Hello,
I am analyzing a within-subject design with two factors (A and B)
Following up on previous interesting discussions on this list about the
full factorial design in SPM, I analyzed the appropriate first-level
contrast images with either the SPM full factorial design or with a
classical repeated measures ANOVA.
I am attaching the scatter plot for the F statistics for the main
effects of A and B and for the AxB interaction in both repmea ANOVA and
SPM (one dot = one voxel).
What strikes me here is the fact that while for the main effect the two
models generally agree (correlation between F values in the repmea and
SPM models = 0.93 and 0.95 for main effects of A and B, respectively),
the departure is striking for the AXB interaction (correlation = -0.1)!
The large departure between the F test for the AxB interaction in repmea
ANOVA and SPM factorial design appears to be due to large differences in
the mean squares (MS) computed in the two models. The following table
gives the repmea ANOVA - factorial SPM correlation between the various
relevant numerical quantities (num = numerator; den = denominator):
MS num MS den
A: 0.985 0.852
B: 0.998 0.874
AxB: 0.182 0.929
Note that the second column of this table equals the correlations
between the three error terms for the repeated measures ANOVA model on
the one hand, and the ResMS.img error term used in SPM, on the other.
Did I get something terribly wrong here?
Finally, my question: should the SPM factorial design be trusted when it
comes to testing the significance of an interaction?
One additional question: say I want to consider the repeated measures
ANOVA approach. Would it be appropriate to compute a separate FWE
correction for each of the tested effects based on the residual mean
square map for the effect at hand?
Thank you,
Bruno
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Bruno L. Giordano, PhD
Voice Neurocognition Laboratory (CCNi, Glasgow Univ) &
Music Research Department (McGill Univ, Montréal)
Www: http://www.brunolgiordano.net
|