[pymvpa] null classification performance in the presence of strong univariate signal??

[Nick Oosterhof]

On Sep 11, 2014, at 5:15 PM, David Soto <d.soto.b at gmail.com> wrote:

> but I am not clear why you mention that I have 2 samples
> only for MVPA

I meant 2 samples per subject. 
Let me try to be clear about terminology and how I understood your analysis:

univariate: 
- within-subject (first-level analysis): you ran a GLM in each subject individually (with all N preprocessed volumes as input), and get 2 beta values (samples) per subject. 
- group (second-level analysis): you ran a paired t-test over subjects using the beta values

multivariate:
- within-subject: you did not do MVPA here.
- group: you took the beta values from the within-subject GLM, with 2 samples per subject, 19 subjects (chunks).

> still did not figure out how the univariate t-test gives 
> strong signal in frontoparietal cortex but MVPA nothing

that is still a bit puzzling indeed, though it could, in principle, be due to differences in overall amplitude (I think). The idea would be that if there is large variability in overall signal magnitude over subjects, the SVM hyperplane trying to separate patterns in each condition in the training set often falls so that in the test subject both test patterns are on the same side, leading to change performance.

as a first step to understand what is going on, I would suggest to run within-subject MVPA using a searchlight (as I wrote earlier).