You are on page 1of 9

Food Quali!y and Prefkncr Vol. 8, No. 1, pp.

63-71, 1997
Copynght 0 1996 Elsevier Science Ltd
Printed in Great Britain. All rights reserved
ELSEVIER PII: SOSSO-3293(96)00016-X 0950-3293/97 8 17.00+.00

EVALUATION
OFSENSORY PROFILING
AND PROJECTIVE
MAPPING DATA
Einar Risvik,a* Jean A. McEwanb & Marit Redbottena
“MATFORSK, Norwegian Food Research Institute, Osloveien 1, 1430 As, Norway
“Dept. of Sensory Science, Campden, Campden and Chorleywood Food Research Association, Chipping Campden,
Gloucester GL55 6LD, UK

(Received 2 January 1996; accepted 26 March 1996)

technique has proven to be a strong tool for the food


ABSTRACT industry in product development and marketing, as it
provides information of major importance to describe
differences between food products.
This paper deals with the method of projective mapping and In market research the tradition for qualitative tech-
its use in sensory analysis. Projective mapping is a method niques is strong in work with consumers to understand
which allows naive consumers to map products on a two the perceived association or relationship between objects.
dimension space, based on similarities and di$rences in the The elicitation of vague, ambiguous and unstructured
products. An experiment was carried out on seven blueberry ideas is an approach often used in market research,
soups, whereby sensory profiling was undertaken using a referred to as projective techniques (Oppenheim, 1966).
trained panel and projective mapping. Preference assessment This group of techniques is not always well defined, but
is in frequent use, although difficult to quantify as there
was undertaken using a small group of consumers.
is no quantitative structure to the data collection, the
Results indicated that the three replicates of the mapping
information being purely qualitative.
exercise produced visually very similar maps, at least on the
The relationship between conventional profiling tech-
first two dimensions. However, it was found that consumers
niques and projective techniques is not immediately
perceived the samples in somewhat di$erent ways as high- apparent, as one collects one set of data from trained
lighted by R V coeficients. assessors and the other from untrained consumers. A
The consensus mapping dimensions were compared to combination of the two will provide both a description of
those from the projle data, and it was apparent that the best sample differences and also a possibility for inter-
similarity was found when comparing the first dimension, pretation of these differences as regards to consumer
thus suggesting good agreement on the obvious aspects of the perception. To investigate this further an approach
product. The internal preference map also revealed major where the two techniques were combined, was suggested.
product contrasts along this dimension with some weak evi- A similar approach has been tested before in the pub-
lication by Risvik et al. ( 1994). In the present publication
dence of segmentation.
this approach is forwarded further in several aspects.
The overall conclusion of this paper raises further fun-
First of all the data are collected from two different
damental questions about consensus spaces with consumers,
sources; the descriptive data from a trained panel, and
and the question of dimensionality of consumer perception as
the mapping data from an untrained group of con-
compared to trained panels. Copyright 0 1996 Elsevier sumers. Secondly the profiling data show greater com-
Science Ltd plexity, that is in the systematic and interpretable
structure, with four significant dimensions (from cross-
validation) in a principal component analysis (PCA). In
INTRODUCTION addition data for preference were collected from the
consumers.
Both profiling and projective mapping techniques,
Data from conventional sensory profiling (Amerine et al., as used here, demand the utilisation of sophisticated
1965; Williams & Carter, 1977; Jellinek, 1985; Powers, statistical techniques in order to provide quantifiable
1984; Risvik, 1985; Stone & Side;, 1985; Meilgaard et al., results. Output from multivariate analysis of sensory
1987) give a quantitative description of sensory attri- profiles is very similar to the format of the data provided
butes as perceived by a trained group of subjects. This by projective mapping techniques (Risvik et at., 1994).
The purpose of this study was to investigate further a
*To whom correspondence should be addressed. possible way of allowing consumers to evaluate samples in

63
64 E. Risvik et al.

