained.
 Space can get very crowded, difficult to interpret.
 Segment-level analysis alleviates crowding.

• Measure attributes or not – see earlier comparison of

factor analysis vs. MDS.
• Metric vs. Nonmetric MDS
o Rank orders of pairs used for nonmetric MDS (which
pair is most similar? . . . Least similar?
o All output is metric, meaning that the following do not
change the solution:
 Rotation about the origin.
 Reflection of axes.
 Entire solution can be stretched or compressed.
o Metric data contains more information.

• Vector vs. Ideal Point (unfolding) Preference
Representations:
o Vector models assume more (or less) of a dimension is better
(economy, performance, prestige).
o Preference ordering of brands is found by dropping
perpendicular lines from the brands to the consumer’s vector.
o Ideal point models assume some intermediate level of a
dimension is preferred (sweetness).
o Brands closest (in a Euclidean distance sense) to a
consumer’s ideal point are more preferred.
o Ideal point may be found by having respondent evaluate his/her
ideal brand along with the others, or ideal points may be
estimated for reach respondent using specialized estimation
procedures.
• Internal vs. External Analysis:
o Internal analyses estimate the brand coordinates.

• In external analyses, the coordinates of brands

are assumed given or known, for example, from
previous analyses.
• With internal analyses, brand coordinates can be
reparameterized using attribute information to aid
in the interpretation of the derived dimensions.
• Better to do all analyses in one step rather than
multiple steps!

How to do an MDS?
The researcher performs three basic steps to do a

multidimensional scaling analysis:
• Gathers measures of similarity or preference

across the entire set of objects to be analyzed,
• Uses MDS techniques to estimate the relative
position of each object in multidimensional
space, and
• Identifies and interprets the axes of the
dimensional space in terms of perceptual and/or
objective attributes.

Stage 1: Objectives of MDS
• To identify latent dimensions (primary needs)

affecting consumer behavior.
• To obtain comparative evaluations of objects

when the specific bases of comparison are
unknown or undefined.

Stage 1: Objectives of MDS continued . . .
Perceptual mapping, and MDS in particular, is

most appropriate for achieving two
objectives:
1) . . . as an exploratory technique to identify
unrecognized dimensions affecting behavior,
and
2) . . . as a means of obtaining comparative
evaluations of objects when the specific
bases
of comparison are unknown or undefined.

Similarities vs. Preferences Data?
Both bases of comparison can be used to develop
perceptual maps, but with differing
interpretations:
• Similarity-based perceptual maps – represent
attribute similarities and perceptual dimensions of
comparison but do not reflect any direct insight
into the determinants of choice.
• Preference-based perceptual maps – reflect
preferred choices but may not correspond in any
way to the similarity-based positions, because
respondents may base their choices on entirely
different dimensions or criteria from those on
which they base comparisons.
• Similarities vs. Preference data?

o Either can be used in MDS, though
interpretation will differ (how?).
o Common to use similarities in MDS,
overlay preference vectors using
regression analysis.

Rules of Thumb 10–1
Objectives of MDS
• MDS is an exploratory technique well suited for:
o Identifying unrecognized dimensions used by respondents in making
comparisons between objects (brands, products, stores…)
o Providing an objective basis for comparison between objects based on
these dimensions
o Identifying specific attributes that may correspond to these dimensions
• An MDS solution requires Identification of all relevant

objects (e.g., all competing brands within a product
category) which set the boundaries for the research
question.
• Respondents’ provide one or both types of perceptions
o Perceptual distances where they indicate how similar/dissimilar objects are
to each other, or
o “Good-bad” assessments of competing objects (preference comparisons)
which assist in identifying combinations of attributes that are valued
highly.
Rules of Thumb 10–1 continued . . .
Objectives of MDS
• MDS can be performed at the individual or group level:
• Disaggregate (individual) analysis:
o Allows for construction of perceptual maps on a respondent-
by-respondent basis
o Assessment of variation among individuals.
o Provides a basis for segmentation analysis.
• Aggregate (group) analysis:
o Creates perceptual maps of one or more groups
o Helps understand overall evaluations of objects and/or
dimensions employed in those evaluations.
o Should be found by using the average evaluations of all
respondents within a group.

