Print

Print
Hi Kathy,

As you have a value for the covariate for each condition, for for each subject, my suggestion is to instead, use the full paired t-test as described in the manual (this one: http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Paired_Difference_.28Paired_T-Test.29) and add one extra EV for the additional covariate.

To create this model, use the command Glm (or Glm_gui in the Mac). When the model is saved, it creates the design.mat, design.con, etc.

As this model will run in randomise, you'll need to make a change in the design, though: in the column "group", fill with values such that they define one group per subject. That is, if you have 10 subjects, there should be 10 such "groups". These are in fact exchangeability blocks, and are passed to randomise with the option "-e design.grp". Permutations will then happen only within EB, which in this case is only within subject.

Hope this helps.

All the best,

Anderson




On 19 February 2015 at 21:07, Kathy Do <[log in to unmask]> wrote:
Hi everyone,

I want to conduct a paired t-test with one group and two conditions (high, low). I am interested in adding in a covariate (behavioral measure) that differs by each condition. I originally set up my analysis per the FSLwiki instructions for a single group paired t-test (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Paired_Difference_.28Paired_T-Test.29). However, now following this suggestion by Jeanette (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;85ebf34.1402), I plan to:
1) Compute the within-subject paired difference first using fslmaths and enter that in as my EV1.
2) Conduct a 1-sample t-test using the above values, entering in my covariate as EV2.

A few questions:
1) If I wanted to look at the within-subject paired difference for (high>low) and (low>high) separately, I assume I would have to compute this difference using fslmaths command twice (e.g. one for the high>low and one for the low>high) and conduct two separate group analyses, correct?

Example…
Model 1: High>Low
fslmaths high.feat/cope1/stats/cope1.nii.gz -sub low.feat/cope1/stats/cope1.nii.gz diff_highlow.nii.gz

Enter in “3D cope images from FEAT directories” (i.e. diff_highlow.nii.gz) for EACH subject as an input.

Model 2: Low>High
fslmaths low.feat/cope1/stats/cope1.nii.gz -sub high.feat/cope1/stats/cope1.nii.gz diff_lowhigh.nii.gz

Enter in “3D cope images from FEAT directories” (i.e. diff_lowhigh.nii.gz) for EACH subject as an input.

Is there a way to consolidate both of these in one model, or is it easier to have them be separate?

--

2) Do I need to use fslmerge (i.e. fslmerge diff_highlow diff_subA diff_subB diff_subC) to create a single 4D image for all subjects, or can I just enter the cope images I computed for the within-subject paired difference into the group analysis? I’m a bit confused as I assume each diff_subA represents a single input in the group analysis, but by using the fslmerge command, that creates a single 4D volume with all the subjects’ images - so what would each subjects' input in the group analysis be?

At what point would I conduct the following command (and why is it necessary):
randomise -i <4D_input_data> -o <output_rootname> -d <design.mat> -t <design.con>  -m <mask_image> -n 500 -T -D

Also, what is the design.mat, design.con and mask_image referring to?

--

3) Because the covariate differs by each condition, would I similarly compute the within-subject paired difference for (highCOV - lowCOV) and (lowCOV - highCOV)? Would I then demean across ALL subjects (not within condition)?

Example…
Design:
1 -12.31
1 -11.31
1 11.50
1 15.50

C1 (group mean): 1 0
C2 (cov): 0 1

--

Hopefully this all makes sense! This is all new to me, so I’d appreciate any help setting this design up. Thanks so much!

Best,
Kathy