Hello Anderson,
Thank a lot for the clear explanation. I check for correlation between my events and orthogonalized them when necessary. Should I model positive and negative contrast in a single design with –twotail and -corrcon options? Or use two separate designs. My concern is that the first case would be huge and take lot of time and memory resources..
I promise it´s my last question J
Thousands of thanks
Yacila
Von: FSL - FMRIB's Software Library [mailto:[log in to unmask]] Im Auftrag von Anderson M. Winkler
Gesendet: Dienstag, 17. Mai 2016 08:34
An: [log in to unmask]
Betreff: Re: [FSL] AW: [FSL] AW: [FSL] AW: [FSL] Bonferroni correction of ICA
Hi Deza,
Please, see below:
On 16 May 2016 at 13:59, Deza Araujo, Yacila Isabela <[log in to unmask]> wrote:
Dear Anderson,
I would like to perform more designs to avoid collinearity,
There is no free lunch: splitting across designs sadly won't fix issues related to collinearity. If A and B are variables that can explain the outcome, both should go in the same design, instead of two designs in which one of them is ignored.
Consider an extreme case: A and B that are so highly correlated to the point of being nearly identical. If both are used in the same design, the result will be that neither is found significant, as the unique contribution of either alone is minimal to explain the data in any way, and we cannot disambiguate the effects of each. Now, put each on its own separate design, and run both, correcting across. Because they are nearly identical, and because PALM takes care of such dependency, both will yield separately nearly identical results, and after correction, there will be almost no change over the uncorrected, as the dependency is almost perfect. It's almost as if having run the same test twice, and PALM isn't sensitive to that (in sharp contrast to Bonferroni), and the ultimate result would have been as if the second regressor had been ignored. But this isn't what you want: as both regressors can explain the data, it isn't correct that one is left out.
It is possible to run multiple competing designs, particularly when different modalities would use different set of EVs (e.g., consider the case of having ICV in the analysis of surface area and also global FA in the analysis of TBSS), but evading collinearity wouldn't be a good justification.
That said, it's possible to use NPC to jointly analyse these two designs (-npccon). Compared to an F-test (single design with both EVs), NPC allows disambiguation of direction, whereas the F-test is always two-tailed.
also corrected across modalities and contrasts. Ithink I should go with PALM again. I did not regress global mean, but my contrast are not orthogonal. Can I still go with Bonferroni? Following your formula, it will be: -log10(0.05/N) = 1.301+log10(4) = 1.90.
The 1.90 is right, but if the contrasts are not orthogonal, they share information (variance) and Bonferroni will be conservative. It's still possible to use, at the cost of power.
My question now would be: What is the rational for using PALM and not randomise? Why is it better to consider each IC as a separate modality?
I know both of your papers describing PALM, but I would like to know where exactly to go deeper in case some reviewers ask for this.
PALM and randomise do similar things, with PALM having a super-set of features (and is more experimental). The hope is that in the future all features in PALM will be in randomise. Currently randomise sees one modality at a time, which is fine if they are all independent (Bonferroni). But if dependencies exist, the loss of power can be substantial.
All the best,
Anderson
Thanks a lot again
Yacila
Von: FSL - FMRIB's Software Library [mailto:[log in to unmask]] Im Auftrag von Anderson M. Winkler
Gesendet: Montag, 16. Mai 2016 01:42
An: [log in to unmask]
Betreff: Re: [FSL] AW: [FSL] AW: [FSL] Bonferroni correction of ICA
Hi Yacila,
Please, see below:
On 15 May 2016 at 17:39, Deza Araujo, Yacila Isabela <[log in to unmask]> wrote:
Dear Anderson,
I ran PALM with the following parameters:
# Configuration file for PALM.
# Version alpha86, running in MATLAB.
-i DR.dr/dr_stage2_ic0006.nii
-i DR.dr/dr_stage2_ic0007.nii
-i DR.dr/dr_stage2_ic0008.nii
-i DR.dr/dr_stage2_ic0011.nii
-i DR.dr/dr_stage2_ic0012.nii
-i DR.dr/dr_stage2_ic0013.nii
-d design.mat
-corrmod
-t design con
-m DR.dr/mask.nii
-T
-n 5000
-corrcon
-logp
-demean
-o palm_results
This looks all correct.
The results are saved as “mcfwep”, with their respective contrast and modality and saved as log(p) =1.301. Where and how can I threshold them for visualization?
To view the results, use FSLview or any image viewer of your choice. Use as threshold 1.301, and what is above is significant after correcting for the multiple modalities and contrasts. All good.
