Multivariate Analysis.

Hervé Abdi1
The University of Texas at Dallas

Introduction
As the name indicates, multivariate analysis comprises a set of techniques
dedicated to the analysis of data sets with more than one variable. Several
of these techniques were developed recently in part because they require the
computational capabilities of modern computers. Also, because most of them
are recent, these techniques are not always unified in their presentation, and the
choice of the proper technique for a given problem is often difficult.
This article provides a (non-exhaustive) catalog in order to help decide when
to use a given statistical technique for a given type of data or statistical question
and gives a brief description of each technique. This paper is organized according
to the number of data sets to analyze: one or two (or more). With two data sets
we consider two cases: in the first case, one set of data plays the role of predictors
(or independent) variables (IV’s) and the second set of data corresponds to
measurements or dependent variables (DV’s); in the second case, the different
sets of data correspond to different sets of DV’s.

One data set

Typically, the data tables to be analyzed are made of several measurements
collected on a set of units (e.g., subjects). In general, the units are rows and
the variables columns.
Interval or ratio level of measurement: principal component analysis (PCA)
This is the oldest and most versatile method. The goal of PCA is to decom-
pose a data table with correlated measurements into a new set of uncorrelated
(i.e., orthogonal) variables. These variables are called, depending upon the con-
text, principal components, factors, eigenvectors, singular vectors, or loadings.
Each unit is also assigned a set of scores which correspond to its projection on
the components.
The results of the analysis are often presented with graphs plotting the
projections of the units onto the components, and the loadings of the variables
1 In: Lewis-Beck M., Bryman, A., Futing T. (Eds.) (2003). Encyclopedia of Social Sciences

Research Methods. Thousand Oaks (CA): Sage.

Address correspondence to
Hervé Abdi
Program in Cognition and Neurosciences, MS: Gr.4.1,
The University of Texas at Dallas,
Richardson, TX 75083–0688, USA
E-mail: [email protected] http://www.utdallas.edu/∼herve

1
 (the so-called “circle of correlations”). The importance of each component is
expressed by the variance (i.e., eigenvalue) of its projections or by the proportion
of the variance explained. In this context, PCA is interpreted as an orthogonal
decomposition of the variance (also called inertia) of a data table.
Nominal or ordinal level of measurement: correspondence analysis (CA), multiple correspondence analysis (MCA)
CA is a generalization of PCA to contingency tables. The factors of CA give
an orthogonal decomposition of the Chi-square associated to the table. In CA,
rows and columns of the table play a symmetric role and can be represented in
the same plot. When several nominal variables are analyzed, CA is generalized
as MCA. CA is also known as dual or optimal scaling or reciprocal averaging.
Similarity or distance: multidimensional scaling (MDS), additive tree, cluster analysis
These techniques are applied when the rows and the columns of the data
table represent the same units and when the measure is a distance or a similarity.
The goal of the analysis is to represent graphically these distances or similarities.
MDS is used to represent the units as points on a map such that their Euclidean
distances on the map approximate the original similarities (classic MDS, which
is equivalent to PCA, is used for distances, nonmetric MDS for similarities).
Additive tree analysis and cluster analysis are used to represent the units as
“leaves” of a tree with the distance “on the tree” approximating the original
distance or similiarity.

Two data sets, Case one: one independent variable set and one dependent variable set

Multiple linear regression analysis (MLR)

In MLR, several IV’s (which are supposed to be fixed or equivalently are
measured without error) are used to predict with a least square approach one
DV. If the IV’s are orthogonal, the problem reduces to a set of univariate re-
gressions. When the IV’s are correlated, their importance is estimated from
the partial coefficient of correlation. An important problem arises when one of
the IV’s can be predicted from the other variables because the computations
required by MLR can no longer be performed: This is called multicolinearity.
Some possible solutions to this problem are described in the following section.
Regression with too many predictors and/or several dependent variables
Partial least square (PLS) regression (PLSR)
PLSR addresses the multicolinearity problem by computing latent vectors
(akin to the components of PCA) which explains both the IV’s and the DV’s.
This very versatile technique is used when the goal is to predict more than one
DV. It combines features from PCA and MLR: The score of the units as well
as the loadings of the variables can be plotted as in PCA, and the DV’s can be
estimated (with a confidence interval) as in MLR.
Principal component regression (PCR)
In PCR, the IV’s are first submitted to a PCA and the scores of the units
are then used as predictors in a standard MLR.

