Multi-dimensional

scaling
Introduction
Multidimensional scaling (MDS)
• MDS, also known as perceptual mapping, is a
procedure that enables a researcher to determine
the perceived relative image of a set of objects
(firms, products, ideas, or other items associated
with commonly held perceptions)
• It helps the researcher identify the key dimensions
underlying respondents’ evaluations of objects and
the position these objects in this dimensional space
• It is an exploratory technique
 IDENTIFY KEY DIMENSIONS COMPARISON OF PHYSICAL
UNDERLYING CUSTOMER QUALITIES (E.G. FOOD TASTES OR
EVALUATION OF PRODUCTS, VARIOUS SMELLS)

Applications
SERVICES OR COMPANIES

of MDS

PERCEPTIONS OF POLITICAL ASSESSMENT OF CULTURAL

CANDIDATES OR ISSUES DIFFERENCES BETWEEN DISTINCT
GROUPS
Broad logic of MDS
• Transform consumer judgments of similarity or
preference (e.g. preference for stores or brands)
into distances represented in multidimensional
space
Steps in MDS

Gather similarity measures • Ask respondents for similarity or preference between

or preference measures objects based on single measure for each pair of
across the objects objects (rank orders, quantitative values)

• Use MDS to estimate relative position of the objects

Create Perceptual Map in MD space based on similarity measures,
preference measures or both

Identify and interpret the

• After the perceptual map is defined, identify what
axes of the dimensional
the dimensions (e.g. two axes) mean (Type of
space in terms of perceptual
chocolate bar Vs Price)
&/or objective attributes
Some outputs of the MDS
Explain what you observe
Explain what you observe
Explain what you observe
MDS versus other
interdependence techniques
Factor Analysis Cluster Analysis MDS

• Define structure by • Defines structure by • Estimate relative

grouping variables grouping objects position of the
into variates that according to their objects in MD space
represent underlying profile on a set of based on one
dimensions in the variables (the cluster measure
original set of variate) • A solution can be
variables • Objects in close obtained for each
• Variables that highly proximity are individual
correlate are grouped together • It does not use a
grouped together variate (i.e. it tries to
figure out what the
independent
variables are)
 Advantages Disadvantages
Some broad Individual as the unit of Researcher does not
advantages analysis – The relative
position is determined
know/have control on
the variables used by the
and based on how the respondent to make
respondent perceives comparisons (i.e.
disadvantages proximity between conceptual framework
of MDS objects
Lack of a variate –
arises from the
respondent)
Reduces the influence of
the researcher in
specifying variables to
locate and cluster objects
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Decide between Select between

Identify Objects to be
Identify the problem Similarities or Aggregate Vs
Evaluated
Preferences Data Disaggregate Data
• Identify unrecognized • All relevant firms, • Similarities: Similar • Disaggregate: As
evaluative products, services or Vs. Dissimilar – Not many maps as
dimensions objects to be much insight on the subjects
• Comparative included determinants of • Aggregate: (1)
evaluation of objects • E.g. “Study of Soft choice Average values
• E.g. which are the Drinks” = Sweet Soft • Preference: Makes before analysis (2)
most preferred Drinks + Sugar Less the choice of the Cluster maps after
brands and how has Soft Drinks respondents explicit, analysis (3) Combine
the respondent • Omit irrelevant ones but may not be based aggregate with
determined that can bias the MDS on similarities (i.e. individual weights to
preference? can apply to dimensions of
• E.g. which are the dissimilar objects as respondents
most similar brands well)
and what makes
them similar?
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Metric Vs
Data
Approach Objects Non
collection
Metric

Decompositional Approach (Attribute-free) Compositional Approach (Attribute-based)

