Print

Print
Hi Michelle,

It isn't possible to do group comparisons in this design because, as you say, this would be a between-subjects effect. In this design, only within-subject effects (task condition), and interactions (task condition by group) are possible, but not group alone.

Regardless, there is no need for the nuisance EVs there. Any effect age or sex could have in the main effect of task or in the interaction task by group is already taken care of by the subject-specific EVs.

If, however, you create a different design in which between-subject effects can be tested, then age and sex could be entered. If think this was discussed in an earlier message, please see in this spreadsheet: https://dl.dropboxusercontent.com/u/2785709/outbox/mailinglist/design_michelle.ods

All the best,

Anderson


On 13 December 2016 at 20:31, Michelle VanTieghem <[log in to unmask]> wrote:
Hi Anderson, 

To clarify, group is between subject (not within subject), and emotion is within subject task condition. Therefore, I do think that covariates of age and sex would make a difference. Particularly when I have a very wide age range across a developmental sample. Essentially I would like to do a repeated measures ANCOVA to test for group x Emotion interaction while controlling for individual differences in age, sex, accuracy on the task, ect.. Can you please advise how this can be modeled in the design? I attached my design in the previous email (that generated errors: matrix is singular). 



Thanks! 
Michelle 

On Tue, Dec 13, 2016 at 4:17 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Michelle,

So this is a repeated measurements design in which each subject was assessed three times. Since age and sex doesn't change for the 3 visits, these don't need to be modelled, as the subject-specific EVs already take care of these nuisance. So, just remove age and sex from the model.

All the best,

Anderson


On 12 December 2016 at 23:09, Michelle VanTieghem <[log in to unmask]> wrote:
Hi Anderson, 

Great thanks --that fixed the ftoz conversion error. 

I would like to include between subject covariates of age and sex in my design1 (group x emotion). However, when I tried to add these covariates into the design matrix, I am getting the following error:

An exception has been thrown
Runtime error:- detected by Newmat: matrix is singular

MatrixType = Crout  # Rows = 75; # Cols = 75
Trace: Crout(lubksb); GeneralSolv; Gsmanager::flame_stage1_onvoxel; Gsmanager::flame_stage1; Gsmanager::run.

 I think this means that I did not set up the design properly. See attached for design.mat and contrast file. Can you please advise on how to add covariates to the model, so that I can control for between subject differences in Age and sex? 

Thanks!
Michelle 


On Mon, Dec 12, 2016 at 11:13 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Michelle,

One correction: the dof1 should be the dof of the numerator, so you'd swap the first and second values.

All the best,

Anderson


On 9 December 2016 at 11:11, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Michelle,

Yes, the conversion is right. Re: large design, I'm not sure if there is a script that would do it... sorry.

All the best,

Anderson


On 8 December 2016 at 18:02, Michelle VanTieghem <[log in to unmask]> wrote:
Hi Anderson, 

So If I have 24 rows of my design1_contrast.con file, and then 24 columns of my ftest file, then the numerator of df is 24, correct? 

ftoz -zout fzstat fstat1 134 24 

Would give me the correctly converted Z statistic for my F-tests? 

Also, you recommended using FEAT instead of the command line.  Did you mean setting it up in the FEAT gui? I haven't been doing that because the FEAT gui freezes when I try to enter this model, because it has so many EVs. I cannot get a work around so that this doesn't happen. If you have any suggests I would greatly appreciate! 
Alternatively, is there a way to hand-make the .fsf file without having to open the GUI to enter my design? Can I use a .fsf file to only perform cluster thresholding on the f-test files?

Thanks!

On Thu, Dec 8, 2016 at 5:33 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Michelle,

F-tests have the same DOF as the t-test (for the denominator). For the numerator, the t test has always df=1 (usually omitted) and the F-test has DOF that is the same as the number of lines in the contrast (e.g., for a simple 3 group comparison with or without nuisance variables, it's 2).

All the best,

Anderson


On 7 December 2016 at 18:31, Michelle VanTieghem <[log in to unmask]> wrote:
Hi Anderson, 

Also, how do I determine the DOF for my F-tests? 
For the z-tests I was using # rows in the design - # of columns (207 - 77 = 134)
However, I'm having trouble thinking about what it should be for Ftests since this is a mixed effects design... 
thanks,
Michelle 

On Wed, Dec 7, 2016 at 8:39 AM, Michelle VanTieghem <[log in to unmask]> wrote:
Hi Anderson, 

Do you mean setting it up in the FEAT gui? I haven't been doing that because the FEAT gui freezes when  I try to enter this model, because it has so many EVs. Is there a way to use an .fsf file without having to open the GUI to enter my design? Can I use a .fsf file to only perform cluster thresholding on the f-test files?

Thanks!
M

On Wed, Dec 7, 2016 at 2:55 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Michelle,

You'd use ftoz to convert to a z-score map. That said, if you'll use the random field theory for inference, it might be much simpler and less error-prone to simply use FEAT. It does everything for you and you don't have to worry about these commands.

All the best,

Anderson


On 6 December 2016 at 21:33, Michelle VanTieghem <[log in to unmask]> wrote:
Hi Anderson,

Ok great thanks. I have run the f-tests. Now I would like to use smoothest and cluster in order to identify significant clusters of activation for F4, the Group x Emotion interaction. However, cluster, as far as I can tell from the information on FSL wiki, is used with zstat files.
Can I do the same exact thing, but replace zstat files (of a specific contrast) with the fstat files (to test the omnibus interaction)?
When testing F-tests in cluster, does make sense to use a t- threshold (e.g. 3.1)?
I would think there should be F threshold, but that doesn't appear to be an option in cluster....
Here is an example of what I would like to do:

smoothest -d 134 -r res4d.nii.gz -m mask.nii.gz >> cluster_correction/smoothness_output_wholebrain_mask.txt

fslmaths fstat4.nii.gz -mas mask.nii.gz cluster_correction/fstat4_mask.nii.gz

cluster -i fstat4_mask.nii.gz -t 3.1 -o fstat4_clusters_3.1 -d 0.043 --volume=234781 -p .05 --othresh=fstat4_thresh_3.1.nii.gz > cluster_info_wholebrain_3.1_fstat4.txt

I appreciate any input!
Thank you,
Michelle




--
Michelle VanTieghem
PhD student in Psychology
Developmental Affective Neuroscience Lab
Columbia University 



--
Michelle VanTieghem
PhD student in Psychology
Developmental Affective Neuroscience Lab
Columbia University 




--
Michelle VanTieghem
PhD student in Psychology
Developmental Affective Neuroscience Lab
Columbia University 





--
Michelle VanTieghem
PhD student in Psychology
Developmental Affective Neuroscience Lab
Columbia University 




--
Michelle VanTieghem
PhD student in Psychology
Developmental Affective Neuroscience Lab
Columbia University