Multiple Factor Analysis: main features and

application to sensory data

Jérôme Pagès
AGROCAMPUS Rennes - 65 rue de St Brieuc - CS 84215 - F 35042 Rennes Cédex
[email protected]

Abstract. Data table in which a single set of individuals is described by several

groups of variables are frequently encountered. In the factor analysis framework,
taking into account different groups of variables in a unique analysis firstly raises
the problem of balancing the different group. This problem being solved, beyond
classical outputs from factor analysis, it is necessary to have at one’s disposal
specific tools in order to compare the structure upon individuals induced by the
different groups of variables. That is the aim of Multiple Factor Analysis (MFA),
factor analysis devoted to such data table. This paper presents the method, its main
properties and an application to sensory data.
Keywords. Factor analysis, principal components analysis, canonical analysis

1. Data table: denotations, examples

Multiple Factor Analysis (MFA; Escofier & Pagès 1988-1998 ; Pagès 2002) deals
with data table in which a set of individuals is described by several sets of
variables. Within one set, variables must belong to the same type (quantitative or
categorical) but, even for active ones, groups of variables can belong to different
types. We focus the hereafter presentation on quantitative variables, with only
some comments about qualitative ones.

sets 1 j J
variables 1 K1 1 k Kj 1 KJ
1

individuals i xik

I
Figure 1. Data table.
xik : value of variable k for individual i. If k is a continue variable, xik is a real
number ; if k is a categorical variable, xik is a number of category. The jth set is
denoted by j or Kj.
If we consider the whole table: individuals are denoted i (i =1, I); they constitute
the cloud NI lying in the K-dimensional space RK ; the K variables constitute the
cloud NK lying in the I-dimensional space RI.
In we consider the only (sub-)table j : individuals are denoted ij (i =1, I); they
constitute the cloud NIj lying in the Kj-dimensional space RKj ; the Kj variables
constitute the cloud NKj lying in the I-dimensional space RI.
Examples
Survey. An individual is a person ; a variable is a question. Questions are gathered
according to the different themes of the questionnaire. Each theme defines one set.
Sensory analysis. An individual is a food product. A first set of variables includes
sensory variables (sweetness, bitterness, etc.); a second one includes chemical
variables (pH, glucose rate, etc.).
Ecology. An individual is a observation place. A first set of variables describes soil
characteristics ; a second one describes flora.
Times series. Several individuals are observed at different dates. In such a case,
there is often two ways of defining sets of variables : generally, each set gathers
variables observed at one date ; but, when variables are the same from one date to
the other, each set can gather the different dates for one variable.

2. General factor analysis : denotations and main

relationships
General Factor Analysis (GFA) is here understood according to Lebart et al
(1997). Main features and denotations of GFA can be summed up by 6 points.
1) Given a table X having I rows and K columns, two clouds are considered : the
one of rows (NI lying in RK) ; the one of columns (NK lying in RI) ; to simplify this
short synthesis, weights of individuals and weights of variables are supposed
uniform.
2) The maximum inertia directions of NI and NK are highlighted : let us (resp. zs) a
unit vector along the rank-s principal direction of NI (resp. NK) in RK (resp. RI).
These vectors satisfy (λs being the rank s eigenvalue of X’X):
X ′Xus = λs us us = 1 XX ′z s = λs zs zs = 1
3) NI and NK are projected onto their maximum inertia directions us and zs; co-
ordinates of NI (resp. NK) onto axis s constitute the I-factor of rank s (resp. K-
factor) denoted Fs (resp. Gs):
Fs = Xus Gs = X ′z s
In PCA, Fs is named principal component.
4) The inertia directions of NI and NK are related (that is named duality) by:

2
 Fs F Gs G
zs = = s us = = s
Fs λs Gs λs

zs is often named standardised I-factor (in PCA: standardised principal

component).
5) The projection of row i (resp. column k) onto rank s axis in RK (resp. RI) can be
calculated from the co-ordinates of NK (resp. NI) onto rank s axis in RI (resp. RK)
by the way of the transition formulae:
1 1
Fs = XGs Gs = X ′Fs
λs λs

1 1 1
Fs (i ) =
λs
∑ xik Gs (k ) Gs (k ) =
λs
∑ I xik Fs (i)
k i

Since Benzecri (1973), these formulae are especially known in the case of
correspondence analysis (often under the name of barycentric properties); they are
seldom quoted in the case of PCA but are implicit during the interpretation.
6) Principal Components Analysis (PCA), Correspondence analysis (CA) and
Multiple Correspondences Analysis (MCA) can be viewed as particular cases of
GFA.

