Methodological Report

Guidance for Consumer Testing Data

Analysis & Reporting - Supplement to
Step 4
Understanding the Drivers of Trait Preferences and the Development of Multi-
user RTB Product Profiles, WP1, Step 4

Montpellier, France, October 2021

Geneviève FLIEDEL, Centre de coopération Internationale en Recherche Agronomique

pour le Développement (CIRAD), Montpellier, France
Isabelle MARAVAL, CIRAD, Montpellier, France
Aurélie BECHOFF, Natural Resources Institute, University of Greenwich, Chatham, UK
Eglantine FAUVELLE, CIRAD, Montpellier, France (Validator)
This report has been written in the framework of RTBfoods project.

To be cited as:

Geneviève FLIEDEL, Isabelle MARAVAL, Aurélie BECHOFF, Eglantine FAUVELLE,

(2022). Guidance for Consumer Testing Data Analysis & Reporting - Supplement to Step 4.
Understanding the Drivers of Trait Preferences and the Development of Multi-user RTB
Product Profiles, WP1, Step 4. Montpellier, France: RTBfoods Methodological Report, 50 p.
https://doi.org/10.18167/agritrop/00660

Ethics: The activities, which led to the production of this manual, were assessed and
approved by the CIRAD Ethics Committee (H2020 ethics self-assessment procedure).
When relevant, samples were prepared according to good hygiene and manufacturing
practices. When external participants were involved in an activity, they were priorly informed
about the objective of the activity and explained that their participation was entirely voluntary,
that they could stop the interview at any point and that their responses would be anonymous
and securely stored by the research team for research purposes. Written consent (signature)
was systematically sought from sensory panellists and from consumers participating in
activities.

Acknowledgments: This work was supported by the RTBfoods project

https://rtbfoods.cirad.fr, through a grant OPP1178942: Breeding RTB products for end user
preferences (RTBfoods), to the French Agricultural Research Centre for International
Development (CIRAD), Montpellier, France, by the Bill & Melinda Gates Foundation
(BMGF).

Image cover page © LAJOUS P. for RTBfoods.

Page 2 of 49
 This document has been reviewed by:

Aurélie BECHOFF 09/2020 and 12/01/2022

Final validation by:

Eglantine FAUVELLE 12/01/2022

Page 3 of 49
 CONTENTS
Table of Contents
1 Study context and general objectives ....................................................................................... 7
2 Methodology ............................................................................................................................ 7
2.1 Sampling .................................................................................................................... 7
2.2 Consumer testing ....................................................................................................... 7
2.3 Data analysis .............................................................................................................. 8
3 Results ..................................................................................................................................... 9
3.1 Overall liking of the product samples .......................................................................... 9
3.2 Segmentation of consumers into groups of similar overall liking ............................... 10
3.2.1 Demographic data of the consumers interviewed ..................................................... 11
3.2.2 Consumption attitudes .............................................................................................. 12
3.3 A Just About Right test (JAR) ................................................................................... 12
3.4 Check All That Apply (CATA) test ............................................................................. 14
3.5 Sensory mapping of the sensory characteristics ....................................................... 15
4 Discussion and conclusion ..................................................................................................... 16
5 Appendices: Tutorials for consumer testing data analysis ...................................................... 18
5.1 Annex 1 One-way ANOVA & multiple comparisons .................................................. 18
5.1.1 Selecting the dataset for running a one-way ANOVA................................................ 18
5.1.2 Conducting a one-way ANOVA................................................................................. 19
5.1.3 Interpreting the one-way ANOVA results .................................................................. 20
5.1.4 Concluding on the one-way ANOVA analysis ........................................................... 22
5.2 Annex 2 Agglomerative Hierarchical Clustering (AHC) ............................................. 23
5.2.1 What is Agglomerative Hierarchical Clustering? ....................................................... 23
5.2.2 Selecting the dataset to run an Agglomerative Hierarchical Clustering in XLSTAT ... 23
5.2.3 Setting up an Agglomerative Hierarchical Clustering ................................................ 23
5.2.4 Interpreting the results of an Agglomerative Hierarchical Clustering ......................... 25
5.2.5 Create a histogram chart .......................................................................................... 27
5.2.6 Calculating standard errors....................................................................................... 29
5.2.7 Adding standard errors to the histogram ................................................................... 30
5.3 Annex 3 Create a PivotTable to analyse JAR data ................................................... 31
5.3.1 Create a PivotTable .................................................................................................. 31
5.3.2 Building out your PivotTable ..................................................................................... 31
5.3.3 Display a value as a calculation and a percentage ................................................... 33
5.3.4 Create a PivotChart .................................................................................................. 35

Page 4 of 49
 5.3.5 Create a PivotChart .................................................................................................. 35
5.3.6 Create a Histogram Chart ......................................................................................... 36
5.4 Annex 4 Principal Component Analysis .................................................................... 39
5.4.1 What is Principal Component Analysis ..................................................................... 39
5.4.2 Dataset for running a Principal Component Analysis ................................................ 39
5.4.3 Setting up a Principal Component Analysis in Excel using XLSTAT ......................... 39
5.4.4 Principal Component Analysis in XLSTAT - launching the computations .................. 42
5.4.5 Interpreting the results of a Principal Component Analysis in Excel using XLSTAT .. 42

List of Tables
Table 1: Number of consumers interviewed in the rural and urban areas of the two regions .......... 7
Table 2: Quality characteristics identified during the previous activities 3 & 4 and selected for
building the CATA table .................................................................................................................. 8
Table 3: Mean overall liking scores for the four product samples tested ......................................... 9
Table 4: Demographic differences of the consumers with respect to cluster division .................... 11
Table 5: Percentage of consumers who scored the three specific sensory characteristics............ 12
Table 6: Frequency of citations of each quality characteristic by all the consumers ...................... 15

List of Figures
Figure 1: Clustering of the consumers based on their overall liking scores of the product ............ 10
Figure 2: Mean overall liking of the product samples by consumer cluster type (%) ..................... 10
Figure 3: Percentage of consumer cluster type by gender ........................................................... 12
Figure 4: Percentage of consumers who scored the three specific quality characteristics ............ 14
Figure 5: Mapping of the sensory characteristics and the overall liking of the product samples .... 16

Page 5 of 49
 ABSTRACT
This is a guidance document that gives a step-by-step description of the different analyses that
should be included for consumer testing data analysis and reporting, referred to as WP1 Step 4
within RTBfoods project. The objectives of this document are to provide country partners with a guide
and a template to follow when reporting on the activity. The main document is divided into Study
context; Methodology that includes sampling, consumer testing, data analysis; Results including
overall liking (9-point scale hedonic testing), segmentation (Ward cluster analysis), Just-About Right
(JAR), Check-all-that-applies (CATA), Sensory mapping (Principal Component Analysis (PCA)); and
Conclusions. In addition, there are appendices that give detailed guidance on the use of the XLStat
software to conduct the various statistical tests. This appendix is downloadable at
https://mel.cgiar.org/reporting/download/report_file_id/25496
Key Words: template, statistical analysis, methodology, consumer testing, roots, tubers and
bananas

Page 6 of 49
1 STUDY CONTEXT AND GENERAL OBJECTIVES
The main aim of this Step 4 “Consumer testing” is to understand the consumers’ demand for the
quality characteristics of Root, Tuber and Banana products.
Another aim is to provide WP2 with a clear and visual mapping of the most liked products associated
with high quality characteristics and high Overall liking scores, and of the least liked products
associated with low quality characteristics and low Overall liking scores.
The activity consists in inviting a large number of consumers to test the 4-5 products made in the
previous processing step from varieties with very different quality characteristics.

