Dear Hans,
> I have 2 groups of subjects and I would like to compare them with a
> t-test. I am interested in using Randomise since it generates the
> near-exact null distribution for the test statistic in a non-
> parametric way. My concern is about the first level variances (the
> varcopes). Are these variances accounted for in randomise or does it
> just consider between subjects variance?
Randomise doesn't explicitly use the varcopes, but that does _not_
mean that it only considers between subject variance. The estimates
you get from the first level are precisely estimates, i.e. they
represent that particular subjects response with some first level
error added on to it. Hence, when you are using only the copes you
will enter in the responses of the different subjects (there is your
between subject variance), and each of these responses will have some
error/uncertainty (there is your with subject variance). So, even when
performing an ordinary least squares (OLS) analysis using only the
copes the first level error is taken into account.
The times when it is important to know something about the first level
errors (i.e. the 1st level varcopes) is if/when the first level errors
are vastly different between different subjects. This could happen
e.g. of you want to look at correct responses in some memory task and
subject 1 has 100 correct responses whereas subject 2 has 1 correct
response. Then we can estimate subject 1's response with much higher
precision and in a 2nd level analysis we would want to giver higher
weight to that subject by weighting the response with the 1st level
varcope.
> The other option that I can use is the standard FSL mixed effects
> group analysis that will account for cope, varcope and dof plus the
> between subjects variance.
If you are talking about 2nd level fMRI data I think either way is
fine. We mainly use Randomise for the cases where we cannot assume
that the 2nd level errors are normal distributed, such as e.g. VBM or
TBSS. It is often also the case that Randomise is a little more
sensitive when one has low degrees of freedom, in which case GRFT
tends to be a little conservative.
Good luck Jesper
|