Hello all,
I am interested in using the mri_glmfit simulation to control for multiple comparisons in data I have run on the surface in AFNI. Before doing this, I have a few questions:
1. What does the simulation with the mc-z flag do, exactly? It claims to be comparable to AFNI's AlphaSim, but it takes a maximum cluster area for each iteration, which is not exactly what AlphaSim does. Here is my guess:
Given a surface, a given smoothness of the data, and a given per-vertex threshold, for each iteration the simulation populates that surface with random data taken from a normal distribution, thresholds the data, and applies the smoothness of the actual data (supplied as an input parameter). It then computes the maximum cluster size in area for that "image". Doing this n iterations gives a distribution of maximum cluster sizes that occur for random data of a given smoothness, and taking cluster sizes above a certain percentile rank controls for the FWE at a level equal to that percentile rank (e.g., 95th% controls for FWE = .05). AlphaSim does something similar, although instead of taking maximum cluster sizes at each iteration it computes all given cluster sizes. AlphaSim also allows for different cluster connectivity radius, but it seems Freesurfer computes only for neighboring vertices. All in all, if this is correct, it seems like a good implementation.
2. It is my understanding that one could bypass running the glm in Freesurfer and only compute the simulation, as the simulation only needs information about the surface, and the smoothness of the data (which are supplied by the user). To do so, you have to "fake out" Freesurfer to bypass glm, but that turns out to be pretty painless.
3. In a future distribution, is it possible to modify this procedure to also output maximum cluster sizes in terms of number of nodes, rather than area?
Can you please let me know if I am mistaken in any of these assumptions? Thanks in advance.
Anthony
1. Correct on both counts. When I wrote the simulation, I was only trying to replicate the random fields analysis. But with a simulation, you have more freedom that I am not yet exploiting. 2. This is what we are already doing with mc-z 3. I'm working on this as well. It turns out that the random fields approximation works a lot better when using the number of vertices.
Also, I've run mc-z simulations under a bunch of thresholding and FWHM conditions for whole-hemisphere cortex labels. These will be integrated in new version of FS, but I've put them here ftp://surfer.nmr.mgh.harvard.edu/transfer/outgoing/flat/greve/mult-comp-cor.tar.gz as well. There's a README file in there. This will make running your own time-consuming simulations unnecessary (when using the cortex mask at least).
doug
Anthony Dick wrote:
Hello all,
I am interested in using the mri_glmfit simulation to control for multiple comparisons in data I have run on the surface in AFNI. Before doing this, I have a few questions:
- What does the simulation with the mc-z flag do, exactly? It claims to
be comparable to AFNI's AlphaSim, but it takes a maximum cluster area for each iteration, which is not exactly what AlphaSim does. Here is my guess:
Given a surface, a given smoothness of the data, and a given per-vertex threshold, for each iteration the simulation populates that surface with random data taken from a normal distribution, thresholds the data, and applies the smoothness of the actual data (supplied as an input parameter). It then computes the maximum cluster size in area for that "image". Doing this n iterations gives a distribution of maximum cluster sizes that occur for random data of a given smoothness, and taking cluster sizes above a certain percentile rank controls for the FWE at a level equal to that percentile rank (e.g., 95th% controls for FWE = .05). AlphaSim does something similar, although instead of taking maximum cluster sizes at each iteration it computes all given cluster sizes. AlphaSim also allows for different cluster connectivity radius, but it seems Freesurfer computes only for neighboring vertices. All in all, if this is correct, it seems like a good implementation.
- It is my understanding that one could bypass running the glm in
Freesurfer and only compute the simulation, as the simulation only needs information about the surface, and the smoothness of the data (which are supplied by the user). To do so, you have to "fake out" Freesurfer to bypass glm, but that turns out to be pretty painless.
- In a future distribution, is it possible to modify this procedure to
also output maximum cluster sizes in terms of number of nodes, rather than area?
Can you please let me know if I am mistaken in any of these assumptions? Thanks in advance.
Anthony
Doug-
Let me be the first to say thank you for saving me massive amounts of simulation time. One bug though: the right hemisphere files say they were created using the left hemisphere, and mri_surfcluster rejects them. Would it be verboten to just alter the csd files, or do you have versions that were created off of fsaverage's right hemisphere?
matacao:mult-comp-cor porterj$ head fsaverage/rh/cortex/fwhm19/abs/th33/mc-z.csd # simtype null-z # anattype surface fsaverage lh # FixGroupSubjectArea 1 # merged 0 # contrast NA # seed 1271355821 # thresh 3.300000 # threshsign 0.000000 # searchspace 74490.928733 # nullfwhm 19.000000
Jim Porter, M.A. Graduate Student Clinical Science& Psychopathology Research University of Minnesota
On 7/22/64 1:59 PM, Douglas N Greve wrote:
- Correct on both counts. When I wrote the simulation, I was only
trying to replicate the random fields analysis. But with a simulation, you have more freedom that I am not yet exploiting. 2. This is what we are already doing with mc-z 3. I'm working on this as well. It turns out that the random fields approximation works a lot better when using the number of vertices.