an overall and simple way. To achieve this the consumers Sensory analysis
‘were asked to provide a two dimensional projection of a
set of samples. There were no instructions as to how the Descriptive sensory analysis was undertaken at MAT-
samples should be separated in this space, but examples FORSK, Norwegian Food Research Institute. A trained
of two-dimensional geographical maps were used for panel of 12 assessors, all with previous experience of sen-
illustration. It is possible, retrospectively, to generate a sory testing, participated in the test. The samples were
quantification of the separation, by reading the co-ordi- not swallowed, and the assessors were asked to rinse well
nates for the samples as a two-dimensional representation between each sample using either still or carbonated
of them. These data could then enter the analysis and be mineral water and non-salted crackers.
compared to the data from the profiling exercise and The profiling analyses were performed in a sensory
from preference measurements. laboratory, according to accredited procedures and
For the profiling part of the study a trained panel was equipped according to international standards (ASTM,
utilised to profile a set of samples. 1968a,6; ISO, 1985, 1988, 1992). The data were collected
A small group of consumers, all non-rejecters of blue- on a direct registration system (CSA, Compusense,
berry soup, was selected to simulate a focus group set- Guelph, Canada) with a 15 cm unstructured line scale,
ting. In a typical situation this would be a part of a anchored with ‘low intensity’ on the left and ‘high
larger scenario where the focus groups only are seen as intensity’ on the right. The samples were marked with
one constituent in the entire experimental set-up. The three-digit random codes and given randomised serving
consumers would then be selected according to criteria orders. The samples were served in two sensory repli-
for participation, which could include screening in order cates. Samples and codes used only for the presentation
to get a homogeneous group of consumers. of the results are presented in Table 1. The sensory pro-
A group of consumers with similarities in their filing procedure was performed according to interna-
response patterns would simplify data interpretation as a tional standards (ASTM, 1968a,b, 1985, 1988, 1992).
single group response could provide a convenient con- Sensory attributes for the profile were selected accord-
sensus for the group. In this case the consumers were not ing to standard procedures (Risvik, 1985). After training
screened according to any criterion except the willingness and screening, 12 attributes describing colour, taste, fla-
to participate in the test. This implies that the response vour, visual texture and mouthfeel were selected. The
pattern for each consumer is expected to be different. A attributes are presented in Table 2.
solution based on consensus is therefore not expected to In the projective mapping experiment a separate
provide meaningful information. group of 8 consumers (typical size for focus group inter-
To meet the major constraints from a previous pub- views) with no prior experience or knowledge of the
lication with this approach (Risvik et al., 1994), a set of method was utilised. The consumers were not selected to
seven samples was selected, with a complex internal represent a given consumer segment or consumer group,
structure. as would normally be the case in focus group interviews.
This paper adopts the approach of a co-ordinate The assessors were asked to place the samples on a two-
system to quantify information from the projective map- dimensional space (an A3 sheet of white paper), accord-
ping exercise and utilises the RV coefficient (Schlich ing to how they perceived them to be related to each
& Guichard, 1989), PCA (Escoufier, 1976) and other. The towns map of Kruskal and Wish (1978) was
internal and external preference mapping (McEwan, used as a training example. To avoid introducing the
1995) for comparing information from the different ideal of ‘right’ solution, the concepts of individual differ-
approaches. ences and the rotation of perceptual maps were intro-
duced (see Risvik et al., 1994 for more detailed
description). The introduction was limited to 5 min. in
MATERIALSAND METHODS order to fit possible focus group interview requirements,
so that this could be performed in a similar time frame to
a traditional qualitative projective technique, utilised in
Samples
TABLE 1. Samples and Codes Used Only for the Presentation
Seven commercially formulated dried soups (three pro-
of the Results
ducers with one product each, and two producers with
two slightly different formulations of the same product) Sample Producer Refomdation Code in the
were selected and referred to as producers l-5 and plots
reformulation type 1 and 2. The samples were kept in 1 1 2 12
closed cardboard containers stored at room temperature 2 2 1 21
and 40% r.h. prior to testing. The soups were prepared 3 3 1 31
according to standardised procedures provided by the 4 3 2 32
5 4 1 41
producers and served at 60°C in white plastic cups,
6 4 2 42
50 ml in each cup. 7 5 2 52
Evaluation of Sensory Projling 65