3. Usual methods or MFA ?

In the context of factor analysis (PCA or MCA according to the type of variables),
to study the kind of data table described figure 1, usual practice consists in
introducing only one set of variables as active, the others being illustrative. This
ensures homogeneity of active variables, characteristic which goes hand in hand
with a clear two-steps problematic: 1) looking for main factors describing data
variability according to one theme (the one corresponding to the active variables)
2) relating each of the illustrative variables to the previous factors.
This basic methodology is excellent. But it should be noted that, in this strategy,
the only multidimensional structure really handled is the one of the active
variables; the illustrative variables intervene independently one to the other.
According to this point of view, one can want to introduce several sets of variables
simultaneously as active elements, in order to take them simultaneously into
account in the definition of distance between individuals. Introducing, as active
elements, several sets of variables (or, according to an other point of view,
distinguishing sets among active variables) firstly implies to balance these sets and,
secondly, enriches problematic, that is to say induces new questions about data.
MFA brings solutions to these problems in a way described hereafter.

3
4. Balancing the sets of variables
If all the sets of variables are introduced, as active elements, without balancing
their influence, a single set can contribute quite by itself to the construction of the
first axes. In such a case, the user want to analyse all the sets and, in fact, analyses
only one of them.
Thus, the global analysis, in which several sets of variables are simultaneously
introduced as active ones, requires balancing the influences of these sets. The
influence of one set j derives from its structure, in the sense of its inertia
distribution (of the two clouds NIj and NKj it induces) in the different space
directions. For example, if a set presents a high inertia in one direction, this
direction will strongly influence the first axis of the global analysis.
This suggests to normalise the highest axial inertia of each set. Technically, it is
done by weighting each variable of the set j by 1/λ1j, denoting λ1j the first
eigenvalue of factor analysis applied to set j.
This weighting can be easily interpreted : considering the two clouds (NIj et NKj)
induced by the set j of variables, MFA weighting normalises each of these two
clouds by making its highest axial inertia equal to 1. This weighting does not
balance total inertia of the different sets. Thus, a set having a high dimensionality
will have a high global influence in that sense that this set will contribute to
numerous axes. But such a set has no reason to contribute particularly to the first
axes. Correlatively, a one-dimensional set can strongly contribute to only one axis,
but this axis can be the first one.

5. MFA as a general factor analysis

The core of MFA is a general factor analysis applied to all active sets of variables
(global analysis). MFA works with continuous variables as principal component
analysis does, the variables being weighted ; MFA works with categorical
variables as multiple correspondences analysis does, the variables being weighted.
Weighting, which balances highest axial inertia of sets, allows to work
simultaneously with continuous and categorical variables as active elements.
Likewise, it is possible to introduce simultaneously as active elements, set(s) of
standardised variables and set(s) of un-standardised variables.
The aim is to bring out main factors of data variability, individuals being
described, in a balanced manner, by several sets of variables (those introduced as
active).
According to this point of view, MFA provides classical outputs of general factor
analysis, that is to say, for each axis :
• Co-ordinates, contributions and squared cosines of individuals ;
• Correlation coefficient between factors and continuous variables ;

4
• For each category, co-ordinate (and test-value in the sense of Spad software
2002) of the centre of gravity of individuals belonging to this category.
Remark about categories inside MFA.
In MFA, categories are represented by exact centres of gravity (as in PCA and
differently from MCA). For each category, we can calculate the inertia of the
corresponding centre of gravity in per cent of the total inertia ; this ratio is named
contribution. It is proportional to the contribution usually defined in MCA for the
active variables and possesses the following property : its sum, for all the
categories of the variable k and for axis s, equals to the correlation ratio between
the variable k and the factor s. This ratio can be calculated for all categorical
variables (active and illustrative).

6. Superimposed representation of the J clouds of individuals.

Questions
To each set j, we associate the cloud NIj of individuals lying in the space RKj. This
cloud, named “ partial ”, is the one analysed in the factor analysis restricted to set j;
it contains “ partials ” individuals, denoted i j (individual i according to the set j).
A classical question is : are there structures common to these clouds NIj j=1, J ?
That is to say : are there some resemblances, from one cloud to the other, among
distances between homologous points ?
To answer these questions, we are looking for a superimposed representation of
clouds NIj which :
• fits well each of the clouds NIj ;
• highlights the resemblances between the different NIj, that is to say displays
homologous points as close one to the other as possible (homologous i.e.
referring to the same individual).
This problematic is similar to the one of Generalised Procustes Analysis (GPA ;
Gower, 1975). The way used by MFA to obtain such a superimposed
representation is described herafter.