2 METHODOLOGY
2.1 Sampling
The 4-5 products made by the processors from varieties with very different quality characteristics
during the Step 3 “Processing diagnosis”, were tested by a large number of consumers.
Mention here the name of the cultivars / number of products / regions / urban and rural areas /
number of consumers (women and men) in each location and region …
It should be useful to include here a map of the different locations and a table with regions / areas /
villages, small towns, big cities / number of consumers / number of women / men interviewed.
Table 1: Number of consumers interviewed in the rural and urban areas of the two regions

Region 1 Region 2
Big
Total
Village Village Village Village Small Village Village Village Village Small city
1 2 3 4 town 1 2 3 4 town

Number of
Consumers

Women

Men

2.2 Consumer testing

A method including a hedonic test, a just-about-right (JAR) test, and a check-all-that-apply (CATA)
test was used. Consumers (n = 300) from different locations in rural and urban areas were asked
individually to look/touch/smell/taste each Product sample, one after the other, in a random order,
and score the overall liking using a nine-point hedonic scale (from 1. “Extremely dislike, to 9.
“Extremely like”).
Consumers were also asked to assess how they perceive the intensity of 2-4 characteristics
identified as important in the previous Activities 3 & 4, using the 3-point JAR “Just About Right” scale
(1 = “Too low”, too weak, not enough, 2= “Just About Right” and 3 = “Too high, too strong, too much”)
for each of the Product samples.
Mention here the JAR quality characteristics chosen. Explain why you made this choice?
Consumers were then asked to select the quality characteristics that better describe each Product
sample, among a list of 20-25 sensory characteristics -the most liked and the least liked collected
during the previous Activities 3 & 4- using a “Check-All-That-Apply” (CATA) approach. Finally,
consumers were invited to give their opinion and preferences on the Product samples.

Page 7 of 49
Propose a table with the 20-25 CATA quality characteristics. Mention what quality characteristics
were identified during Step 2, those identified during Step 3, and those identified during both
Activities by using three different colours. Explain why you made this choice.
Note that the CATA table should contain a balance of the most liked and the least liked quality
characteristics related to the appearance, odour, texture between fingers, taste, texture in mouth,
aroma, and aftertaste of the final products. The quality characteristics should be mainly sensory
characteristics rather than emotional characteristics to be useful for WP2. Complete the table below.
Table 2: Quality characteristics identified during the previous activities 3 & 4 and selected for building
the CATA table

Quality characteristics of
the ready to eat product
Appearance
-
-
odour
-
-
Texture when touching
-
-
List of the most liked
Taste
characteristics
-
-
Texture in mouth
-
-
Aroma
-
Aftertaste
-
Appearance
-
-
Odour
-
-
Texture when Touching
-
-
List of the least liked
Taste
characteristics
-
-
Texture in mouth
-
-
Aroma
-
Aftertaste
-

2.3 Data analysis

An analysis of variance (ANOVA) was carried out to identify significant differences in Overall liking
scores between the 4-5 Product samples as tested by 300 consumers. The region or gender effect
can be studied. Multiple pairwise comparisons were applied using the Tukey test, with a confidence
interval of 95% at p < 0.05 (n=300 consumers). For each Product sample, the number of consumers
who judged each specific characteristic either Just All Right (JAR), Too weak or Too strong was

Page 8 of 49
counted, and the percentage of consumers (out of 300) was determined. A Principal Component
Analysis (PCA) was used to describe the relationships between frequencies of citation of CATA
sensory characteristics and the mean Overall liking scores for each Product sample. All statistical
analyses were performed using XLSTAT 2019 software (Addinsoft).

3 RESULTS
3.1 Overall liking of the product samples
Please refer to the Excel file “Consumer testing WP1 Workshop Benin”.
The first worksheet entitled “Raw data” contains all the data collected during our Consumer testing
in Benin during the WP1 Workshop.
The Overall liking scores for each Product sample tested by consumers in Benin (here n=40
consumers, but you should have n=300 consumers in the two regions in your country) were extracted
from the “Raw data” worksheet and organized in one column in a new worksheet entitled "ANOVA
data", one product below the other, with the corresponding Overall liking scores in a second column.
The ANOVA analysis is conducted on another worksheet “ANOVA” using Overall liking scores as
dependent variable and Product samples as Qualitative explanatory variable, and using a Turkey
test with a confidence interval of 95% (see Tutorial in Appendix A: One-way ANOVA & multiple
comparisons).
See below the analysis of our data.
The overall liking of the product significantly differed between the four samples at a significant level
of p<0.05 (one-way ANOVA) (Table 1).
Table 3: Mean overall liking scores for the four product samples tested

Mean
Product
Overall liking scores* Groups**
Samples
(n consumers)
426 8.2 A
329 6.9 B
851 4.5 C
153 2.6 D
*Overall liking was rated on a nine-point scale from 1 = dislike extremely, to 9 = like extremely.
**Different letters correspond to the products, which are significantly different. Tukey test (p<0.05).
The most liked product samples were the 426 and the 329 samples with a mean overall liking score
close to 8 (like very much) and 7 (like moderately) respectively. The least liked was the 153 sample
with a mean overall liking score between 2 (dislike very much) and 3 (dislike moderately). The 851
sample got a medium score, between 4 (dislike slightly) and 5 (neither like nor dislike).
Comment on your results in relation with the varieties or the process used.
The same type of one-way ANOVA analysis can be used with another factor than the Product
samples – for example gender factor, or regions’ factor or rural/urban areas’ factor. But in these
cases, you need to use a one-way ANOVA analysis for each product separately to see if there is a
significant difference between men and women consumers, or between the two regions, or between
the rural areas (consumers interviewed in the 8 villages) and the urban areas (consumers in small
towns and in the big city) respectively, regarding the Overall liking of each product.

Page 9 of 49
 3.2 Segmentation of consumers into groups of
similar overall liking
The overall liking scores for each product tested by all the consumers were extracted from the “Raw
data” and organized in a new Excel worksheet entitled “AHC data”, with the first column for the
consumers and the four other columns for the Overall liking scores of the four Product samples (426,
851, 153, and 329). Please refer to the Excel file “Consumer testing WP1 Workshop Benin”).
The AHC analysis is conducted on the Overall liking scores of the four Product samples and will
appear in a new worksheet entitled “AHC” (see Tutorial in Appendix B: Agglomerative Hierarchical
Clustering).
The aim of an Agglomerative Hierarchical Clustering (AHC) analysis is to create homogeneous
clusters of consumers who have similar Overall liking scores. It is useful to classify consumers who
have been interviewed randomly, into similar groups.
In our example, by using an Agglomerative Hierarchical Clustering analysis of the mean overall liking
scores, we identified three groups of consumers that we have named “426 & 329 likers”, “426 likers”
and “153 dislikers”. These three clusters contained 25%, 15% and 60% of all the consumers
interviewed respectively.
There were significant differences (P < 0.001) in the overall liking of the three clusters (Figure 1 and
Figure 2).

Figure 1: Clustering of the consumers based on their overall liking scores of the product

Where: error bars represent the standard error.

