SPM Project wrote:
> Hi Will, how are you doing? I wanted to forward you this posting I had put up last week. You
> seem to be very knowledgeable about fMRI analyses (I've read many of your postings), and I was
> hoping you could shed some light on my situation for me, and if you had any suggestions on how to
> better analyze this.
>
> Also, I know people apply AR(1) serial correlations on the individual fixed-effects level. But
> do we apply it on a large fixed-effects level, with a number of subjects (under 12 subjects of
> course)?
Yes, you would.
Each subjects data should go into a separate session, and SPM will
make session-specific adjustments for the autocorrelation.
>
> Many thanks, Vy
>
> ---------- Forwarded message ---------- From: Vy N <[log in to unmask]
> <mailto:[log in to unmask]>> Date: Mar 4, 2006 11:02 AM Subject: No signifcant activations??
> Did I do something wrong..? To: [log in to unmask] <mailto:[log in to unmask]> Cc: Vy N
> <[log in to unmask] <mailto:[log in to unmask]>>
>
> Hey SPM Experts,
>
> I just ran some preprocessing and analysis on a small data set (7 patients and 4 controls - all
> children) and I am stumped! After doing so, I ran some contrasts and didn't get any significant
> activations whatsoever.
>
> The study has a pre- and post- therapy scan for the patients, whereas the controls are only
> scanned once. Painful stimuli is administered, followed by nonpainful stimuli, on a symptomatic
> body part as well as a nonsymptomatic body part . I was hoping to see significant activations
> when performing the contrast for (SP - NSP) - (NSP - NSNP), where S = symptomatic; NS =
> Nonsymptomatic; P = painful; and NP = nonpainful.
Should'nt this be: (S P - NS P) - (S NP - NS NP), the interaction. I expect you have done this and
that the above is a typo.
And I was hoping that after therapy, these
> significant activations would be reduced (or at least resemble the activation pattern in
> controls).
>
> It is a block design, with 4 sessions. Sessions 1 and 3 are on the Symptomatic Body Part, and
> Sessions 2 and 4 are on the Nonsymptomatic Body Part. Each session consists of 10 painful
> stimuli administration alternating with 10 nonpainful stimuli administration for 4 repetitions
> (80 TRs total).
>
> Preprocessing steps include: 1) realignment to first volume of first run - created mean image 2)
> segmentation 3) coregistering the grey matter with the mean epi 4) Normalization of grey matter
> to grey matter of CCHMC Pediatric template, then applying the parameters to the functional (3x3x3
> voxel sizes) 5) smoothing at an 8 mm kernel
>
> When I was just looking at each individual subject, I redid the same steps as above, but during
> realignment, I created all images + mean image to avoid getting the "images do not all have same
> orientation & voxel size" error message that I got before. And of course, no normalization was
> done.
>
> Because of the few number of subjects, I put all the data (7 patients pre-therapy/7 patients
> post-therapy/4 controls) into one big fixed effects model. Since there were four sessions, there
> was a total of 72 sessions (for anyone who had memory problems, I might be able to help - just
> ask!).
>
> Design Matrix: 1) fMRI -> specify design 2) TR = 4 3) Scans per session: 80 80 80 80 ..etc for a
> total of 72 sessions 4) How to specify - Scans 5) Are sessions replications: No 6) Basis Set: Hrf
> 7) Model interactions: No 8) # of conditions: 4 11) Duration: 10
>
> Motion parameters were included as regressors. The data was specified using the smoothed
> images, and I selected "none" for remove global effects. A high pass filter of 128 s was chosen
> (the default).
>
> For individual analyses, AR(1) serial correlations = yes. For the big fixed effects model, I
> chose "no" for the AR(1) serial correlations. The estimation took a lovely 19 hours! Isn't that
> grand?
>
> The contrasts were created for (SP - NSP) - (NSP - NSNP) for patients pre-therapy, controls, and
> patients post-therapy. The motion parameters were taken into account when creating the
> contrasts.
>
> For the p-value adjustments of the patient contrasts in the fixed effects model, I saw the
> results under a p-value of 0.001, 0.005 and 0.05 (scattered activations) - I figured they were
> noise for the most part, since when I applied either an FWE or FDR of 0.05, there were no
> significant activations pre- or post-therapy! Even when I did a contrast for the main effect of
> SP and applied FDR of 0.05, there were no activations! For the controls, there were significant
> activations for the difference contrast...Remember, AR(1) = no...
>
> Individually, in which AR(1) serial correlations = yes, there were no significant activations
> under an FWE/FDR of 0.05 for any subject!
>
> Can someone explain this?? Am I doing something wrong? Is there something else I should try?
> Any help would be greatly appreciated...
>
Its difficult to understand what's going on here but
I would try and isolate the problem by doing some more specific contrasts.
You mention the overall interaction in your group (ie. using the contrast I have written above but
repeated so it covers every session). An F-test here will
look for an average effect over all patients pre and post therapy and subjects. These could be
positive or negative interactions. Perhaps the effects have been averaged out.
You should look at group specific contrasts and also subject specific contrasts ie.
zero out all columns except for subject X and then see if you get the same effect
as when you did a separate analysis for that subject (with identical pre-processing).
Best,
Will.
> Many thanks in advance, Vy
--
William D. Penny
Wellcome Department of Imaging Neuroscience
University College London
12 Queen Square
London WC1N 3BG
Tel: 020 7833 7475
FAX: 020 7813 1420
Email: [log in to unmask]
URL: http://www.fil.ion.ucl.ac.uk/~wpenny/
|