An Example of Attribute Based MDS Using
Discriminant Analysis
Problem : A chocolate company
wants to draw a perceptual map
using an attribute based
procedure, of its consumers
perceptions regarding its own
brand and two competing brands.
Assume that it is Nestle against
Cadburys and Amul, for example.
DATA
Data was collected from 15
respondents (5 of each brand), on
five attributes, namely Price,
Quality, Availability, Packaging
and Taste. The variables are
measured using different scales,
but a higher value indicates a
favourable rating in each
variables measurement.
Input Data (High = Better)
BRAND
PRICE
QUALITY
AVAILABILITY
PACKAGING
TASTE
12
34
500
18
11
35
234
15
10
36
250
14
13
22
345
12
12
23
432
13
10
14
234
15
11
17
231
11
15
23
45
10
13
14
35
12
12
15
25
10
10
22
75
12
24
80
13
28
90
10
11
17
96
12
11
18
59
Means Output by Brand
Group Statistics
BRAND
1
T otal
PRICE
QUALIT Y
PACKAG
T AST E
AVALBLT Y
PRICE
QUALIT Y
PACKAG
T AST E
AVALBLT Y
PRICE
QUALIT Y
PACKAG
T AST E
AVALBLT Y
PRICE
QUALIT Y
PACKAG
T AST E
AVALBLT Y
M ean
11.6000
30.0000
4.2000
14.4000
352.2000
12.2000
16.6000
2.8000
11.6000
114.0000
11.4000
21.8000
3.4000
8.6000
80.0000
11.7333
22.8000
3.4667
11.5333
182.0667
Std. Devi ation
1.14018
6.89202
.83666
2.30217
114.76149
1.92354
3.78153
.83666
2.07364
108.41125
1.14018
4.49444
1.34164
2.40832
14.33527
1.38701
7.48522
1.12546
3.22638
151.30266
Vali d N (l i stwi se)
Unwei ghted
Wei ghted
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
5
5.000
15
15.000
15
15.000
15
15.000
15
15.000
15
15.000
Univariate F tests
Tests of Equality of Group Means
Wilks'
Lambda
PRICE
.936
QUALITY
.418
PACKAG
.722
TASTE
.423
AVALBLTY
.314
F
.413
8.349
2.313
8.195
13.131
df1
df2
2
2
2
2
2
12
12
12
12
12
Sig.
.671
.005
.141
.006
.001
Discrim Functions
Eigenv alues
Function Eigenvalue % of Variance Cumulative %
1
4.749 a
81.4
81.4
2
1.083 a
18.6
100.0
Canonical
Correlation
.909
.721
a. First 2 canonical discriminant functions were used in the
analysis.
Significance Test
Wilks' Lambda
Test of Function(s)
1 through 2
2
Wilks'
Lambda
.084
.480
Chi-square
24.827
7.336
df
10
4
Sig.
.006
.119
Standardised Coeffs.
Standardized Canonical Discriminant Function Coefficients
Function
PRICE
QUALITY
PACKAG
TASTE
AVALBLTY
1
.207
.988
-.398
-.136
.999
2
.701
-.454
-.293
.986
-.122
Var. Loadings on Functions
Structure M atrix
Function
AVALBLTY
QUALITY
PACKAG
TASTE
PRICE
1
.664*
.517*
.268*
.431
-.044
2
.294
-.336
-.203
.668*
.235*
Pooled within-groups correlations between discrim inating
variables and standardized canonical discrim inant functions
Variables ordered by absolute size of correlation within function.
*. Largest absolute correlation between each variable and
any discrim inant function
Centroids of Brands on Functions
Functions at Group Centroids
Function
BRAND
1
2
3
1
2.745
-1.596
-1.149
2
.123
1.073
-1.196
Unstandardized canonical discriminant
functions evaluated at group means
Plot of Brands on 2 Dimensions
Canonical Discriminant Functions
2
2
1
1
0
-1
BRAND
Group Centroids
3
-2
2
-3
1
-4
-2
Function 1
Putting Variables/Attribute Vectors on the
Above Map
Vectors which represent the original attributes
can be located on the above map. If there are
more than 3 brands, we may get more than 2
dimensions, and may have to draw more than
one plot of the above type.
To plot the attributes on the map above, we can
use the standardized coefficients of the original
variables in the discriminant function. For
example, for Taste, the standardized coefficients
are -.136 and .986 on Dimensions 1 and 2
respectively. So we can locate this point (-.136,
.986) on the map, and draw an arrow from the
origin to that point.
This will be labeled the Taste vector, and
similarly, all other vectors can be located,
one for each of the five attributes - Price,
Quality, Availability, Packaging and Taste.
The length of the arrow represents its
effect in discriminating on each dimension.
Longer arrows pointing more closely
towards a given group centroid represent
variables most strongly associated with
the group (or Brand, in this case). Vectors
pointing in the opposite direction from a
given group centroid represent lower
association with a group.
Variables with longer vectors in a given
dimension, and those closest to a given
axis (dimension represented by the
discriminant function) are contributing
more to the interpretation of that
dimension. Looking at all variables that
contribute to a given axis (dimension), we
can label the dimension as a combination
of those variables.
In this case, the interpretation in terms of
the variables and their correlation to
dimensions 1 and 2 can be found from the
graph which follows (on next page).
Plot of Brands and Attribute Vectors
1.5
Cadbury
Taste
1
Price
Dimension 2
0.5
Nestle
0
-2
-1
Packaging
1 Availability
-0.5
Quality
-1
Amul
-1.5
Dimension 1
As seen from the graph, Nestle, Cadbury
and Amul, the three brands have their
unique positions on the map. In addition,
on the same map, we now have plotted
values of the attributes on the same 2
dimensions (each discriminant function
represents a dimension). As we can see,
Dimension 1 seems to be a combination of
Availability (closest to the x-axis) and
Quality. This is also evident from the
standardized discriminant coefficients for
Availability (.999) and Quality (.988) on
Dimension 1, from the earlier output table.
Dimension 2 seems to comprise of Taste and
Price, the two vectors (arrows) that are closest
to the vertical axis. This is also evident from the
standardized coefficients, of .986 and .701
respectively, for Taste and Price on Dimension
2, from the earlier output table.
Packaging is not useful in defining any of the two
dimensions, as its arrow is not close to either of
the two dimensions.
Brands and their Association with
Attributes/Dimensions
Nestle seems to be stronger on Dimension 1
(Availability and Quality), and Cadbury on
Dimension 2 (Taste and to a lesser extent,
Price). Amul scores low on both dimensions
compared to its competitors.