Dear Jack,
> As I was reading Andreas question's, I felt concerned with this
> problem. The original study consist of 250 scans * 12 subjects * 2
> sessions (6000 scans) As I thought previously that SPM99b may have the
> same limitations as winSPM (~800 scans), I first performed a "data
> reduction" namely by making 1 mean image out of 10 after realignment,
> and before performing the rest of the analysis (~500 scans remaining).
>
> Unfortunately as there is a large amount of covariates, there is a
> tremendous loss of degree of freedom with some influence on the level of
> statistical inference (~250 residual ddl).
>
> We have interesting results that looks very stable using even high
> uncorrected thresholds (less statistically significant). Unfortunately
> the used thresholds remain weak (alfa < 0.001 uncorrected for the main
> subtraction, alfa < 0.025 for the factorial design - gloups !). Sure
> that the right way should be to perform a true random effect study, but
> it may be that some of the interesting results of the factorial study
> may not survive to this further loss of df.
You can perform second-level analyes on epoch-specific data. It sounds
as if you have emulated this sort of analysis. The sphere of inference
then generalizes to all possible epochs in that session[s] for that
subject[s].
>
> This will probably sounds absurd for statisticians, but I was wandering
> whether if could be possible to reset the df at 10* the number of
> scans - the covariates. As it should only be used while performing the
> {T} statistic, it could be of limited work to change it ? This may
> sounds as tinkering, but is it really that absurd ? Thanks for any
> comment.
I am afraid it is probably 'absurd'. Generally activations should be
detected with 64 or more df. If you have ~250 then this is more than
enough. If you have too many df then any slight [biologically
meaningless] departure from the null hypothesis will give you
significant results.
I would focus on anatomically constrained interpretations of your data.
If you get a result at p < 0.025, in a region predicted in advance,
then this is sufficient to reject the null hypothesis.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