Also, I've run mc-z simulations under a bunch of thresholding and FWHM conditions for whole-hemisphere cortex labels. These will be integrated in new version of FS, but I've put them here ftp://surfer.nmr.mgh.harvard.edu/transfer/outgoing/flat/greve/mult-comp-cor.tar.gz as well. There's a README file in there. This will make running your own time-consuming simulations unnecessary (when using the cortex mask at least).
doug
Anthony Dick wrote:
Hello all,
I am interested in using the mri_glmfit simulation to control for multiple comparisons in data I have run on the surface in AFNI. Before doing this, I have a few questions:
- What does the simulation with the mc-z flag do, exactly? It claims
to be comparable to AFNI's AlphaSim, but it takes a maximum cluster area for each iteration, which is not exactly what AlphaSim does. Here is my guess:
Given a surface, a given smoothness of the data, and a given per-vertex threshold, for each iteration the simulation populates that surface with random data taken from a normal distribution, thresholds the data, and applies the smoothness of the actual data (supplied as an input parameter). It then computes the maximum cluster size in area for that "image". Doing this n iterations gives a distribution of maximum cluster sizes that occur for random data of a given smoothness, and taking cluster sizes above a certain percentile rank controls for the FWE at a level equal to that percentile rank (e.g., 95th% controls for FWE = .05). AlphaSim does something similar, although instead of taking maximum cluster sizes at each iteration it computes all given cluster sizes. AlphaSim also allows for different cluster connectivity radius, but it seems Freesurfer computes only for neighboring vertices. All in all, if this is correct, it seems like a good implementation.
- It is my understanding that one could bypass running the glm in
Freesurfer and only compute the simulation, as the simulation only needs information about the surface, and the smoothness of the data (which are supplied by the user). To do so, you have to "fake out" Freesurfer to bypass glm, but that turns out to be pretty painless.
- In a future distribution, is it possible to modify this procedure
to also output maximum cluster sizes in terms of number of nodes, rather than area?
Can you please let me know if I am mistaken in any of these assumptions? Thanks in advance.
Anthony
Yea, that was a bug in the simulation program I used. I did confirm that the simulations were done on the correct hemisphere. Feel free to edit the files. I'll fix the master when I get a chance.
doug
James Porter wrote:
Doug-
Let me be the first to say thank you for saving me massive amounts of simulation time. One bug though: the right hemisphere files say they were created using the left hemisphere, and mri_surfcluster rejects them. Would it be verboten to just alter the csd files, or do you have versions that were created off of fsaverage's right hemisphere?
matacao:mult-comp-cor porterj$ head fsaverage/rh/cortex/fwhm19/abs/th33/mc-z.csd # simtype null-z # anattype surface fsaverage lh # FixGroupSubjectArea 1 # merged 0 # contrast NA # seed 1271355821 # thresh 3.300000 # threshsign 0.000000 # searchspace 74490.928733 # nullfwhm 19.000000
Jim Porter, M.A. Graduate Student Clinical Science& Psychopathology Research University of Minnesota
On 7/22/64 1:59 PM, Douglas N Greve wrote:
- Correct on both counts. When I wrote the simulation, I was only
trying to replicate the random fields analysis. But with a simulation, you have more freedom that I am not yet exploiting. 2. This is what we are already doing with mc-z 3. I'm working on this as well. It turns out that the random fields approximation works a lot better when using the number of vertices.
Also, I've run mc-z simulations under a bunch of thresholding and FWHM conditions for whole-hemisphere cortex labels. These will be integrated in new version of FS, but I've put them here ftp://surfer.nmr.mgh.harvard.edu/transfer/outgoing/flat/greve/mult-comp-cor.tar.gz as well. There's a README file in there. This will make running your own time-consuming simulations unnecessary (when using the cortex mask at least).
doug
Anthony Dick wrote:
Hello all,
I am interested in using the mri_glmfit simulation to control for multiple comparisons in data I have run on the surface in AFNI. Before doing this, I have a few questions:
- What does the simulation with the mc-z flag do, exactly? It
claims to be comparable to AFNI's AlphaSim, but it takes a maximum cluster area for each iteration, which is not exactly what AlphaSim does. Here is my guess:
Given a surface, a given smoothness of the data, and a given per-vertex threshold, for each iteration the simulation populates that surface with random data taken from a normal distribution, thresholds the data, and applies the smoothness of the actual data (supplied as an input parameter). It then computes the maximum cluster size in area for that "image". Doing this n iterations gives a distribution of maximum cluster sizes that occur for random data of a given smoothness, and taking cluster sizes above a certain percentile rank controls for the FWE at a level equal to that percentile rank (e.g., 95th% controls for FWE = .05). AlphaSim does something similar, although instead of taking maximum cluster sizes at each iteration it computes all given cluster sizes. AlphaSim also allows for different cluster connectivity radius, but it seems Freesurfer computes only for neighboring vertices. All in all, if this is correct, it seems like a good implementation.
- It is my understanding that one could bypass running the glm in
Freesurfer and only compute the simulation, as the simulation only needs information about the surface, and the smoothness of the data (which are supplied by the user). To do so, you have to "fake out" Freesurfer to bypass glm, but that turns out to be pretty painless.
- In a future distribution, is it possible to modify this procedure
to also output maximum cluster sizes in terms of number of nodes, rather than area?
Can you please let me know if I am mistaken in any of these assumptions? Thanks in advance.
Anthony
freesurfer@nmr.mgh.harvard.edu
-
Anthony Dick -
Douglas N Greve -
James Porter