Hi Anderson,

Many thanks for confirming and I am very sorry for confusing matters by passing on mixed output. That was certainly not what I intended. You are absolutely right to flag this, and as shown by the different prefixes we accounted for different nuisance variables in a couple of different models.

Thank for your continued support and in addition to citing PALM we will certainly acknowledge your help.

Regards,

Jack

From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Anderson M. Winkler [[log in to unmask]]
Sent: 26 January 2017 09:47
To: [log in to unmask]
Subject: Re: [FSL] PALM Output: vox_vstat output

Hi Jack,

These are just sparse voxels. Some of the comparisons aren't right, though: from your results I have files names two_sample_fa_iq_age_tfce_vstat_fwep_c1.nii and two_sample_fa_iq_vox_vstat_fwep_c1.nii. Since the prefixes are different, they must be coming from different runs, and with different nuisanve variables (one including age, the other not). This could explain oddities.

Regardless, there are just very few sparse voxels where the simple voxelwise led to more powerful results than TFCE. Everywhere else, it's the other way round as expected.

When I run for L1 (attached the options I used), the same happens: TFCE usually with lower p-values (so, higher -log10(p)), with simple voxelwise with higher, less significant p-values. There might be one or another voxel following the opposite trend, but this is fine.

I'd say opening the images provides more information than just the command cluster. The latter is also useful, but doesn't replace viewing the maps.

It seems to me these results are fine.

All the best,

Anderson

On 25 January 2017 at 13:46, Jack Rogers <[log in to unmask]> wrote:

Hi Anderson,

To follow up, I uploaded the axial diffusivity skeleton (L1) rather than FA as intended. My apologies. This would explain our discrepant findings (there are some nice TFCE-FWER corrected results for AD).

If you could confirm that the PALM command that we use is appropriate given the multi-site variability that would be great. From there, I think everything else is fine.

palm -i all_FA_skeletonised.nii -m mean_FA_skeleton_mask.nii -d design_two_sample_ttest_IQ_Age.mat -t design_two_sample_ttest_IQ_Age.con -eb block.csv -vg auto -n 5000 -T -C 3.1 -fdr -o two_sample_fa_iq_age -logp

Thank you again for all of your help with this data.

Jack

From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Jack Rogers [[log in to unmask]]
Sent: 25 January 2017 09:48

To: [log in to unmask]
Subject: Re: [FSL] PALM Output: vox_vstat output

Hi Anderson,

Thank you for this very useful feedback. Although I am afraid I am a little confused by a couple of your points as I don't mean the opposite (I don't think!) - when thresholding at 1.3 it seems to me that that there is nothing there for the TFCE corrected contrast (I find the peak logp to be 0.854) whilst for the voxelwise image there are voxels, albeit very few, at peak logp 1.3 and 1.66.

Might you be kind enough to pass on the model you ran on the data (.mat/.con) and the palm command used please. Sorry to labour this issue - I am just aware that we used PALM, appropriately, to account for multi-site variance but as this differs from a lot of the DTI literature that uses randomise with TBSS I want to ensure that the comparison is comparable.

Many thanks,

Jack

From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Anderson M. Winkler [[log in to unmask]]
Sent: 25 January 2017 08:46
To: [log in to unmask]
Subject: Re: [FSL] PALM Output: vox_vstat output

Hi Jack,

Thanks for sending. Looking again at the original post, do you mean the opposite? That is, TFCE more significant than simple voxelwise? This is what I can see in your results (although it seems the nuisance variables are different), and also when I run here the same, with both age and sex as covariates.

So, in answering the two original questions:

1) more lenient than the TFCE output following permutation testing

Neither is really lenient in that both control the error rate exactly. That said, it's more the other way round: TFCE is more powerful than simple voxelwise, and I can attest this in your results.

2) should not be reported as meaningful in our paper?

Both are meaningful and both can be reported.

Hope this solves the issue.

All the best,

Anderson

On 23 January 2017 at 11:21, Jack Rogers <[log in to unmask]> wrote:

Hi Anderson,

Many thanks for your reply. I have uploaded the input data and design/contrasts that correspond to the command line below as requested. This command line (and variants thereof) is run for all designs/contrasts of interest and across the different DTI-metrics (dependent on input -i). The block.csv file includes information about the different sites, in response to our earlier correspondence about controlling for site effects (this prompted our use of PALM - happy to pass on our email exchange). I have also uploaded the two different output files I refer to in my previous query - one TFCE-FWE corrected and the other VOX-FWE corrected.

palm -i all_FA_skeletonised.nii -m mean_FA_skeleton_mask.nii -d design_two_sample_ttest_IQ_Age.mat -t design_two_sample_ttest_IQ_Age.con -eb block.csv -vg auto -n 5000 -T -C 3.1 -fdr -o two_sample_fa_iq_age -logp

It would be very useful to know if the command line above is appropriate. We are comparing two different groups (patients vs controls) so very simple design matrix (N= 300) with IQ and Age included as covariates of no interest.

Best wishes

Jack

From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Anderson M. Winkler [[log in to unmask]]
Sent: 21 January 2017 13:54
To: [log in to unmask]
Subject: Re: [FSL] PALM Output: vox_vstat output

Hi Jack,

Could you upload input data, design/contrasts, and let me know the command line used? You should have just received an automated email with instructions.

Thanks.

All the best,

Anderson

On 20 January 2017 at 11:06, Jack Rogers <[log in to unmask]> wrote:

Dear Experts, Dear Anderson,

I have a query regarding the PALM output. Namely, in a couple of instances we see sub-threshold (<1.3) differences for TFCE FWEP output but significant voxels in the VOX_VSTAT FWEP output. Am I right in thinking that this output represents a voxel-by-voxel comparison across the skeleton and so is 1) more lenient than the TFCE output following permutation testing and 2) should not be reported as meaningful in our paper?

Many thanks
Jack