Dear Thomas,
> >> How can we do the subtraction A - B when A and B are acquired in
> >> different sessions from one subject ? I heard that there could be a
> >> problem with the global scaling if A and B are not acquired in the same
> >> session.
> >
> >At the second level this should not be a problem (the data are globally
> >normalised at the first level). To compare A and B, relative to rest,
> >simply perform a two sample t test on the contrasts [1 -1] and [1 -1]
> >from the first level (i.e. A-R - B-R = A-B).
> >
>
> thank you for your help. One question regarding the subtraction A-R -
> B-R: What is the difference here between performing the two-sample t
> test on the contrasts [1 -1] and [1 -1] (from the first level analysis)
> or performing it on the first level contrasts [1 0] and [0 1] ?
It is possible that the contrasts [1 0] and [1 0] are inestimable and
would therefore be wildy variable. This is because the sum of the two
regressors modelling A and R may sum to unity and be collinear with the
constant term. For example paramter estimates for A, R and the
constant could be
2.2 -2.2 100
or they could be
102.2 97.8 -100
- the fit would be the same.
Using a contrast (i.e. [1 -1] = 4.4 in both cases) that is estimated
efficiently reduces the error variance at the second level and avoids
these problems.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|