[pymvpa] Analysing data from a 1-run block design

[Jan Derrfuss]

Hi Nick,

Thank you very much for your quick reply!

>> I was wondering if there's a recommended way to analyse fMRI data
>> from a block design where all data were acquired in one run. I've
>> modelled each block individually and now have one parameter
>> estimate for each block. There were two conditions (A and B) with 6
>> blocks each. Conditions alternated (A - B - A - ... - B) and were
>> separated by 15-s rest phases. Half of the participants started
>> with A, the other half with B.
>>
>> In particular, I was wondering a) how much of a problem the single
>> run is,
>
> Not ideal, but also not necessarily a game-breaker.

Cool.

>> b) how chunks should be assigned,
>
> You mention 15-s rest phases. Does that mean the design was:
>
> R A B R A B R A B R A B R A B R A B R
>
> If that is the case, then each ?A B? piece (block) would have a
> unique chunk value. Given the duration of 15 s between blocks, this
> should be enough to assume independence between blocks.

No, it was R A R B R B R A ...

>> c) what type of preprocessing should be applied
>
> Was the order randomised for each block or not? If it was not
> randomised, then detrending becomes seriously important. But in any
> case I would suggest to apply detrending. z-scoring may also be a
> good idea, in particular if you have not normalised the data
> otherwise (such as dividing each voxel?s time course by the mean
> value over that voxel?s time course)

I assume "block" refers to the "A B" blocks you mention above? I see 
your point, but as there was a rest phase between every single block, 
this shouldn't be an issue, right?

For now, I arbitrarily assigned 4 blocks to each chunk (so, 3 chunks 
overall) and detrended and z-scored. PyMVPA told me that this would be 
"discouraged" given the small number of samples, so I wasn't sure what 
the best thing to do would be.

>> d) if there's a classifier that would be expected to work well
>> under these conditions.
>
> I may be wrong, but I don?t expect very significant differences
> between the typical classifiers; I would suggest to try SVM or
> regularised LDA.
>
> However, there are not too many samples, only 12, so the
> classification accuracy can only have 13 possible values (i/12 for i
> from 0 to 12 inclusive).
>
> Alternatively you could do a split-half correlation analysis, which
> may give a more continuous measure of pattern descriminability.

Thank you for these suggestions!

Cheers,
Jan