Figure 2: Mean overall liking of the product samples by consumer cluster type (%)

Page 10 of 49
You can calculate the standard error (SE) of the mean Overall liking per cluster and per product (see
Tutorial in Appendix B: Agglomerative Hierarchical Clustering) by using a formula (calculation in a
new worksheet entitled “AHC standard error”).
You can also calculate the standard error (SE) as follows: calculate the standard deviation (STDEV)
using a PivotTable (see Tutorial in Appendix B: Agglomerative Hierarchical Clustering) and apply a
formula to calculate the SE (calculation in a new worksheet entitled “AHC SE-PivotTable).

3.2.1 Demographic data of the consumers interviewed

Among the n consumers interviewed, x% were women and y% were men. z% were 18-25 years
old… Most of them were employed as …
Comment if the three clusters differ in terms of sociological characters such as gender, age or level
of studies (Table 4).
Table 4: Demographic differences of the consumers with respect to cluster division

C1 C2 C3
Total 426 & 329 426 153
likers likers dislikers
Number of consumers
Gender (n)
Women (%)
Age Men (%)
18-25 years old (%)
26-35 years old (%)
36-45 years old (%)
Ethnicity 56-65 years old (%)
Fon (%)
Marital status Other
Single (%)
Married (%)
Occupation Widower (%)
Student (%)
Artisanship (%)
Civil servant (%)
Trading business (%)
Employed (%)
Education Unemployed (%)
No education (%)
Degree and less (%)
High level degree (%)
Consumption Graduated (%)
frequency Daily (%)
Several times a week
(%)
One time a week (%)
Several times a month
Consumption form (%)
One time a month (%)
Dry (%)
Added with water (%)
Added with water and
Occasion of ingredients (%)
consumption Sprinkled on beans (%)
Cooked into piron (%)
Breakfast (%)
Lunch (%)
In between meal (%)
Dinner (%)

Page 11 of 49
 3.2.2 Consumption attitudes
Most of consumers interviewed were used to consume the product daily (x% of answers). y% are
used to consume the product several times a week, … or once or several times a month (z%)…
The first form of consumption was …. (x% of answers), the second was …. (y%), and then...
The product is mainly consumed at lunch time for more than x% of consumers interviewed. Only y%
of interviewees consume it …

Percentage of
Gender Total
consumers
Class Women Men
C1 7.5% 17.5% 25.0%
C2 5.0% 10.0% 15.0%
C3 35.0% 25.0% 60.0%
Total 47.5% 52.5% 100.0%
Figure 3: Percentage of consumer cluster type by gender

47.5% of all the consumers interviewed were women and among them, 73.7% were 153 dislikers
(Cluster 3) and only 10.5% of them were 426 likers (Cluster 2). At the opposite, 47.6% of the men
out of a total of 52.5% of the consumers interviewed, were 153 dislikers (Clusters 3) and 33.3% were
426 & 329 likers (Cluster 2).
Using a PivotTable (see Tutorial in Appendix B: Create a PivotTable), allows you to present your
results and disaggregate the data by gender (as in the example above, refer to “AHC Class-Gender”
worksheet in the Excel file “Consumer testing WP1 Workshop Benin”).
You can also present your results by region, or by rural/urban areas (in villages or in the big city).

3.3 A Just About Right test (JAR)

Just about right (JAR) scale was used to determine the optimum level of intensity as perceived by
the consumers for some important sensory quality characteristics of the Product samples. Such
“descriptor diagnostic” may help understand why consumers like or dislike this Product sample.
Consumers were asked to give their perception of the Colour, Dryness and Sourness of each
Product sample, by using a 3-point JAR scale (1 = “Too low, too weak, not enough”, 2= “Just About
Right” and 3 = “Too high, too strong, too much”).
JAR data were extracted from the “Raw data” in the Excel file “Consumer testing WP1 Workshop
Benin”, and organized in a new worksheet “JAR data” with the first column for the consumers, the
second column for the four Product samples, one product below the other, and the three other
columns for the JAR scores of the three sensory characteristics “Colour”, “Dryness” and “Sourness”.
Using a PivotTable (refer to Appendix B: Create a PivotTable to analyse JAR data), allows you to
calculate (in a new worksheet entitled “JAR”) the percentage of consumers who scored each Product
sample as “Too low, too weak, not enough”, or as “JAR Just About Right”, or as “Too high, too strong,
too much”.
Table 5: Percentage of consumers who scored the three specific sensory characteristics
Colour
Product samples Too clear JAR Too dark
153 0.0% 7.7% 92.3%
329 18.9% 73.0% 8.1%
426 17.5% 80.0% 2.5%

Page 12 of 49
 851 7.5% 32.5% 60.0%
Dryness
Product samples Not enough dry JAR Too dry
153 16.7% 52.2% 31.1%
329 22.4% 70.2% 7.3%
426 1.3% 93.2% 5.5%
851 50.1% 46.2% 3.7%
Sourness
Product samples Not enough sour JAR Too much sour
153 25.0% 30.0% 45.0%
329 42.1% 47.4% 10.5%
426 17.5% 75.0% 7.5%
851 52.5% 32.5% 15.0%

You can now create a Pivot Chart or a Histogram (see Tutorial in Appendix B: Create a PivotTable
to analyse JAR data) to visualize the JAR data (in the worksheet “JAR”).

Page 13 of 49
Figure 4: Percentage of consumers who scored the three specific quality characteristics
A majority of consumers was satisfied with the three sensory characteristics of the 426 & 329 product
samples: Colour was scored “Just About Right” by 80 and 73% of consumers respectively, the
Dryness was scored JAR by 93.2 and 70.2% of consumers respectively, and the Sourness was also
scored JAR by 75 and 47.4% of consumers respectively.
The 851 & 153 product samples were perceived “Too dark” by 60 and 92.3% of consumers
respectively. Moreover, the 851 Product sample was found “Not enough dry” by 50.1% of
consumers, and “Not enough sour” by 52.5% of consumers. The 153 Product sample was scored
“Too much sour” by 45% of consumers. Its lower mean overall liking (2.5) could be explained mainly
because it was found “Too dark” and “Too much sour” by a majority of consumers.
Sourness was considered “Just About Right” by a high percentage of consumers (75%) only for the
426 Product sample, and by about 50% for the 329 Product sample.
You can also comment your results by gender by calculating the percentage of women/men who
scored the three specific quality characteristics.
You can also comment your results by region, or by rural/urban area (in villages or in the big city),
by calculating the percentage of consumers who scored the three specific quality characteristics in
these regions or areas.

3.4 Check All That Apply (CATA) test

The objective of the CATA test is to show the relationships between hedonic Overall liking scores
for each Product sample and the frequencies of citation of each CATA sensory characteristic by all
the consumers.
After scoring the Overall liking and the perception of some specific sensory characteristics,
consumers were invited to choose the most appropriate terms among 20-25 sensory characteristics
that better describe each Product sample.
CATA data were extracted from the “Raw data” in the Excel file “Consumer testing WP1 Workshop
Benin”, and organized in a new worksheet “CATA data” with the first column for the four Product
samples, one product below the other, the second column for the Overall liking scored by the
consumers for each product, and the other columns (here 20) for the citations of each quality
characteristic by the consumers to describe each Product sample.
The count of citations (in green) and the count of a mean Overall liking (in orange) per product and
per characteristic were calculated in a new worksheet entitled “CATA citation frequencies” in the
Excel file “Consumer testing WP1 Workshop Benin”.
Then a summary of the frequency of citations (in green) with a total of citations, and the mean Overall
liking (in orange) per product and per characteristic, were reported in a new worksheet entitled
“CATA 2”.

