Dear Jon,
It is not particuarly unusual to get different results when
comparing fixed and random
effects analyses. Random effects analyses tend to be discussed as 'more
stringent, more
conservative, more...everything' analyses - the implicit assumption
being that you're realy
only changing significances. So it's no surprise that you'd expect, when
lowering the
threshold, to see your activation peaks pop up again.
However, RFX analyses are qualitatively different from FFX analyses,
because the error
maps are different between the two. As the error maps are used to
generate your SPMs, it's no
wonder that you may see differences. Add to this the fact that as you
drop the significance,
your criteria for rejecting false positives changes and...well, you get
the idea.
See: http://www.mailbase.ac.uk/lists/spm/1999-10/0000.html for a bit
more detail.
Best,
Dave McG.
> Thanks for getting back to me.
>
> Masking using the segmented cortex as a template does sound
> like an interesting proposition - and perhaps more valid than a
> post-hoc ROI analysis. Though I do have some ideas
> about which areas might show a BOLD response, I was just
> having an initial "look see" at the data using random effects.
>
> Prior to performing the random effects (RE) analysis, I had a
> look with fixed effect (FE) analysis - which gave several significant
> areas of activation (corrected-p). The finding that
> these areas vanished with the RE analysis wasn't unexpected,
> but when I dropped the significance level I didn't obtain
> the same pattern of activation (as in the FE analysis). Is this
> to be expected? I guess it may have been a consequence of the
> FE results being biased by a single subject (?) - or am I
> missing the point?
>
> Is there a good reference for RE/FE analysis?
>
> Thanks again,
>
> Jon.
> _____________________________________________________
> Jonathan Brooks Ph.D. (Research Fellow)
> Magnetic Resonance and Image Analysis Research Centre
> University of Liverpool, Pembroke Place, L69 3BX, UK
> tel: +44 151 794 5629 fax: +44 151 794 5635
>
> On Sat, 3 Jun 2000, Foucher Jack wrote:
>
> >Dear Jon, Karl and others
> >
> >>
> >> Dear Jon,
> >>
> >> > I have recently been experimenting with a random effects analysis
of
> >> > several trials in an fMRI study. I have 18 subjects, so I guess
this is
> >> > the right analysis to use. Basically, I was wondering whether I
should
> >> > use a corrected height threshold in the generation of the SPMs?
If I do
> >> > I get no supra-threshold clusters.
> >> >
> >> > Is it appropriate to use an uncorrected p, given that the random
> >> > effects analysis is quite stringent? Is there any general
consensus on
> >> > the best approach to choosing p-levels in this situation?
> >>
> >> I am afraid exactly the same inference criteria apply to first and
> >> second level analyses. I would think about any anatomical priors
that
> >> could be used to provede small volume correction to the p values.
One
> >> point you might want to take forward is that inter-subject
differences
> >> may include anatomical variations in the response profile.
Increasing
> >> the smoothing (of the con??? or beta???.img) prior to the 2nd level
> >> analysis might improve your sensitivity (e.g. 8mm FWHM).
> >>
> >> I hope this helps - Karl
> >
> >I hope that the following proposition won't seems to odd !
> >
> >Let say that you have no clear idea of which cortical region should
activate
> >except that you are looking for cortical activation only.
> >Why not using a mask constructed on the basis of gray matter
segmentation for
> >corrected p ?
> >You will just have to choose a threshold from your gray matter
segmentation map
> >in order to put the above values to 1 and the others to 0. Than you
could
> >eliminate the basal ganglia by a logical routine setting to 0 every
voxel
> >between a certain coordinate.
> >I tried this approach once just for testing the feasibility, and it
allowed me
> >to use less conservative threshold (although the gain was not
proportional to
> >the volume loss - see Matthew's pages on SVC for reasons to that).
> >
> >Does this sounds reasonable ? Did anyone else tried this, and what
difficulty
> >did they have to cope with ? The most critical point I faced was the
threshold
> >do use (as far as I remember, 200 seems to be resonable). Any advice
on that
> >point ?
> >
> >Since I did try this at the time of SPM99 beta and using Matthew's
SVC routines
> >(that gave the uncorrected p to use for equivalent corrected p), I
faced a
> >problem of representation : how to display the only activation
falling within
> >the mask (except by removing them 'by hand' from the T.img using the
same mask
> >- I haven't tried that) ?
> >
> >Thanks for any input
> >Sincerely
> >
> >
> >
> >Jack
> >
> >_________________________________________________________________
> >| Jack Foucher Universite Louis
Pasteur |
> >| Institut de Physique Biologique UPRES-A 7004 du
CNRS |
> >| 4 rue Kirschleger Tel: 33 (0)3 88 77 89
90 |
> >| 67085 STRASBOURG Fax: 33 (0)3 88 37 14
97 |
> >|
France
|
> >| Faster E-mail:
[log in to unmask] |
> >| Other
[log in to unmask] |
>
>|_______________________________________________________________
|
> >
> >
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|