Presented at University of Stellenbosch, September 2010

1. Segmentation of BIB
Consumer Liking Data of
12 Cabernet Sauvignon Wines
Chris Findlay, PhD
Chairman
Compusense Inc.
Guelph, Canada

2. Background
• Consumer testing of beverage alcohol has a number of
serious challenges. The effect of consumption of alcohol
is a limiting factor in obtaining complete block data.
• Collecting consumer data over several days affects the
quality of the consumer response. By the third day, most
consumers are behaving like experts, a conclusion that is
supported by the decrease in first position effect.
Segmentation of BIB Consumer Liking Data
August 26, 2010

3. The Effect of Order and Day on Consumer Liking
12 White Wines, 115 Consumers, CBD 12:12 over 3 Days
70
60
50
40
1
2
3
30 4
20
10
0
1 2 3 All
Segmentation of BIB Consumer Liking Data
August 26, 2010

4. Ensuring Quality Data
• Careful selection of appropriate consumers
• Sensory selection of test products
• Diligent application of Best Practices
• Typically, segmentation of consumer liking data
requires a complete block.
Segmentation of BIB Consumer Liking Data
August 26, 2010

5. The Study
• In this study, 12 Cabernet Sauvignon wines
were evaluated by over 600 red wine consumers
in a 12 present 3 Balanced Incomplete Block
design. Each consumer tasted 3 of the wines in
a single 10 minute session, with demographic
questions providing a break between samples.
• A total of 11 sessions were conducted at 5 LCBO
store locations.
Segmentation of BIB Consumer Liking Data
August 26, 2010

6. Factor Scores plot : dimension 1 versus 2
2.58 PC 1 vs PC 2
Sensory Map #1
W3
Astringent
W11
W6 W7 W2 Bitter
W12 Raisin
FruityWoody
Eucalyptus
Leather
Vanilla Sour
Pepper
-2.58 W10 Floral T obacco W5 2.58
W4 W8 Green
Sweet Asparagus
Coffee Smoke
W9
W1
-2.58

7. Factor Scores plot : dimension 3 versus 4
2.58
PC 3 vs PC 4
Sensory Map #2
W3 W2
W10
W9 Leather
W1
Eucalyptus
Sweet
Asparagus
W6 Coffee
Floral
Green Bitter
Fruity
-2.58 W12 Pepper Woody
Astringent 2.58
Raisin T obacco Vanilla
Smoke
W5
Sour W11
W7
W4
W8
-2.58

8. Data collection at LCBO stores
Segmentation of BIB Consumer Liking Data
August 26, 2010

9. RAW RESULTS
Mean Liking across All Consumers
W6
W1
Overall Liking
W4 (n=614)
W10
W12
W8
W7
W11
W9
W5
W3
W2
3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0
Segmentation of BIB Consumer Liking Data
August 26, 2010

10. Segmentation Procedure
• Substitute missing data with for each panelist
1. Panelist Mean for the 3 products tested
2. Product mean for each product
3. Grand mean for all products
• Cluster using Qannari method using Senstools
3.3.1 (OP&P, Utrecht)
• Apply the cluster solutions to the original data
to create segments
Segmentation of BIB Consumer Liking Data
August 26, 2010

11. Segmentation of BIB Consumer Liking Data
August 26, 2010

12. T1 Clustering results
Increase Relative Stress By Decreasing Number Of Clusters
0.20
The average liking response for each panelist was inserted
0.15 into the missing data points. The Qannari Clustering method
(Senstools 3.3.1) provided a four cluster solution.
0.10
0.05
0.00
20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1
Segmentation of BIB Consumer Liking Data
August 26, 2010

13. Clusters
28% 23%
Cluster 1 - 172 Cluster 2 - 142
Cluster 4 - 105
Cluster 3 - 195
17%
32%
Segmentation of BIB Consumer Liking Data
August 26, 2010

14. W2
W4
W10
W1 Cluster 2 Liking
Cluster 1 Liking
W3 (n=142 or 23%)
W6 (n= 172 or 28%)
W1
W7
W11
W8
W5
W9
W8
W12
W4
W11
W12
W2
W6
W10
W9
W5
W7
W3
3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0
3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0
W9 W6
W6 W8
Cluster 3 Liking Cluster 4 Liking
W7 (n= 195 or 32%) W3 (n = 105 or 17%)
W12 W5
W3 W1
W2 W11
W5 W4
W10 W9
W1 W12
W8 W10
W11 W7
W4 W2
3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0
Segmentation of BIB Consumer Liking Data
August 26, 2010

