You can always use a Bonferroni-like strategy here for correction.
However, I would not worry too much about correction here, unless your
pseudo-blocks are really many. The reason is that your Type error I
here is always bound at the 5% level; this is not the case if you omit
correction for the repeated tests across voxels. So the consequences
for not correcting in these two cases are dramatically different. I
read all the time papers that do not correct for repetition across
voxels, so your paper would still be in a superior Nth percentile or
so. Methodologically your problem will rather be that you will not be
able to tell if the drug had an effect on baseline.
I see no merit whatever in the ICA strategy. Extracting the signal
from data and using it as regressor is entirely circular. This is
worse than not using corrections, as it makes testing meaningless.
Best wishes,
Roberto Viviani
University of Ulm, Germany
> I have a within subjects, placebo controlled IV infusion
> pharmacological fmri study. Subjects will be scanned on two
> occasions.
>
> I have seen a few papers that have analysed pharmacological fmri
> studies (i.e. fmri with IV infusion) using a psuedo-block design. In
> this approach each scan is divided into blocks (e.g. 1min in
> length) and the pre-infusion block is used as a baseline. At the
> first level this would give N-contrasts (block 1 - baseline, block
> 2 - baseline .......) per individual.
> What would be the most robust way to analyse these data? I could
> just use a paired sample t-test but that does not seem appropriate
> as I would not (neccessarily) be correcting for the number of
> contrasts. Are there more robust alternatives?
>
> Alternatively, could one use ICA to identify a pharmacological
> regressor (I am not collecting blood samples or any other surrogate
> marker of drug pharmacokinetics but I know from previous studies the
> approximate shape and time course of the drug) and then use this
> pharmacoregressor at the first level - giving one contrast per
> individual? My concerns are that I would be guessing on the time
> course (although an educated guess as I could examine the time
> course in a region I know to be rich in the receptor of interest)
> and that this seems circular? Would it be better to pilot this
> approach in a small number of subjects and generate an average
> pharmacoregressor?
>
> Thanks
>
|