help
help > RE: ICA: Parameter Choice and 2nd Level Effects ?
Oct 17, 2016 07:10 PM | Shady El Damaty - Georgetown University
RE: ICA: Parameter Choice and 2nd Level Effects ?
Hi Alfonso,
I'm not quite sure what you the notation min(56.1*Nsubjects,Nvoxels) means or how you got 7.5 components. Is this the minimum of 56.1*133 as the number of voxels increase? The square root of 56 is ~7.5.. also (56.1*133)/1000 = ~7.5.
Thanks for the hint about validation by partitioning the dataset into smaller subject pools. I'll probably end up trying that to avoid developing a whole new toolbox for fmri ica model validation :P
Originally posted by Alfonso Nieto-Castanon:
I'm not quite sure what you the notation min(56.1*Nsubjects,Nvoxels) means or how you got 7.5 components. Is this the minimum of 56.1*133 as the number of voxels increase? The square root of 56 is ~7.5.. also (56.1*133)/1000 = ~7.5.
Thanks for the hint about validation by partitioning the dataset into smaller subject pools. I'll probably end up trying that to avoid developing a whole new toolbox for fmri ica model validation :P
Originally posted by Alfonso Nieto-Castanon:
Hi
Shady,
Unfortunately there is no such formula for the multiple-subject case. To be a bit more precise, the actual degrees of freedom of the multiple-subject data will just be in this case min(56.1*Nsubjects,Nvoxels) (because the components across subjects cannot possibly be perfectly aligned) but what this really tells you is only that the multiple-subject data is not going to show the very sharp drop that you found in the single-subject data (or at least it is not going to show it until ~7.500 components), but unfortunately this tells you nothing about how many of these components are really replicable or significant. For that you need something like Bartlett's test, model comparison measures, etc. which gets us back to our starting point. In practical terms, one easy option to identify significant/reliable components when you have sufficient subjects (as you do), is to repeat the ICA analyses across multiple smaller subsets of subjects, and then simply select those components that appear consistently across many or most of those smaller analyses.
Hope this helps
Alfonso
Originally posted by Shady El Damaty:
Unfortunately there is no such formula for the multiple-subject case. To be a bit more precise, the actual degrees of freedom of the multiple-subject data will just be in this case min(56.1*Nsubjects,Nvoxels) (because the components across subjects cannot possibly be perfectly aligned) but what this really tells you is only that the multiple-subject data is not going to show the very sharp drop that you found in the single-subject data (or at least it is not going to show it until ~7.500 components), but unfortunately this tells you nothing about how many of these components are really replicable or significant. For that you need something like Bartlett's test, model comparison measures, etc. which gets us back to our starting point. In practical terms, one easy option to identify significant/reliable components when you have sufficient subjects (as you do), is to repeat the ICA analyses across multiple smaller subsets of subjects, and then simply select those components that appear consistently across many or most of those smaller analyses.
Hope this helps
Alfonso
Originally posted by Shady El Damaty:
Hi Alfonso,
Ok that makes a bit of sense -- I tried calculating the degrees of freedom using the equation you listed as follows:
TR = 2.28
NumScans = 150
Filter Width = 0.09-0.008 = 0.082
NumScans * FilterWidth/NyqFreq = 150*0.0820/((1/2.28)*.5 = 56.1
So how do I incorporate the number of subjects (N=133) to calculate the final number of components?
PS -- are there any resources I can consult to understand where you got that equation from? this is all very interesting!!
Originally posted by Alfonso Nieto-Castanon:
Ok that makes a bit of sense -- I tried calculating the degrees of freedom using the equation you listed as follows:
TR = 2.28
NumScans = 150
Filter Width = 0.09-0.008 = 0.082
NumScans * FilterWidth/NyqFreq = 150*0.0820/((1/2.28)*.5 = 56.1
So how do I incorporate the number of subjects (N=133) to calculate the final number of components?
PS -- are there any resources I can consult to understand where you got that equation from? this is all very interesting!!
Originally posted by Alfonso Nieto-Castanon:
Hi
Shady,
Those MDL estimates based on single-subject data are likely not terribly meaningful, as the plots seem to suggest. They are probably all turning out to be 56 because that may be the degrees of freedom of your single-subject data after filtering (i.e. approximately the number of frequency samples within your band-pass window = NumberOfScans * FilterWidth / NyquistFrequency). This only partially relates to the "optimal" number of components for the group-level dimensionality reduction or ICA steps, as the number of subjects in your study plays a very important role in limiting the number of components that you can reliably estimate.
Hope this helps
Alfonso
Originally posted by Shady El Damaty:
Those MDL estimates based on single-subject data are likely not terribly meaningful, as the plots seem to suggest. They are probably all turning out to be 56 because that may be the degrees of freedom of your single-subject data after filtering (i.e. approximately the number of frequency samples within your band-pass window = NumberOfScans * FilterWidth / NyquistFrequency). This only partially relates to the "optimal" number of components for the group-level dimensionality reduction or ICA steps, as the number of subjects in your study plays a very important role in limiting the number of components that you can reliably estimate.
Hope this helps
Alfonso
Originally posted by Shady El Damaty:
Wow! great response Alfonso --
There are all sorts of issues I ran into trying to implement this on my own (memory issues nonwithstanding) so I have been using existing implementations. I've managed to estimate MDL for each individual subject's denoised data using the icatb_estimate_dimension.m function in the GIFT toolbox. However I'm not too confident in the results since the MDL estimate hasn't been performed using the group level data but rather individual subjects (which all turned out to be 56...). I've attached a plot of the MDL/AIC estimate with shaded error bars across subjects. Does this look right to you?
Thank you so much for your continued support!!
There are all sorts of issues I ran into trying to implement this on my own (memory issues nonwithstanding) so I have been using existing implementations. I've managed to estimate MDL for each individual subject's denoised data using the icatb_estimate_dimension.m function in the GIFT toolbox. However I'm not too confident in the results since the MDL estimate hasn't been performed using the group level data but rather individual subjects (which all turned out to be 56...). I've attached a plot of the MDL/AIC estimate with shaded error bars across subjects. Does this look right to you?
Thank you so much for your continued support!!
Threaded View
| Title | Author | Date |
|---|---|---|
| Shady El Damaty | Sep 18, 2016 | |
| Alfonso Nieto-Castanon | Sep 29, 2016 | |
| David Pagliaccio | Jan 10, 2018 | |
| Julia Binnewies | Oct 12, 2016 | |
| Julia Binnewies | Oct 18, 2016 | |
| Shady El Damaty | Oct 18, 2016 | |
| Jeff Browndyke | Oct 18, 2016 | |
| Shady El Damaty | Oct 18, 2016 | |
| Shady El Damaty | Oct 12, 2016 | |
| Alfonso Nieto-Castanon | Oct 14, 2016 | |
| Shady El Damaty | Oct 16, 2016 | |
| Alfonso Nieto-Castanon | Oct 17, 2016 | |
| Shady El Damaty | Oct 17, 2016 | |
| Shady El Damaty | Oct 21, 2016 | |
| Alfonso Nieto-Castanon | Oct 25, 2016 | |
| Jeff Browndyke | Oct 19, 2016 | |
| Alfonso Nieto-Castanon | Oct 20, 2016 | |
| Jeff Browndyke | Oct 20, 2016 | |
| Shady El Damaty | Sep 28, 2016 | |