Re: [SPM] fMRI: first-level with SPM12 in predefined ROIs followed by second-level with SwE toolbox

[Thomas Nichols]

Dear Mélanie,

I'm CCing the SwE list which is the best place for advice relating to the toolbox.

We are conducting an fMRI analysis with two groups (patients and controls). Individual GLM were designed in SPM12 with HRF and derivatives as regressors. For the second-level within- and between-group analyses, we use the SwE toolbox to model repeated measures (HRF and derivatives).

In fact we are planning to run the individual analyses by restricting the search space to regions of interest defined from physiological/anatomical hypotheses, rather than conducting whole-brain
analyses.

As far as I remember from Matthew Brett’s advice (quite long ago), for such analysis (similar to an SVC correction) it is recommended to use ROIs that are large and (possibly) convex enough so that the hypotheses for subsequent FWE correction are valid. Could you please help me by answering the following questions?

The first thing to realise is that SwE doesn't use random field theory; so there is no worry about 'crinkly' ROIs.  It *is* true that if you have very convoluted ROIs that clusters can never get very big, and so cluster infernece might not be so powerful, but the resampling approach SwE uses should be valid for any type of analysis mask.

 

-    Can we select Brodmann areas as ROIs (for instance from WFU Pickatlas) ?

Again, any kind shape is fine.

 

-    If we have several ROIs, we should merge them in a single mask. If so, is it a problem if they are not contiguous ?

Yes best is to use a single mask and use that as the analysis mask, so that the multiple testing correction happens implicitly.  

There is no problem if they are not contiguous. (That is true for SPM and RFT, but, again, also true for SwE).

-    If it is impossible to merge them in a single mask, for instance because they correspond to different a priori hypotheses, then we need to run separate analyses. Then we will have to Bonferroni-correct all p-values by dividing by the number of analyses. But should we consider the corrected (typically .05) or the uncorrected (.001) p ?

If you had to run separate analyses (not recommended), indeed, you could obtain a Bonferroni correction by multiplying p-values by the number of ROI masks you considered.

 

-    Is there a minimum size acceptable for an ROI ? I would presume
that an ROI needs to be larger than the RESEL size, or at least the
filter FWHM that was used to smooth the data, if we plan FWE inference ?

For SwE, no limit; but if the size is very small you basically have no advantage from cluster size inference.

(For SPM, I don't know a hard and fast rule, but I would tend to try to use ROIs at least the size of 1 RESEL, yes).

 

We use SwE for second-level analyses and I must confess I am not sure at all which inference method would be the most appropriate following individual analyses in restricted search space.

SwE's design configuration uses the exact same analysis masking options as SPM; you can supply the single, combined ROI mask at that stage.

 

-    If I choose parametric inference then FWE is not used at all(as far as I understood); I can choose voxelwise FDR correction, and then maybe I do not need to bother about ROI size at all ?

Using ROIs to limit the search space will always reduce the severity of the multiple testing problem, regardless of the method you use to correct for multiple testing.  So using a constrained analysis mask / ROIs, when you can specify them objectively and a priori, is always a good idea, whether using FWE or FDR correction.

Regarding SwE specifically:  You are correct, parametric does not give you FWE-corrected-values, only FDR correction (voxelwise). 

 

-    If I go for non-parametric wild bootstrap, then is FWE used ? can I choose voxelwise or TFCE inference ? I am sorry but this is not clear for me from SwE manual…

Sorry this isn't clearer.  Yes, that's right... the Wild Bootstrap is required to obtain voxelwise, clusterwise or TFCE FWE-corrected inferences.

Thanks for your questions!

-Tom