Anna,
> I would like to ask your opinion about a study design combining the
> RFX analysis with SnPM.
>
> In the PET/SPECT studies I have been analysing we usually only acquire
> 2 replications per condition. My question therefore is which of the
> following is a better analysis procedure to use.
>
> 1. Create adj mean images out of the 2 replications per condition
> and then perform a SnPM analysis.
>
> 2. Do an individual subject analysis using the replications per
> condition (but then only have 3df per subject) and then entering the
> con*.img into a SnPM analysis.
>
> 3. Do an individual SnPM analysis and then an SnPM analysis on the
> hole group
>
> 4. Use all scans and replications and use FDR ?
There are many different issues flying around here, so let me
tackle them one by one.
RFX:
The proper way to do random effects in SPM99 is to reduce all effects
to one image per subject, and then anlayze those images with one-sample
t test (or two-sample t if you have two groups.) Equivalently you can
reduce the data to one scan per condition per subject, and then use a
paired t test.
To get these images with PET data the easiest thing is probably to use
adj mean, but you could also fit one big model where each subject
is their own group and obtain the same result (with the right
contrasts).
Note that because of the greater measurement error (intrasubject
variabilty) in (S)PE(C)T, the results from the proper RFX analysis and
a traditional model (where you include each replication) will probably
be similar.
SnPM vs SPM (for RFX):
If all you're using SPM for is a one sample, two sample or paired t-test,
then SnPM is dead easy to use (and quick!). I always compare my small
group SPM results to SnPM (and usually find greater sensitivity with no
variance smoothing, and always find greater sensitivity with variance
smoothing).
(Remember that SnPM only gives you Familywise-Error-Rate-corrected
thresholds and FWER-corrected p-values, just like SPM99's random field
theory inference).
FDR vs FWER:
FDR is a completely different approach to the multiple comparisons
problem. It is generic with respect to how you create a statistic
image as long as you can create p-values. Unfortunately, SnPM currently
doesn't create uncorrected p-values, so you can't apply FDR on
SnPM results.
So, to answer your question I'd say, go for #1; this will use the most
powerful group methods that give FWER-corrected inferences. However, I
would also see how this compares with doing #1 with SPM RFX and looking
at FDR-corrected inferences; the up side with this second approach is
that you'll be applying a more powerful multiple comparisons proceedure,
but you'll be dependent upon the usual parametric assumptions.
Hope this helps!
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|