Page 14 of 49
The frequency of citations given by consumers to describe each Product sample were calculated
(Table 6).
The sensory characteristics most frequently cited by the consumers were considered the best for
describing the products. They were the following: “Fine” and “Dry” with a frequency of citation
between 100 and 150, followed by “Little Sour” and “White” with a frequency of citation between 75
and 100. The least used terms were “Coarse” and “No taste”.
The 426 Product sample was described as “White” by consumers (35 citations), “Fine” and “Dry” (32
citations respectively), “Little Sour” (26 citations), with a “Good Taste” and “Attractive” (36 and 34
citations respectively). Consumers used the same characteristics to describe the 329 Product
sample with almost the same frequencies of citation.
The 153 Product sample was qualified as “Brown” (35 citations), “Burnt” and “Bitter” (24 and 21
citations respectively. However, it was perceived “Dry” (36 citations) as the three other Product
samples.
The 851 Product sample ….
Table 6: Frequency of citations of each quality characteristic by all the consumers

Quality characteristics 426 851 153 329 Total

Fibers 2 13 10 11 36
Little sour 26 27 13 26 92
Sweet 14 3 4 9 30
Too acidic 1 2 13 4 20
Good taste 36 15 7 32 90
Beautiful 18 1 2 14 35
Fermented 8 6 16 10 40
Fermented odor 5 5 9 4 23
Bitter 0 2 21 1 24
Coarse 2 3 13 1 19
No taste 0 12 5 0 17
Yellow 2 39 2 0 43
White 35 0 3 36 74
Burnt 0 6 24 0 30
Brown 0 3 35 2 40
Attractive 34 7 5 29 75
Heavy 10 5 14 9 38
Fine 32 24 16 28 100
Dry 32 30 36 34 132
Mean overall liking 8.2 4.5 2.6 6.9

You can also comment your results by gender by calculating the percentage of women/men who
used these sensory terms to describe the Product samples.
You can also comment on your results by region or by rural/urban area (in villages or in the big city),
by calculating the percentage of consumers who used these sensory terms to describe the Product
samples in these regions or areas.

3.5 Sensory mapping of the sensory characteristics

Principal component analysis (PCA) was used to summarize the relationships between CATA
sensory characteristics, Product samples, and mean Overall liking of each product scored by all the
consumers.

Page 15 of 49
The PCA analysis is conducted on a new worksheet “PCA” in the Excel file “Raw data Consumer
testing Benin”.
In the Excel sheet “CATA data 2”, select the frequency of citations for all the quality characteristics
(in green) as the observations/variables, the Product samples (in blue) as the observation labels,
and the Mean Overall liking for each Gari sample (in orange) as a supplementary quantitative
variable (refer to Appendix D: Principal Component Analysis PCA).
The PCA plot explained 98.65% of the variance of the sensory characteristics, the first and second
axes accounting for 67.93% and 30.73% respectively. Most of the variance was explained by the
first axis.
The loading of sensory characteristics on PCA plan (Figure 5) shows that axis 1 was mainly
explained positively by the terms such as “White”, “Fine”, “Sweet” related to the most liked Product
samples (426 & 329) and negatively by the terms such as “Burnt”, “Coarse”, “Bitter”, “Fermented
odour”, “Brown”, and “Too acidic” related to the least liked Product sample (153) (refer to Squared
cosines of the variables in Appendix D: Principal Component Analysis).
Axis 2 was mainly explained positively by the terms such as “Yellow” and “Little Sour” related to the
851 Product sample, and negatively by the terms such as “Heavy”, “Dry” and “Fermented” (refer to
Squared cosines of the variables in Appendix D: Principal Component Analysis).
A high Mean Overall liking scored by consumers was related to the high quality characteristics such
as “White”, “Fine”, “Sweet”, “Good Taste (here on the right part of the PCA plan), which were
associated to the most liked Gari samples (426 & 329, made from good cassava varieties).
At the opposite, a low Mean Overall liking by the consumers were related to the low quality
characteristics such as “Burnt”, “Coarse”, “Bitter”, “Fermented odour”, “Brown”, and “Too acidic”
(here on the left part of the PCA plan), which were associated to the least liked Gari sample (153,
bad variety).

Figure 5: Mapping of the sensory characteristics and the overall liking of the product samples

You can also comment your results by gender, by regions or by rural/urban areas (in villages or in
the big city), by mapping the cited sensory characteristics and the overall liking of the product
samples scored by women/men, or consumers in each region or area respectively.

4 DISCUSSION AND CONCLUSION

Page 16 of 49
The four Product samples were perceived differently by consumers.
The least liked 153 Gari sample got the lowest mean overall liking score (2.5), mainly because it was
found “Too dark” and “Too much sour” by the Beninese consumers (JAR test).
The terms that better describe the product samples were …
The conclusion should be focused on the results to deliver to WP2.
You should pinpoint the high quality characteristics related to a high Mean Overall liking by the
consumers (here on the right part of the PCA plan), and associated to the most liked Product samples
(good varieties).
Conversely, you should pinpoint the low quality characteristics related to a low Mean Overall liking
by the consumers (here on the opposite left part of the PCA plan), and associated to the least liked
Product samples (low quality varieties).

Page 17 of 49
5 APPENDICES: TUTORIALS FOR CONSUMER
TESTING DATA ANALYSIS
5.1 Annex 1 One-way ANOVA & multiple
comparisons
This tutorial shows how to set up and interpret a one-way Analysis of Variance (ANOVA) followed
by Tukey’s HSD multiple comparisons test in Excel using the XLSTAT software.

5.1.1 Selecting the dataset for running a one-way ANOVA

The Excel sheet entitled “Raw data” correspond to the Consumer testing fieldwork conducted by the
WP1 partners during our WP1 workshop in April 2018, Cotonou Benin (see Excel file “Consumer
testing WP1 Workshop Benin”). During this fieldwork, four Gari products coded as follows: 426, 851,
153, 329 were tasted (one after the other) by 40 Beninese consumers who gave an overall liking
score for each of the products.
Using the ANOVA function of XLSTAT, we want to find out whether the four Gari samples differ
significantly in terms of their overall liking as scored by the 40 Beninese consumers and, if they do
differ, which product is the most liked. We use a one-way ANOVA because there is only one factor
- the Gari samples - and no repetition.
From the Excel sheet “Raw data”, create a new Excel sheet that you will name “ANOVA data”. The
first column will be used for the factor, also called Qualitative explanatory variable (here the “Gari
samples”), and the second column for the Dependent variable (here the "Overall liking”). The
dependent variable (here the "Overall liking”) measured during our Consumer testing fieldwork is
“dependent” on the independent variable also called factor or explanatory variable (here the “Gari
samples”).
Organize the “ANOVA data” as shown in the Excel file in two columns, with each Gari sample, one
below the other, in the first column of the Excel sheet, with the corresponding Overall-liking score in
the second column (from Raw data: column AA for 426 Gari sample, column AY for 851, BW for 153
and column CU for 329 Gari sample).
This one-way ANOVA analysis can be used with another factor than the Gari samples – for example
gender (men or women in the first column) or the regions or the rural/urban areas, and Overall
liking as a dependent variable (in the second column).
But if we want to know whether gender, regions, or rural/urban areas will affect the Overall liking,
then we have to use a one-way ANOVA analysis for each product separately to see if there is a
significant difference between men and women consumers, between the two regions, or between
rural areas (consumers interviewed in the 8 villages) and urban areas (consumers in small towns
and in the big city) respectively, regarding the Overall liking of each product.

