Hi Michael,

Please, see below:

On 25 June 2014 01:32, Michael Spilka <[log in to unmask]> wrote:

Dear FSLers,

I know this is not a new topic but I wanted to seek input regarding a higher-level repeated-measures ANOVA I am trying to set up in feat:

I have 28 participants who completed 5 different runs (analyzed separately at the first-level), and I want to test whether there is any difference between these runs in order to determine whether or not I can combine them for further analyses (I have multiple groups who completed the same runs that I would like to compare eventually).

I have followed the instructions for a a repeated-measures ANOVA on the wiki (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FEAT/UserGuide#ANOVA:_1-factor_4-levels_.28Repeated_Measures.29) but have several questions:

1) How does one determine what the reference run/scan is when adding EVs and setting up contrasts? It is clear to me in the non-repeated-measures example on the wiki (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FEAT/UserGuide#ANOVA:_1-factor_4-levels), where there is a clear grand mean, but this is not the case for repeated-measures. How does the order of the input runs relate to the way the contrasts are set up?

There's no grand mean, but one mean per subject. 

I have attached an example set-up with only 3 participants and 3 different runs (named Anger, Fear, and Happy). Can someone confirm that this set-up would indeed model individual participant means and test the omnibus F-test for any treatment effect of emotion? I am specifically wondering about whether I have correctly omitted the reference run based on the order of my input data, and whether the contrasts are specified correctly. If so, would the contrasts be the same had the ordering of my inputs been different?

Looks fine to me.

2) While I have followed the set-up illustrated in the wiki, it seems like I should be able to set up the same model and contrasts using the cell means approach, in which I would have an extra EV to model the "happy" run, and my contrasts would then look like [1 0 -1], [0 1 -1], and [1 -1 0]. When I tried to run the analysis in this way, I got a warning about my design matrix being rank deficient and my F-test being invalid. Is this warning common when trying to set up a repeated-measures ANOVA this way or is it more likely something specific to my data?

It won't work in this case, because there would be 2 sets of EVs that, once added up, would be equal to vectors of all-ones. For a repeated measurements design as this, following the example is probably the simplest route.

 

3) Since I plan to include 28 participants in this analysis, is there a less unwieldy approach that does not require an EV for each subject?

Nope I'm afraid. One EV is needed per subject. However, if the concern is creating this using the GUI, the design can be created in an external program pasted into a text editor and used to create the design.mat (use another design.mat as a template, it's quite straightforward).

Hope this helps!

All the best,

Anderson