Hi Volkmar, Selena, everyone,
> what do you mean with "different"? The effect sizes and T values should
> be the same, but your p values will be different due to the different
> number of degrees of freedom in your analysis.
If the DF are different due to pooling over subjects, can't the pooled
variance estimate (and hence t-statistic) differ too? Or if separate
subject variances are estimated without pooling, shouldn't the DF used
for each Mean Square be the same as the single subject case?
Although, even if separate variances are estimated, my understanding
of the mechanics of this is that it would use SPM's ReML approach
(e.g. as would be the case if one specified a factorial design with
unequal variance over a subject factor). In this case I would still
not expect the t-values to be the same, because the variances are not
completely partitioned, due to the fact that ReML pools a single
inhomogeneous correlation structure over multiple voxels...
Perhaps I am just very confused?
Ged.
|