Hi Anderson,
I have been doing the longitudinal and cross sectional analysis that you recommended. I also realized that my waitlist group has more participants who takes CNS Stimulant medication. I added an EV to my cross-sectional design to control for medications effect. Since these meds can change the brain and the effect of intervention, I am wondering if I have to also add an EV for my longitudinal design to control for their effect. If so, how should I do that?
Just so you know, I have 22 participants in treatment group (3 take meds) and 20 in my waitlist group (10 take meds).
Many thanks,
Sara
________________________________
From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Anderson M. Winkler [[log in to unmask]]
Sent: Sunday, October 13, 2019 5:59 PM
To: [log in to unmask]
Subject: Re: [FSL] Study design and analysis
Hi Sara,
Please see below:
On Thu, 10 Oct 2019 at 14:06, [log in to unmask]<mailto:[log in to unmask]> <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hi Anderson,
Thank you very much for your response. It makes a lot of sense. However, I have a few questions putting these statistical analysis in the context of DTI and Resting State analysis.
Resting State:
For Analysis 1, should I calculate the differences you explained for each component in my group ICA and then run all of them together in a single call to PALM? If so, I would have a total of 68 inputs for PALM (17 components * (2 calculated differences for treatment group + 2 calculated differences for waitlist group). Right? If so, It is a lot of analysis and I doubt I could find any significant result corrected over modalities.
You can do all in a single call if there is enough memory but it isn't needed. As each group has its own run, that reduces to two calls with 34 inputs each. For the conjunction, the permutations don't need be in sync, and don't need be corrected for the fact that two are being done, so this could also mean separate calls. So four calls, each with 17 inputs.
Since my group assignment has been random, I can definitely consider Analysis 2. In this case, I can compare both scan 2 and scan 3 of the two groups for all 17 components (a total of 34 inputs) in one single call to PALM, using Scan1 as a baseline regressor with options -corrmod (and -corrcon). So far, it seems the best option to me. I am wondering if using NPC would increase the power of Analysis 2. However, I reckon that it is probably not possible to run NPC over 17 components, right?
I may be missing something... why would it be interesting to do NPC over scans 2 and 3? Seems to be these could be quite different, no? NPC over components isn't a great idea because the ICs have (by definition) weak spatial overlap to make NPC meaningful, though it wouldn't be wrong to combine.
In any case, when NPC is done, it acts upon all inputs, so if you enter, say, 17 inputs for scan 2 and 17 inputs for scan 3, NPC will combine all 34, and not make two sets of combinations (over ICs, one per scan), nor 17 sets of combinations (over scans, one per IC). So, here either you choose one scan, or you run them separately and correct with Bonferroni, or write a separate wrapper around PALM that will do the correction non-parametrically (involved but not impossible).
DTI:
For Analysis 2, I can use NPC over 4 modalities (FA, MD, RD, AD) and over the two group comparisons (scan 2 and scan 3) in one call to PALM. Is that right or is it best to run them separately and not use NPC.
Because they way I understood how the experiment happened I would not do NPC over scans 2 and 3. However, doing NPC over the 4 DTI measures separately for scan 2 and for scan 3 seems fine, and PALM can do. For correction, it's a similar case as for ICA above.
Hope this helps!
All the best,
Anderson
Thanks a lot in advance,
Sara
________________________________
From: FSL - FMRIB's Software Library [[log in to unmask]<mailto:[log in to unmask]>] on behalf of Anderson M. Winkler [[log in to unmask]<mailto:[log in to unmask]>]
Sent: Thursday, October 10, 2019 7:44 AM
To: [log in to unmask]<mailto:[log in to unmask]>
Subject: Re: [FSL] Study design and analysis
Hi Sara,
This is very interesting design that can be analysed in a number of different ways. Let's recap:
Group 1 (Treatment): scan1 -> treatment -> scan2 -> nothing -> scan3
Group 2 (Waitlist): scan1 -> nothing -> scan2 -> treatment -> scan3
Is the above correct? If so, here are two possible analysis:
Analysis 1 (longitudinal):
- For group 1, show whether scan2 > scan1 (or the reverse, correcting for the fact that you looked into both directions), and whether scan3 > scan1. You can do a conjunction of the two results, showing where the effect persisted for the two follow-up scans versus scan1. You can also do a NPC over these two differences, to investigate "any" change over time. However, if the point is to investigate whether the result remain stable between scan2 and scan3, that is probably not the best test as it will pick differences happening anywhere over time, from scan1 to scan3.
