<Carles> <Soriano-Mas> wrote:
> I want to create a study-specific anatomical template. I have two groups
> of subjects (patients and controls). If I have twice more patients than
> controls (70 vs. 35), will you include all the subjects (107) for
> creating the template? Will this result in a worse normalization of
> controls?
I think it might well bias the registration in favour of the group
with more subjects. So what I'd do instead is make a weighted average
by creating separate patient and subject averages, and then averaging
the two of these. This is equivalent to:
(35*p1 + 35*p2 + ... + 35*p70 + 70*c1 + ... + 70*c35) / 105
i.e. the controls get weighted twice as heavily because there are half
as many. It's a bit like pretending that you have extra controls who
just happen to be identical to your existing ones...
I personally think this is a better approach than throwing away the
extra subjects of the larger group, but others might disagree, as the
above might bias things in favour of the control group, due to the
lower variability of the extra "pretend" controls compared to the
extra genuine patients.
If you throw extras away (randomly!), then you'll know that you aren't
biased in favour of either group, but then you'll only have 70
subjects, which may or may not be enough... Sorry to give such a
confusing answer, but I think the short answer is: it isn't clear cut.
Best,
Ged.
P.S. Somewhere in this reply, if John was writing it, he'd probably
have mentioned the fact that you *should* be using SPM5, where custom
templates are much less important ;-)
|