rick reynolds wrote:

> This is not quite correct. The t-distribution approaches the
> normal
> distribution as the degrees of freedom approaches infinity
> (subjects
> minus 1 in your case), but you have only 12 subjects. 60000
> voxels is
> multiple comparisons, but of the same non-normal test.
>
> Keep in mind that these results depend on the variance of your
> data,
> which comes not only from BOLD responses to your stimuli, but
> from how
> the experiment was designed and how the data was processed.

Sure. So if I had an infinite number of subjects, I would expect exactly 5% of my voxels to exceed the t (z)=1.96 threshold associated with p = .05. If I have a finite number of subjects, the expected distribution becomes wider; its more likely that I will find a more extreme t-value. So with 12 subjects, I'd expect 5% of my voxels to exceed the t = 2.18 threshold by random chance.

But AFNI accounts for that. 3dTTEST know how many subjects I have, and what my degrees of freedom are, so when it calculates p-values, it doesn't use a z-distribution, it uses a t-distribution with N-1 degrees of freedom. So it will assign a p < .05 to any voxel with a t > 2.18. Under the null hypothesis, exactly 5% of my voxels should exceed this value.

So if only 1.8% of my voxels exceed t = 2.18, something very funny is going on. That would mean that there are fewer voxels in the tails of the distribution than would have been expected given the degrees of freedom and the standard error of each voxel. Correct?