afniuser2014 Wrote:
-------------------------------------------------------
> I am trying to applying a cluster size based
> thresholding to my accuracy maps after MVPA
> classification (in surface space).
>
> To this end, I am using slow_surf_clustsim.py.
>
> As I did not apply any smoothing to my data, I
> think that it would be correct to use a blur
> factor = 0 (that is, no smoothing applied). Do you
> think this is correct?

I think not. If you've done MVPA with a searchlight analysis, then the maps will definitely be smooth. The reason is that neighboring searchlights share many voxels, so their MVP results (classification accuracies, or correlations) will be correlated.

If you want to use slow_surf_clustsim.py, you would have to estimate the smoothness of the noise of the maps (*not* the signal). Thus, it is not really valid to apply SurfFWHM on the MVPA results. It *may* be the case that the smoothness of the signal is greater than that of the noise, so that estimates would result in too conservative multiple comparison correction; in that case, you would at least not increase the probability of a type-1 error. But it could in principle also go the other way, and in that case, you *would* increase the probability of a type-1 error. Whether that's true or not depends on the characteristics of the signal and the noise, which are unknown. For that reason, I would really discourage using SurfFWM on MVPA results directly.

One approach that may work is PyMVPA's afni_surface_alphasim.py, which uses residuals by subtracting the group average from each individual's map. Note that his approach is rather experimental and has not been validated properly.

> My other question is that, from the function
> instructions I am not sure if I have to use the
> flag "-on_surface yes" for my analyses or not.

As the data is on the surface, you should use -on_surface yes.

> Finally, the last question is if I can run
> slow_surf_clustsim.py in the two hemispheres
> together. My two hemispheres accuracy maps are
> already merged in one single niml.dset, so I am
> providing a spec file containing the two
> hemispheres to run the simulations. Do you think
> that this is correct?

That sounds fine to me.

You may also consider using whole-brain correction using functionality in CoSMoMVPA. In particular there is an demonstration of surface-based multiple comparison correction. For group analysis, assuming you have one value (classification accuracy or correlation) per participant and per node, you can stack output datasets from each individual participant; set all .sa.targets value to 1, and .sa.chunks to (1:nsubjects)' and run cosmo_montecarlo_cluster_stat. This approach works for both surface-based and volume-based datasets. You can use either Threshold-Free Cluster Enhancement (the default, and recommended), or if you really want to, a fixed uncorrected threshold.