Page 18 of 49
 5.1.2 Conducting a one-way ANOVA
Once XLSTAT is opened, select the XLSTAT / Modeling data / ANOVA command (see below).

Once you have clicked on the button, the ANOVA dialog box appears (see below).
In XLSTAT, it is possible to select the data in two different ways for the ANOVA (column or table).
We will choose the form of column (dialog box below), one column for the dependent variable,
another for the explanatory variable.
Click in the window Quantitative Dependent Variable, and select the data in the “ANOVA data”
worksheet as follows: Select the second column that contains the “Overall liking” scores
$B$1:$B$161.
Click in the window Qualitative Explanatory variable, and select the data in the “ANOVA data”
worksheet as follows: Select the first column that contains the Gari samples codes $A$1:$A$161.
While selecting the column title for both variables, also tick the option Variable labels (dialog box
below).
In this example, we want to display the results on another sheet “ANOVA” where the data are stored,
so we chose the Sheet option (dialog box below).

In the Outputs tab (Means sub-tab), select the Pairwise comparisons option to run a Tukey's
test.

Page 19 of 49
The computations begin once you have clicked on the OK button.

Once you have clicked on the Continue button, the computations resume and the results are
displayed in a new worksheet called “ANOVA”.

5.1.3 Interpreting the one-way ANOVA results

The first important results displayed by XLSTAT are the goodness of fit statistics.
The goodness of fit coefficients includes the coefficient of determination R², the adjusted R² and
several other statistical coefficients (see below).
The coefficient of determination R2 (here 0.563) tells how much of the variability of the modeled
variable (here the “Overall liking”) is being explained by the explanatory variables (here the “Gari
samples”); in our case, we have 56.3% of the variability of the Overall liking explained by the Gari
samples.
Regression of variable Overall liking

Goodness of fit statistics (Overall liking):

Observations 160.000
Sum of weights 160.000
DF 156.000
R² 0.563
Adjusted R² 0.555
MSE 3.788
RMSE 1.946
MAPE 60.725
DW 2.109
Cp 4.000
AIC 217.029
SBC 229.330
PC 0.459

Page 20 of 49
The next table down is the Analysis of variance. This is an important result to look at because it
tells whether there are significant differences between the Gari samples in terms of their Overall
liking (see below). The null hypothesis H0 is that there is no difference between the Gari samples. If
we reject the hypothesis H0 this means that there are significant differences in the mean Overall
liking between the Gari samples.
Analysis of variance (Overall liking)

Source DF Sum of squares Mean squares F Pr > F

Model 3 762.619 254.206 67.114 < 0.0001

Error 156 590.875 3.788
Corrected Total 159 1353.494
Computed against model Y=Mean(Y)

The test used here is the Fisher's F test. A probability Pr > F less than 0.0001 means that we would
take a 0.01% risk to conclude that the null hypothesis (no effect of the type of Gari samples on the
Overall liking) is wrong.
We can therefore conclude with confidence that there is a significant influence of the type of Gari
samples on the Overall liking as scored by the Beninese consumers.
The bar chart of the standardized coefficients shows the relative impact of the categories or
explanatory variables (here the Gari samples).

Now that we have obtained an answer to our initial question: is there a significant difference between
the Gari samples, the next step is to classify the difference between the samples.
As shown in the next table, the Tukey's HSD (Honestly Significantly Different) test is applied to all
pairwise differences between means. The analysis of the differences between the categories is done
with a confidence interval of 95% (this means that Pr > Diff <0.05 will indicate a significant
difference between the Gari samples).

Page 21 of 49
Gari samples / Tukey (HSD) / Analysis of the differences between the categories with a
confidence interval of 95% (Overall liking)

Contrast Difference Standardized difference Critical value Pr > Diff Significant

426 vs 153 5.675 13.041 2.597 < 0.0001 Yes
426 vs 851 3.750 8.617 2.597 < 0.0001 Yes
426 vs 329 1.350 3.102 2.597 0.012 Yes
329 vs 153 4.325 9.938 2.597 < 0.0001 Yes
329 vs 851 2.400 5.515 2.597 < 0.0001 Yes
851 vs 153 1.925 4.423 2.597 0.000 Yes
Tukey's d critical value: 3.673

Based on the p-values below (Pr > Diff), all the pairs appear to be significantly different. This can
also be confirmed by the 95% confidence intervals (last four columns). If an interval does not
contain zero, then we can reject the null hypothesis that there is no significant difference between
the two means.
The overall scores for the different Gari samples are then ordered in hierarchical order. Different
letters (A, B, C, D indicate significant differences in the mean Overall liking of the Gari samples). The
table gives a summary of all pairwise comparisons of Gari samples using the Tukey’s HSD test.
Summary of all pairwise comparisons for Gari samples (Tukey (HSD)

Category LS means
Groups
Gari samples Overall liking
426 8.250 A
329 6.900 B
851 4.500 C
153 2.575 D

Other multiple comparison tests can be applied such as the Dunnett's test that compares each
category with a control category.

5.1.4 Concluding on the one-way ANOVA analysis

The conclusion is that the four Gari samples significantly differ in terms of their Overall liking as
scored by the Beninese consumers.

Page 22 of 49
 5.2 Annex 2 Agglomerative Hierarchical Clustering
(AHC)
This tutorial will help you set up and interpret an Agglomerative Hierarchical Clustering (AHC) in
Excel using the XLSTAT software.

5.2.1 What is Agglomerative Hierarchical Clustering?

Agglomerative Hierarchical Clustering (AHC) is a classification method whose principle is simple.
The process starts by calculating the dissimilarity between the N objects. When two objects minimize
a given agglomeration criterion, they are clustered together, thus creating a class comprising these
two objects. Then the dissimilarity between this class and the N-2 other objects is calculated using
the agglomeration criterion. When the two objects or classes of objects minimize the agglomeration
criterion, they are then clustered together. This process continues until all the objects have been
clustered.
These successive clustering operations produce a binary clustering tree (dendrogram), whose root
is the class that contains all the observations. This dendrogram represents a hierarchy of partitions.

5.2.2 Selecting the dataset to run an Agglomerative Hierarchical

Clustering in XLSTAT
From the Excel sheet “Raw data”, create a new Excel sheet that you will name “AHC data”. The first
column will be used for the consumers and the four columns for the Overall liking of the four Gari
samples (426, 851, 153, and 329).
The aim of AHC is to create homogeneous clusters of consumers who have similar Overall liking
scores.

5.2.3 Setting up an Agglomerative Hierarchical Clustering

Once XLSTAT is activated, go to XLSTAT / Analysing data / Agglomerative Hierarchical
Clustering.

The Hierarchical Clustering dialog box will appear. Then select the data on the Excel sheet.

