Hello Yufeng:

In general, I prefer not to calculate "statistics of statistics". That is,
I think it is better to use, as input to the across-subject statistical
analysis, some physical parameter that is directly related to the hypothesis
that is being tested. For example: amplitude of the response, area under
the IRF curve, % change relative to baseline, etc. As opposed to using
statistics such as correlation coefficients, t-stats, F-stats, z-scores, etc.,
which are only indirectly related to the hypothesis under consideration.
But this is more of a philosophical position; I won't complain too loudly
if people ignore my advice.

That said, one must be careful when applying the t-test to the statistical
output. The t-test assumes that the underlying populations are normally
distributed. However, under the null hypothesis, the F-statistic has
an F-distribution, not a normal distribution. So, I would not advise applying
the t-test directly to the F-statistics. (Since the t-statistics are approx-
imately normally distributed, particularly for the large dof's associated with
fMRI experiments, this should not be a problem when calculating a t-test of
t-stats).

One alternative to using the partial F's as described in your method (1) is
to use the -glt option to add up the IRF coefficients for each of the stimuli
separately, and then use the t-stats corresponding to their respective sums.
Another alternative is to compare the partial F-stats using a nonparametric
test (such as the Wilcoxon signed-rank test) for a paired comparison of the
stimuli (A and B). This avoids the normality assumption. For more details,
please see the documentation in file Nonparametric.ps.

As to the question of which method (1 or 2) is better: In order to compare
results across subjects, the usual procedure is to "normalize" the data by Tlrc
conversion, followed by Gaussian blurring. My personal preference is to delay
this step as late as possible in the data processing stream. That is, try to
perform as much of the analysis as possible within-subject, before the Tlrc
conversion step. This has the additional advantage that it allows you to
examine the difference between stimuli A and B on a subject-by-subject basis.
So, I guess I would come down in favor of method 2.

Doug Ward