5
 K Kj
R R
Kh
R

ij
ih
i j
NI
h
NI

NI

Figure 2. Representation of the J partial clouds NIj in the space RK

i : individual described by all the variables ; i j : individual described by the
variables of the group Kj. NIj can be viewed as a projection of NI onto RKj, the
subspace of RK spanned by the variables of Kj.
Principle
RK can be viewed as the direct sum of the RK j : RK = ⊕ RK j . Using this property, it
is possible to put all the NIj in the same space (cf. figure 2). In MFA, the clouds NIj
are projected upon the axes of the global analysis, as illustrative elements (cf.
figure 3). In fact these elements are not exactly illustrative since their data
contribute to axes construction ; moreover, this representation is possible only for
the clouds NIj corresponding to active sets.

K Kj
R R

Kh j
R NI
ij
i

h ih
NI
NI us

Figure 3. Principle of the superimposed representations provided by MFA

Each partial cloud NIj is projected onto main axes of the mean cloud NI
The co-ordinate of ij along axis s is denoted : Fs(ij). These co-ordinates can be
gathered in the vector Fsj such as : Fsj (i)= Fs(ij) .
Restricted transition formula

6
The superimposed representation of MFA is not optimal, in the sense that the Fsj
do not satisfy a global criterion. But it posseses the very useful following property.
It can be easily shown that the co-ordinate Fs(ij) can be calculated from the
coordinates of the variables Gs(k), k∈Kj, by the way of the following relationship :
1 1
Fs j (i ) = Fs (i j ) =
λs
∑ xik Gs (k )
λ1j k∈K j

We recognise here the usual transition formula (see § 2) but restricted to the
variables of the group Kj.
Ratio to measure the global similarity between axial representations of the clouds
NIj
When the different sets induce similar structures on individuals, homologous
points {i j, j=1, J} are close one to the other. This global property is measured, per
axis, through the ratio described below.
Let’s consider all the points of all the clouds NIj (j = 1, J) and a partition of these I
×J points in I classes, such as the J homologous points {i j, j=1, J} corresponding
to the same individual i belong to the same class. When axis s brings out a
structure common to the different sets of variables, the homologous points i j,
corresponding to the same individual i, are close one to the other and this partition
has a low within-inertia (along axis s). The ratio (between-inertia) / (total-inertia)
can be calculated for each axis. This ratio is close to 1 when the axis represents a
structure common to the different sets.
Be careful: 1) this ratio does not decrease with axis rank order since it is not the
criterion optimised by MFA; 2) it cannot be summerised for several axes.
Detailed examination of axial representations of NIj
The distance between each point i j and the corresponding mean point i gives an
idea about the position of i (among I) in the cloud NIj compared to the one in the
cloud NI. These distances can be examined visually, or by selecting the projections
of i j having the highest contributions to the within inertia. This allows to detect :
• Individuals having their homologous points close one to the other (low within
inertia) ; they illustrate the common structure represented by axis s ;
• Individuals having their homologous points far one from the other (high within
inertia) ; they constitute exceptions to the common structure represented by
axis s.
Case of categories
In factor analysis, when the individuals are numerous, it is the case in surveys for
example, they aren’t studied directly but by means of categorical variables, active
and/or illustrative (students, old people, etc.). Thus :

7
• In PCA, each category k is represented by the centre of gravity of individuals
that belong to this category k ;
• In MCA, the co-ordinates of points representing the categories are only
proportional to those of the corresponding centres of gravity (application of the
correspondence analysis centroid property to indicator matrix)
In MFA, the categories are represented by their associated centres of gravity. This
allows to work with categories as with individuals. Particularly, each category (e.g.
student in a survey) can be represented by a global point (centre of gravity of the
students) and by one partial point for each set of variables (e.g. the centre of
gravity of partial points representing the students according to set j).

7. MFA as a multicanonical analysis

Principle of multicanonical analysis
In the simultaneous study of several sets of variables, the main question is : are
there factors common to the different sets of variables ?
In the simple case of two sets, this question refers to canonical analysis (Hotelling
1936). When there are more than two groups, the reference method is
multicanonical analysis. There are several multicanonical analyses. The most used
is the one of CARROLL (1968), that works in two steps :
• Looking for a sequence of variables {zs ; s=1,S} (named general variables),
normalised and not correlated one to the other, related to the sets of variables as
strongly as possible ;
• For each general variable zs and for each set j, looking for the linear
combination of the variables of set j (combinations named canonical variables)
related to general variable zs as strongly as possible.
MFA can be interpreted in this framework.
Measure of relationship between one variable and a group of variable
To do this, it is necessary to firstly define a measure of relationship (denoted )
between one continuous variable z and a set of variables Kj= {vk, k=1,Kj}
(z, {vk, k=1,Kj}) = (z, Kj) = inertia of all variables vk projected upon z.
When the vk are reduced continuous variables, weighted by mk :
(z, Kj) = ∑k mk[r(z,vk)]²
This measure is implicitly used in PCA : the first principal component is the linear
combination (of variables) the most related to the whole set of variables (it
maximises (z, K)).
If (z, Kj)=0, variable z is not correlated to any variable of the set Kj.

