Print

Print
Hi Stephen,

Please see a starting point for your design at this link, to be used in the 2nd level in randomise:
https://dl.dropboxusercontent.com/u/2785709/outbox/mailinglist/design_bailey.ods

Each subject has 4 conditions (2x2), named A, B, C and D, and you'd delete, for each subject, those rows from the design for which the condition is missing. You'd run using the option "-e design.grp", with one block per subject, such that observations will be shuffled within-block only, i.e., within-subject.

This is a difficult case in which the assumptions of either permutation or parametric (be with 2 or 3 levels in FEAT) will not hold well: the different number of observations for each subject will make the variances not homogeneous, although I tend to believe the effect will not be too severe.

Nonetheless, if you want, you can further define one variance group for each number of observations per subject (e.g., all subjects with complete data would go in one VG, all subjects with 3 observations in another VG, and so forth). Then you'd use PALM with the option "-vg". This is FMRI data where the residuals are usually symmetric so if you use PALM you can further add the options "-ee -ise", which will do permutations and sign-flippings.

Hope this helps.

All the best,

Anderson


On 7 October 2016 at 18:34, Bailey, Stephen Kent <[log in to unmask]> wrote:

Hello FSL Gurus –

 

I have a higher-level Feat analysis question:  We have an experiment with four fMRI sessions, where our main contrast (passage reading) is varied across two factors: two types of text and two presentation formats. So, we would like to run a 2x2 ANOVA across all subjects. However, many of our subjects have only 1-3 of the four sessions, because of motion or other issues.

 

Is the appropriate way to model this by entering all first-level Feat directories into a second-level model, adding EVs for the conditions, and N subject-specific EVs? (i.e. TextType, Format, Subj1, Subj2 ….) ?

 

If this is correct, in what cases should I implement a 3-level design?

 

Thanks!

 

Stephen Bailey

Education and Brain Sciences Research Lab

Vanderbilt University