TABLE 2. Attributes and Descriptions Used by the Trained Panel to Profile the Soup Samples. (NCS* = Natural Colour System)

Attribute Description of attributes !!scale

Intensity of smell Total intensity of smell Low-high intensity


Blueberry aroma Intensity of characteristic smell Low-high intensity
Whiteness Intensity of whiteness according to the NCS* system Low-high intensity
Colour tone Colour tone according to the NCS* system Yellow/red-red/blue
Colour intensity Intensity of the colour tone according to the NCS* system Low-high intensity
Clarity Degree of transparency Low-high clarity
Taste intensity Intensity of taste Low-high intensity
Blueberry flavour Intensity of blueberry aroma Low-high intensity
Sweet Intensity of sweetness Low-high intensity
Acidic, fruity Intensity of acidic and fruity aromas Low-high intensity
Bitter, astringent Intensity of bitterness and astringency Low-high intensity
Watery Degree of watery texture Low-high intensity

market research for this purpose (Oppenheim, 1966). Each attribute was then centred by subtracting the
Data were collected in three separate replicates with one attribute mean and normalised by dividing by the stan-
week intervals, over a period of three weeks. dard deviation. This was undertaken as it was expected
After the 3rd replicate of the mapping exercise the that each consumer would use their own scaling. These
consumers were asked to record their preference for the 7 data (7 rows by 16 columns) were then submitted to
samples. This was measured on a 15 cm open ended PCA, and biplots produced. Each replicate was analysed
hedonic scale, anchored with a ‘grimacing face’ on the separately. The dimensions from the profile and map-
left and a ‘smiling face’ on the right. The results were ping exercises were then compared using the RV coeffi-
measured in mm from the left side. cient (Schlich, 1993). This method provides a simple way
of measuring the similarity between two sets of variables
Data analysis which have been measured on the same samples. The
RV coefficient is analogous to the R-squared goodness of
The conventional profiling data were analysed using fit measure used in regression analysis. It takes values
covariance PCA, comprising a matrix of average scores between 0 and 1, where 1 represents perfect similarity.
with 7 rows (samples) by 12 columns, (attributes). The The consumer preference data, which involved only 8
results of the PCA were presented as biplots (Gabriel, people, were analysed in the first instance by internal
1971), which represent both the samples (Principal preference mapping which is similar to PCA on the con-
Component scores (PC-scores)) and attributes (PC- sumer data alone (MacFie & Thomson, 1988). The pre-
loadings) on the same plot. On the biplots a diagnostic ference map could then be compared to the mapping
biplot was used where r = 0.5 (r is a scaling factor where biplots to determine if preference could be interpreted as
the value 0 equals a variate biplot, 0.5 a diagnostic biplot the main source of information driving the consumers in
and 1 a principal components biplot). The diagnostic the mapping exercise.
biplot tends to be used when there is equal attention to External and internal preference mapping (Schiffman
samples and attributes uoliffe, 1986). Crossvalidation et al., 1981; Schlich & McEwan, 1992; Schlich, 1995;
(GENSTAT 4 Numerical Algorithms Group, Rotham- McEwan, 1995) was then undertaken to see how the
stead, UK) was used to determine the number of preference data related to the conventional profiling
dimensions. In crossvalidation one random sample was results.
omitted from the calculations every time. This was repe- Correlations were calculated between the preference
ated until all samples had been excluded from the calcu- data and profile biplot PCs, and between the preference
lations. The crossvalidation solution was then calculated data and mapping biplot dimensions.
as the average solution, which implies that this solution
was stabilised from strong effects from single deviant
samples (Stone, 1974).
RESULTS AND DISCUSSION
The mapping data were collected by converting the
consumer maps on A3 paper to co-ordinates in with two
axes (x and y). The mapping data were quantified using a Conventional profiling
coordinate system placed in the lower right hand corner
of the A3 page. Coordinates were read in mm on the hori- One purpose related to this study involved a product
zontal and vertical axis as x and y variables, respectively. change for one producer. The project started with a
The data from each replicate were treated separately. descriptive profile of each of the participating samples in
For analysis, the x and y coordinates for each con- order to characterise the differences between the pro-
sumer were treated as two attributes, thus for 8 con- ducts. In this way the producer could see which attri-
sumers 16 attributes (8x2) were available. butes separate the current product from the products of
66 E. Risvik et al.