8
Due to MFA weighting, (z, Kj)≤1 ; (z, Kj)=1 when z is the first principal
component of Kj.
General variables
The first factor of MFA (as defined in §5) maximises projected inertia of all the
sets of variables, that is :
∑j (z, Kj) maximum
In that sense, MFA factors (denoted Fs §6) can be considered as general variables
of a multicanonical analysis (in CARROLL’s method, relationship between one
variable and one set of variables is measured by means of multiple correlation
coefficient).

Canonical variables
The coherence between the multicanonical point of view and the superimposed
representation point of view suggests to use the previously defined Fsj as
canonical variables. It can be shown (Pagès & Tenenhaus, 2001) that Fsj is the
first component in the PLS regression between the general variable zs and the data
table Xj. This result reinforces the superimosed representation: it induces that the
Fsj j=1,J must be correlated one to the other since each Fsj expresses the same
structure Fs in the group Kj.
Canonical correlation coefficients in MFA
In MFA, factors of global analysis (denoted Fs) are the common factors and factors
of partial points (denoted Fsj) represent common factors in each set j of variables.
In order to judge if factors of global analysis really are common to the different
sets, it is possible to calculate, for each set j and each factor s, the correlation
coefficient between general variable Fs and canonical variable Fsj. If this
coefficient (named canonical correlation coefficient and always positive) is high,
then the structure brought out by variable Fs does “ exist ” in the set j. If not, it
does not. The synthesis of all these correlation coefficients shows factors common
to all the sets, common to some sets, specific to only one set.

8. Global study of sets of variables

It is often interesting to globally study the sets of variables, the question being : do
these sets define similar structures upon individuals (i.e. similar distances between
individuals from one cloud Nij to the other) ? We find again the problem of
superimposed representations and the one of common factors but now the
investigation about the similarities between sets is global.
Here, we are looking for a display in which each set of variables is represented by
a unique point. In such a display, two sets must be close one another if they induce
similar structures on the individuals.

9
To each set of variables Kj, we associate the I×I matrix Wj of scalar products
between individuals (Wj=Xj X′j). Each scalar product matrix Wj can be represented
by one point in the I²-dimensional Euclidean space (denoted RI² ). Thus, in this
space, one set is represented by one point: the J points constitute the set cloud,
denoted NJ. In this cloud NJ, the distance between two points Wj and Wl decreases
as the similarity between the structures (defined upon individuals) induced by the
sets Kj and Kl increases. For this reason, it is interesting to get a representation of
the cloud NJ.
The Statis method (Lavit, 1988) is based on such a representation, obtained by
projecting NJ onto its main inertia directions. But these directions cannot be
interpreted (they are linear combinations of couples of individuals) (Pagès 1996).
The representation provided by MFA is obtained by projecting NJ upon vectors (in
RI²) induced by I-factors of global analysis (one factor may be considered as a set
including a single variable; it is possible to associate to this set a scalar product
matrix and thus a vector in RI²).
The normalised factor of rank s in RK, previously denoted zs, induces ws = zszs′ in
RI². Some properties of zs induce corresponding properties for ws :
zs′ zt = 0 ⇒ 〈 ws , wt 〉 = 0
zs = 1 ⇒ ws = 1

j
NK
Wj D
z zz'D

coordinate
Projected inertia

Space RI Space RI
2

Figure 4. The representation of the groups and its links with the one of variables
The main interest of this projection space is that its axes (upon which NJ is
projected) are interpretable and, above all, possess the same interpretation that axes
of global analysis (in the same manner, due to factor analysis duality, axis of rank s
upon which individuals are projected and axis of rank order s upon which
variables are projected possess the same interpretation).
This representation has the following property: it ca be shown (Escofir & Pagès
1998 p 167) that co-ordinate of set j upon axe of rank s is equal to (zs, Kj). Thus:

10
• Set co-ordinates are always comprised between 0 and 1;
• A small distance between two sets along axe s means that these two sets
include the structure expressed by factor s each one with the same intensity. In
other words, set representations shows which ones are similar (or different)
from the point of view of global analysis factors.
This representation has been introduced as an aid to the interpretation of
representations of individuals and variables. But it possesses its own optimality:
axes upon which NJ is projected, taking into account the usual orthogonality
constraint but also the constraint to be of order 1 (i.e. induced by one direction in
RI), make the sum (and not the sum of squares) of co-ordinates maximum (for axis
s, this sum is equal to the sth eigenvalue of the global analysis). Thus, from the RI²
point of view, the contribution of the set j to axis s is equal to the set j co-ordinate
divided by the sum of co-ordinates (this contribution is equal to the sum of
contributions to axis s, in RI, of variables belonging to the set j).
The set study can be completed by squared cosines computed in RI².

