Вы находитесь на странице: 1из 24

MULTIVARIATE ANALYSIS OF VARIANCE

(MANOVA)

MANOVA
Previously, we explored the use of
analysis of variance to compare
groups on a single dependent
variable.
In many research situations, however,
we are interested in comparing groups
on a range of different characteristics.
This is quite common in clinical
research, where the focus is on the
evaluation of the impact of an
intervention on a variety of outcome
measures (e.g. anxiety, depression,

MANOVA
Multivariate analysis of variance
(MANOVA) is an extension of analysis
of variance for use when you have
more than one dependent variable.
These dependent variables should be
related in some way, or there should
be some conceptual reason for
considering them together.
MANOVA compares the groups and
tells
you
whether
the
mean
differences between the groups on the
combination of dependent variables

MANOVA
MANOVA creates a new summary
dependent variable, which is a linear
combination of each of your original
dependent variables.
It then performs an analysis of
variance using this new combined
dependent
variable
(composite
dependent variable).
MANOVA will tell you if there is a
significant difference between your
groups on this composite dependent
variable;

MANOVA
Why not just conduct a series of
ANOVAs
separately
for
each
dependent variable?
By conducting a whole series of
analyses you run the risk of an
'inflated Type 1 error'.
This means that the more analyses
you run, the more likely you are to find
a significant result, even if in reality
there are no differences between your
groups.
The advantage of using MANOVA is

MANOVA
If separate ANOVAs are conducted on
each dependent variable, then any
relationship between DVs is ignored
we lose information about any
correlations that might exist between
DVs.
By running all DVs together, MANOVA
takes into account of the relationship
between outcome variables.
MANOVA has the power to detect
group differences within IV along a
combination of dimensions.

MANOVA
MANOVA is a much more complex set
of procedures, and it has a number of
additional assumptions that must be
met.
If you have a number of dependent
variables, you can still perform a
series of ANOVAs separately for each
dependent variable.
If you choose to do this, you might like
to reduce the risk of a Type 1 error by
setting a more stringent alpha value.
One way to control for the Type 1 error

MANOVA
To do this, you divide your normal
alpha value (typically .05) by the
number of tests that you intend to
perform.
If there are three dependent variables,
you would divide .05 by 3 (which
equals .017 after rounding) and you
would use this new value as your cutoff.
Differences between your groups
would need a probability value of less
than .017 before you could consider

MANOVA
To do this, you divide your normal
alpha value (typically .05) by the
number of tests that you intend to
perform.
If there are three dependent variables,
you would divide .05 by 3 (which
equals .017 after rounding) and you
would use this new value as your cutoff.
Differences between your groups
would need a probability value of less
than ,017 before you could consider

MANOVA
MANOVA can be used in one-way, twoway and higher-order factorial designs
(with multiple independent variables)
and when using analysis of covariance
(controlling for an additional variable).

MANOVA (Example)
In this example, the difference
between males and females on a
number of measures of wellbeing is
explored. These include a measure of
negative mood (Negative Affect scale),
positive mood (Positive Affect scale)
and perceived stress (Total Perceived
Stress scale).

MANOVA (Example)
Research question: Do males and
females differ in terms of overall
wellbeing? Are males better adjusted
than females in terms of their positive
and negative mood states and levels
of perceived stress?
What you need: One-way MANOVA
One categorical, independent
variable (e.g. GENDER); and
Two
or
more
continuous,
dependent
variables
(e.g.
negative affect, positive affect,

MANOVA (Example)
MANOVA can also be extended to twoway
and
higher-order
designs
involving two or more categorical,
independent variables.
What it does: Compares two or more
groups in terms of their means on a
group of dependent variables. Tests
the null hypothesis that the population
means on a set of dependent variables
do not vary across different levels of a
factor or grouping variable.

MANOVA Assumption Testing


Sample size
You need to have more cases in each
cell than you have dependent
variables.
Having a larger sample can also help
you 'get away with' violations of some
of the other assumptions (e.g.
normality).
The minimum required number of
cases in each cell in this example is
three giving you the minimum total
sample size of 18.

MANOVA Assumption Testing


Normality
Although the significance tests of
MANOVA are based on the multivariate
normal distribution, in practice it is
reasonably robust to modest violations
of normality
According to Tabachnick and Fidell
(2007, p. 251), a sample size of at
least 20 in each cell should ensure
'robustness'.

MANOVA Assumption Testing


Outliers
MANOVA is quite sensitive to outliers
(i.e. data points or scores that are
different from the remainder of the
scores).
You need to check for univariate
outliers (for each of the dependent
variables separately) and multivariate
outliers.

MANOVA Assumption Testing


Linearity
This assumption refers to the presence
of a straight-line relationship between
each pair of your dependent variables.
This can be assessed in a number of
ways, the most straight forward of
which is to generate a matrix of
scatterplots between each pair of your
variables, separately for each group
(male and female separately in this
example)

MANOVA Assumption Testing


Multicollinearity and singularity
MANOVA works best when the
dependent variables are only
moderately correlated.
With low correlations, you should
consider running separate univariate
analysis of variance for your various
dependent variables.
When the dependent variables are
highly correlated, this is referred to as

MANOVA Assumption Testing


Multicollinearity and singularity
(cont.)
While there are quite sophisticated
ways of checking for multicollinearity,
the simplest way is to run Correlation
and to check the strength of the
correlations among your dependent
variables.
Correlations up around .8 or .9 are
reason for concern. If you find any of
these, you may need to consider
removing
one
of
the
strongly

MANOVA Assumption Testing


Homogeneity of variance-covariance
matrices
Test of this assumption is generated as
part of MANOVA output.
The test used to assess this is Box's M
Test of Equality of Covariance
Matrices.

MANOVA
Run MANOVA using the data provided.

MANOVA (Sample report)


A one-way between-groups multivariate
analysis of variance was performed to
investigate
sex
differences
in
psychological
wellbeing.
Three
dependent variables were used: positive
affect, negative affect and perceived
stress. The independent variable was
gender. Preliminary assumption testing
was conducted to check for
normality, linearity, univariate and
multivariate outliers, homogeneity of
variance-covariance
matrices, and multicollinearity, with no

MANOVA (Sample report) (cont.)


There was a statistically significant
difference between males and females
on the combined dependent variables, F
(3, 428) = 3.57, p = ,014; Wilks' Lambda
= 98; partial eta squared = .02. When
the results for the dependent variables
were
considered
separately,
the
only
difference
to
reach
statistical
significance, using a Bonferroni adjusted
alpha level of ,017, was perceived stress,
F (1, 430) = 8.34, p = ,004, partial eta
squared = .02. An inspection of the

Вам также может понравиться