Page 23 of 49
There are several ways of selecting data with XLSTAT. In our example, we will select columns. This
explains why the columns are labelled.
Click in the window Observations/variables table, and select the data in the “AHC data” worksheet
as follows: Select the columns B, C, D, E containing the “Overall liking” scores for the four Gari
samples $B$1:$E$41.
Click in the window Row labels, and select the data in the “AHC data” worksheet as follows: Select
the first column that contains the Consumers $A$1:$A$41.

In the Options tab, select the Automatic - Entropy truncation. The automatic truncation is based
on the entropy and tries to create homogeneous groups of consumers having similar Overall liking
of the Gari samples.

The computations begin once you have clicked on OK.

Page 24 of 49
Once you have clicked on the Continue button, the computations resume and the results are
displayed in a new worksheet called “AHC”.

5.2.4 Interpreting the results of an Agglomerative Hierarchical

Clustering
The first result to look at is the levels bar chart. The shape reveals a great deal about the structure
of the data. When the increase in dissimilarity level is strong, we have reached a level where we are
grouping groups that are already homogenous. Automatic truncation uses this criterion to decide
when to stop aggregating observations (or groups of observations, here the consumers).

The chart below is the dendrogram. It represents how the algorithm works to group the observations
(the consumers), then the sub groups of observations. As you can see, the algorithm has
successfully grouped all the observations. The dotted line represents the automatic truncation,
leading to three groups (three Clusters).

Page 25 of 49
 C3 C1 C2

The three groups do not necessarily have the same size. In this example, the C3 cluster (displayed
in blue colour, 24 consumers) is more homogeneous (it is flatter on the dendrogram) than the C1
cluster (displayed in red colour, 10 consumers), and the C2 cluster (displayed in green colour, 6
consumers).
This is confirmed when looking at the Within-class variance. C1 and C2 exhibit higher variance
than C3 (see the table below).
Results by class

Class C1 C2 C3
Objects (40 consumers) 10 6 24
Sum of weights 10 6 24
Within-class variance 11.044 21.967 4.402
Minimum distance to centroid 1.364 2.804 0.604
Average distance to centroid 2.804 4.128 1.868
Maximum distance to centroid 6.623 6.287 5.036

The following table shows how the consumers (here the 40 consumers interviewed in Benin, April
2018) have been classified into each cluster (C1, C2 or C3).
Results by object

Observations
Class
(Consumers)
1 1
2 2
3 3
4 3
5 3
6 3
7 1
8 3
9 1
10 2
11 3
12 3
13 1
14 1
15 3

Page 26 of 49
 16 3
17 3
18 3
19 2
20 3
21 3
22 1
23 3
24 3
25 1
26 3
27 3
28 3
29 2
30 3
31 2
32 3
33 1
34 2
35 3
36 1
37 3
38 1
39 3
40 3

5.2.5 Create a histogram chart

The means of the Overall liking per cluster and per Gari sample are given in the following table
(displayed in the Excel sheet “AHC”).
From the class centroid table (here below), you can create a histogram chart to visualize the data.
1. Label the clusters C1, C2, C3 in the “Class” column
2. Select your dataset table
3. Click Insert > Chart.
4. In the Insert Chart dialog box, under All Charts, click Histogram, and click OK.
Class centroids

426-Overall 851-Overall 153-Overall 329-Overall

Class
liking liking liking liking

C1 (10) 7.300 1.800 1.300 6.200

C2 (6) 7.833 3.333 4.833 3.833

C3 (24) 8.750 5.917 2.542 7.958

Page 27 of 49
 Chart Title
10,000
9,000
8,000
7,000
6,000
5,000
4,000
3,000
2,000
1,000
0,000
426 -Overall liking 851-Overall liking 153 -Overall liking 329-Overall likng

C1 (10) C2 (6) C3 (24)

Now you can customize this chart by selecting:

- Design, Switch row/column, Ok

Page 28 of 49
 - Delete the grids by clicking on them and press the “delete” button
- Delete the “Chart Title” by clicking on it and press the “delete”
button
- Format the legend: “Legend format”, “Right”
- Format the y-axis: “y-axis”, “format axis
- “Axis options”, bounds, minimum 1, maximum 9
- “Number”, decimal places 0
- Click on the graph, click on the + on the right, title of the axes
- Insert “connectors” “line” to create a line and move it midway on
the graph at y-axis = 5 (neither like nor dislike)
- Name the clusters based on the Overall liking of the Gari samples
as in the example below and calculate the percentage of the
consumers in each cluster.

9
8
Mean overall liking

7
6 426
5 851
4 153

3 329

2
1
C1(25%) C2(15%) C3(60%)
426 & 329 likers 426 likers 153 dislikers

Consumer clusters

5.2.6 Calculating standard errors

Create a new Excel sheet entitled “AHC Standard error”. Add the column “Class” found in the “AHC”
worksheet (and here above) to the five columns of the “AHC data” worksheet.
Organize the data in the Excel sheet as follows:
- Click on “Sort” the data
- Select “Custom sorting”, Expand selection
- Sort on “Class” data, on the Cell value, from the smallest to the biggest.

Page 29 of 49
Calculate the standard error (SE) of the Mean Overall liking for each Gari sample scored by the
Consumers in each cluster.
The standard error of the mean (SE) can be expressed as the standard deviation (STDEV) of the
Mean Overall liking divided by the square root of the number of consumers in each cluster.
SE = STDEV /√(n)
Where n is the number of consumers in each cluster: n (C1) = 10; n (C2) = 6; n (C3) = 24.
SE Class 426 851 153 329
C1(10)
0.78951462 0.35901099 0.21343747 0.55377492
426 & 329 likers
C2(6)
0.47726070 0.91893658 1.13773654 1.13773654
426 likers
C3(24)
0.10851434 0.25478512 0.27570178 0.17527584
153 dislikers

5.2.7 Adding standard errors to the histogram

- Click on a histogram bar (ex: blue bar, Gari
sample 426)
- Click on “Design”, “Add a chart element”
- Select “Error bar”
- Click on “More error bar options”
- Select “Custom”, “Specify a value”
- For the positive standard error of each Gari
sample (e.g. code 426, blue bar), select the
data in the column corresponding to the Gari
sample 426 in the table SE above
- For the negative standard error, select the
same data in column corresponding to the Gari
sample 426 in the table SE above
- Click on “Ok”
- Repeat the same procedure for each histogram
bar (each Gari sample).

Page 30 of 49
 5.3 Annex 3 Create a PivotTable to analyse JAR data
A PivotTable is a powerful tool to calculate, summarize, and analyse data that lets you see
comparisons, patterns, and trends in your data.

5.3.1 Create a PivotTable

1. From the Excel sheet “Raw data”, create a new Excel sheet entitled “JAR data”. The first
column will be used for the Consumers, the second column for the Gari samples, and then
three columns for the JAR scores of the three selected descriptors (Colour, Dryness,
Sourness).
Organize the “JAR data” as shown in the Excel sheet “JAR data” with each Gari sample, one
below the other in the second column, and the corresponding Consumers and JAR scores
for the three descriptors.

2. Select the dataset “JAR data” you want to create a Pivot Table from.
Note: Your data should not have any empty rows or columns. It must have only a single-row
heading.
3. Select in Excel, Insert > PivotTable.

4. Under Choose the data that you want to analyse, select a table or range.

5. In Table/Range box, select all the data from “JAR data”: $A$1:$E$161.
6. Under Choose where you want the PivotTable report to be placed, select New
worksheet to place the PivotTable in a new worksheet.
7. Select OK.

