WIREs Computational Statistics Multiple factor analysis
one whereas the sum of the squared correlations of
all the dimensions for a given variable is equal to one
(and therefore it is always possible to transform one
set into the other).
HOW TO FIND THE IMPORTANT
ELEMENTS: CONTRIBUTIONS, ETC.
Contributions of Observations, Variables,
and Tables to a Dimension
In MFA, just like in standard PCA, the importance of
a dimension (i.e., principal component) is reflected by
its eigenvalue which indicates how much of the total
inertia (i.e., variance) of the data is explained by this
component.
To better understand the relationships between
components, observations, variables, and tables
and also to help interpret a component, we can
evaluate how much an observation, a variable, or
a whole table contribute to the inertia extracted
by a component. In order to do so, we com-
pute descriptive statistics, called contributions (see
Refs 78,89–91 and Ref 75, p. 437ff.). The stability of
these descriptive statistics can be assessed by cross-
validation techniques such as the bootstrap whose
results can be used to select the relevant elements for
a dimension.
Contribution of an Observation to a Dimension
As stated in Eq. (19), the variance of the factor
scores for a given dimension is equal to its eigenvalue
(i.e., the square of the singular value) associated with
this dimension. If we denote λ
, the eigenvalue of a
given dimension, we can rewrite Eq. (19) as
λ
=
i
m
i
× f
2
i,
(24)
where m
i
and f
i,
are, respectively, the mass of the ith
observation and the factor score of the ith observation
for the th dimension. As all the terms m
i
× f
2
i,
are
positive or null, we can evaluate the contribution
of an observation to a dimension as the ratio of
the squared weighted factor score by the dimension
eigenvalue. Formally, the contribution of observation
i to component , denoted ctr
i,
, is computed as
ctr
i,
=
m
i
× f
2
i,
λ
. (25)
Contributions take values between 0 and 1, and for
a given component, the sum of the contributions
of all observations is equal to 1. The larger a
contribution, the more the observation contributes
to the component. A useful heuristic is to base the
interpretation of a component on the observations
that have contributions larger than the average
contribution. Observations with high contributions
and whose factor scores have different signs can then
be contrasted to help interpreting the component.
Alternatively (as described in a later section) we can
derive pseudo t statistics (called bootstrap ratios) in
order to find the observations important for a given
dimension.
Contributions of a Variable to a Dimension
As we did for the observations, we can find
the important variables for a given dimension by
computing variable contributions. The variance of the
loadings for the variables is equal to one when the α
weights are taken into account (cf., Eq. (13)). So if we
denote by a
j
the α weight for the j th variable (recall
that all variables from the same table share the same
α weight cf., Eq. (11)), we have
1 =
j
a
j
× q
2
j,
(26)
where q
i,
is the loading of the jth variable for the
th dimension. As all terms a
j
× q
2
i,
are positive or
null, we can evaluate the contribution of a variable to
a dimension as its squared weighted loading for this
dimension. Formally, the contribution of variable j to
component , denoted ctr
j,
, is computed as
ctr
j,
= a
j
× q
2
j,
. (27)
Variable contributions take values between 0 and
1, and for a given component, the contributions of
all variables sum to 1. The larger a contribution of
a variable to a component the more this variable
contributes to this component. Variables with high
contributions and whose loadings have different signs
can then be contrasted to help interpreting the
component.
Contribution of a Table to a Dimension
Specific to multiblock analysis is the notion of
a table contribution. As a table comprises several
variables, the contribution of a table can simply be
defined as the sum of the contributions of its variables
(a simple consequence of the Pythagorean theorem
that states that squared lengths are additive). So the
contribution of table k to component is denoted
ctr
k,
and is defined as
ctr
k,
=
J
[k]
j
ctr
j,
. (28)
© 2013 Wiley Periodicals, Inc.