The summary statistics are indeed con images and not t-images.
t-images conflate effect size and error and it is therefore not
meaningful to do further statistics on them (for instance compute
their mean across subjects). The only thing that you can do with them
is use them for computing p-values. With the summary statistic
approach within-subject variance is not taken into account at the
second level. The inference is made in relation to between-subject
variance. This makes it possible to draw conclusions with respect to
the population from which the subjects were taken. A fully
hierarchical approach makes it possible to take into account both
within and between subject variance but it is computationally
expensive (or at least was when people worked on those things) and it
has been shown that the summary statistic approach is a good
approximation to it.
for more details and references.
On Fri, Jul 22, 2011 at 5:16 PM, Tal Linzen <[log in to unmask]> wrote:
> Hi Vladimir,
> One point of confusion -- contrary to what I had thought, the contrast
> images (con_0001.img) are not t-statistics, but the regression coefficients
> themselves (in simple contrasts such as [0 1]). Does that make sense?
> Shouldn't I take the t statistic maps instead, to account for the
> within-subject variance? I thought that the mixed-effects model should take
> into account both the within-subject variance and the between-subject
> variance, and in the summary statistic approach this would amount to running
> a t test on the t statistics rather than the raw beta values.