of sample 52, characterised by its watery/bitter and


acidic flavour, clarity and colour intensity. Sample 12,
characterised by high negative scores along PC 2, is more
white, more acidic and more intense in blueberry aroma
and flavour. Samples 31, 32 and 41, 42 are both pairwise

8 INTENSITY OF SMELL l
COLfXJRSTRENGTH.
0 WHITENESS close to each other but opposite to 12 and 52 in these first
two PCs. This indicated that they are more alike and
. BLUEBERRY FLAVOUR that PC 1 and PC 2 are not very important for inter-
pretation of attributes separating the reformulations of
sample 3 and 4.
PC 3 indicates a separation of samples 41 and 42 from
sample 2 1. These are separated mainly on the intensity of
-2 12 I I smell and taste. PC 4 is related to sweetness only, and
-1 0 1 2
separates the samples 21 and 41 from the others. This PC
PC 1 55%
only accounts for 3% of the information in the data.
FIG. 1. Principal components Gabriel biplot (r=0.5) 1 and
for 7 samples of blueberry soup. Profiling and preference data

other producers and draw conclusions as to how this can To obtain an impression of how the preferences relate to
be changed in order to become more alike to another the profiling information, it was decided to perform
product. The interest, in this case, was to see how sam- both external and internal preference mapping. In the
ples 12 and 52 were different from the others, as these internal preference mapping the consumer preference
represented the main competitors in the market (see Figs data were analysed alone, and then compared to the
1 and 2). As is apparent from the PCA these two samples mapping biplots to see how the preference was driving
were the most different from the others and could easily the results.
be characterised with the first two principal components. The structure of the products, observed in Fig. 3, is
A PCA on the conventional profiling data revealed an very similar to the profiling map in Fig. 1 (PC 2
underlying four-dimensional perceptual structure, where reversed). It is possible to see sample 52 driving the first
a total of 99% of the information was explained in the PC, and 12 vs. 21 and 31 along PC 2. All the preferences
first four principal components. A graphical representa- are pointing in one direction, away from sample 52 on
tion of this is given in Figs 1 and 2. All four principal PC 1. It may be possible to consider the consumers as
components could be interpreted in a meaningful way, separated in two groups, one towards the top left, where
although the fourth PC accounted for only 3% of the sample 12 is most preferred, the other towards the bot-
information. Crossvalidation indicated that 4 PCs had tom left, where samples 21 and 31 are located. The sam-
predictive power. ples form a similar structure to the profiling (Fig. 1)
which have already been discussed.
Interpretation of profiling data The external preference mapping exercise is performed
with the preference data superimposed on top of the
PC 1 and PC 2, each separate out one sample and pro- scores from the PCA of the profiling data, as shown in
vide characteristics to explain how this sample is different Fig. 1. It was decided that only the first two PCs gave
from the others. PC 1 is very important for interpretation meaningful information, these are shown in Fig. 4.

1 3132
0.5 - . BITTER
BLUE;&;; l’ BLUEBERRY
. CLAIRITY FLAVOUR

I 0 SWEET 21
-1 0 1
PC3 9%

FIG. 2. Principal components 3 and 4 for 7 samples of blueberry soup.