9. Relationship between global analysis and separated

analysis of each set.
It is always important and interesting to relate MFA results to those of separated
analysis of each set. To do this, I-factors of separated analysis (called “ partial ”
factors) are projected as illustrative quantitative variables.
It is equivalent to perform MFA from variables or from all the partial factors (each
one being, within its set, “ pre-weighted ” proportionally to its eigenvalue). Thus:
• For each partial factor, the ratio between its projected inertia along axis s and
the axis s eigenvalue may be interpreted, in case of active sets, actually as a
contribution to MFA axis s;
• MFA may be considered as a method providing an optimal representation of
separated analyses axes.
This last point is useful for applications: MFA is a convenient tool in order to
compare several factor analyses having the same individuals. Example: in order to
compare, for the same variables, normalised PCA an un-normalised PCA,
variables must appear twice in data base, firstly within a normalised set and
secondly within a un–normalised set; MFA is there an optimal tool in order to
display factors of the two analyses (a direct PCA is not desirable because set
weighting is necessary).

10. The orange juice example

10.1 Data

11
Six pure orange juices (P1 to P6) were selected from the main brands on the
French market. These juices were pasteurised in two ways: thus, three of them
must be stored in refrigerated conditions (R) while the others can be stored at
ambient temperature (A). Here is the list of the six orange juices: Pampryl at
ambient temperature (P1), Tropicana at ambient temperature (P2), refrigerated
Fruivita (P3), Joker at ambient temperature (P4), refrigerated Tropicana (P5),
refrigerated Pampryl (P6).
Ninety-six students in food science, both trained to evaluate foodstuffs and
consumers of orange juice, described each of these six products on the basis of
seven attributes: intensity and typical nature of its smell, intensity of the taste, pulp
content, sweetness, sourness and bitterness.
The serving order design was a juxtaposition of Latin squares balanced for carry-
over effects (MacFie, Bratchell, Greenhoff and Vallis, 1989).
In addition to the sensory investigation, chemical measurements (pH, citric acid,
overall acidity, saccharose, fructose, glucose, vitamin C and sweetening power -
defined as : saccharose + .6 glucose + 1.4 fructose) were carried out.
The data are gathered in a table using the format shown in figure 5. The complete
data table is in the appendix.
Chemical var. Sensory var.
1 k K1=9 1 k K2=7
1

Orange juices i xik xik

I=6
Figure 5. The orange juices data table
For juice i : xik is the panel average of the sensory variable k or the chemical
measurement k
The ouputs described below come from Spad 2002 software.

10.2 Separate analyses

Table 1 gathers inertia from separate and global analysis. The first eigenvalue of
the separate PCA of chemical variables is slightly higher than the one of PCA of
sensory variables. The balancing is useful to avoid the domination of chemical
variables in the construction of the first axis.
The sequence of eigenvalues is similar from one analysis to the other: the two
groups of variables have a strong first direction of inertia. Moreover, the
homologous factors of the two separate PCA are correlated one another (table 2).

12
This data set is interesting from a methodological point of view: the similarity
between the two groups of variables justifies their simultaneous analysis; the
differences between the two groups are sufficiently important to justify the use of
a specific method to highlight common and specific features.

PCA chemical var. PCA sensory var. MFA

Axes Eigenvalue % Eigenvalue % Eigenvalue %
1 6.2135 69.04 4.7437 67.77 1.7907 61.24
2 1.4102 15.67 1.3333 19.05 0.4764 16.29
3 1.0457 11.62 0.8198 11.71 0.2938 10.05
4 0.3173 03.53 0.0840 01.20 0.2009 6.87
5 0.0133 00.15 0.0192 00.27 0.1623 5.55
Table 1: Eigenvalues (= inertia) from separate PCA and from MFA

PCA chemical var.

F1 F2
PCA sensory F1 -0.78 -0.25
variables F2 0.08 -0.74
Table 2: Correlations between separate PCA factors