5.3.2 Building out your PivotTable

The Field List has a field section in which you pick the fields you want to show in your PivotTable,
and the Areas section (at the bottom) in which you can arrange those fields the way you want.

Page 31 of 49
Use the areas section (at the bottom) of the Field List to arrange fields the way you want by dragging
them from the Field list to the target area. To delete a field from the PivotTable, drag the field out of
its areas section.
Fields that you place in different areas are shown in the PivotTable as follows:
• Columns area fields are shown as Column Labels at the top of the PivotTable.
• Rows area fields are shown as Row Labels on the left side of the PivotTable
Values area fields are shown as summarized numeric values in the PivotTable
NOTE: Typically, nonnumeric fields are added to the Rows area (here Gari samples), numeric fields
are added to the Values area (here Consumers), and Online Analytical Processing (OLAP) (here
JAR sensory descriptors) are added to the Columns area, one after the other.

The Field List should appear when you click anywhere in the PivotTable.

By default, PivotTable fields placed in the Values area are displayed in SUM format: Consumers
appear as Sum of Consumers and not as Number of Consumers. You can change the default
calculation by clicking on the field name “Sum of Consumers” in the Values area, then selecting the
Value Field Settings, and in the “Summarize values by” tab, clicking on Count, and Ok.

Page 32 of 49
 5.3.3 Display a value as a calculation and a percentage
Instead of using a calculation to summarize the data, you can also display it as a percentage of a
field.
In our example, we propose to change the Count of Consumers to show them as a % of the total of
Consumers.

Page 33 of 49
Once you have opened the Value Field Settings dialog box, go to “Show
values as” tab.
Choose the type of calculation you want to summarize the data in the
selected field
“Show values as” option, and here select “% of row total”.
You can remove the “blank column” by clicking on “Column labels” and
unticking the “blank column” box.

The PivotTable appears with the percentage of consumers who have chosen one of the three options
(1= Too clear, 2 = JAR Just About Right as I like, 3 = Too dark) to assess the Colour descriptor for
each Gari sample.

Select the pivot table and copy it with “Paste special

“Values” into a new worksheet. The table is no longer active
when you paste it with “Paste special”.

You can therefore rename the row and column labels in the PivotTable (as shown below). The next
step is the creation of a histogram to visualize the data.

Page 34 of 49
 5.3.4 Create a PivotChart
You have two possibilities: create a PivotChart or Insert a Histogram.

5.3.5 Create a PivotChart

1. Select a cell in your table.

2. Select Insert > PivotChart

3. Select Bars
4. Select OK.

Page 35 of 49
You can change the colour of the bars, add the
values of % of consumers in each bar, remove the
X-Axis legend, remove the grid… by right-clicking in
the chart.

5.3.6 Create a Histogram Chart

1. Select the pivot table and copy it with “Paste special “Values” beside in the worksheet JAR.
2. Delete the Grand Total and the column labels.
3. Select the data in your table.
4. Select Insert > Chart
5. Select Histogram 2D bar
6. Select OK.

Page 36 of 49
Now you can customize this chart by selecting:
- Click on Y-Axis scale on the Chart
- Design, Select Data, Modify

- Click on Modify
- Axis Label, Select Gari samples codes as shown below
- Click on OK.

- Delete the grids by clicking on them and press the “delete” button
- Right Click on the Legend in the Chart, “Legend format”, “Right”
- Right Click on the x-axis scale in the Chart and press the “delete” button
- Click on the graph, click on the + on the right, and add Title of the axes
- Click on the graph, click on the + on the right, and add data labels on the bar

Page 37 of 49
- You can choose a title for the Chart and for x-axis and for y-axis
- You can modify the colour of each bar by right clicking on the bar and change the colour

Page 38 of 49
 5.4 Annex 4 Principal Component Analysis
This tutorial will help you set up and interpret a Principal Component Analysis (PCA) in Excel using
the XLSTAT software.

5.4.1 What is Principal Component Analysis

Principal Component Analysis is one of the most frequently used multivariate data analysis. It
investigates multidimensional datasets with quantitative variables.
Principal Component Analysis is a very useful projection method to analyse numerical data
structured in n observations / p variables. PCA projects observations from a P-dimensional space
with p variables to a K-dimensional space (where K<P) so as to conserve the maximum amount of
information (information is measured here through the total variance of the scatter plots) from the
initial dimensions. It allows to:
• Quickly visualize and analyse correlations between the p variables,
• Visualize and analyse the n observations (initially described by the p variables) in a 2- or 3-
dimensional map, in order to identify uniform or atypical groups of observations.
PCA dimensions are also called axes or Factors.
For example, if a table of n observations (here the four Gari samples) are described by p variables
(here the CATA quality characteristics or CATA descriptors), and if p is quite high, it is impossible to
grasp the structure of the data and the nearness of the observations by merely using univariate
statistical analysis methods or even a correlation matrix.
If the information associated with the first 2 or 3 axes represents a sufficient percentage of the total
variability of the scatter plot, the observations will be able to be represented on a 2-3-dimensional
chart, thus making interpretation much easier.

5.4.2 Dataset for running a Principal Component Analysis

Open two new worksheets in your Excel file named “CATA data” and “CATA citation frequencies”
respectively. Copy from your “Raw data” all the CATA data for the four Gari samples and paste them
in these two new worksheets.
In “CATA citation frequencies” worksheet, calculate the Mean overall liking for each Gari sample
(orange line) and the Count of the citations per quality characteristic and per Gari sample (green
lines). Alternatively, you can create a Pivot table to calculate those frequencies from “CATA data”.
Copy these summarized data (Mean overall liking and Count of citations per descriptor for each Gari
sample) in a new sheet “CATA data 2”.

5.4.3 Setting up a Principal Component Analysis in Excel using

XLSTAT

Selecting the data

Once XLSTAT is activated, select the XLSTAT / Analysing data / Principal component analysis
command (see below).

Page 39 of 49
The Principal Component Analysis dialog box will appear.
Select the data in the Excel sheet entitled “CATA data 2”.
The Data format chosen is Observations/variables.
Click in the Observations/variables window and select in the Excel sheet “CATA data 2”, the data of
citation frequencies for all the quality characteristics (in green).
Click in the Observation labels window and select the Products (Gari samples) labels (column in
blue) in the Excel sheet “CATA data 2”.
Sheet: Activate this option to display the results in a new worksheet.

Principal Component Analysis: what type to choose - Pearson or covariance

The PCA type to select here is “Covariance”. Covariance is used when we want to analyse the
variance of the variables (here the CATA quality characteristics). Covariance matrices allocate more
weight to variables with higher variances.

Principal Component Analysis: supplementary variable

XLSTAT allows for additional variables (qualitative or quantitative ones) called supplementary
variables. This can be used in several contexts. Here are two examples:
• If you want to investigate how a quantitative dependent variable relates to the others (i.e.
independent variables) that should be used to build the PCA.
• If you want to see how different categories of observations behave in the PCA space (Men
vs Women for example). In this case, a qualitative supplementary variable (gender) may be
used to colour observations differently for men and women.

Page 40 of 49
In our example, choose a supplementary quantitative variable in the Supplementary data tab (here
below): The Mean Overall liking for each Gari sample (column in orange) in the Excel sheet “CATA
data 2”.