Evaluation of Sensory Projiling 67

TABLE 3. Explained Variance for Centred Projective Map-


ping Data

Replicate 1 Replicate 2 Replicate 3

Dimension 1 41% 46% 46%


Dimension 2 27% 22% 20%
Dimension3 19% 13% 15%
Dimension4 11% 9% 10%
Total 98% 90% 91%

consumers were in two dimensional space. Two dimen-


I sions were considered restrictive, as it did not allow for
I 0
the possibility of consumers using completely different
DIMENSION 1 42%
dimensions.
FIG. 3. Internal preference map. In this paper, it was decided to treat each consumer’s
two dimensions as separate attributes. In this way the
concept of each consumer using completely different
O.’31 52 dimensions (attributes) was allowed for. Using this
6 21 rational, 16 attributes (8 x 2 consumer dimensions)
instead of two were submitted to a PCA analysis, allow-
ing a possible six dimensions to evolve. The authors felt
” I, that this was a much more realistic approach to obtain-
‘,( ,8 I
ing true product attribute dimensionally.
3 ,,”++ ,’ :
The explained variance for the first 4 dimensions is
7 :
given in Table 3.
-1 - a,!I 6
On average the principal components explain more
than 90% of the data after 4 components. Looking at the
sample structure (Figs 5-7) it is possible to recognise the
main structure in the sample space from the PCA of the
-2 , 12 I I profiling data. Dimension 1 separates sample 52 from the
-1 0 1 2
others, while dimension 2 has samples 12 and 21 as the
PC 1
extremes, see Figs 5-10. Dimension 2 is reversed in the
FIG. 4. Preference data superimposed on top of the scores from third replicate.
the first two dimensions of the PCA of the profiling data. Dimensions 3 and 4 are difficult to interpret, as they
are different from replicate to replicate. Similarities with
The structure in Fig. 4 is very similar to the internal dimensions 3 and 4 from the PCA of the profiling data do
preference map in Fig. 3 (PC 2 reversed). Again the exist, in that the dimensions separate samples 3 and 4
consumer preferences go in two directions, the first group and their reformulations. The different formulations of
along PC 1 comprises samples 31, 21 and 32 char- samples 3 and 4 are in orthogonal axes to each other in
acterised by sweetness, colour tone and clarity. The sec- the mapping exercise, while for dimension 3 and 4 in the
ond grouping along PC two is characterised by acidic profiling these are close together pairwise, although
flavour and intensity of blueberry smell and taste. When separated along dimensions 3 and 4 respectively.
these results were shown to product developers from one The results of the mapping suggest that there may well
of the companies involved, they were immediately able be some individual differences between the consumers,
to explain the observed differences in terms of the recipe beyond the major difference described by dimension one
variations made, verifying the findings. and two. To test this, a comparison of profiles and map-
ping sample spaces were calculated. The RV coefficients
Mapping structure for this are shown in Table 4.
It is clear that the best fit between the mapping space
One important aspect of the analysis of the mapping and profile space was on PC 1, particularly on the third
data that merits further discussion, is the fact that replicate. This suggests that the major differences
each consumer’s dimensions were considered as separate between the products are well agreed on. The other three
attributes in one multivariate analysis, rather than two dimensions show some relationship, but it would be dif-
attributes. The approach used in the paper on chocolate ficult to say with any degree of confidence what the
by Risvik et al. ( 1994) was to take each consumer’s similarities are.
two dimensional map, submit it to generalized Pro- Table 5 looks at the how similar the consumer’s two
crustes analysis (GPA) and observe how similar the dimensional maps were, for each of the three experimental
68 E. Risvik et al.

1 21

42
52 41
s
M c
cu *
8 12
ti

-1-
-1
I 33 42
0.2 1.2 I I
-1 0 1
PC1 41%
Pc319%
FIG. 5. Principal components 1 and 2 for projective mapping
by 8 consumers of 7 blueberry soups. First replicate. FIG. 8. Principal components 3 and 4 for projective mapping
by 8 consumers of 7 blueberry soups. First replicate.