15. Cluster 1
Astringent Sensory Contrast
Bitter Positive
Sour
Sweet
Leather
Raisin
Fruity
Eucalyptus
Woody
Pepper
Vanilla
Floral
Coffee
Asparagus
Smoke
Tobacco
-3 -2 -1 0 1 2 3
Segmentation of BIB Consumer Liking Data
August 26, 2010

16. Cluster 2
Astringent Sensory Contrast
Bitter Positive
Sour
Sweet
Leather
Raisin
Fruity
Eucalyptus
Woody
Pepper
Vanilla
Floral
Coffee
Asparagus
Smoke
Tobacco
-3 -2 -1 0 1 2 3
Segmentation of BIB Consumer Liking Data
August 26, 2010

17. Cluster 3
Astringent Sensory Contrast
Bitter Positive
Sour
Sweet
Leather
Raisin
Fruity
Eucalyptus
Woody
Pepper
Vanilla
Floral
Coffee
Asparagus
Smoke
Tobacco
-3 -2 -1 0 1 2 3
Segmentation of BIB Consumer Liking Data
August 26, 2010

18. Cluster 4
Astringent Sensory Contrast
Bitter Positive
Sour
Sweet
Leather
Raisin
Fruity
Eucalyptus
Woody
Pepper
Vanilla
Floral
Coffee
Asparagus
Smoke
Tobacco
-3 -2 -1 0 1 2 3
Segmentation of BIB Consumer Liking Data
August 26, 2010

19. Statistical Challenge
• We need a valid approach to
segmentation of consumer BIB data
• Possibly a combination of sensory best
practice, experimental design and
advanced statistical analysis
Segmentation of BIB Consumer Liking Data
August 26, 2010

20. • Let’s consider a sensory space
Factor Scores plot : dimension 3 versus 4
Factor Scores plot : dimension 1 versus 2
2.58
2.58
W3 W2
W3
W10
Astringent
W11 W9
W6 W7 W2 Bitter Leather
W12 Raisin W1
Eucalyptus
Sweet
FruityWoody
Eucalyptus
Leather Asparagus
Vanilla Sour Coffee
Pepper W6 Floral
Green Bitter
Fruity
-2.58 W10 Floral T obacco W5 -2.58 2.58 W12 Pepper Woody
Astringent 2.58
Raisin Tobacco Vanilla
W4 W8 Green Smoke
Sweet Asparagus
W5
Coffee Smoke Sour W11
W7
W9
W4
W8
W1
-2.58
-2.58
• Can we find logical contrasts to test
Segmentation of BIB Consumer Liking Data
August 26, 2010

21. • Quadrangles and Triangles
Factor Scores plot : dimension 1 versus 2
Factor Scores plot : dimension 3 versus 4
2.58
2.58
W3 W2
W3
W10
Astringent
W11
W6 W7 W2 Bitter W9 Leather
W12 Raisin W1
FruityWoody Eucalyptus
Sweet
Eucalyptus
Leather
Sour Asparagus
Vanilla Pepper W6 Coffee
Floral
Green Bitter
Fruity
-2.58 W10 Floral T obacco W5 -2.58 2.58 W12 Pepper Woody
Astringent 2.58
Raisin T obacco Vanilla
W4 W8 Green Smoke
Sweet Asparagus
W5
Sour W11
Coffee Smoke W7
W9
W4
W8
W1
-2.58
-2.58
Segmentation of BIB Consumer Liking Data
August 26, 2010

22. Is there a Preference?
• To state a true preference a consumer must see
a real difference.
• Otherwise it’s just a guess.
• Considerations:
– Do all attributes influence liking equally?
– For all consumers?
– In all contexts?
Are you ready for a new era of sensory & consumer research?
April 23, 2010

23. Is there a Better Clustering Method?
• Model-based clustering is being developed to improve
the validity of the clusters.
• Ongoing research by Dr. Paul McNicholas and his team
at the University of Guelph and Compusense is
attempting to combine sensory design and effective
clustering to provide robust solutions to the challenge of
BIB segmentation of high fatigue products.
• Please refer to the Sensometrics.org for proceedings of
the 10th Conference held in Rotterdam, July 2010.
Are you ready for a new era of sensory & consumer research?
April 23, 2010

24. Conclusion
• The selection of sensory contrasts may be used to strengthen the
design of BIB experiments. If random incomplete blocks are chosen
it is possible that groupings of quite similar products may be
received by some assessors and widely different products by others.
• Prior knowledge of the sensory properties is essential in assigning
blocks that will emphasize the inherent differences in the products
and improve the quality of data from the study.
Segmentation of BIB Consumer Liking Data
August 26, 2010

25. For more information, please contact
Chris Findlay, PhD
Chairman
+1 800 367 6666 (North America)
+1 519 836 9993 (International)
cfindlay@compusense.com