Principal Component Analysis in XLSTAT, configuring charts

In the Charts tab, in order to display the labels on all charts, and to display all the observations
(observations charts and biplots), the filtering option is unchecked.
It is possible to simultaneously represent both observations and variables in the factor space. The
term biplot is reserved for simultaneous representations which respect the fact that the projection of
observations on variable vectors must respect the order and the relative distances of the
observations for that same variable, in the input data.

For the graphic representation, a transformation is required in order to make the interpretation
precise.
Distance biplot: A distance biplot is used to interpret the distances between the observations. The
position of two observations projected onto a variable vector can be used to determine their relative
level for this variable. Lastly, the length of a variable vector in the representation space is
representative of the variable's level of contribution to building this space.

Page 41 of 49
The computations begin once you have clicked on OK. You are asked to confirm the number of rows
and columns. Then Continue

5.4.4 Principal Component Analysis in XLSTAT - launching the

computations
Then you should confirm the axes for which you want to display plots. In this example, the percentage
of variability represented by the first two factors is 98.65%.

5.4.5 Interpreting the results of a Principal Component Analysis in

Excel using XLSTAT

How to interpret Eigenvalues in Principal Component Analysis

This table and the corresponding chart are related to a mathematical object, the eigenvalues, which
reflect the quality of the projection of the data. In our example, we can see that the first eigenvalue
equals 1025.6 and represents 67.9 % of the total variability. This means that if we represent the data
on only one axis, we will still be able to see 67.9 % of the total variability of the data. With two axes,
we will see 98.7 % of the total variability, which is very good.

Page 42 of 49
Each eigenvalue corresponds to a factor, and each factor to a one dimension. The eigenvalues and
the corresponding factors are sorted by descending order of how much of the initial variability they
represent (converted to %).
Ideally, the first two or three eigenvalues will correspond to a high % of the variance, ensuring us
that the maps based on the first two or three factors are a good quality projection of the data. In this
example, the first two factors allow us to represent 98.7 % of the initial variability of the data. This is
a very good result, but we'll have to be careful when we interpret the maps as some information
might be hidden in the next factors.
Broadly speaking, factor = PCA dimension = PCA axis
Eigenvalues

F1 F2 F3
Eigenvalue 1025.594 463.972 20.309
Variability (%) 67.926 30.729 1.345
Cumulative % 67.926 98.655 100.000

How to interpret results related to variables in PCA

The first map is called the correlation circle or variable chart (below on axes F1 and F2). It shows
a projection of the initial variables in the factor space. When two variables are far from the centre,
then, if they are: close to each other, they are significantly positively correlated (r close to 1); if they
are orthogonal, they are not correlated (r close to 0); if they are on the opposite side of the centre,
then they are significantly negatively correlated (r close to -1). Supplementary variables can also be
displayed in the shape of vectors.
When the variables are close to the centre, some information is carried on other axes by looking at
the correlation circle on axes F1 and F3.

Page 43 of 49
The correlation circle is useful in interpreting the meaning of the axes. These trends will be helpful
in interpreting the next map.
Contributions: Contributions are an interpretation aid. The variables which had the highest influence
in building the corresponding PCA axes are those whose contributions are highest.
To confirm that a variable is well linked with an axis, take a look at the squared cosines table: the
greater the squared cosine, the greater the link with the corresponding axis. The closer the squared
cosine of a given variable is to zero, the more careful you have to be when interpreting the results in
terms of trends on the corresponding axis. Looking at this table we can see that the variable “Fibres”
would be best viewed on a F1/F3 map.

Page 44 of 49
Contribution of the variables (%)

F1 F2 F3
Fibres 0.573 0.484 46.145
Little sour 1.578 3.692 0.891
Sweet 1.512 0.308 11.406
Too acidic 1.112 2.281 2.509
Good taste 13.838 0.008 1.427
Beautiful 4.874 0.813 4.546
Fermented 0.321 2.259 1.108
Fermented odour 0.221 0.253 1.227
Bitter 4.608 6.065 0.494
Coarse 1.501 1.565 2.605
No taste 1.247 2.451 0.120
Yellow 4.115 48.273 2.520
White 25.775 5.164 9.568
Burnt 7.403 4.357 3.002
Brown 12.518 17.447 0.808
Attractive 15.675 0.991 4.081
Heavy 0.046 2.053 1.241
Fine 3.053 0.598 4.497
Dry 0.027 0.939 1.803

Squared cosines reflect the representation quality of a variable on a PCA axis. Squared cosine
analysis is used to avoid interpretation errors due to projection effects. If the squared cosines of a
variable associated to an axis is low, the position of the variable on this axis should not be interpreted.

Page 45 of 49
Squared cosines of the variables

F1 F2 F3
Fibres 0.336 0.128 0.536
Little sour 0.483 0.511 0.005
Sweet 0.805 0.074 0.120
Too acidic 0.507 0.470 0.023
Good taste 0.998 0.000 0.002
Beautiful 0.914 0.069 0.017
Fermented 0.235 0.749 0.016
Fermented odour 0.615 0.318 0.068
Bitter 0.626 0.373 0.001
Coarse 0.664 0.313 0.023
No taste 0.529 0.470 0.001
Yellow 0.158 0.840 0.002
White 0.911 0.083 0.007
Burnt 0.785 0.209 0.006
Brown 0.613 0.386 0.001
Attractive 0.967 0.028 0.005
Heavy 0.046 0.929 0.025
Fine 0.895 0.079 0.026
Dry 0.055 0.871 0.073
Mean overall liking 0.969 0.010 0.021

Values in bold correspond for each variable to the factor for which the squared cosine is the largest
Factor scores: Activate to display the coordinates of the observations (factor scores) in the new
space created by PCA.
Factor scores

Observations F1 F2 F3
426 34.438 -3.925 -6.064
851 -20.988 34.535 0.010
153 -41.261 -24.775 -0.589
329 27.811 -5.835 6.642

Factor scores are the observations coordinates on the PCA dimensions. They are displayed in a
XLSTAT table.
As for the results related to variables, XLSTAT displays contributions of observations (i.e. their
contribution in building the PCA axes) as well as squared cosines (i.e. their representation quality
on the different axes).

How to interpret results related to observations in PCA

The observations charts represent the observations in the PCA space. It enables you to look at
the observations on a two-dimensional map, and to identify trends.

Page 46 of 49
It is also possible to display biplots, which are simultaneous representations of variables and
observations in the PCA space. Here as well the supplementary variable (here Mean overall liking)
can be plotted in the form of vectors.

Contributions: The table below shows the contributions of the observations in building the principal
components.
Squared cosines: The table below displays the squared cosines between the observation vectors
and the factor axes.

Page 47 of 49
Contribution of the observations (%)

F1 F2 F3
426 28.909 0.830 45.261
851 10.737 64.263 0.000
153 41.500 33.073 0.427
329 18.854 1.834 54.312

Squared cosines of the observations

F1 F2 F3
426 0.958 0.012 0.030
851 0.270 0.730 0.000
153 0.735 0.265 0.000
329 0.908 0.040 0.052
Values in bold correspond for each observation to the factor for which the squared cosine is the
largest.

Page 48 of 49
Institute: Cirad – UMR QualiSud
Address: C/O Cathy Méjean, TA-B95/15 - 73 rue Jean-François Breton - 34398
Montpellier Cedex 5 - France
Contact Tel: +33 4 67 61 44 31
Email: [email protected]
Website: https://rtbfoods.cirad.fr/

Page 49 of 49