Hi Yuki,

Sorry for the late answer. 

Re Manuel Garcia-Garcia's poster. I unfortunately do not have a copy readily accessible. There is a manuscript on that work which I am hopeful to see come out sooner rather than latter, at least in the form of a preprint. I will make sure to share it with you should this possibility materialize. 

Re choice of region growing size, it does indeed impact markedly stability estimates. For small regions you may be able to work at the voxel level. I would go with the number of regions: the pipeline runs in a reasonable time with up to 3000 regions. Beyond that it is starting to be really slow. I don't think the CPAC implementation features any dimension reduction prior to BASC, so you will have to work at the voxel level. 

Re choice of the number of clusters, the rationale behind MSTEPS applies as well to regions as it does to full brain analysis.  

Re the choice of the window length, it is based on the variability of the correlation coefficient between regional time series, wthich peaks at about sqrt( # time frames) -- that is described in a old relatively obscure paper (Bellec et al. Statistica Sinica 2008). This argument does not depend on the space of analysis (region vs brain). Note that the actual parameter is treated as a nuisance parameter and actually randomized in the NIAK BASC implementation (may be as well in CPAC). 

Best,

Pierre