You are on page 1of 15

Factorial Manova

Harmaya University
March, 2019
Multivariate Factorial Design
 Number of DV’s = q
 Number of factors (IV’s) = 2, plus their
interaction: A, B, AxB
T=B+W

Bfactor 1 Binteraction
Bfactor 2
 Three different hypotheses to test.
Significance Tests
 Use Wilk’s  to test each main factor and
the interaction:
W
h 
Bh  W
 h denotes the particular null hypothesis
being tested.
If effects are significant?
 Follow-ups should be used to try and
understand the source of differences
 Especially if there are interactions!
 Look at the DVs individually, their
correlations, and use stepdown tests
 Consider discriminant analysis if the
interaction is significant (may also be
helpful if only main effects sig).
Example (Tatsuoka, 1988)
 Speed and accuracy scores from 60 subjects learning
typing
 Factor A: method (J=2)
 Traditional versus experimental method
 Factor B: condition (K=3)
 Three levels, distributed towards massed practice (1,2,3)
 Statistical Approach
 Data cleaning, checking distributions, correlations and
descriptive statistics (r=.749 for the two DVs)
 Plots of means
 HOC
 Omnibus and Follow-ups
 Stepdown and Discriminant analysis
Speed (words per minute)
45

40

35

30

25
traditional
experim ental
20

15

10

0
1 2 3
condition
Accuracy
30

25

20

traditional
15
experim ental

10

0
1 2 3
condition
Research Questions
 Main effect for Method:
 Averaging across conditions, is there a difference in
the two methods (on the composite score created for
the DV’s?)
 Main effect for Condition:
 Averaging across the two methods, is there a
difference in the three conditions (on the composite
scores created for the DV’s?)
 Interaction effect:
 Does the effect of condition differ as a function of
method (and vice versa)?
HOC?
 Results of a Box-M test indicated that
homogeneity of covariance for the 2x3=6
cells (groups) was safely assumed, Box-
M=23.41, F(15, 15949)=1.413, p=.131.
Omnibus tests?
 Manova brief summary
Effect  p partial 2
MxC .63609 .000 .202
Cond. .13415 .000 .634
Method .90771 .077 .092
Findings…
 There is a significant interaction between method
of learning and practice time, Wilk’s =.63609,
F(4, 106)=6.73, p=.000, partial 2 = .202.
 Since the interaction is significant, we go on to
try and fully interpret the interaction. This could
be through univariate tests focusing on
interaction; stepdown tests for the interaction; and
interaction contrasts.
 In many situations with factorial interaction,
better to proceed to discriminant analysis.
If there were no interaction…
 If there were no interaction, we would report:
 Averaging across methods, there is evidence for a differential
learning effect across the three conditions, Wilks =.13415,
F(4,106)=45.85, p=.000, partial 2 = .634.
 Averaging across conditions, there is no statistical evidence for
a difference between experimental versus traditional methods,
Wilks =.13415, F(2,53)=2.69, p=.077, partial 2 = .092.
 Note: These interpretations make no sense for our
problem because of the significant interaction!
 For follow-ups (if no interaction), we would consider the
pattern of means, the univariate and step-down tests, and
perhaps the contrast results.
Stepdown Approach
 Since the DV’s are correlated, we’ll use the Roy-Bargman stepdown
F-tests to try to understand the interaction effect. We’ll assume
speed is more important (given automatic spell checks). The order of
importance is the order specified on the MANOVA statement.
 First analysis is from a univariate analysis of variance just on speed.
Notice result is the same as in the univariate section just above it.
 There are statistically significant interaction effects between method
and condition on speed, F(2,54)=6.24, p=.004, 2=.188.
 Second analysis is from a univariate analysis of covariance, with
DV=accuracy, COV=speed.
 The effect of speed is removed from the test on interaction differences
in accuracy across combinations of method and condition.
 After adjusting for the effects of speed, a statistically significant
interaction effect on accuracy remains across the combinations of
method and practice condition, stepdown F(2,53)=7.34, p=.002,
2=.270.
Effect size in the Stepdown tests…
 For first variable, use 2 from the univariate
analysis.
 2 = SSB/SST (where SST=SSB+SSW)
 For second variable (effect adjusted for the first
variable:
 Compute SSB=Hypoth MS * Hypoth DF
 Compute SSW=Error MS * Error DF
 Compute SST = SSB + SSW
 2 = SSB/SST
Contrasts…
 Orthogonal Contrasts included for this
example, but could use any kind.
 See written handout for calculation and
interpretation.

You might also like