Uses of the MANOVA Procedure
Top of Page
StatMan Toolbar
MANOVA procedures are multivariate, significance test analogues of various univariate ANOVA
experimental designs. MANOVA, as with its univariate counterparts typically involve random
assignment of participants to levels of one or more nominal independent variables; however, all
participants are measured on several continuous dependent variables.
There are three basic variations of MANOVA:
Hotelling's T: This is the MANOVA analogue of the two group T-test situation; in other
words, one dichotomous independent variable, and multiple dependent variables.
One-Way MANOVA: This is the MANOVA analogue of the one-way F situation; in other
words, one multi-level nominal independent variable, and multiple dependent variables.
Factorial MANOVA: This is the MANOVA analogue of the factorial ANOVA design; in other
words, multiple nominal independent variables, and multiple dependent variables.
While all of the above MANOVA variations are used in somewhat different applications, they all
have one feature in common: they form linear combinations of the dependent variables which best
discriminate among the groups in the particular experimental design. In other words, MANOVA is a
test of the significance of group differences in some m-dimensional space where each dimension is
defined by linear combinations of the original set of dependent variables. This relationship will be
represented for each design in the following sections.
Why does MANOVA have to be so complicated? Why can't you just do separate ANOVA's for each
dependent variables rather than form linear combinations of all dependent variables? There are
several primary concerns to consider. First, very often and most likely, the dependent variables will
be correlated with each other; thus, your findings from separate ANOVA's will be redundant and
difficult to integrate. Second, the family-wise error rate becomes high; the odds of finding
something significant simply because of chance rises with repeated use of the same sample of
data.
Two Group Discrimination: Hotelling's T
Top of Page
StatMan Toolbar
The first diagram below represents the situation of two groups on a univariate dimension (y) where
population differences are "real." The second diagram represents "real" population differences
between group centroids--points defined by the y1 and y2 means for each group--of the two groups
on y1 and y2.
Now the task becomes the
formation of linear combinations
of y1 and y2 in order to (a) correct
for redundancy and (b) maximally
discriminate amongst the groups.
The discriminant weights of this
composite (d1 and d2) are
mathematically derived such that
if we were to calculate a
composite score for each
participant, then the weights yield
composite scores that will give the
maximum F Ratio value.
Essentially, this minimizes withingroup variance and maximizes
between-group variance. The new
composite dimension (a linear
combination) can be plotted
against y1 and y2 to examine the
univariate frequency distributions of the groups were overlap is minimized:
With Hotelling's T we can construct only one dimension, and the test is basically checking whether
the differences between the centroids are significantly different on the composite dimension. The
statistic is distributed as an F, with q and n1+n2-q-1 degrees of freedom, where q is the number of
dependent variables.
The One-Way MANOVA
Top of Page
StatMan Toolbar
This MANOVA application is simply an extention of Hotelling's T to more than two groups. We have
a single independent variable with more than two levels, and each participant is measured on
several dependent variables. Rather than being limited to constructing a single dimension we can,
the maximum number of dimensions that can be constructed in always the smaller of the following
two values: (a) the number of groups minus one, or (b) the number of continuous dependent
variables. The test statistic in this situation (e.g. Wilk's Lambda) determines whether the whole set
of composites can significantly discriminate the groups.
If more than one composite is constructed by the MANOVA procedure, all composite dimensions
are uncorrelated. Therefore, information offered by the first composite is independent of that offered
by the second, and so on. This occurs because the construction of each composite is built by using
the residual of the previous composite.
Factorial MANOVA
Top of Page
StatMan Toolbar
This MANOVA application is truly a super-factorial design. Participants are nested under multiple
treatment combinations (multiple independent variables) and are measured using several
continuous dependent variables. It maintains the advantages that factorial designs have over
simple one-way designs (i.e. interaction information) plus it can form composite dimensions specific
to each effect in the design which optimally separate the groups being evaluated for each effect. In
other words, factorial MANOVA involves the calculation of several sets of composite variables and
each set is specific to a particular effect.
For factorial MANOVA, we can form a total number of uncorrelated linear combinations of all
dependent variables for each effect that is equal to the smaller of the number of degrees of
freedom for that effect (Number of cells minus one) or the number of dependent variables. For
example, in a 3 x 4 design with 20 dependent variables we can examine the following:
The A Effect: As in any factorial design, we examine the A main effect by collapsing across
all other variables. Since we have three levels of A (df A = 2), we can form a maximum of 2
independent linear composite dimensions (like those in Hotelling's T). Then calculate
Lambda to evaluate whether these two dimensions as a set significantly discriminate the A
groups.
The B Effect: By the same logic as above, we can form a total of 3 linear combinations.
Again, calculate Lambda to determine whether the B groups differ significantly on all three
dimensions.
The AB Interaction: Since the degrees of freedom for the interaction equals six, we can
calculate six independent composite dimensions on which we can calculate Lambda to
determine if we have a significant interaction.
Solving the RMD Problem: The Poor Technique
Top of Page
StatMan Toolbar
The MANOVA procedure does have another application. The Repeated Measures design and the
Mixed design both assume compound symmetry (i.e. the variances are equal among groups and
that the covariances among levels are equal). When this is violated, each independent variable can
be treated as a separate dependent variable (using dummy coding). A MANOVA is then performed
in order to look for effects, and a significant linear combination of variables would indicate a
significant difference between the original levels. This can also be extended to the violation of
compound symmetry for several dependent variables; in this case, the procedure is known as a
doubly multivariate analysis.
StatMan Topic Index
Previous Topic
Top of Page
Take Quiz
Next Topic
