Preference mapping Methodology and examples Relating sensory

1
Preference mappingMethodology and examples

• Relating sensory to consumer data
• Tormod Næs

2
Overview

• Data structures used
• Linear prefmap
• Methodology
• Advantages and disadvantages
• Examples
• Ideal point mapping
• Methdology
• Advantages and disadvantages
• Examples

3
Preference mapping

• Interested in understanding the relationship
  between sensory data and consumer preferences for
  a number of samples/products
• Product development
• Comparison with competitors
• Which samples are preferred
• Which attributes characterise these samples
• Which consumers prefer which samples (segments)
• Get ideas for new products

4
Focus on sensory analysis and consumer studies

• Sensory analysis (descriptive sensory analysis)
• A number of assessors test a number of
  objects/products
• Analytical instrument, no preferences
• Intensity of a number of attributes (objective,
  integrated (experts))
• Training, calibration of scales
• Ends up with intensity scores between limits for
  all products, attributes and assessors. One
  usually takes averages over assessor.

5
Descriptive sensory analysis - intensity of
attributes (usually using a scale with lower and
upper ends)
One usually takes averages over assessors for
these studies
N Samples (objects, wines)
L Assessors
Averages
K Attributes
6
Consumer studies

• Consumer studies
• A number of consumers test a number of
  objects/products
• Degree of liking (or purchase intent, acceptance)
• No training, limited/no calibration of scales
  (sometimes words are used, like very much)
• Ends up with degree of liking for all products
  and consumers

7
Consumer testing/degree of liking
(NL matrix)
N objects
L consumers
8
Data structures

• Descriptive sensory data for a number of samples
• NK matrix of data (N samples and K attributes)
• Intensities, scores on a scale (typically 1-9)
• Consumer preferences for the same samples
• NM matrix of data (N samples, M consumers)
• Preferences (degree of liking, preference,
  purchase intent etc.)
• In both cases, the numbers are on a specified
  scale

9
Relate and interpret
N samples
M consumers
K sensory attributes
Preference mapping
10
Analysis

• Can compute and analyse the average preferences
  (taken over consumers)
• Can be useful, but usually not enough
• Sometimes of little value to consider an
  average consumer
• Two segments of consumers may disagree strongly
• Better to use multivariate analysis.
• Produce maps/plots by PCA, PCR or PLS
• Interpret maps
• Information about which samples are preferred by
  which consumers
• Focus first on linear methods

11
Example of average vs. individual values
Example based on 8 noodle samples
Tang, Heymann and Hsieh (2000) Food Quality and
Preference
12
Different techniques

• MDPREF, PCA of consumer data. The sensory
  attributes are regressed on the principal
  components. Internal preference mapping
• PREFMAP, PCA of sensory data. The consumers are
  regressed onto the principal components. External
  preference mapping

13
Sensory data
PREFMAP
PCA plots of sensory data
Relate each consumer to these directions
Consumer loadings
14
Consumer data
MDPREF
PCA plots of consumer data
Relate each sensory attribute to these directions
Sensory loadings
15
Comparison

• Both MDPREF and PREFMAP are principal component
  regressions
• X and Y are switched
• PLS can be used instead of PCR, more difficult to
  test significance of relations
• The first step of MDPREF can be done without
  sensory data

16
Comparison

• Both PREFMAP and MDPREF give score plots and
  loading plots
• Score plots show the relationship between the
  samples
• The consumer loadings correspond to the different
  consumers, directions of increased preference
• The sensory loadings correspond to the different
  sensory attributes

17
Different loading plots

• Sensory loadings are plotted as usual
• Consumer loadings are sometimes scaled to have
  unit length
• This focuses on direction only, not on strength
• Sometimes arrows/lines are used as well
• Correlation loadings, later

18
Advantages and disadvantages

• External preference mapping is easier to extend
  to non-linear models, see later
• External preference mapping may overlook
  important variables in component 3, 4 etc.
• Internal mapping is a solution to this
• is guaranteed to concentrate on the preference,
  which is the most important aspect

19
Miscellaneous

• Points in the centre?
• Weak relation, no preference direction
• Open areas?
• If model is believed to be valid, ideas for new
  products

20
Analysis of dry fermented lamb sausages

• Understand the sensory variability among the
  sausages in the Norwegian market
• Understand relationships between sensory and
  chemistry
• Investigate preferences for the different
  sausages
• Segments
• Which samples are most liked
• Which attributes are most liked