2
 Ridge regression (RR)
RR accommodates the multicolinearity problem by adding a small constant
(the ridge) to the diagonal of the correlation matrix. This makes the computa-
tion of the estimates for MLR possible.
Reduced rank regression (RRR) or redundancy analysis
In RRR, the DV’s are first submitted to a PCA and the scores of the units
are then used as DV’s in a series of standard MLR’s where the original IV’s are
used as predictors (a procedure akin to an inverse PCR).
Multivariate analysis of variance (MANOVA)
In MANOVA the IV’s have the same structure as in a standard ANOVA,
and are used to predict a set of DV’s. MANOVA computes a series of ordered
orthogonal linear combinations of the DV’s (i.e., factors) with the constraint
that the first factor generates the largest F if used in an ANOVA. The sampling
distribution of this F is adjusted to take into account its construction.
Predicting a nominal variable: discriminant analysis (DA)
DA, which is mathematically equivalent to MANOVA, is used when a set
of IV’s are used to predict the group to which a given unit belongs (which is a
nominal DV). It combines the IV’s in order to create the largest F when the
groups are used as a fixed factor in an ANOVA.
Fitting a model: confirmatory factor analysis (CFA)
In CFA, the researcher first generates one (or a few) model(s) of an underly-
ing explanatory structure (i.e., a construct) which is often expressed as a graph.
Then the correlations between the DV’s are fitted to this structure. Models are
evaluated by comparing how well they fit the data. Variations over CFA are
called structural equation modelling (SEM), LISREL, or EQS.

Two (or more) data sets, Case two: two (or more) dependent variable sets

Canonical correlation analysis (CC)

CC combines the DV’s to find pairs of new variables (called canonical vari-
ables, CV, one for each data table) which have the highest correlation. However,
the CV’s, even when highly correlated, do not necessarily explain a large portion
of the variance of the original tables. This make the interpretation of the CV
sometimes difficult, but CC is nonetheless an important theoretical tool because
most multivariate techniques can be interpreted as a specific case of CC.
Multiple factor analysis (MFA)
MFA combines several data tables into one single analysis. The first step
is to perform a PCA of each table. Then each data table is normalized by
dividing all the entries of the table by the first eigenvalue of its PCA. This
transformation—akin to the univariate Z-score—equalizes the weight of each
table in the final solution and therefore makes possible the simultaneous analysis
of several heterogenous data tables.

3
 Multiple correspondence analysis (MCA)
MCA can be used to analyze several contingency tables; it generalizes CA.
Parafac and Tucker3
These techniques handle three-way data matrices by generalizing the PCA
decomposition into scores and loadings in order to generates three matrices of
loading (one for each dimension of the data). They differ by the constraints they
impose on the decomposition (Tucker3 generates orthogonal loadings, Parafac
does not).
Indscal
Indscal is used when each of several subjects generates a data matrix with
the same units and the same variables for all the subjects. Indscal generates
a common Euclidean solution (with dimensions) and expresses the differences
between subjects as differences in the importance given to the common dimen-
sions.
Statis
Statis is used when at least one dimension of the three-way table is com-
mon to all tables (e.g., same units measured on several occasions with different
variables). The first step of the method performs a PCA of each table and gen-
erates a similarity table (i.e., cross-product) between the units for each table.
The similarity tables are then combined by computing a cross-product matrix
and performing its PCA (without centering). The loadings on the first compo-
nent of this analysis are then used as weights to compute the compromise data
table which is the weighted average of all the tables. The original table (and
their units) are projected into the compromise space in order to explore their
communalities and differences.
Procustean analysis (PA)
PA is used to compare distance tables obtained on the same objects. The
first step is to represent the tables by MDS maps. Then PA finds a set of
transformations that will make the position of the objects in both maps as close
as possible (in the least square sense).

References
[1] Borg I., & Groenen P. (1997). Modern multidimensional scaling. New York:
Springer-Verlag.
[2] Escofier, B., & Pagès, J. (1988). Analyses factorielles multiples. Paris:
Dunod.
[3] Johnson R.A., & Wichern D.W. (2002). Applied multivariate statistical anal-
ysis. Upper Saddle River (NJ): Prentice-Hall.
[4] Naes T., & Risvik E. (Eds.) (1996). Multivariate analysis of data in sensory
science. New York: Elsevier.
[5] Weller S.C., & Romney A.K., (1990). Metric scaling: Correspondence analy-
sis. Thousand Oaks (CA): Sage.