External Email - Use Caution
Dear FreeSurfer experts,
I am reading your work 'False positive rates in surface-based anatomical analysis' and having several questions about multiple test correction in mri_glmfit-sim:
1. The permutation is permuting the design matrix, but how to permute it in a one-sample t-test? According to my understanding, all the subjects are listed as one group in the design matrix.
2. What's the difference between clustering-forming thresholds (CFT) and cluster-wise p value (CWP)? Whether the clustering-forming thresholds is set to the p map from the permutation test? Then how the CWP is calculated?
3. In the paper ''False positive rates in surface-based anatomical analysis, you suggested ''For surface area, one would need CFT 0.001 and FWHM>10 mm". Whether I could understand this as "the bigger smooth kernel the better"? The default smooth kernel in FSFAST is 5mm, which one should I use?
4. For the quality assurance in 'tkregister-sess', if I can use the QA value as a exclude criteria during analysis? If so is there any reference I can cite?
Looking forward to your kind suggestions.
Best regards,
Lizhi
On 4/25/19 10:54 AM, Zhi Li wrote:
External Email - Use Caution
Dear FreeSurfer experts,
I am reading your work 'False positive rates in surface-based anatomical analysis' and having several questions about multiple test correction in mri_glmfit-sim:
- The permutation is permuting the design matrix, but how to permute
it in a one-sample t-test? According to my understanding, all the subjects are listed as one group in the design matrix.
In that case, the signs are flipped (ie, instead of having a column of ones, you get a column of +1 and -1, the sign being randomly permuted).
- What's the difference between clustering-forming thresholds (CFT)
and cluster-wise p value (CWP)? Whether the clustering-forming thresholds is set to the p map from the permutation test? Then how the CWP is calculated?
As the name implies, the CFT is used to form the cluster. Think of you sig map as a landscape with mountains and valleys. If you were to fill the landscape with water, you'd get some islands. The number and size would depend on how high the water was. If you think of each island as a cluster, then the water level is the CFT. In a purely random landscape, some islands would form purely by chance. The propbablility of getting an island (cluster) of a certain size or bigger is the CWP. You have to choose the CFT. The CWP is computed from permutation.
- In the paper ''False positive rates in surface-based anatomical
analysis, you suggested ''For surface area, one would need CFT 0.001 and FWHM>10 mm". Whether I could understand this as "the bigger smooth kernel the better"? The default smooth kernel in FSFAST is 5mm, which one should I use?
If you use permutation, then you can set the FWHM to anything. The optimal size depends on the size of your blobs.
- For the quality assurance in 'tkregister-sess', if I can use the QA
value as a exclude criteria during analysis? If so is there any reference I can cite?
There is no reference. Usually, I look at the worst data sets and work my way to the better data sets to determine where the threshold should be.
Looking forward to your kind suggestions.
Best regards,
Lizhi
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
External Email - Use Caution
Thank you very much for your detailed explanation. About the second question, I am still confusing on the permutation and CWP map calculation:
1. Whether the CFT is applied to the uncorrected p-value map first (e.g., the sig.nii.gz generated by 'mri_glmfit')? Then we could get several clusters survived from the CFT, right?
2. But how the permutation works and generates the CWP map? I found that the CWP within a cluster is the same, but zero out clusters (perm.th30.pos.sig.cluster.nii.gz).
3. Besides, may I know what the value is in the file 'perm.th30.pos.sig.voxel.nii.gz'?
Thanks and regards,
Lizhi
On Thu, 25 Apr 2019 at 12:05, Greve, Douglas N.,Ph.D. < DGREVE@mgh.harvard.edu> wrote:
On 4/25/19 10:54 AM, Zhi Li wrote:
External Email - Use CautionDear FreeSurfer experts,
I am reading your work 'False positive rates in surface-based anatomical analysis' and having several questions about multiple test correction in mri_glmfit-sim:
- The permutation is permuting the design matrix, but how to permute
it in a one-sample t-test? According to my understanding, all the subjects are listed as one group in the design matrix.
In that case, the signs are flipped (ie, instead of having a column of ones, you get a column of +1 and -1, the sign being randomly permuted).
- What's the difference between clustering-forming thresholds (CFT)
and cluster-wise p value (CWP)? Whether the clustering-forming thresholds is set to the p map from the permutation test? Then how the CWP is calculated?
As the name implies, the CFT is used to form the cluster. Think of you sig map as a landscape with mountains and valleys. If you were to fill the landscape with water, you'd get some islands. The number and size would depend on how high the water was. If you think of each island as a cluster, then the water level is the CFT. In a purely random landscape, some islands would form purely by chance. The propbablility of getting an island (cluster) of a certain size or bigger is the CWP. You have to choose the CFT. The CWP is computed from permutation.
- In the paper ''False positive rates in surface-based anatomical
analysis, you suggested ''For surface area, one would need CFT 0.001 and FWHM>10 mm". Whether I could understand this as "the bigger smooth kernel the better"? The default smooth kernel in FSFAST is 5mm, which one should I use?
If you use permutation, then you can set the FWHM to anything. The optimal size depends on the size of your blobs.
- For the quality assurance in 'tkregister-sess', if I can use the QA
value as a exclude criteria during analysis? If so is there any reference I can cite?
There is no reference. Usually, I look at the worst data sets and work my way to the better data sets to determine where the threshold should be.
Looking forward to your kind suggestions.
Best regards,
Lizhi
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Greve, Douglas N.,Ph.D. -
Zhi Li