I would like to run 3 more tests, so I think I should still correct for multiple comparisons (0.05/4) right?, but I am not sure how to frame the results part, for example.
By more tests do you mean more modalities (ICs), more designs, or more contrasts? If more modalities or more than one design, need to run PALM again. If more contrasts, ideally PALM should be executed again, but if the contrasts are orthogonal to each other, and if the input images have not been normalised by some global measure (which can introduce negative correlations), Bonferroni can be used. In that case, you'd take the "mfwep" (corrected across modalities only), and use as threshold -log10(0.05/N) = 1.301+log10(N), where N is the number of contrasts.
Hope this helps.
All the best,
Anderson
I would be really happy if you can give some advice in this regard
Thanks a lot!
Yacila
Von: FSL - FMRIB's Software Library [mailto:[log in to unmask]] Im Auftrag von Anderson M. Winkler
Gesendet: Freitag, 15. April 2016 11:09
An: [log in to unmask]
Betreff: Re: [FSL] AW: [FSL] Bonferroni correction of ICA
Hi Yacila,
Use the dr_stage2_ic* files. These are the same that would be used with randomise.
The design can be the single group with covariate, that is, this one:
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Average_with_Additional_Covariate
However, you should definitely remove C1, because you know already that, in the places that define the network, the mean is different than zero, so there is no point in testing the intercept. But leave EV1 there, only remove C1 (or use -demean).
Since you'll be testing C2 only (that will become the new C1), it is possible to use actual permutations, so the -ise isn't necessary.
All the best,
Anderson
On 14 April 2016 at 20:09, Deza Araujo, Yacila Isabela <[log in to unmask]> wrote:
Dear Anderson,
Sorry for bugging the FSL mailing list, but I want to be sure that the command that I will use is correct.
You suggested input each components as an image modality, and I think I did not get this point:
As inputs, should I use: -i dr_stage2_ic000.nii ? (because it is already a 4D file with all the subjects-specific component 0 inside) Or -i IC2.nii, -i IC3.nii, etc (from the group analysis)?
It will be a single group T-test with covariate, so I think I should use the –ise. Am I right?.
Thanks a lot!!
Yacila
Von: FSL - FMRIB's Software Library [mailto:[log in to unmask]] Im Auftrag von Anderson M. Winkler
Gesendet: Mittwoch, 6. April 2016 09:40
An: [log in to unmask]
Betreff: Re: [FSL] Bonferroni correction of ICA
Hi Yacila,
Please, see below:
On 5 April 2016 at 17:24, Deza Araujo, Yacila Isabela <[log in to unmask]> wrote:
Dear FSL experts,
I was wondering if the argument of selecting networks „a-priori“ is a good one to support high-but-significant p values. Specifically: I have 7 networks and 4 tests (2-sided) which need a p <0.00089 for being significant, which is very unlikely to get, especially after randomise.
I read somewhere that I can correct only for the number of tests (0.05/4*2) if the networks were previously selected because the implication of these networks with our behavioral measure has already been reported. Is this correct?
Yes, if the selection of the components is completely independent from the data that you have at hand, and the selection is decided upon before you pry into the data, then this is fine.
And what about randomise? I have a big sample (170) and after 5000 permutations I think it is only possible to have p values within certain threshold. Am I correct?
If you consider each IC an imaging modality, then it's possible to do the correction in a less conservative way than Bonferroni using permutation tests in PALM. Input each component with one "-i", use a common design for all ("-d") and use the option "-corrmod". To correct for the two tails, enter both contrasts as usual (positive and negative), and include also the option "-corrcon". The relevant outputs are the files named "mcfwep", or "mcfdrp" if you used FDR.
A strongly recommended option is to use -logp, so that the p-values are saved as -log10(p), which can then be thresholded at -log10(0.05) = 1.301.
All the best,
Anderson
Thanks you very much
--
Yacila Deza Araujo, M.Sc. Neuropsych.
PhD Student
Technische Universität Dresden
Faculty of Medicine Carl Gustav Carus
Department of Psychiatry & Psychotherapy
Section of Systems Neuroscience
Würzburger Straße 35
01187 Dresden
Germany
Phone: +49 (0) 351 463 - 43300 <tel:%2B49%20%280%29%20351%20463%20-%2043300>
Fax: +49 (0) 351 463 - 42202 <tel:%2B49%20%280%29%20351%20463%20-%2042202>
mailto: [log in to unmask]