10.3 Representation of individuals (= products) and variables (fig. 6 and 7)

This MFA builds a product space starting from factors common to the sensory and
instrumental data, in which the influences of these two groups of variables are
balanced. These MFA representations (of products and variables) can be read like
those from a PCA: the co-ordinates of a product are its values for the common
factors; the co-ordinates of a variable are its correlations with these factors.
The first axis is highly correlated to variables belonging to the two groups ; that
was awaited due to the group balancing. It opposes the juices 1, 4 and 6 to the
juices 2, 3 and 5.
According to the usual transition formulae, the juices 1, 4 and 6 have a high level
of acidity (and a low pH), and a high [(glucose + fructose)/saccharose] ratio; they
were perceived as sour, bitter and not very sweet. Symmetrically, the juices 2, 3
and 5 have opposite characteristics: a low level of acidity (and a high pH), and a
low [(glucose + fructose)/saccharose] ratio; they were perceived as being not sour
or bitter, but sweet.
The juices 2, 3 and 5 come from Florida; the first axis is linked to geographic
origin.
The second bissector is more interesting than the second axis. This bissector
correspond to pulpy. It opposes the refrigerated juices to the others.
Let us also point out:

13
• the opposition between fructose and glucose on the one hand and saccharose
and pH on the other hand, connected with the hydrolysis of saccharose,
facilitated in an acid medium;
• the correlation between acidity, pH and sourness;
• the absence of correlation between sweetening power and sweetness : a high
level of sweetness is associated with a low level of sourness (this refers to the
concept of gustatory balance). Thus the strong correlation between saccharose
and sweetness is not due to the direct influence of saccharose but to a high pH.

10.4 Factors from separate analyses (Fig. 8)

Factors from separate analyses can be represented by the way of their correlations
with factors of MFA. Figure 8 shows that the first factor of MFA is highly
correlated with the first factor of each separate analysis. These factors of separate
analyses are not so highly correlated one another (cf. table 2) but, in this analysis,
the first MFA factor, being a kind of compromise between them, is highly
correlated to the two. The same observation can be made for the second factor.
Thus, the first MFA map is roughly similar to the one of each separate analysis
(figure 8 gives an idea of the slight rotation to obtain one map from an other).

10.5 Superimposed representations (Fig. 9)

Figure 9 derives from figure 6 by adding partial points (C and S). Whatever the set
of variables considered, the first axis opposes the products 1, 4 and 6 to the
product 2, 3 and 5. This is a other way to highlight common factor.
The resemblance between the two partial clouds can be globally evaluated by two
series of measures (cf. table 3 and 4). The two first factors from MFA can be
considered as factors common to the two groups of variables.

Factor from MFA

F1 F2 F3 F4 F5
G1 : chemical var. 0.95 0.88 0.49 0.46 0.82
G2 : sensory var. 0.95 0.90 0.45 0.90 0.26
Table 3. Canonical correlation coefficient
At the intersection of row j and column s: r(Fsj , Fs) (cf. § 7)

F1 F2 F3 F4 F5
ratio 0.90 0.80 0.22 0.51 0.35
Table 4. Ratio [(between-inertia) / (total-inertia)]; cf. § 6

14
This representation allows a precise comparison of the clouds NIj. Thus, the figure
9 suggests that the product 5 is highly characteristic from a sensory point of view
though it is not the case from a chemical point of view; conversely, product 2 is
characteristic from the chemical point of view but that does not induce a particular
sensory evaluation.
This can be directly verified in the data (cf. appendix), preferably in the
standardised data that can be compared from one variable to the other. Thus, the
standardised data table shows that product 5 has absolute values higher for the
sensory attributes than for the chemical variables. It is the reverse for he product 2.
In these data, in which the two first factors from MFA are highly correlated to the
corresponding ones from each separate analysis, the superimposed representation
gives a good idea of the representation from separate analysis. This can be
illustrated by the comparison between the representation of partial individuals in
MFA and the representation of individuals from separate PCA (cf. Fig. 10). Thus,
for exemple, the opposition between the products P2 and P4 is much bigger from a
chemical point of view than for a sensory point of view.

10.6 Representation of categories (Fig. 11)

In this data table, the individuals are very few : here, the interest of the
representation of categories is mostly technical. But, when the individuals are
numerous, as in a survey for example, this representation is essential. Each
category lies in the center of gravity of the individuals which posses this category.
This is applied to mean and partial point.
In Figure 11, the mean points show immediatly that the factor 1 corresponds to
origin (Florida / elswhere) and that the second bissector corresponds to way of
storing (refrigerated or ambient). The partial points shows that the opposition of
the two origins along axis 1 is equally clear from the two points of view
(chemical/sensory). Along the second bissector, the opposition of the two ways of
storing mostly appears from the sensory point of view (pulpyness).

10.7 Representation of groups of variables (Fig. 12)

These data being composed of two groups only, the representation of groups as
points of RI² (cf. § 8) has not practical interest. To enrich the technical comments,
here a third group is added (as a supplementary one): it derives from the chemical
group, in which we have removed the variables pH2 (because it is quite equivalent
to pH1) and vitamin C (because it is not related to sensory evaluation). Thus we
get an empiric idea about the stability of the chemical group.
The three groups are strongly related to Factor 1 (in the sense of measurement ;
cf. § 7): this factor correspond to a direction having a strong inertia for the three
clouds of variables (in other words, many variables of each groups are related to
this factor).

