[pymvpa] classification based on individual parameter estimates from FSL

[Nick Oosterhof]

On Jul 31, 2014, at 10:49 PM, David Soto <d.soto.b at gmail.com> wrote:

> Hi, I keep plugging away with this pretty basic classification 
> [...]
> I get a whole-brain classification accuracy of around 68%
> (though did not assess significance)
> Then I run a searchlight analyses and looking at the classification accuracy maps it appears like a chance distribution with mean 50% and the max classification accuracy
> around 56%- I wonder how it be that none of the searchlights reaches the level of wholebrain classification ? and if this is the case then can it be the wholebrain classification meaningful at all?

That is quite possible because the whole-brain classification uses many more features than each searchlight.

Assuming there is sufficient signal in the data (which there seems to be in your case) which is not limited to a small subset of features (voxels), generally one sees better classification with more features. This was already reported by Cox et al 2003, and later by e.g. [disclaimer: shameless self promotion] Oosterhof et al 2011. (there are some cases where this might not be true)

There's often tradeoff between spatial selectivity and classification accuracy. In one extreme you use all features for a single classification analysis (i.e. your whole-brain classification), in the other extreme you use one feature at a time (i.e., univariate analysis). A searchlight analysis is somewhere in between, finding a compromise between getting high classification accuracy and good spatial selectivity. But also for a searchlight it holds that neighborhood (sphere or disc) size can affect both classification accuracy and spatial selectivity.