Dear all,
I have not received a reply to my previous question, and need one
urgently, so I'll try again.
I have an fMRI experiment with 12 subjects, three runs, and three
activation blocks within each run, alternating with three rest blocks.
Each run represents a separate task, and each block represents a
different stimulus type. Using SPM, I modelled each subject with a
separate fixed effects model, treating each run as a separate session,
and each activation block as a separate trial. I then created a random
effects model incorporating the twelve subjects.
I thought that a random effects analysis would allow me to model
correctly the variance across runs, since I would have to concatenate
runs if I ran a fixed effects model across subjects. However, I now
need to perform a conjunction analysis, which, as far as I can tell, is
not possible at the second level in SPM-99. I am considering
switching to SPM 2b, but I
am wondering how much more conservative running a conjunction
level at the second level is than running one at the first. I am
concerned that I do not have enough power to generate the desired
effects in a second-level conjunction analysis. When using a one-
sample t-test at the second level, I observe activation at p<0.01 and 30
voxels. Why are the p-values so much less conservative in RFX than
fixed effects models?
One option is to perform a conjunction analysis at the fixed effects
level, but I'm not sure that concatenating all the runs is statistically
acceptable. In addition, I only want to analyze two activation blocks out
of the three (two out of three stimulus types) in each run at this point,
and I'm
concerned that modelling the third stimulus block in each run in the
design matrix would significantly change the variance and degrees of
freedom such that the activation observed when comparing the other
two stimulus types would not be entirely reliable. I do not encounter
this problem at the second level, because the degrees of freedom are
determined by the number of subjects rather than the number of
scans.
If I do perform a conjunction at the fixed effects level, what does the
orthogonalization order mean? As I understand it, "orthogonal" refers
to independent contrasts, and I wish to do a contrast between two
activation conditions that have separate baselines, which I assume
makes the contrasts independent. I'm not sure why the
orthogonalization order should influence the activation pattern for the
conjunction of two independent contrasts. As I understand it, SPM-99
calculates the conjunction based on the lowest t-value of the
orthogonal contrasts within a conjunction. If this is true, why should
the orthogonalization order matter?
Additionally, I need clarification on what the p-value of the conjunction
represents. If I have two contrasts that I want to input into a
conjunction, and the p-values are, in each individual contrast, p<.1
and p<.1, I think that the resulting p-value for the conjunction would be
0.1 x 0.1. However, what is the signifance of the activation for each
individual contrasts that I input into the conjunction? In other words,
would activation produced by both contrasts within the conjunction
only be significant at p<.1? This would certainly influence my
interpretation of the results.
Since I can't perform a conjunction analysis in random effects, is it
appropriate to mask one contrast by the other, and does this have the
same effect as a conjunction? From the archives, it appears that the
masking option assumes that contrasts are not orthogonal, but I
might be wrong about this.
My final question: if I am performing a conjunction across contrasts of
conditions performed in the same run, versus a conjunction across
contrasts of conditions subjects performed in different runs, are both
of these conjunctions still composed of orthogonal contrasts?
Does anyone have any thoughts or advice?
Thanks in advance,
Jenna Gelfand
UCLA Brain Mapping Center
|