Hi Anderson,

Thank you for the detailed answer. I did think about the problem of controls having a correlation with NP test performance but I was thinking of dropping those tests where correlation is positive for both groups. I did not write that in the scheme explicitly. 

I like to do everything with one big model though. But I have some further questions about that. 

1. We have a whole bunch of NP tests (I think 16 each with several component). Do you think one can use this big a number (approximately 52) of variables in the model? 

2. Disease stage is a tricky one too. What exactly I have here is the time after injury. It is zero for controls and greater than zero for patients. Now, is it redundant to have the subjects divided into groups on the basis of injury presence and absence and then use the time since injury again as a variable (all controls will have it zero and this becomes a variable applicable to the patients group only!). 

I have another question about your suggestion of using whole-brain (voxelwise) for correlation. Is there any option in FSL to do such kind of analysis? I know i can use SPM for this but I am curios if I can do it in FSL too?

Thanks,

Eyesha 

---

Date: Fri, 29 Apr 2016 10:35:21 +0100
From: [log in to unmask]
Subject: Re: [FSL] Statistical analyses- which mean values to use
To: [log in to unmask]

Hi Eyesha,

Please, see below:

On 29 April 2016 at 03:23, Eyesha <[log in to unmask]> wrote:

Dear experts,

I have a question about statistical analyses of DTI data.

I am comparing DTI parameters between two groups (controls and patients) and eventually I also want to look at the correlation between performance on neuropsychological tests  and DTI parameters. This is my plan of action (after performing the basic DTI processing) after much thought:

A- Run the randomise tool to obtain t-statistics and p-values for voxel-wise comparison between the two groups, corrected for multiple comparisons,

B-      Threshold the output at the desired p-value. I am choosing 0.05 and further dividing it by 4 to account for multiple analyses (4 DTI parameters: FA, MD, RD and L1).

C-      Use the thresholded image to find tracks with significantly different values.

D-      Get the mean value in each of the significant track (here I use only that part of the tract where significant differences were found, i.e. only the part of the tract that was included in the thresholded image and not the entire track).

E-      Use the mean values found in D to investigate the correlation with NP test performance in the patient group.

F-      Use regression to account for variables like age, sex and disease stage.

If the whole sample (patients and controls) were used for this follow up test, this would be a circular analysis, unless NP were perfectly orthogonal to group. Using only one of the groups, this isn't really an issue.

However, it's still problematic in other grounds: suppose there is a correlation in that particular area found in A above in the patient group. How can one know if that is a noteworthy result if one hasn't looked into the control group? What if there is similar correlation between the DTI measures and NP in the control group in that particular region? Then there is nothing interesting to be reported as far as NP goes, and although the correlation may be significant in patients, it's still a false positive. Conversely, if there is a different correlation between NP and DTI measures according to group, then the difference between groups in terms of DTI varies according to NP, and the initial analysis of group differences is likewise misleading.

To deal with this, perhaps be best thing is to drop these multiple models/stages and run a single one, that has:

EV1: intercept

EV2: group (+1/-1)

EV3: NP

EV4: NP*group (interaction)

EV5, EV6, etc: age, sex, disease stage, etc.

Then test for the interaction with contrasts [0 0 0 1 0 ...] and [0 0 0 -1 0 ...]. If significant, then it's something interesting. If not significant, you can investigate the effect of group and NP with additional contrasts then.

 

The point where I got stuck was if I should use only the tracts with significant difference for further analyses or all the tracts. And if I should just use the mean value in the entire tract from the skeletonized parametric image, or  un-skeletonized mean value in the entire tract or the skeletonized value from the significantly different region of the tract? I see people doing all of these things but I am not convinced for all.

Here I would use the whole brain (voxelwise).

Hope this helps.

All the best,

Anderson

 

Thank you for your expert opinion and help.

Eyesha