We revisited the unbalanced issue in VBM recently with a much larger sample
database, where we also looked at the sources of non-Gaussianity, the impact
of smoothing, and the approach of transforming the images to limit
non-Gaussianity:
http://dx.doi.org/10.1016/j.neuroimage.2006.11.037
We also looked at the idea of calibrating a statistic from a normal sample
with the one subject problem in mind. We take the statistic suggested by
Rick Henson's post, and calibrate its degrees of freedom to take kurtosis
and skewness into account.
http://dx.doi.org/10.1002/hbm.20332
For both approaches, you need a reasonably large sample of normal subjects.
This is not surprising, giving the size of MR images.
Best,
R. Viviani
----- Original Message -----
From: "Ged Ridgway" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Thursday, April 12, 2007 1:05 PM
Subject: Comparison of one subject with a group of subjects in VBM
: Hi Shary,
:
: I think this is dodgy, but (I hope you don't mind) I've CCed the list
: to get other opinions...
:
: I seem to remember seeing something very similar (e.g. using a cohort
: of controls to give a z-score to an individual subject) but I can't
: remember the reference off the top of my head, and having failed to
: find it on my first page of Google Scholar results, I'm afraid I'll
: leave you to look ;-)
:
: The reason I think it's a bad idea, is because Salmond et al found
: that unbalanced group sizes made the analysis very fragile to
: non-Gaussianity etc. Since a group with a single subject is as
: unbalanced as you can get, it seems possibly that you might get many
: false positives from only a slight departure from Gaussianity. But I'm
: not expert, so see what other people have to say...
:
: http://dx.doi.org/10.1016/S1053-8119(02)91153-5
:
: Best,
: Ged.
:
: P.S. I think you're right about your interpretation of the conjunction
: null below.
:
: Shahryar (Shary) Rafi-Tari wrote:
: > Hi Ged,
: >
: > I have a question about comparing one subject (patient) with a a group
: > of controls (n=20). Here are the questions:
: >
: > 1. Can I do t-test with the one patient in group 1 and the 20 healthy in
: > grouop 2?
: >
: >
: > 2. If I needed to compare three groups: group 1 with n=1 , group 2 with
: > n=1 and group 3 n=20. Groups 1 and 2 are made of one patient each.
: > Group 3 is made of healthy controls. In this case can I simply do full
: > factorial (one factor with 3 levels and three cells). Then, I do
: > conjunction null available in SPM5. If I find a region of difference
: > between the two patients and the controls, I can say that there is a
: > significant difference and overlaping areas of differences between these
: > two patients and the controls.
: >
: >
: > 3. Do you know of any good reference for the comparison of single
: > subject and a group of subjects?
: >
: >
: > Thanks,
: > Shary
: >
: >
:
|