21
1.5
31
0.6

42

8 0.5 0
2; 52
$ 0
41 41
52 ;: -0.2
z 12
0 n

-0.5 12
4232

I I -1.2
I 33 I
-1 0 1 -1 0 1
PC1 46% PC3 13%

FIG. 6. Principal components 1 and 2 for projective mapping FIG. 9. Principal components 3 and 4 for projective mapping
by 8 consumers of 7 blueberry soups. Second replicate. by 8 consumers of 7 blueberry soups. Second replicate.

12 2'

l-
0.5
42

52
o- 32
42
52
32

41

I 41 2' I 31
-1= I -1 cl I
-2 -1 0 1 -1 0 1
PCl46% PC3 15%

FIG. 7. Principal components 1 and 2 for projective mapping FIG. 10. Principal components 3 and 4 for projective mapping
by 8 consumers of 7 blueberry soups. Third replicate. by 8 consumers of 7 blueberry soups. Third replicate.

replicates. For each replicate, it can be observed that Mapping and preference
individual differences are often large, as represented by
the low RV coefficients. However, there are some simi- In order to determine if consumer preference was related
larities, for example consumers 4 and 6 have RVs of directly to the consensus mapping biplot, simple linear
greater than 0.6 over the three replicates. correlations between each consumer’s preference scores
Evaluation of Sensory Pro&g 69

TABLE 4. RV Coefficients for Profile and Mapping Dimen- On PC 1 of the profile biplot, it was apparent that the
sions sample positions correlated very highly to those on
dimension one of the mapping biplot on all 3 replicates
Replicate IPC 2 PCS 3 PCS 4 PCS
(p < 0.01). This was not the case for PC 2 ,3 and 4. It
1 0.694 0.559 0.647 0.626 would therefore seem, that whilst the consumers could
2 0.777 0.474 0.539 0.661 pull out the major sample differences in a similar way to
3 0.900 0.571 0.536 0.649
the trained profiling panel, they were not so comparable
on the smaller, or more subtle differences. Two con-
and dimensions one to four were calculated. The results sumers seem to have picked up the differences in dimen-
are presented in Table 6, which clearly show that there sion 3 which are differences in the taste and smell
were few significant correlations between these data sets. intensities and the colour tone.
The most obvious reason for this is that the consumers
had different perceptions of the samples, even though the
consensus mapping plot related well to the profile biplot.
DISCUSSION AND
This was confirmed in Table 5, where the RV coefficients
between consumers are shown. CONCLUSIONS

Mapping and conventional profiling


The results of this research illustrate that projective
It has already been commented that the profile biplots mapping can be a useful tool in combination with
and the mapping biplots were very similar on the preferences and descriptive profile data. Data from
first dimension/principal component. To measure projective mapping are simple to collect and analyse,
this similarity in more detail, the linear correlation and will therefore provide simple results for interpreta-
coefficients were calculated between each dimension/ tion of consumer perception of products in a product
PC on the profile biplot and the mapping biplot development process. However, there are several com-
(Table 7). ments to be made on the research.

TABLE 5. RV Coefficients between Consumers’ 2-D Maps, Raw Data Input

Replicate 1
Consumer 1 2 3 4 5 6 7 8

1 0.315
2 0.209
3 0.209 0.488
4 0.376 0.496 0.243
5 0.496 0.416 0.415 0.783
6 0.341 0.350 0.276 0.707 0.644
7 0.773 0.282 0.513 0.383 0.543 0.330
8 0.152 0.359 0.117 0.114 0.091 0.273 0.146
Replicate 2
Consumer 1 2 3 4 5 6 7 8

1
2 0.309
3 0.669 0.303
4 0.591 0.149 0.416
5 0.523 0.512 0.364 0.542
6 0.718 0.359 0.512 0.802 0.771
7 0.568 0.368 0.424 0.457 0.702 0.553
8 0.086 0.511 0.270 0.105 0.082 0.198 0.130

Replicate 3
Consumer 1 2 3 4 5 6 7 8

1
2 0.216
3 0.720 0.226
4 0.430 0.562 0.478
5 0.537 0.315 0.680 0.644
6 0.286 0.501 0.430 0.636 0.671
7 0.163 0.438 0.250 0.231 0.411 0.126
8 0.750 0.274 0.702 0.334 0.329 0.094 0.124
70 E. Risvik et al.

