Missing data in ROI - does this affect group analysis?

[clare....@gmail.com]

Hi Nick and CoSMo MVPA team,

I hope you're well! I had a question about group level fMRI RSA analyses, when the participant-level analysis has been performed using a searchlight within an ROI mask. 

We have identified that, due to slice positioning, some participants' raw data are missing from some voxels within the ROI mask. The analysis still runs fine, and when we view the unthresholded z-maps, we can just see that results are absent for these parts of the mask (i.e. the mask appears 'cut off').

The script for the group level analysis (using commands including cosmo_stack, cosmo_cluster_neighborhood) also runs fine.

My question is: how does the missing participant data affect interpretation of group-level analyses? For these participants, are the voxels with missing values just treated as NaN and not included for that participant in the group analysis? Is there any risk that this missing data will interfere with the group levels analyses, and any steps we should take to protect against it?

Thanks for any clarification!

All the best,

Clare

[Nick Oosterhof]

Hi Clare,

If data for missing participants is encoded with NaN, then the t-test functionality in cosmo_stat can take that into account by adjusting the dfs (degrees of freedom) when using the ‘z’ output (z-scores). For example if there are n=30 participants in a one-sample t-test, then normally df=29 would be used. But if 2 participants have a NaN value for a particular feature (voxel), then df=27 would be used there. This functionality can be used for multiple features (voxels), including the scenario of whole-brain group analysis. What is nice here (in my opinion) is that for different features (voxels) a different number of dfs can be used.

Note:
- this was inspired by AFNI’s ttest++ with the -toz option
- this functionality is used by default in cosmo_montecarlo_cluster_stat

Does that answer your question?

best,
Nick