15
Regarding the first two MFA factors, these three groups are very similar: the
clouds of individuals they induce (previously denoted NIj ) are very similar. In
particular, removing pH2 and vitamin C has had a weak influence.

11. Conclusion
MFA allows to take into account several sets of variables as active elements in a
unique factor analysis. Its main features are:
• the balancing of the sets of variables;
• outputs specific of the partition of the variables in differents sets ; mainly 1)
the superimposed representations of individuals and of categories 2) the
groups representation.

F2 - 16.29 %

P2 tropicana amb.

1.0

0.5
P1 pampryl amb.

P4 joker amb. F1 - 61.24 %

0

P5 tropicana refr.

-0.5
P3 fruvita refr.

P6 pampryl refr.
-1.0

-2 -1 0 1 2

Figure 6. First factorial map from MFA : mean individuals

Refr. : refrigerated ; amb. : stored at ambient temperature. Tropicana and Fruvita
come from Florida

16
 F2 - 16.29 %

0.8

0.4
Vitamin C
pH1
Bitter
pH2
Acidity
Sw eet
0
Citric acid
Saccharose

Sour Odor typicity

-0.4 Fructose

Glucose Taste intensity

Odor intensity

-0.8 Pulpy
Sw eetening pow er
F1 - 61.24 %
-0.8 -0.4 0 0.4 0.8

Figure 7. First factorial map from MFA : chemical and sensory variables

17
 F2 - 16.29 %

G1F2
0.8

0.4
G2F3

G2F4
G1F3
0

G1F4 G2F1
G1F1 G1F5

G2F5
-0.4

-0.8
G2F2
F1 - 61.24 %
-0.8 -0.4 0 0.4 0.8

Figure 8. Representation of the factors from separate analyses

GjFs : rank-s factor of separate PCA of group j

F2 - 16.29 %
C
P2 tropicana amb.

1.0
S S

0.5 S
P1 pampryl amb.
C C
P4 joker amb. F1 - 61.24 %
0

P5 tropicana refr.
C
-0.5
C P3 fruvita refr. S
C S
P6 pampryl refr.
-1.0
S
-2 -1 0 1 2

Figure 9. Superimposed representation of partial individuals

The mean points of figure 6 are here joined to corresponding partial points. S:
sensory; C: chemical

18
 PCA on chemical data
F2 - 15.67 % P2 tropicana amb.
1.50

P1 pampryl amb.

0.75

P4 joker amb.
F1 - 69.04 %
0

P5 tropicana refr.

-0.75

P6 pampryl refr. P3 fruivita refr.

-1.50
3.0 1.5 0 -1.5 -3.0

PCA on chemical data PCA on sensory data

F2 - 15.67 % F2 - 19.05 %
0.8 -0.8

Vitamin0.4C -0.4

Acidity Sweet
F1 - 69.04 %
pH1 F1 - 67.77 %
Citric acid
0.8 0.4 -0.4 -0.8 -0.8 -0.4 0.4 0.8
Bitter
pH2
Fructose Odor typicity
Saccharose 0.4
-0.4 Sour
Glucose Taste intensity
Pulpy
0.8
-0.8 Odor intensity
Sweetening power

F2 - 19.05 %

-1.50 P2 tropicana amb.

-0.75
P4 joker amb.

P1 pampryl amb.
P3 fruivita refr.
0
F1 - 67.77 %

P5 tropicana refr.
0.75

1.50

P6 pampryl refr.

-1.5 0 1.5 3.0

Figure 10. Separate PCA of each group of variables

19
 S F2 - 16.29 %
Ambient

0.4
C C

Florida
S
0

Elsew here than Flor.

S

0.4 C C

Refrigerated

S
-1.0 -0.5 0 0.5 F11.0- 61.24 %

Figure 11. Display of mean and partial categories

Each category lies in the center of gravity of the individuals belonging to it

1.00
F2 - 16.29 %

0.75

0.50

Sensory
0.25
Chem. partim

Chemical

F1 - 61.24 %
0
0 0.25 0.50 0.75 1.00

Figure 12. Representation of groups of variables.

