|You are working with the text-only light edition of "H.Lohninger: Teach/Me Data Analysis, Springer-Verlag, Berlin-New York-Tokyo, 1999. ISBN 3-540-14743-8". Click here for further information.
Factor analysis has a long history with its root in psychology. The
first development was reported by Hotelling in 1933. Later on, other authors
extended this method, finally resulting in a range of various methods which
are difficult to overlook. An excellent book on factor analysis was published
by Paul Horst [Horst 1965].
One of the possible definitions has been given by Malinowski:
"Factor analysis is a multivariate technique for reducing matrices
of data to their lowest dimensionality by the use of orthogonal factor
space and transformations that yield predictions and/or recognizable factors"
The principle behind factor analysis is quite simple, although the
concrete realization depends on the requirements of the specific situation,
and may be quite demanding: in principle, a data matrix X is split
up into a product of two data matrices:
X = U V
The matrices U and V are called scores and loading matrices,
respectively. This can be visualized by the following figure:
Using this approach involves factors which may be useless for tangible
problems, since one of the major goals of factor analysis is to provide
factors which are related in some way to "real" factors. So the matrices
V are rather abstract matrices which provide little to no improvement
concerning the interpretation of the original matrix. It is therefore necessary
to come up with a way to find factors which are closely related to reality.
Since the theoretical split into the matrices
U and V has
many solutions, we need some additional criteria to guide the decomposition
of the original data matrix X.
The following general rules for finding suitable factors are usually
1) The number of factors should be kept as low as possible. The number
of factors should reflect the complexity of the data space. Finding the
correct number of factors can be a tedious task.
2) The factors should be calculated in such a way that they could be
related to some real effects of the experiment from which the data has
been drawn, i.e. the factors should be physically meaningful. In order
to achieve this, the factors have to be rotated.
Last Update: 2006-Jšn-17