TABLE 6. Correlation Coefficients between Consumer Preference and the Mapping Biplot Dimensions for the Three Replicate
Mappings (Sign Level ofp < 0.10 = *)

Replicate 1
Consumer Dimension 1 Dimension 2 Dimension 3 Dimension 4

1 0.563 0.287 0.103 -0.154


2 0.570 0.493 -0.036 -0.387
3 0.085 a.517 0.723* -0.130
4 0.069 -0.540 -0.154 -0.331
5 0.860* 0.145 0.444 0.179
6 0.383 0.086 0.255 0.101
7 0.353 -0.598 0.029 0.478
8 0.125 -0.580 -0.134 -0.029

Replicate 2
Consumer Dimension 1 Dimension 2 Dimension 3 Dimension 4

1 0.642 -0.198 -0.386 0.524


2 0.487 0.298 +x750* 0.237
3 0.309 -0.448 a.018 0.101
4 -0.128 -0.413 -0.750* -0.253
5 0.939* 0.235 0.102 -0.026
6 0.622 -0.439 0.270 0.538
7 0.165 -0.225 -0.069 -0.402
8 -0.039 -0.529 -0.49 1 -0.155

Replicate 3
Consumer Dimension 1 Dimension 2 Dimension 3 Dimension 4

1 0.852* a.013 0.302 -0.058


2 0.478 -0.486 0.367 0.406
3 0.553 0.720* -0.229 0.309
4 0.086 0.292 -0.252 0.383
5 0.854* -0.376 -0.124 0.110
6 0.819* 0.293 0.209 -0.435
7 0.403 0.116 a.81 l* -0.063
8 0.281 0.383 a.402 0.072

TABLE 7. Correlation between Mapping and Profile Dimen- should be remembered that while the trained panel
sions (Sign Level ofp < 0.01*) were more discriminating, the results re-iterate ques-
tioning the complete relevance of the trained panel
ProWing dimensions
1 2 3 4 results in relation to consumer preference. The trained
Mapping
replicates panel is the essential link to the product development
team, however, the consumer is the best link to under-
1 AI.883* 0.630 0.787* -0.094
standing preferences. Thus, projective mapping may
2 -O.882* 0.292 0.402 a.098
3 x).949* -0.467 XI.367 -0.375 offer a link between preference and trained panel studies.
Another point to consider when interpreting these
results, is that the randomly selected consumers were not
chosen to represent a segment or target group. This gives
The use of scaling to unit variance for the projective rise to the possibility that heterogeneity between the
mapping data, but not the profiling data has been ques- consumers confounded information which may have
tioned. This is a perfectly valid point, but the following been obtained when looking at a specific consumer seg-
comment is made. It should be expected of a trained ment. Clearly, future research should address this issue,
panel that they use the scale in a way that represents together with looking at a more representative sample
similar scale use for all attributes. This cannot be con- size.
sidered to be the case with naive consumers, who used The authors consider that the projective mapping
two dimensions, comprising of several attributes, not procedure has a potential application with children,
specified. In the latter case scale standardisation is where less developed language abilities may be an
essential. obstacle to preventing traditional scaling exercises. The
The research illustrated, not surprisingly, that trained method could be turned into a fun game. Subsequent
panels and consumers perceived the samples in different data analyses will reveal aspects of importance for inter-
ways, though the major dimension was similar. It pretation of preference.
Evaluation of Sensory Profiling 71

