General Linear Model

(GLM) Multivariate Analysis

AIM
GLM multivariate analysis
takes into account the
interrelation among
dependent variables and
analyzes the variables
simultaneously.

Checklist of
Requirements
Depending on the research question
and hypotheses to be tested, the
experiment can include or exclude
classification on independent
variables.
When independent variables are
included, there can be two or more
levels for each independent variable.
There should be two or more
dependent variables.

Assumptions
The observations must be independent
(i.e., responses among groups of
respondents should not be correlated).
The variancecovariance matrices must
be equal for all treatment groups.
The set of dependent variables must
follow a multivariate normal distribution
(i.e., any linear combination of the
dependent variables must follow a
normal distribution).

Example 1 GLM
Multivariate Analysis:
One-Sample Test

 Suppose that a sports psychologist has recorded

the following running times (in seconds) from
five men in four different events: 50-yd dash,
100-yd dash, 200-yd dash, and 300-yd dash.
The hypothesis to be tested is that the observed
sample comes from a population with specified
values for the means of the running events. For
illustrative purposes, the standard values are
taken to be 6 sec for the 50-yd dash, 12 sec for
the 100-yd dash, 25 sec for the 200-yd dash,
and 40 sec for the 300-yd dash. Because GLM
automatically tests the hypothesis that a set of
means is equal to zero, the normative values
must be subtracted from the observed scores,
and the hypothesis that the differences are zero
is tested.

The running times for the four

events are presented in the
following:

PROCEDURE

SPSS Output

Results and
Interpretation

 The Descriptive Statistics table presents means and standard

deviations for the four running events after their normative
values have been subtracted. The sample exceeds (i.e., is poorer
than) the norm for all four running events.
The Multivariate Tests table tests the hypothesis that the four
sample means do not differ from the specified set of constants.
Of the three multivariate tests (Pillais, Hotellings, Wilks, and
Roys), Pillais trace is the most powerful (i.e., the ability to
detect differences if they exist) and the most robust (i.e., the
significance level based on it is reasonably correct even when
the assumptions are violated). Because the observed significance
level is small (p < .05), the null hypothesis that the sample
means do not differ from the specified constants is rejected,
multivariate Pillai F(1,4) = 232.56, p< .05. When the null
hypothesis of no difference is rejected, it is often informative to
examine the univariate Tests of Between-Subjects Effects to
identify which variables yielded significant differences. The
univariate test results (Intercept) show that the means for all
running events differ significantly from their specified standards
(p< .01).

Example 2 GLM
Multivariate Analysis:
Two-Sample Test

 Following from the previous example,

suppose that the researcher also
recorded running times for the four
events for five women. The running
times for the two samples for the
four events are presented in the
following. The hypothesis that men
and women do not differ on the four
running events will be tested.

PROCEDURE
Create the four variables of FIFTY,
ONEHUND, TWOHUND, and
THREEHUN.
From the menu bar, click Analyze,
then General Linear Model, and
then Multivariate. The following
Multivariate window will open:

SPSS Output

Results and
Interpretation

Results and Interpretation

The Estimated Marginal Means table presents
the mean running times for men and women,
for each of the four running events after the
normative values have been subtracted. Both
samples exceed the norms for all four running
events.
The Multivariate Tests table (Pillais,
Hotellings, Wilks, and Roys) tests the
hypothesis that men and women do not differ
significantly on overall running time. The
significance level is based on the F distribution
with 4 and 5 degrees of freedom. The observed
significance levels for all four multivariate tests are
small (p < .001), so the null hypothesis that
men and women performed similarly on the
running events is rejected (e.g., multivariate Pillai
F(4,5) = 41.74, p < .001).

Results and Interpretation

The Tests of Between-Subjects Effects table
presents the univariate test of sex difference
for the individual running events. The obtained
F-values are equivalent to those obtained from
the one-way analysis of variance. In the case
where the comparison is between two
groups, the F-values are the square of the
two-sample t-values. The results indicate
significant sex differences in running times
for three of the four running events (FIFTY: male:
M = 2.60, female: M = 7.80; ONEHUND: male: M
= 7.20, female: M = 17.20; THREEHUN: male: M
= 23.60, female: M = 33.60); there is no
significant sex difference for the 200-yd dash.

Example 3 GLM: 2
2 Factorial Design

 In Example 2, the researcher used the GLM

two-sample test to test for sex differences in
running times across four running events.
Suppose that in addition to sex differences, the
researcher was also interested in whether the
subjects ethnicity (white vs. nonwhite) would
make a difference to their running times. In
particular, the researcher was interested in the
interaction between subjects sex and ethnicity
in influencing their running times. The running
times for the four groups (men-white, mennonwhite, womenwhite, women-nonwhite) for
the four events are presented in the following.

