[pymvpa] Group analysis - surface searchlight

[Nick Oosterhof]

On 12 Apr 2015, at 00:53, Mike E. Klein <michaeleklein at gmail.com> wrote:

> I’m trying to run a sample top-level analysis on surface-based searchlight data using Nick’s afni_surface_alphasim script. I’ve never tried this before, and am getting a ZeroDivisionError (bottom of output). […]
> 
> Here is my bash input/output:
> […]
> ++ 3dbucket: AFNI version=AFNI_2011_12_21_1014 (May  1 2014) [64-bit]
> ** FATAL ERROR: Dataset /Volumes/deep_thought/fmri/mvpa2/april_analyses/subs/06/suma_surfaces_deoblique/zeroed_soundVrest_NEW-minimal_ico32-128_mh_50vx.niml.dset differs in size from first one

Actually the issues turned out to be that a an anatomical surface file had to be specified. I’ve modified the code now to throw a more informative error message if that case is detected, and sent a PR [1] which will probably be merged soon. Also the padding actually had to go to 20483 to make it work, and increasing the smoothing bandwidth made it run significantly faster. So briefly, the options I added that work were:

-N 20483 --surface_file intermediate_surf.asc -S 1.4

where “-N” indicates the node padding (=(32^2*10+2)*2-1), and “-S” the smoothing bandwidth.

Also, you may want to consider Threshold Free Cluster Enhancement [2]. Not in PyMVPA yet, but if you are happy to go into Matlab/Octave world, it is provided in CoSMoMVPA [3] through a monte carlo simulation [4]. There is an example script [5] illustrating its use (from "%% Run Threshold-Free Cluster Enhancement (TFCE)” onwards). Note: you would have to pad the nodes manually and convert to ASCII format, then use cosmo_surface_dataset to load each dataset and cosmo_stack to make it a single dataset. 

[1] https://github.com/PyMVPA/PyMVPA/pull/285
[2] http://www.fmrib.ox.ac.uk/analysis/techrep/tr08ss1/tr08ss1.pdf
[3] cosmomvpa.org
[4] https://github.com/CoSMoMVPA/CoSMoMVPA/blob/master/mvpa/cosmo_montecarlo_cluster_stat.m
[5] http://cosmomvpa.org/contents_demo.html#demo-surface-tfce