Foods, ed. J. R. Piggott, pp. 381409. Elsevier Applied Sci-


ACKNOWLEDGEMENTS ence, London.
McEwan, J. (1995) Preference mapping for product optimiza-
tion. In Multivariate Analysis of Data in Sensory Science, eds T.
The authors are grateful to the sensory panel and staff at Nrees & E. Risvik. Elsevier,
MATFORSK and the consumers who participated in Meilgaard, M., Civille, G. V. & Carr, B. T. (1987) Sensory
the reported experiments. The authors would also like to Evaluation Techniques, I and II. CRC Press, Florida.
thank Pascal Schlich, INRA, Dijon, France for supplying Oppenheim, A. M. (1966) Questionnaire Design and Measurement.
a program for the RV coefficient. London, Heinemann,
Powers, J. J. (1984) Current Practices and Applications of
Descriptive Methods. In Sensory Analysis of Foods, ed. J. R.
Piggot, pp. 187-266. Elsevier Applied Science, London.
REFERENCES Risvik, E. (1985) Sensory Analysis. Tecator AB, Hoganas,
Sweden.
Risvik, E. (1994) Sensory properties and preferences. Meat
Amerine, M. A., Pangborn, R. M. & Roessler, E. B. (1965) Science 36, 67-77.
Principles of Sensory Evaluation of Food, eds G. F. Stewart et al., Risvik, E., McEwan, J. A., Colwill, J. S., Rogers, R. and Lyon,
pp. 108-l 14. Academic Press Inc., New York and London. D. H. (1994) Projective mapping: a tool for sensory analysis
ASTM (1968) Basic Prinsiples of Sensory Evaluation, Special and consumer research. Food @ali9 and Preference 5, 263-269.
Technical Publication No 433, American Society for Testing
Schiffman, S. S., Reynolds, M. & Young, F. W. (1981) Intro-
and Materials, Philadelphia.
duction to Multidimensional Scaling. Academic Press, New York.
ASTM (1968) Manual of Sensory Testing Methods, Special Tech-
Schlich, P. (1993) Contributions a la sensometrie. These prt-
nical Publication No 434, American Society for Testing and
sentee pour obtenir le grade de docteur en sciences de l’Uni-
Materials, Philadelphia.
versiti Paris XI Orsay.
Escoufier, R. P. and Robert, P. (1976) A unifying tool for linear
Schlich, P. and Guichard, E. (1989) Selection and classification
multivariante statistical methods: the RV-coefficient. Applied
of volatile compounds of apricot using the RV coefficient.
Statistics 25, 257-265.
Journal of Agriculture and Food Chemistry 37, 142-150.
Gabriel, K. R. (1971) The biplot graphic display of matrices
with application to principal component analysis. Biometrika Schlich, P. and McEwan, J. A. (1992) Cartographie des pre-
58, 453467.
feences: un outil statistique pour l’indusstrie agro-
IS0 6564 (1985) Sensory Analysis-Methodology Flavour alimentaire. Sciences des Aliments 12, 3399355.
profile. Schlich, P. (1995). Preference mapping: relating consumer
IS0 8587 (1988) Design of Test Rooms for Sensory Analysis of preferences to sensory or instrumental measurements. In
Food. BioJavour ‘95, eds Etievant, P. & Schreier, P. Analysis pre-
ISO-Draft DIS 11036 (1992) Sensory Analysis-Methodology- cursor Studies/Biotechnology, France.
Texture profile analysis. Stone (1974) Cross-validatory choice and assessment of statis-
Jellinek, G. (1985) Sensory Evaluation of Food: Theory and Practice, tical predictions. Journal of the Royal Statistics Society Bulletin
Ellis, Horwood, Chichester. 11-113.
Joliffe, I. T. (1986) Principal Components Analysis. Springer-Ver- Stone, H. & Sidel, J. L. (1985) Sensory Evaluation Practices, Aca-
lag, New York. demic Press, London.
Kruskal, J. B. & Wish, M. (1978) Multidimensional Scaling. Sage Williams, A. A. & Carter, C. S. (1977) A Language and Pro-
Publications, California. cedure for the Sensory Assessment of Cox’s Orange Pippin
MacFie, H. J. H. & Thomson, D. M. H. (1988) Preference Apples. Journal of the Science of Food and Agriculture 281,
Mapping and Multidimensional Scaling. In Sensory Analysis of 204214.

You might also like