Hi Bryson,

That looks mostly fine to me, but note that the design.mat file you feed into the dual-regression will only affect how it runs randomise. And you can run randomise outside of the dual_regression script, which is nice for limiting it to the components you care about.

Once you make the file(s) I mentioned (e.g., your dr_stage2_avg_ic0004), you’ll set up a two-group design in Glm (e.g., design_2sample.mat) and use that in randomise. So, if your input image has 80 volumes (40 subjects in each group), your design_2sample.mat would have 80 rows and at least two columns (one for each group). You would not set up exchangeability blocks with this design (i.e., ignore the group column). 

randomise -i dr_stage2_avg_ic0004 -o randomise_out -d design_2sample.mat -t design_2sample.con

Just make sure the volumes/subjects in your dr_stage2_avg_ic0004.nii.gz correspond to the rows in your design_2sample.mat file.

Cheers,
David


On Nov 10, 2014, at 11:20 AM, Bryson Dietz <[log in to unmask]> wrote:

Hello Eugene and David,

Here is my analysis, to be sure everything I am doing makes sense:

1) Use MELODIC on each run for each subject (in each group) to remove unwanted noise
2) Apply a group concat-ICA on the denoised data from both groups (registered to standard space)
3) Feed the Melodic_IC output from step 2 into Dual-Regression (simplified GLM shown below)
<image.png>
4) Use the method mentioned by David to average the runs (dr_stage2_output) for each subject.
5) Run randomize on each component

I guess I would not need a GLM for randomize, since I only have one input (i.e. dr_stage2_avg_ic0004).

Will this method show me statically significant differences between groups for each IC (that I run randomize on)?

Thanks again for taking the time to help me out here,

Bryson

On Mon, Nov 10, 2014 at 10:41 AM, Eugene Duff <[log in to unmask]> wrote:
Hi - 

 I had missed you were using randomise: this makes the multi-level approach tricky, but the model you suggested above should work (ensuring you set exchangability blocks as subjects).  This will be equivalent to taking the means.

Cheers,

Eugene 

On 9 November 2014 23:07, Bryson Dietz <[log in to unmask]> wrote:

Great, thanks Eugene and David. I appreciate your responses.

If I may be so BOLD (had to throw in a pun) as to ask a follow up question; are these two methods equivalent? Or is one more valid than the other?

Thanks again,

Bryson

Hi Bryson,

Yes, you could average the dr_stage2 output from the runs/sessions. I posted some code to help with this procedure in this thread:

JISCMail - FSL Archive - Re: NETWORK INTEGRATION SCORE (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;6b8d55bf.1109)

Hope this helps.

Cheers,
David


On Nov 8, 2014, at 2:53 PM, Bryson Dietz <[log in to unmask]> wrote:

Hello,

I just have a question regarding my resting-state analysis (3 runs, single-session, 6 subjects (for each group), 2 groups). Initially I made my GLM, essentially the same as mentioned below (I pasted the image from the link below in this email).

https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1010&L=fsl&F=&S=&P=159920

Upon reading his response from Stephen Smith,

https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1010&L=fsl&F=&S=&P=163366

It appears that this analysis ignores variance between subjects. In the above link Stephen mentions "Hence I would recommend just averaging the 3 sessions' images (output by the dual regression) for each subject"

I would just like to clarify, does Stephen mean to average the sessions (in my case runs) output from the dual-regression run from the above link (with the GLM shown below)? Also, I am assuming that I can simply exchange sessions with runs.

I hope that I have made my question clear!

Thanks,

Bryson