• Measure overall impression or evaluation of • Attributes are pre-defined

object • (1) Post-hoc graphical analysis (i.e. after an
• Typically associated with MDS outcome has been achieved): importance vs
• Advantages: Ease of use for respondents; performance
Perceptual maps can be created at • (2) Develop a dimensional structure on the
disaggregate and aggregate level basis which the objects can be represented
• Disadvantages: Researcher or manager (discriminant analysis, factor analysis)
objectively unaware of the basis used by • (3) Correspondence analysis – based on
respondents qualitative or nominally scaled data as input
• Advantages: Dimensions made explicit;
better visual control
• Disadvantages: Other dimensions may be
missed out (strong theoretical justification
needed); More data to be collected (multiple
measures); Results not available at individual
level
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Metric Vs
Data
Approach Objects Non
collection
Metric

Selecting objects Number of objects

• Relevance to • Fewer objects – ease

objectives of response
• Omit the rest • More objects – stable
MD solution
• Recommended:
• No. of objects >
4*no. of dimensions
desired
• Fewer objects can
lead to “false fits”
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Metric Vs
Data
Approach Objects Non
collection
Metric

Non Metric Data Metric Data

• Rank ordering pairs • Output preserves

• Distance not related ratio and interval
to measure of data in the input
similarity • More consolidated
• May lead to outputs
degenerate solutions
(wide variations in
perceptual maps of
respondents)
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Metric Vs
Data
Approach Objects Non
collection
Metric

Similarities Preferences

• Comparison of paired objects: Rank • Direct Ranking of Most to Least

similar pairs or rate similarity Preferred Objects (no ties): Useful
between paired objects for fewer objects
• Confusion Data (aggregation): • Paired Comparisons: Overall
• Every individual groups several preference based on frequency of
objects (stimuli) based on assumed being the most preferred object in a
similarities pair
• No. of stacks pre-fixed/as desired • Distance indicates difference in
by respondent preference
• Identify frequency of pairs in a
stack across all individuals
• Derived measure: Pre-fixed
dimensions for comparison of objects
(stimuli) (least desirable for MDS)
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Variation in Variation in Variation in

dimensionality Importance Time
• Individuals • Individuals • Perceptions
will use place and ratings
different different can change
dimensions weights for with time
to compare different
objects dimensions
Example matrix with pair-wise
comparisons (symmetric about the
diagonal)
• A, B, C, D can be firms, brands, products, houses,
political figures, etc.
Exercise
• Identify a problem to be addressed that can use
MDS
• List out the objects (4-5)
• Develop a data collection instrument (i.e. make a
form)
• Collect 2 responses
• Try to create a perceptual map for each respondent
(without use of software – i.e. manually)
• Explain your logic used in creating the map
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Process of MDS (broad overview)

• Start with few dimensions – e.g. two dimensions
• Place the objects randomly over the two dimensions and
measure distances (𝑑𝑑𝑖𝑖𝑖𝑖 ) between objects (random initial
configuration, 𝑆𝑆𝑘𝑘 )
• Compare the distances with the distances (𝑑𝑑� 𝑖𝑖𝑖𝑖 ) provided in the
similarity judgments (𝑠𝑠𝑖𝑖𝑖𝑖 )
• “Badness of fit” is measured - STRESS
• Comparison is on the basis of RANK ORDER agreement between
the Euclidean distances of the plotted objects and the original
ranks of the objects
• If the distances do not follow the ranks, move the objects and
repeat the process
• Process terminates when satisfactory fit (low stress) is obtained in
the same dimensions and/or more dimensions are tested
 Stage 4 – Deriving
Stage 2 – Stage 3 –
Stage 1 – Objectives of the MDS Solution
Research Design Assumptions of
MDS & Assessing
of MDS MDS
Overall Fit

Best fit
• Stress measure is the proportion of variances of the
disparities not accounted for by the MDS model
• Kruskal’s stress test measure

• Lower value of stress = better fit

• Similar problems like 𝑅𝑅2 (more dimensions, stress
reduces)
Dataset 1 – self created (based on
similarities or dissimilarities)
Dataset 2 – HBAT.sav
• Similarity judgments using a scale for each pair of
firms
• 1=“not at all similar”
• 9=“very similar”

• Total respondents = 18
Dataset 3 - R