[pymvpa] Balancing strategy

[Chris Johnson]

On 11/25/2013 04:39 PM, Yaroslav Halchenko wrote:
> 
> On Mon, 25 Nov 2013, Chris Johnson wrote:
>> I have an experimental (fMRI) paradigm wherein the conditions were
>> balanced across runs, and performed a searchlight MVPA analysis with
>> leave-one-run-out cross validation. In some subjects, however, we found
>> excessive head motion and had to eliminate scans
> 
> before jumping to rebalancing:
> 
> was such periods motion associated just with few consecutive volumes?

They're typically isolated. To be clear, we're using a sparse scanning
paradigm (e.g.
http://www.frontiersin.org/Brain_Imaging_Methods/10.3389/fnins.2013.00055/abstract),
so this is not surprising.

In some cases there are multiple in a row, but the effect is that each
modeled event impacts only two scans, leaving the events relatively
unconstrained. Attempting to remove these scans from the GLM actually
ends up causing artifacts in surrounding scans due to the reduced
constraints.

We thus decided that it would be best to use standard GLM modeling and
simply remove the events that are measured in motion-outlier scans.

> if so -- and if you are using some GLM modeling first to extract betas
> for your MVPA analysis -- you could also 'drop' those volumes by adding
> additional single-time-point regressors into your GLM model, so that
> would account for those volumes effects, while still providing you
> estimates for the explanatory variables of interest.  This way you might
> avoid eliminating complete trials from your further MVPA and thus not
> requiring re-ballancing
> 


-- 
Chris Johnson

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 620 bytes
Desc: OpenPGP digital signature
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20131126/e30910db/attachment.sig>