Stage 2: Research Design of MDS
Identification of all relevant objects to be

evaluated:
• Errors of omission, inclusion of irrelevant

objects.
• More than 4x as many objects as
dimensions – results in many comparisons
for 2+ dimensions.
• Overfitting can be a big problem if too few
objects
are used.

Stage 2: Research Design of MDS
continued . . .
Perceptual mapping techniques can be classified into

one of two types based on the nature of the responses
obtained from the individuals concerning the object:
• Decompositional method – measures only the overall
impression or evaluation of an object and then attempts
to derive spatial positions in multidimensional space that
reflect these perceptions. This technique is typically
associated with MDS.
• Compositional method – an alternative approach that
uses several multivariate techniques in forming an
impression or evaluation based on a combination of
specific attributes.
RESEARCH DESIGN OF MDS

o Perceptual maps can be generated through
decompositional or compositional approaches . . .
• Decompositional approaches are the “traditional” and most
common MDS method requiring only overall
comparisons of similarity between objects.
• Compositional approaches are used when the research
objectives involve comparing objects on a defined set of
attributes.
o The number of objects to be evaluated is a tradeoff
between
• a small number of objects to facilitate the respondents’ task
• four times as many objects as dimensions desired (i.e., 5
objects for one dimension, 9 objects for two dimensions . . . )
to obtain a stable solution.
Stage 3: Assumptions of MDS Analysis
Assumptions = none, other than the

correct type of data is collected for the
procedure used (metric vs. nonmetric
MDS).

Stage 3: Assumptions of MDS Analysis
continued . . .
Multidimensional scaling has no restraining
assumptions on the methodology, type of data, or form of
the relationships among the variables. But there are three
perception requirements:
1. Variation in dimensionality — respondents may vary in the
dimensionality they use to form their perceptions of an
object (although it is thought that most people judge in
terms of a limited number of characteristics or dimensions).
2. Variation in importance — respondents need not attach the
same level of importance to a dimension, even if all
respondents perceive this dimension.
3. Variation over time — judgments of a stimulus in terms of
either dimensions or levels of importance are likely to
change over time.
Stage 4: Deriving the MDS Solution and
Assessing Overall Fit
• Determining an object’s position in the

perceptual map
• Selecting the dimensionality of the perceptual
map using a stress measure (lower stress
means better fit).
• Fit will improve when the number of dimensions
increases.
• Index of fit, like R2, can be used.

Stage 4: Deriving the MDS Solution and
Assessing Overall Fit continued . . .
• Perceptions or Similarities Data?

• Determining an Object’s Position in
the Perceptual Map.
• Dimensionality of the Perceptual Map.

DERIVING AND VALIDATING AN MDS SOLUTION
• Stress measures (lower values are better) represent an MDS
solution’s fit.
• Researchers can identify a degenerate MDS solution which is
generally problematic by looking for:
a circular pattern of objects suggesting that all objects are
equally similar, or
a multi-clustered solution in which objects are grouped at
two ends of a single continuum.
• The appropriate number of dimensions for a perceptual map is
based on:
a subjective judgment as to whether a solution with a given
dimensionality is reasonable.
use of a scree plot to identify the ‘elbow’ where there is a
substantial improvement in fit.
use of R2 as an index of fit – measures of .6 or higher are
considered acceptable.
DERIVING AND VALIDATING AN MDS SOLUTION

continued . . .
o External analysis, such as is performed by
PREFMAP, is considered preferable in generating
ideal points relative to internal analysis.
o The most direct validation method is a split-sample
approach.
• Multiple solutions are generated by either
splitting the original sample or collecting new
data
• Validity is indicated when the multiple solutions
match
Stage 5: Interpreting the MDS Results
• Subjective procedures – may be best for

affective, highly intangible, or emotional
dimensions
• Objective procedures

Stage 5: Interpreting the MDS
Results continued . . .
Once the perceptual map is obtained, the two

approaches – compositional and decompositional –
again diverge in their interpretation of the results.
The differences in interpretation are based on the
amount of information directly provided in the
analysis (e.g., the attributes incorporated in the
compositional analysis vs. their absence in the
decompositional analysis) and the generalizability of
the results to the actual decision-making process.

