[pymvpa] Combine the fMRI data with other data in the pattern analysis?

[Nynke van der Laan]

Just to let you know: the script does exactly what I intended! The
accuracies have become higher as I would expect since the covariate on
its own in a logistic regression also predict quite ok).
Thanks for the help,
Best, Nynke

On Wed, Mar 16, 2011 at 2:10 PM, Nynke van der Laan
<nynkevanderlaan at gmail.com> wrote:
> Hello Michael and Yaroslav,
>
> again thanks for the help! Sorry for putting so much workload for you
> on the mailinglist but I think that I now managed to implement it in
> the right way. See below for the total script (could be useful for
> other users with version 0.4 that want to do the same). Would you
> please check it if this is indeed what it is doing?
>
> Just to be sure about how and in which sequence this data is handled:
> The searchlight first makes a subdataset of an amount of voxels
> (features) depending of the chozen sphere size. Is it correct that,
> AFTER THAT, the covariate is added as an additional feature? So for
> example, you have a sphere of 26 voxels, the covariate is added as an
> extra (27th) voxel for in the classification analysis(and thus added
> to each sphere). Then I wonder: doesnt' the fact that the covariate
> has very different values (i.e., in the range of 1-9 instead of the
> much smaller voxel values) complicate the analysis?
>
> Best,
> Nynke
>
> code:
>
> class MyMeasure:
>        def __init__(self, measure, cov): # This is done when you do
> MyMeasure(cv , covariate_ds): the cv is the measure, cov the covariate
>                self._msre = measure
>                self._cov = cov
>
>        def __call__(self, ds):
>                cov_ds = Dataset(samples=N.hstack((ds.samples, self._cov)),
>                                   labels = ds.labels, chunks = ds.chunks)
>                return self._msre(cov_ds)
>
>
>
> covariate = N.array([1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38])
> # The covariate has 1 value for each of the 38 samples
> covariate = covariate.reshape((38,1)) # It is important to reshape in
> order to have the same dimensions for the covariate and the ds.
> m_nynke = MyMeasure(cv, covariate)
> sl = Searchlight(m_nynke, radius=6)
> sl_map = sl(dataset)
>
>
> On Tue, Mar 15, 2011 at 5:33 PM, Nynke van der Laan
> <nynkevanderlaan at gmail.com> wrote:
>> Dear Michael,
>> it still does not work for me (I have version 0.4).
>>
>> When I do:
>> covariate_ds = Dataset([1,2,3,4,5,6,7,8,3,5,4,5,4,4,3,4,5,4,3,4,5,6,7,8,9,7,5,4,5,6,7,8,7,4,3,3,3,4])
>> I get the error: AttributeError: 'list' object has no attribute
>> 'copy'. Also making a Numpy array of it doesn't seem to help.
>>
>> But I do I understand correctly from your post that for version 0.4
>> covariate_ds = [1,2,3,4,5,6,7,8,3,5,4,5,4,4,3,4,5,4,3,4,5,6,7,8,9,7,5,4,5,6,7,8,7,4,3,3,3,4]
>> would also work? (I used that now for the analysis below which
>> probably explains the error I get (see below).
>>
>>
>> I have moderated your script a little to the following:
>> class MyMeasure:
>>        def __init__(self, measure, cov):
>>                self._msre = measure
>>                self._cov = cov
>>
>>        def __call__(self, ds):    # was: _call(self, ds) but gives an error:
>> MyMeasure instance has no __call__ method
>>                cov_ds = mvpa.base.dataset.hstack(ds, cov)
>>                return self._msre(cov_ds)
>>
>> However, now I get the AttributeError: 'module' object has no
>> attribute 'dataset' when I use, which probably originates by the fact
>> that the covariate_ds I use is just a list. How can I do this in 0.4?
>> Did the architecture of the dataset baseclass change in version 0.6 or
>> am I doing something wrong?
>> cv = CrossValidatedTransferError(
>>         TransferError(LinearNuSVMC()),
>>         NFoldSplitter())
>> m_nynke = MyMeasure(cv, covariate_ds)
>> sl = Searchlight(m_nynke, radius=6)
>> sl_map = sl(dataset)
>>
>> Thanks!
>>
>