3dMVM -resid: Split up?

AFNI Message Board
1

Hello!

When running 3dmvm with the -resid option having a large dataset (somehwere over 1500 subjects when having about 125 GB of RAM) it crashed due to memory error. I think it’s R that complains, I saw a smilar post here that said it was memory issues.

The stats are generated but not the residuals, so I am fairly sure that it’s the -resid part that crashes.

Would it be reasonable to make a separate directory called residuals and run 3dmvm twice with the table split in half, to generate residuals form the top and bottom half of the table and then average the two residuals volumes and feed it into 3dFWHMx?

Othwerwise I don’t think I can get the residuals. Running 3dFWHMx on each input file and average across subjects does in the end not yeild the same afc-parameters as if you do 3dMVM -resid. I confirmed this with a smaller sample. Since we are looking at structural maps (not the errts from fMRI data) it is important that we get the reisudals from mvm.

Thanks!

2

Would it be reasonable to make a separate directory called residuals and run 3dmvm twice with the table split in half, to generate
residuals form the top and bottom half of the table and then average the two residuals volumes and feed it into 3dFWHMx?

If you cannot find a computer with a larger memory size, you can try what you’re suggesting. It is not ideal, but the approximation is likely reasonable enough to serve the purpose.

Running 3dFWHMx on each input file and average across subjects does in the end not yeild the same afc-parameters as if you do 3dMVM -resid.

Conceptually smoothness estimates should reflect the spatial relatedness in pure “noise”. Your input data contain a lot of fine structure and signal, so it would be inappropriate to use them for smoothness estimation; otherwise, it would lead to excessive penalty on the final results. Even the residuals from a model likely include some extent of signal/structure because the model is usually imperfect.