JoramSoch
commented
5 months ago Dear @FioritoAnna,
first of all: Yes, the model needs to have the same number of regressors across all runs for the voxel-wise cvLME to be computable (see this page from the manual for more information and these lousy hand-written notes for technical details).
A possible solution (as also explained in the excerpt from the manual) is to add zero regressors, i.e. include regressors fully consisting of zeros in those sessions where you would otherwise have "missing" regressors. Please note that the regressors need to be in the same order in all sessions, such that they are not mismatched when the session-wise design matrices are vertically concatenated.
Another solution would be that you do not include those components that explain 50% of the variance in each session, but those components that explain 50% of the variance across all sessions (not sure whether that's technically feasible) or always include the same number of components, e.g. the first 3 or the first 10 (that seems to be the point of your models 2 and 4).
(Side note: I'm not entirely sure how much sense it makes to apply the cross-validated log model evidence to such confound/noise regression techniques. Let's say you always take the first 5 principal components of that CompCor approach. Those components can be quite different between sessions and their influence on the measured signal (i.e. their beta weights) not necessarily similar across sessions. However, the assumption behind cross-validated estimation of GLM is that regressors from different sessions have similar effects on the measured signal across sessions, such that posterior distributions from the training data (say, S-1 sessions) can be used as prior information for the test data (the remaining session). Whether this works for CompCor regressors or PC-style confound regressors in general, I have no idea and is probably an open empirical question.)
Regarding your Problem 2: It seems as if your subjects have image resolutions. Did you normalize all subjects into the same group-level space? If subjects have different image resolutions, their cvLME images cannot be stacked together. The number of voxels is taken from the first model of the first subject
H = spm_vol(job.models{1}{1}{1}); % LME image header
V = prod(H.dim); % number of voxelsand then applied for loading all models from all subjects:
LME = zeros(N,M,V); % N x M x V array of LMEs
for i = 1:N % subjects
for j = 1:M % models
lme_hdr = spm_vol(job.models{i}{j}{S});
lme_img = spm_read_vols(lme_hdr);
lme_img = reshape(lme_img,[1 1 V]);
LME(i,j,:) = lme_img;
end;
end;Could this be the source of the second error?
Cheers Joram
FioritoAnna
Hi,
Thank you for developing such a great toolbox! I have two questions:
1) I'm currently analyzing fMRI data from a task consisting of four runs. I'm interested in comparing models that combine two options for movement correction (24 motion parameters / ICA-AROMA) and two options for physiological noise correction (first 5 CompCor for CSF and first 5 CompCor for WM / CompCor explaining 50% of variance in CSF and WM). This leads to four models:
1- ICA-AROMA_CompCor50% 2- ICA-AROMA_CompCor5 3- 24reg_CompCor50% 4- 24reg_CompCor5
However, models using CompCor50% have a different number of regressors across the four runs since a different number of CompCor can explain 50% of the variance in each run. This discrepancy seems to be causing the following error message:
Do you have any suggestions for dealing with this problem?
2) I attempted to compare model 2 and 4 to avoid using a model with a different number of regressors across the four runs. I tested this approach with two subjects, and it seems to work better. I can obtain MA_cvLME.nii and MA_cvLME_S1-4.nii in each subject directory. I followed instructions from Manual ("11- How to perform cross-validated Bayesian model selection"). However, I encounter another error message during MS: perform BMC (automatic):
I would greatly appreciate any feedback or suggestions you may have!
Best regards, Anna