Compositional vs. Decompositional Methods
• For compositional methods, the perceptual map can

be directly interpreted with the attributes
incorporated in the analysis. The solution,
however, must be validated against other measures
of perception, because the positions are totally
defined by the attributes specified by the
researcher.
• For decompositional methods, the most important

issue is the description of the perceptual dimensions
and their correspondence to attributes.

Stage 6: Validating the Results
• Validation efforts are problematic.

• In comparing two maps, if the positions
vary, the researcher cannot tell whether
the objects are viewed differently, the
perceptual dimensions vary, or both.

Stage 6: Validating the MDS Results
continued . . .
Any MDS solution must deal with two specific issues
that complicate efforts to validate the results:
• The only output of MDS that can be used for
comparative purposes involves the relative positions of
the objects. Thus, although the positions can be
compared, the underlying dimensions have no basis
for comparison. If the positions vary, the researcher
cannot determine whether the objects are viewed
differently, the perceptual dimensions vary, or both.
• Systematic methods of comparison have not been
developed and integrated into the statistical programs.
The researcher is left to improvise with procedures that
may address general but not specific concerns.
Approaches to Validating the
MDS Results
• Split-Sample Analysis.
• Comparison of Decompositional
vs. Compositional Solutions.

Chapter 10 MDS and
LEARNING OBJECTIVES
Upon completing this chapter, you should be able to
do the following:
• Understand the basics of perceptual mapping
with nonmetric data.
• Select between a decompositional or
compositional approach.
• Explain correspondence analysis as a method of
perceptual mapping.

Correspondence Analysis Overview
• What is it?
• Why use it?

Correspondence Analysis Defined
Correspondence Analysis (CA) . . . is an

interdependence technique that has become
increasingly popular for dimensional reduction
and perceptual mapping. It also is known as
optimal scaling or scoring, reciprocal
averaging or homogeneity analysis.

Correspondence Analysis Defined
When compared to the MDS techniques,
correspondence analysis (CA) has three
distinguishing characteristics . . .
1. It is a compositional technique, rather than a
decompositional approach, because the perceptual map is
based on the association between objects and a set of
descriptive characteristics or attributes specified by the
researcher.
2. Its most direct application is portraying the
“correspondence” of categories of variables, particularly
those measured in nominal measurement scales. This
correspondence is then the basis for developing perceptual
maps.
3. The unique benefits of CA lie in its abilities for representing
rows and columns, for example, brands and attributes, in
joint space.
Stage 1: Objectives of CA
CA can address either of two basic objectives:
1. Association among only row or column categories. CA can be
used to examine the association among the categories of just a
row or column. A typical use is examining the categories of a
scale, such as the Likert 5-point scale (“strongly agree” vs.
“strongly disagree”) or other qualitative scales (e.g., excellent,
good, poor, bad). The categories can be compared to see if two
can be combined (i.e., they are in close proximity on the map) or if
they provide discrimination (i.e., they are located separately in the
perceptual space).
2. Association between both row and column categories. This
portrays the association between categories of the rows and
columns, such as product sales by age group. This use is most
similar to MDS and has propelled CA into more widespread use
across many research areas.
Stage 2: Research Design of CA
Correspondence analysis – requires only a

rectangular data matrix (cross-tabulation) of
nonnegative entries. The most common
type of input matrix is a contingency table
with specific categories defining the rows
and columns.

Stage 3: Assumptions in CA
Correspondence analysis – shares with the

more traditional MDS techniques a relative
freedom from assumptions. The use of
strictly nonmetric data in its simplest form
(cross-tabulated data) represents linear and
nonlinear relationships equally well.

Stage 4: Deriving CA Results and
Assessing Overall Fit
Correspondence analysis uses a cross tabulation table to create a
perceptual map in three steps:
1. Calculate a conditional expectation (the expected cell count) that
represents the similarity or association between row and column
categories.
2. Once obtained, compute the difference between the expected and
actual cell counts and convert them into a standardized measure
(chi-square). Using this as a distance metric, they are comparable
to the input matrices used in MDS.
3. Through a process much like multidimensional scaling, a series of
dimensional solutions (one-dimensional, two-dimensional, etc.) are
created where possible. The “dimensions” simultaneously relate
the rows and columns in a single joint plot. The result is a
representation of categories of rows and/or columns (e.g., brands
and attributes) in the same plot.
Stage 5: Interpretation of the Results
Once the dimensionality has been established,

the researcher is faced with two tasks:
1. Interpreting the dimensions to understand the

basis for the association among categories, and
2. Assessing the degree of association between

categories, either within a row/column or between
rows and columns.