21
Experimental

• 14 sausages from the Norwegian market
• Almost the entire market
• Sensory analysis and chemistry of all
• 6 samples selected for the consumer study

22
Steps in the analysis

• PCA of sensory data
• Selection of samples based on PCA plot
• Box-plots of consumer scores (pr. object)
• External preference mapping
• Eliminating unreliable consumers

23
PCA loading plot, 63 explained
component 2
component 1
Three components, 70
24
Scores plot, 63 explained
Selection of samples from scores plot
Component 2
Component 1
25
Selection of samples for consumer testing

• Main idea is to use a PCA plot for sample
  selection
• Select a small number of samples as evenly as
  possible from the whole region of interest
• Add samples that are absolutely needed for other
  reasons

26
Box plots of each of the 6 sausages
27
Consumer biplot with scaled consumer loadings
28
Sensory biplot
29
Elimination of consumers with weak relation to
sensory

• Can be done by
• regular F-testing for regression analysis / PCR
• Delete those consumers which has no significant
  relationship to the sensory dimensions
• using explained variance (for instance lt50 after
  two components)
• Correlation loadings

30
Reliable consumers only
31
Example 2, preference mapping of white
winesAlice Wilke, Ulrich Fischer 2001
(presentation Pangborn 2001) DLRNeustadt,
Germany.

• 4 white wines
• Weik, Kab. 98
• Kaefferkopf 97
• Neveu, Sptl. 97
• Monrepos 98
• Sensory profile 11attributes
• 200 consumers
• scale between 1 and 4

32
PCA of sensory data
Score plot relations between wines
Sensory loading plot Relations between attributes
33
Consumer loadings
Market for all products, can perform
segmentation (overlap makes it somewhat difficult
to interpret)
34
Consumer loadings, contour plot
35
Segmentation

• If the consumers show very different preferences
  it may be reasonable to segment the population
  into homogeneous groups
• Can be done by visual inspection of loadings
  plots or by using cluster analysis
• Try to interpret the groups/important for
  marketing

36
Segments determined by cluster analysis
37
Average preferences within each segment
Consumers within each segment are similar.
Therefore averages are meaningful
38
Information obtained from external preference
mapping

• Information about which sausages that are
  preferred by the consumers
• Information about which sensory attributes that
  characterise the preferred sausages
• Information about possible market segments

39
Comparisons of methods

• For the lamb sausage data, PLS and PCR gave
  comparable results
• For the lamb sausage data, MDPREF and PREFMAP
  gave the same qualitative information

40
Validation/selecting the number of components

• With very few samples, regular cross-validation
  becomes unimportant (impossible to use)
• The samples do not represent any population, the
  PCA is just a way to visualise the data.
• One must rely on subjective validation. Use prior
  knowledge, information about design, data
  collection etc. to assess the reliability of the
  dimensions

41
Relating segments to external demographic data

• The easiest way is to use tabulation
• For each segment, compute the percentage of
  people with a certain property
• Compare the percentages for the different
  segments
• Can also be done by various regression analyses
• Regressing consumer loadings onto external
  variables( see exercises)
• Can also be done by colour or symbols

42
Lamb sausage data

• Several external variables were measured
• Demographic, age, gender etc.
• Food and other habits
• Attitudes
• Percentages were computed for each segment

43
Examples

• Gender
• Male
• Segment 1 Segment 2 Segment 3 Segment 4
• 36 67 51 62
• Age
• Between 30 and 50
• Segment 1 Segment 2 Segment 3 Segment 4
• 45 56 47 51

44
Serving samples to consumers

• Number of samples to each consumer Trade-off
  between statistical and practical requirements
• 5-7 is usually a reasonable compromise
• Randomisation is important
• Different groups receive the samples in different
  order
• Usually the same samples are served to all
  consumers
• Can sometimes use different samples for different
  consumers
• Split consumers into sugroups and serve different
  samples to the different groups
• Assumption relation between sensory attributes
  and consumer preferences is independent of
  samples used

45
Ideal point mapping

• Nonlinear models using polynomials
• Standard methods
• New method based on fuzzy clustering (FCM)

46
Different models

• Linear models assume that the preferences
  increase in one direction. Different for the
  different consumers
• Ideal point models assume that there is an
  optimal value some place. Different optimal value
  for each consumer

47
Models in two variables
LInear
Y a b1x1 b2x2 e
Polynomial
Y a b1x1 b2x2 b11x12b22x22b12x1x2 e
Xs principal components
48
Different PREFMAP models
Linear model, the more the better
Ideal points models, optimal values
49
Ideal point, simplest case model Y a b1x1
b2x2 b(x12x22) e
50
Models with and without interaction
From J. McEwan. (1996). Preference mapping for
product optimisation. In. Multivariate analysis
of data in sensory science. (eds. T. Næs and E.
Risvik), Elsevier.
51
Ideal point mapping

