Hi, Darren,

I have a similar question about PET.

In your discussion, you said that we can use confiles to perform
2-sample t-test between
patient and control groups even there is only one subject in the patient group.

My thinking is we may collect a large sample of healthy subjects' PET
image, and we could perform 2-sample t-test between the healthy group
and patient group, even we only have one patient.

According to your discussion, I think this idea is workable. If it is,
I believe it has been working. Do you know there is any program
working based on this kind of idea?

Regards,
Eric.

2005/9/13, Darren Gitelman <[log in to unmask]>:
> Dear Maxwell:
> 
> I hope my answers are helpful.
> 
> Quoting Maxwell Boakye <[log in to unmask]>:
> 
> > Darren
> >
> > Can you help me with this
> >
> > What is the best way to compare activation areas common between two tasks.
> >
> > I have 8 subjects doing an active finger movement, 2 runs- analyzed and
> > have con files (8 from each run)
> > I have same 8 subjects doing passive finger movement, 2 runs-analyzed and
> > have con files (8 from each run)
> 
> I'm not quite sure I understand the task organization. Is this a block or
> event-related design in which you either have 1) subjects moving a finger or
> resting or 2) having the finger moved for them or resting? If so then further
> answers below. If not then I'll modify my responses based on your reply.
> 
> 
> > I need to know differences in brain activation between 2 tasks-
> > in normals and compare them to same tasks in
> > spinal cord patients. I currently have done only 1 spinal cord patient.
> > what is the best way to compare that to the data I have collected in
> > normals -for a grant. ie can I do conjuction analysis b/n 1 SCI patient and
> > 8 normals with respect to each task?
> 
> Comparing 8 subjects vs. 1 subject does violate assumptions of the general
> linear model. However, my understanding is that t-tests are generally fairly
> robust to such violations (although less robust for VBM). There can also be
> issues if the error variances are different between groups (random effects
> analyses assume these are close). The variance I believe is pooled across all
> subjects by SPM.
> 
> Here are some SPM list citations about this issue.
> http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind01&L=SPM&P=R319845&I=-3
> http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind04&L=SPM&P=R160144&I=-3
> http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind01&L=SPM&P=R122309&I=-3
> 
> If you want to know the differences between groups then using a 2 sample t-test
> you would compare the active con images from your normal group vs. the active
> con image for the SCI subject, and similarly for the passive images. If you
> want to examine group x condition interactions you would create for each
> subject (both control and SCI) an image from the contrast of the difference
> between active and passive runs. You could then put this into a similar 2
> sample t-test as above.
> 
> If you want to look at common activations you would take the same sets of
> con-images to the second level but you would have to use an ANOVA model with no
> constant term. You would then set up two contrasts [1 0] for the control
> subjects and [0 1] for the SCI subject. Selecting both contrasts would give you
> the conjunction. Particularly given the potential problems noted above you
> should use a conjunction null in order to say that both controls AND the SCI
> subject showed significant effects.
> 
> How about after collecting SCI data in
> > 8 patients
> 
> Same as above but now you have equal subjects so better designed. If you think
> the control and SCI groups have different underlying variances I believe you
> should choose non-sphericity correction, replications over subjects, error not
> correlated.
> 
> >
> > What is the best way to make the following comparisons
> > Within group comparison-common areas of activation b/n passive and active
> > task- is RFX better thsn conjunction analysis for this if I have enough
> > subjects
> 
> for common activations within subjects- you again would use a conjunction at the
> second level as above. However, as there are two images per subject they are
> necessarily correlated so again the conjunction null would be recommended.
> 
> >
> > difference in activation between passive task vs active task (?with 16 files
> > from both runs). can I use 2-sample t-test here?
> 
> Within subject groups you should try to end up with a single image per subject.
> So within each subject you would generate a contrast image of active-passive
> and forward this to the second level. Do a one-sample t-test and the contrast
> is either [1] = active-passive or [-1] which gives you passive-active.
> 
> >
> > Between Group analyses b/n normals and patients:
> > 8 normals vs 8 SCI (active movement con files, ? 16 from both runs)
> > 8 normals vs 8 SCI (passive movement con files) or
> > 8 normals (run1) vs 8 SCI (run 1) etc
> 
> see above.
> 
> > How does this change if I have only 1 or 2 SCI patients and I need the
> > preliminary analysis for grant
> 
> as above, this isn't the best statistical design but seems reasonable to use for
> preliminary data with the above caveats.
> >
> > Finally last question:
> > What is the easiest way to do conjunction analysis for large samples-
> > find the common areas activated by 8 subjects in each task in both runs
> 
> Conjunctions aren't really done this way- i.e., looking for common activations
> across subject groups with each group having only 1 subject in it. i.e., you
> couldn't take 8 subjects and do a conjunction of 8 single image contrasts. In
> this case what you want is a one-sample t-test. If you use the con images
> (these are the parameter estimates) what you are essentially saying is at each
> voxel is the effect size different than 0.
> 
> >
> > my understanding is that there is a way to do this using the confiles?
> 
> see above. you use the con images for analyses at the second level.
> 
> 
> Hope this was helpful. I've fowarded the message to the SPM group as well so
> that if I've made an error someone will hopefully point it out.
> 
> Regards,
> Darren
>