Stage 6: Validation of the Results
The compositional nature of correspondence
analysis provides more specificity the researcher can
use to validate the results. In doing so, the
researcher assesses two key questions concerning
generalizability . . .
• Sample – as with all MDS techniques, an emphasis must be
made to ensure generalizability through split- or multisample
analyses.
• Objects – the generalizability of the objects (represented
individually and as a set by the categories) must also be
established. The sensitivity of the results to the addition or
deletion of a category can be evaluated. The goal is to assess
whether the analysis is dependent on only a few objects and/or
attributes.

• Correspondence analysis (CA) is best suited for
exploratory research and is not appropriate for
hypothesis testing.
• CA is a form of compositional technique which requires
specification of both objects and attributes to be
compared.
• Correspondence analysis is sensitive to outliers which
should be eliminated prior to using the technique.

Rules of Thumb 10–4 continued
• The number of dimensions to be retained in the solution
is based on . . .
 Dimensions with inertia (eigenvalues) greater than .2.
 Enough dimensions to meet the research objectives
(usually two or three).
• Dimensions can be “named” based on the decomposition
of inertia measures across a dimension . . .
 These values show the extent of association for each
category individually with each dimension.
 They can be used for description much like loadings
in factor analysis.
Description of HBAT Primary Database Variables
Variable Description Variable Type
Data Warehouse Classification Variables
X1 Customer Type nonmetric
X2 Industry Type nonmetric
X3 Firm Size nonmetric
X4 Region nonmetric
X5 Distribution System nonmetric
Performance Perceptions Variables
X6 Product Quality metric
X7 E-Commerce Activities/Website metric
X8 Technical Support metric
X9 Complaint Resolution metric
X10 Advertising metric
X11 Product Line metric
X12 Salesforce Image metric
X13 Competitive Pricing metric
X14 Warranty & Claims metric
X15 New Products metric
X16 Ordering & Billing metric
X17 Price Flexibility metric
X18 Delivery Speed metric
Outcome/Relationship Measures
X19 Satisfaction metric
X20 Likelihood of Recommendation metric
X21 Likelihood of Future Purchase metric
X22 Current Purchase/Usage Level metric
X23 Consider Strategic Alliance/Partnership in Future nonmetric 10-49
Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall.

Chapter 10 MDS and Correspondence Analysis

Uploaded by

Copyright:

Available Formats

You might also like

Chapter 10 MDS and Correspondence Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 10 MDS and Correspondence Analysis

Uploaded by

Copyright:

Available Formats

Chapter 10 MDS and

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-1

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-2

• Why use it?

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-3

Multidimensional scaling (MDS) . . . also

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-4

To perform a multidimensional scaling analysis,

1. Gather measures of similarity or preference

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-5

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-7

 Using Multidimensional Scaling (MDS)

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-8

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-9

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-10

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-11

• Measure attributes or not – see earlier comparison of

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-12

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-13

• In external analyses, the coordinates of brands

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-14

The researcher performs three basic steps to do a

• Gathers measures of similarity or preference

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-15

• To identify latent dimensions (primary needs)

• To obtain comparative evaluations of objects

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-16

Perceptual mapping, and MDS in particular, is

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-17

• Similarities vs. Preference data?

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-19

• An MDS solution requires Identification of all relevant

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-21

Identification of all relevant objects to be

• Errors of omission, inclusion of irrelevant

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-22

Perceptual mapping techniques can be classified into

RESEARCH DESIGN OF MDS

Assumptions = none, other than the

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-25

• Determining an object’s position in the

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-27

• Perceptions or Similarities Data?

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-28

DERIVING AND VALIDATING AN MDS SOLUTION

• Subjective procedures – may be best for

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-31

Once the perceptual map is obtained, the two

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-32

• For compositional methods, the perceptual map can

• For decompositional methods, the most important

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-33

• Validation efforts are problematic.

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-34

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-36

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-37

• Why use it?

Copyright © 2010 Pearson Education, Inc., publishing as Prentice-Hall. 10-38