In this display, a group of variables is represented by a single point. Here, appears
a third group, obtained from the chemical one, by giving up two variables (pH2
and vitamin C)

20
12. References
CARROLL J.D. (1968). A generalization of canonical correlation analysis to three or
more sets of variables, Proc. 76th Conv. Amer. Psych. Assoc., pp. 227-228.
ESCOFIER B. & PAGÈS J. (1994). Multiple Factor Analysis (AFMULT package).
Computational statistics & data analysis 18 121-140
ESCOFIER B., PAGES J. (1988-1998). Analyses factorielles simples et multiples ;
objectifs, méthodes et interprétation. Dunod. Paris.
HOTELLING H. (1936). Relations between two sets of variables. Biometrika, 28, p
129-149.
MACFIE H. J., BRATCHELL N., GREENHOFF, VALLIS L. V. (1989). Designs to balance
the effect of order of presentation and first-order carry-over effects in hall tests.
Journal of Sensory Studies 4 129-148.
PAGÈS J. (1996). Eléments de comparaison entre l’Analyse Factorielle Multiple et
la méthode STATIS. Revue de Statistique appliquée XLIV (4) 81-95.
PAGÈS J. (2002). Analyse factorielle multiple appliquée aux variables qualitatives
et aux données mixtes. Rev Statistique appliquée, L (4), 5-37.
SPAD rel. 5.5, (2002). Système pour l’analyse des données. Logiciel diffusé par
Decisia, Levallois-Perret, 92532 France.

21
Appendix : raw data and standardised data

Joker amb.
Tropicana

Tropicana

deviation
Standard
Pampryl

pampryl
Fruivita

mean
amb.

amb.

refr.

refr.

refr.
Raw data
P1 P2 P3 P4 P5 P6
Glucose 25,32 17,33 23,65 32,42 22,7 27,16 24,76 4,57
Fructose 27,36 20 25,65 34,54 25,32 29,48 27,06 4,41
Saccharose 36,45 44,15 52,12 22,92 45,8 38,94 40,06 9,16
Pouvoir sucrant 89,95 82,55 102,22 90,71 94,87 96,51 92,80 6,11
PH brut 3,59 3,89 3,85 3,6 3,82 3,68 3,74 0,12
PH après centrif. 3,55 3,84 3,81 3,58 3,78 3,66 3,70 0,11
Titre 13,98 11,14 11,51 15,75 11,8 12,21 12,73 1,62
Acide citrique 0,84 0,67 0,69 0,95 0,71 0,74 0,77 0,10
Vitamine C 43,44 32,7 37 36,6 39,5 27 36,04 5,18
Saccharose % 0,41 0,54 0,51 0,26 0,49 0,41 0,44 0,10
Intensité odeur 2,82 2,76 2,83 2,76 3,2 3,07 2,91 0,17
Typicité odeur 2,53 2,82 2,88 2,59 3,02 2,73 2,76 0,17
Pulpy 1,66 1,91 4 1,66 3,69 3,34 2,71 0,99
Intensité goût 3,46 3,23 3,45 3,37 3,12 3,54 3,36 0,14
Caractère acide 3,15 2,55 2,42 3,05 2,33 3,31 2,80 0,38
Caractère amer 2,97 2,08 1,76 2,56 1,97 2,63 2,33 0,42
Caractère sucré 2,6 3,32 3,38 2,8 3,34 2,9 3,06 0,30
Joker amb.
Tropicana

Tropicana

deviation
Standard
Pampryl

pampryl
Fruivita

mean
amb.

amb.

refr.

refr.

refr.

Standardised
Data
P1 P2 P3 P4 P5 P6
Glucose 0,12 -1,63 -0,24 1,67 -0,45 0,52 0 1
Fructose 0,07 -1,60 -0,32 1,70 -0,39 0,55 0 1
Saccharose -0,39 0,45 1,32 -1,87 0,63 -0,12 0 1
Pouvoir sucrant -0,47 -1,68 1,54 -0,34 0,34 0,61 0 1
PH brut -1,23 1,26 0,93 -1,15 0,68 -0,49 0 1
PH après centrif. -1,36 1,21 0,94 -1,09 0,68 -0,38 0 1
Titre 0,77 -0,98 -0,75 1,86 -0,57 -0,32 0 1
Acide citrique 0,75 -0,98 -0,78 1,86 -0,58 -0,27 0 1
Vitamine C 1,43 -0,65 0,19 0,11 0,67 -1,75 0 1
Saccharose % -0,28 1,11 0,82 -1,90 0,55 -0,30 0 1
Intensité odeur -0,52 -0,87 -0,46 -0,87 1,75 0,97 0 1
Typicité odeur -1,38 0,35 0,71 -1,03 1,54 -0,19 0 1
Pulpy -1,06 -0,81 1,30 -1,06 0,99 0,64 0 1
Intensité goût 0,68 -0,91 0,61 0,06 -1,67 1,24 0 1
Caractère acide 0,91 -0,66 -1,00 0,65 -1,24 1,33 0 1
Caractère amer 1,52 -0,59 -1,35 0,55 -0,85 0,71 0 1
Caractère sucré -1,50 0,87 1,06 -0,85 0,93 -0,52 0 1

22
23