using range voting analysis techniques for odor profiling · pdf fileusing range voting...
TRANSCRIPT
1
Using Range Voting Analysis Techniques for Odor Profiling Data
Kimberly-Clark Corp. Sensometrics 2008
G Keep, W Raynor
Sensometrics Conference
Odor Profiling
Challenges
• Odor complexities
• Unlimited odor “attributes”
• Panel training demands
● practice
● calibration
• Researcher time demands
One Possible Approach
• Olfaction theory
• Hybrid testing method
● odor identification
● similarity ratings
• Unique analysis
● Range Voting approach
● Compositional data
• Utility
Kimberly-Clark Corp.July 21, 2008
2
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Mnemonic Theory of Odor Perception
Semantic-EpisodicKnowledge
PatternMatcherEncoder
Engram Store
ControlledAssociator
SensoryIntegrator
OtherSensory
Input
OlfactoryInput
Adapted from
Stevenson & Boakes (2003)
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Corresponding Sensory Methodology
Neural odor representations matched to memory encodings.
Better matches leads to more memory activation.
Pattern of activations represent the perception of the odor.
Odors easily confused if memory activations are slight.
Mnemonic Theory Odor Profiling Method
Rate
Identified Odors
Identify
“Matching” Odors
Standard
Odor Referent Set
0 2 4 6 8 10
Similarity Scale
Odor“Notes”
3
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Key Method Factors
• Panelists select referent odor notes independently.
• Rate “similarity” to test odor.
• Adjunct elements:
• Odor Groupings / Families
• Overall Intensity Ratings
0 1 2 3 4 5 6 7 8 9 10
Low Mid High
Traditional Intensity Rating (combined odor)
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Range Voting with Extras
• Winner-take-all elections
• Assign points to candidates
● Bounded 0-100 scale; can assign 0
● Add up – highest score wins
George Box
Copernicus
Albert Einstein
RA Fisher
Galileo Galilei
Jerzey Neyman
0 100
0 100
0 100
Most Famous ScientistRange Voting
Compositional Data
• Proportional parts of a whole
• Sum constrained to a constant
• Relative information
• Non-linear
e.g. 100%0.40
0.25
0.16
0.10
0.09
4
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
The Analysis Plan
“Range VotingAnalysis”
Statistical ProductComparisons
Odor Character
Weights
Pre-set
Odor Groupings
Overall Intensity
Ratings
“Families”
Similarity
Ratings
Odor CharacterRaw Data
Raw Data
Sensometrics Conference
“Range Voting” Analysis Step 1
Kimberly-Clark Corp.July 21, 2008
• Work on individual similarity ratings
• Can use cumulative density function (CDF)
• PROC RANK (Blomnormalization)
• e Blom Score
Logit Transformation
compositions have distributions that are logistic-normal
within panelists
partially remove rater-to-rater variation
5
Sensometrics Conference
“Range Voting” Analysis Step 2
Kimberly-Clark Corp.July 21, 2008
Scale the exponents
• Scale e Blom Score
• Odors sum to 100%
• within product x panelist
Can combine over panelists to obtain average odor referent notes
Sensometrics Conference
Calculate Proportions (Weights)
Kimberly-Clark Corp.July 21, 2008
0.40
0.25
0.16
0.10
0.09 Fruit
Floral
Citrus
Wood
Other• Base on Scale values
• Pool across odor families
• by sample
• Individual notes
• detailed list
Sample A
6
Sensometrics Conference
Hypothesis Testing
Kimberly-Clark Corp.July 21, 2008
• Use total odor intensities
• Shift distribution
• Base on proportion
• Per sample x odor family
• Analyze with General Linear Models (ANOVA)
Sample A
-2 -1 0 1 2 3 4 5 6 7 8 9 10
FruitTotal Odor
Intensity
40%
Wood
10%
Intensity
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
A Useful Method?
• Discriminates Between Products
• Measures Consistently
• Provides Meaningful Info
7
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Discriminating
• Historical data
• Discrim. Index
● omega squared
● 0-100
• n = 204
● families x samples
● over 50 studies0
5
10
15
20
25
30
0 10 20 30 40 50 60 70 80 90 100
Omega Squared
Freq
uen
cy
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Consistency
• Test-retest reliability
● for good discrim:
● r > 0.90
• No test x sample interactions
• Consistent proportions
● RMS < 6%
Sample 1
Sample 3 Sample 4
18
4
32
6
6
21
12
18
42
7
18
15
34
29
2
8
6
6
1425
2818
6
4
19
7
48
16
29
9
44
8
10
26
4
Sample 2
18% 21%
6%
7% 6%
42%
32%
18%
14%
6%4%
29%
Inner = Test 1
Outer = Test 2
3 9 %
2 3 %
1 1 %
8 %
8 %
6 %5 %0 %
Floral
Wood
Fruit
Ozone
Sweet
TropFruit
Citrus
Green
\
44%48%
6%
6%
18%
4%
8%
8%
34%
12%
25% 7%
9%
18%
18%
16%10%
29%26%
15%
28%
40
23
11
8
8
65
44
23
10
6
6
6
6
6%
8%
8%
6%
6%
6%
10%11%
23%
23%
40% 44%
6%
5%
8
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Meaningful Information
• What to compare to?
• Aging stability
• Fragrance target confirmation
• Matching fragrances
• Select fragrance submissions
Generalizability
Utility
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Summary
• Odor profiling methodology
● less labor intensive
• “Range Voting” analysis
● compositional data
• Demonstrated utility
9
Sensometrics Conference
Key References
Stevenson RJ., Boakes RA. 2003. A mnemonic theory of odor perception. Psychological Review 110:340–364.
Brams SJ, Fishburn PC. 2007. 2nd ed. Approval voting. Springer. 198p.
Aitchison J. 1986. The statistical analysis of compositional data. Springer. 416p.
Kimberly-Clark Corp.July 21, 2008
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Acknowledgements
K-C people involved with the method:
• Rhonda Engle
• Barb Bombinski
• Tami Gaubatz
• Rick Zepp
• Patti Braun
• Mary Stiltjes
10
Sensometrics ConferenceKimberly-Clark Corp.July 21, 2008
Range Voting Analysis for Odor Profiling Data
ENGRAM STORE
OLFACTORY PROCESSING MODULE
3. CONTROLLED ASSOCIATOR
2a. PATTERN MATCHER
2b. AUTOMATIC
COMPARATORAND ENCODER
5. SEMANTIC AND
EPISODIC KNOWLEDGE4. SENSORY INTEGRATOR
OTHER
SENSORYINPUT
1. OLFACTORY INPUT
Mnemonic Theory of Odor Processing
Stevenson & Boakes (2003)