Hi,

> Reading previous posts in the forum and some papers, I understand that to perform nuisance signal regression I have to:
> 

Overall there is limited levels of consensus, so no clear ‘musts’ when it comes to nuisance signal regression. 

> 1- For each individual, remove the signal associated with several nuisance covariates (specifically white matter, CSF, global signal, and six motion parameters) using multiple regression analysis.

Well, we recommend ICA-based data denoising (using ICA-FIX or ICA-AROMA) for dealing with head motion and potentially including CSF and WM regressors. There is a recent paper (Pruim et al. Neuroimage. 2015 112:278-87) that discusses all these steps in more detail. We certainly do _not_ recommend Global Signal Regression!


> In order to do this using first-level Feat, after pre-stats and registration steps, in stats I have to: 
> -load filtered_func_data as input 
> -Full Model Setup:load every single nuisance(text files)as “custom” 1-­column regressors(with no convolution, turning off temporal filtering and temporal derivative)
> - What about contrast? Should I create one single contrast with all ones? [1 1 1 1 1 1 1 1 1], since I'm interested in the residual image (that should represent the corrected signal)?

The easiest way is to use fsl_glm on the command line.

> 
> 2- This nuisance signal regression step should therefore produce a 4D residual file (res4d.nii) for each participant from which I should extract mean time series from each ROI of interest. 

Well, if you happen to have multiple ROIs then it would be best to use multiple linear regression, ie. by putting all ROIs into a single 4D file and using dual_regression.. The DR stage 1 outputs then are the different region-specific time courses

> Are these steps correct? Some authors (Kelly et al., 2009) before extracting mean time series perform a voxelwise scaling on the 4D residuals, dividing each voxel’s time series by its SD. "Performing this step ensures that the ensuing FC estimates represent partial correlations rather than regression parameter estimates (e.g., Vincent et al., 2006), and removes potential between-condition differences in the magnitude of BOLD fluctuations (Sorg et al., 2007)." What do you think about it?

You certainly would want to scale the time courses themselves (de_norm option in dual regression, i.e the predictor variables) - the question as to scaling the data depends on whether you think that BOLD magnitude is meaningful or not. In std FMRI analysis (GLM based on tasks) the field has moved from correlation analysis to regression in the 90th, personally I’m surprised that for ‘connectivity’ most people assume that BOLD magnitude is not relevant...


> 
> 3- Moreover, since we acquired three resting state scans for each participant, is it correct to average the 3 final mean time series of each ROI I get from each partecipant? 

No

> Or should I average with a higher level FIXED effect analysis for each partecipant the output of these first level analyses and get a single res4d for each subject? 
> 

You should conduct your analysis separately and then combine the final outcome maps in a fixed-effects analysis at the higher level.

hth

Christian

> Any help whould be greatly appreciated, 
> 
> Claudia