HI Anders,
Keep in mind that the distribution of the statistic is under the null
hypothesis of no signal. If you are examining the distribution (across
voxels) of the t-statistic on actual data from an experimental task, the
distribution may very well not be a t-distribution (e.g., say half your
voxels contain a strong, real response -- then your distribution of t-
values will be skewed positive relative to the t-distribution for null
data of the same dof).
If you input true null data into a processing stream, then by definition
you better get the expected distribution (otherwise, that points to the
existence of a problem/bug in your processing stream).
cheers,
-MH
On Thu, 2010-10-28 at 13:16 +0200, Anders Eklund wrote:
> Dear SPMers,
>
> I've implemented my own single subject fMRI analysis in Matlab (slice
> timing correction, motion compensation, smoothing, detrending, GLM &
> t-test) and get activity maps that seem reasonable. I now want to
> calculate a threshold that is corrected for multiple comparisons and
> have for that purpose used two approaches, Bonferroni correction and
> random field theory. As I understand it, both these methods rely on the
> fact that the test statistics, under the null hypothesis, follows a
> Student's t-distribution (since I calculate a t-test value in each voxel).
>
> For most of my datasets the test statistics seems to approximately
> follow a Student's t-distribution with the same degrees of freedom, but
> it is often slightly wider at the bottom or slightly skewed. For one
> dataset, the test statistics is rather far from the t-distribution.
>
> 1) How much work has been done on validating the assumption that the
> test statistics follow a certain distribution? Is there any paper where
> this is discussed?
>
> 2) Is there any specific preprocessing step, that I might have missed,
> that ensures that the test statistics follows a certain distribution?
>
> 3) Does SPM check the distribution of the test statistics?
>
> Best regards,
> Anders Eklund
>
|