• Ideal point mapping can allow for different
  attributes related to liking and disliking.
• May be of some importance

52
Example of individual anlyses

• From J. McEwan. (1996). Preference mapping for
  product optimisation. In. Multivariate analysis
  of data in sensory science. (eds. T. Næs and E.
  Risvik), Elsevier.
• Orange drink, effect of sweetener and citric
  acid. 62 consumers, 2 PCs

53
max - min
Coefficients and significance
54
Scores and loadings Max and min points indicated
55
Preference mapping
X Sensory attributes
Y Consumer acceptance
Relation
Linear prefmap
Ideal point prefmap

• Advantage
• Stable individual models
• Problem
• Limited to linear modelling

• Advantage
• Non-linear modelling
• Problem
• Unstable individual models due to few samples and
  many model terms

56
Ideal point preference mapping(external mapping)

• Joint model for all assessors
• Unrealistic
• Does not take individual differences into account
• Joint segmentation and estimation of coefficients
• Fuzzy clustering or mixture modelling for
  clustering
• More difficult to implement
• Can handle less than 6 samples and can be used
  for 3 dimensions
• Not established within preference mapping. More
  common in conjoint analysis.

57
Proposed method based on FC

• The general criterion to be optimised is
• Can be used for different samples to different
  people
• Results in membership values and regression
  coefficients
• This criterion enable the used of serving
  different samples to different consumer groups

x sensory attributes y consumer acceptance
58
Algorithm for Fuzzy clustering

• Initialise U
• Update distances for given U.

3. Update membership values for given D.
Næs, T. Kubberød, E. and Sivertsen, H.
Identifying and interpreting market segments
using conjoint analysis (2001). Food Qual. And
Pref. 12,2, 133-144..
59
Case study on low-fat cheese
Objective To understand the preference pattern
for low-fat cheese in Norway
60
The case study on low-fat cheese

• Descriptive sensory analysis
• 19 cheese variants (5 - 17 fat content)
  experimentally produced or commercial available
  at the Finnish, Swedish or Norwegian market
• Consumer study
• 12 cheese samples
• 115 consumers randomly divided in 2 groups
• Each group tasted 6 cheese samples

61
Selection of samples for the two consumer groups
Fermented sour flavour, bitter taste
Sticky, fatty
Firm, hard, rubbery, grainy
Acidic flavour, cream flavour
62
The consumer test

• Criteria
• Consumers of hard or semi-hard cheese
• 25-50 years old
• Reside in the eastern part of Norway
• Health interested
• Procedure
• Blind tasting
• Rating degree of liking on a nine point hedonic
  scale anchored with Dislike Extremely and Like
  Extremely and with a neutral centre point of
  Neither Like nor Dislike (Peryam and Pilgrim,
  1957)
• Attitude to health and light products (Roininen,
  Lahteenmaki, and Tuorila, 1999), food neophobia
  (Pliner and Hobden, 1992) and restrained eating
  (van Strien, Frijters, Bergers, and Defares,
  1986)
• Socio- and demographic data

63
Results Modelling the data

• Choice of the fuzzyfier m
• by comparing average absolute residual for all
  observations in the data set
• The value of 1.1 seemed to give the best results

64
Choosing number of segments
65
First segment (n 47) maximum
Preference maximum Fatty, soft, acidic flavour
and cream flavour
Consumers Health interested Restrained
eaters Eaters of low-fat cheese
66
Second segment (n 30) - saddle point
Fatty, soft and sticky
Grainy, rubbery and hard
Consumers Not food neophobic Cheese experts
67
Third segment (n 38) linear
Fatty, soft, acidic flavour, and cream flavour
Consumers Cheese novices Not users of light
products
68
Value for future product development

• Proposed strategy
• Found the ideal point of preference for one
  segment
• Cheese with optimal sensory attributes
• Found directions of preference for the other
  segments
• Important in the further product development
• Linear preference mapping
• Unable to find non-linear segments

69
Conclusions -FCM

• The proposed strategy
• Keep number of samples served to each consumer
  low and still cover the sensory space well
• The use of fuzzy clustering for segmentation
  proved useful in studying and finding consumer
  segments with either
• Linear preference
• Ideal point of preference
• Non-linear preference (saddle point)