- For group 2, show whether scan3 > scan2 (or the reverse, as above), and whether scan3 > scan1. Again, you can do a conjunction to show where the effect in scan3 the same when compared to both scan1 and scan2. Likewise, NPC can be done, although in this case with a similar interpretation as above, so perhaps here the conjunction is more interesting.
Analysis 2 (cross sectional):
Is the allocation of subjects into groups random? If so, you can take scan2 alone, and do a cross-sectional analysis comparing groups 1 and 2. Power is likely going to increase if you also include scan1 as a baseline regressor, but only as long as the allocation of subjects into groups was random. You may also strengthen the case by showing that, as scan1, there were no group differences. Finally, you can also compare the two groups at scan3, to investigate whether group2 caught up with group1.
Analysis 3 (mixed):
You can use the design_yawu.ods as a template for a mixed design, but the contrasts would be different than in that example, and would be constructed to test the longitudinal effects as above, further compared between groups. However, I believe assembling separate longitudinal analysis as in Analysis 1 above, which can then be tested in PALM with options -corrmod and -corrcon is the safest. The reason is that, by restricting all analyses to longitudinal effects that include only 2 timepoints each (as above), compound symmetry will always hold. When 3 timepoins are used (as in design_yawu.ods), that assumption may or may not hold. Sometimes there is no way around, but in your case you have that option, i.e., you can bypass the assumption of compound symmetry altogether.
So... consider strongly doing analysis 1 and 2, and perhaps do all the analyses 1 in a single call to PALM in which you use multiple inputs and correct with -corrmod (and -corrcon). Feel free to post again if anything isn't clear.
Hope this helps!
All the best,
Anderson
On Tue, 8 Oct 2019 at 15:28, [log in to unmask]<mailto:[log in to unmask]> <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hi Anderson,
I am working on a project looking at the effect of rehabilitation on functional and structural connectivity in children with Developmental Coordination Disorder. I have two groups of children with DCD (Treatment and Waitlist). Each child goes through MRI 3 times (3 months apart). Children in Treatment group received 3-month rehabilitation after their first MRI scan (DTI and Resting state), and then received their second MRI and a follow up scan after 3 more months. Children in waitlist group received their first scan, waited for three month without rehabilitation and then received their second scan and 3-month rehabilitation and their final scan after rehabilitation. In other word, I have two groups, each has three scans with a total of ~ 110 scans after excluding those with high FD.
My research questions are 1) whether rehabilitation has been effective in improving functional connectivity and microstructure properties in these children and 2) whether brain changes still remain after three months (in follow up scan).
I have been struggling to find the best analysis approach for my data. This is my planned analysis:
1. DTI TBSS, running npc with FA, MD, AD, RD as modalities, using a design similar to this: https://s3-us-west-2.amazonaws.com/andersonwinkler/mailinglist/design_yawu.odshttps://s3-us-west-2.amazonaws.com/andersonwinkler/mailinglist/design_yawu.ods
2. ICA with 17 components, running PALM with a similar design as above or using swe for each individual component.
Is there any better way to deal with this data to look at the effect of rehabilitation while controlling for maturation effect and also looking at follow up.
Many thanks,
Sara
________________________________
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
________________________________
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
________________________________
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
________________________________
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
########################################################################
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
|