PROCEDURE

PROCEDURE
Create the four variables of FIFTY,
ONEHUND, TWOHUND, and
THREEHUN.
From the menu bar, click Analyze,
then General Linear Model, and
then Multivariate. The following
Multivariate window will open:

SPSS Output

Results and
Interpretation

Results and Interpretation

The Estimated Marginal Means table presents the
means of the overall running times for the four running
events as a function of subjects SEX, ETHNICITY, and
SEX*ETHNICITY interaction.
The Multivariate Tests table presents the multivariate
tests of significance for the main effects of the betweengroups variables of SEX and ETHNICITY, and the
SEX*ETHNICITY interaction. For all three effects, the
observed significance levels for the four multivariate tests
(Pillais, Wilks, Hotellings, and Roys) are small. Thus, their
associated null hypotheses (no sex difference, no ethnicity
difference, and no sex*ethnicity interaction) are rejected. The
results from the multivariate tests have been converted to
approximate F-values and can be interpreted in the same
way that F-values from oneway ANOVA are interpreted. The
multivariate tests for both SEX (multivariate Pillai F(4,13)
= 392.43, p < .001) and ETHNICITY (multivariate Pillai
F(4,13) = 6.25, p < .01) are statistically significant,
indicating that men (M = 11.1), when compared to
women (M = 22.32), and whites (M = 13.18), when
compared to nonwhites (M = 17.2), differed significantly in
their overall running times.

Results and Interpretation

The Tests of Between-Subjects Effects
can be examined for significant SEX and
ETHNICITY differences for each of the
four running events. For SEX, the results
show significant sex differences for all
four running events (p < .01) (FIFTY:
men: M = 1.10, women: M = 10.30;
ONEHUND: men: M = 5.70, women: M =
19.60; TWOHUND: men: M = 15.50,
women: M = 23.30; THREEHUN: men: M =
22.10,
women:
M
=
36.10).
For
ETHNICITY, the results show significant
ethnicity difference for only the FIFTY-yd dash
(white: M = 5.20, nonwhite: M = 6.20; F(1,16)
= 10.00, p < .01).

Results and Interpretation

Given that the SEX*ETHNICITY interaction
is significant, the Tests of BetweenSubjects Effects table can be examined
to see which of the four running events
are significantly affected by this interaction.
The
univariate
F-tests
for
the
SEX*ETHNICITY interaction are identical
to the F ratios generated from the
SEX*ETHNICITY interaction in a two-way
ANOVA. The univariate results show
significant SEX*ETHNICITY interaction
effect for three of the four running events
(p
<
.01).
Thus,
the
subjects
performances on the 50-yd dash, the 100yd dash, and the 300-yd dash are dependent
on the joint effects of their sex and ethnicity.

Results and Interpretation

To interpret the SEX*ETHNICITY
interactions, it would be useful to
graph the means of the four running
events (FIFTY, ONEHUND, TWOHUND,
and THREEHUN) from the estimated
marginal means table.

Results and
Interpretation

Results and
Interpretation
Using the 100-yd dash as an example,
the SEX*ETHNICITY interaction can
be interpreted as follows. The effect of
subjects sex on the running times in
the 100-yd dash is dependent on the
ethnicity of the subjects, such that for
men, nonwhites ran faster than whites;
for women, the effect is opposite with
whites running faster than nonwhites.

 Post hoc comparisons analysis can

also be used to identify which of the
four groups (male-white, male-nonwhite,
female-white, female-nonwhite)
generated from the SEX*ETHNICITY
interaction is significantly different
from each other in the 100-yd dash. The
following procedures will accomplish
this analysis.

Data Transformation

Data Transformation

Post Hoc Comparisons

Post Hoc Comparisons

Once the four levels (male-white,
male-nonwhite, female-white,
femalenonwhite) have been
created, Scheff post hoc
comparisons can be conducted to
test for differences (simple effects)
between these four levels.

SPSS Output

Results and
Interpretation
Results from the Scheff comparisons
indicate that white females (M = 17.20)
and nonwhite females (M = 22.00) ran
significantly slower than white males (M
= 7.20) and nonwhite males (M = 4.20).
Whereas there was no significant
difference in running times between white
males and nonwhite males, the results
show that nonwhite females ran
significantly slower than white females.