Print

Print
Hi,

See below


On Wed, Aug 28, 2013 at 7:12 AM, Danusia Lisiecka <[log in to unmask]> wrote:
Dear Anderson and Experts,

I would like to follow up on Maya's question and ask if looking at t-contrast she presented in the second model e.g [1 1 -1 -1] would be a valid thing to do.

Same problem as Maya's email, you didn't describe your design, so I can't be sure how you modeled it.  I'm assuming you have 4EVs consisting of 0's and 1's, first is Ax, second Ay, Bx, By.
 
I'm working on a similar statistical model in TBSS (2x2 ANOVA) and I also find main and interaction F-contrasts empty. The F-contrasts are modelled the way it was recommended on GLM webpage (ANOVA: 2 factors 2 levels).

Some directional t-tests in two main contrasts and interaction e.g. A>B contain significant results surviving FWE correction. The differences stand when the two-group model is created and the two groups compared, so EV1=A and EV2=B.

Okay, that makes sense. 

When all 4 groups (Ax, Ay, Bx, By) are put together in one F-contrast the way described on GLM webpage (ANOVA: 1 factor 4 levels) they also give significant results surviving FWE correction, this time in F-contrast.

Technically you have a 2 x 2 ANOVA, but we'll ignore that.  Again, I'm assuming you're using the "cell means" approach from the GLM page (I believe we introduce 2 options for modeling ANOVAs there)
 
The empty F-contrasts appear only when two groups are compared in them like A vs B or x vs y
 
It stays the same with both 500 and 5000 permutations and two different versions of FSL.

My questions are:

a) is it valid in ANOVA in FSL to look at two directional t-contrasts e.g [1 1 -1 -1] and [-1 -1 1 1] instead of F-contrast. I've seen some papers doing this. It is also valid in SPM. Is FSL coded in such a way that it is valid?

Of course it is okay to look at the individual t-tests, but you then need to control for the fact that you ran 2-tests.  Typically we use a Bonferroni correction, so you'd use a 0.025 cutoff istead of 0.05 in randomise.  As we've posted multiple times on this list, an F-test is 2-sided and so the p-values will tend to be twice as large as the t-test p-values for this type of F-test, involving only a single contrast (numerator DF=1 for the F-test, if it were parametric).
 
b) why would the F-contrasts like A vs B, x vs y, AB interaction xy give no results yet F-contrast of Ax, Ay, Bx, By and t-contrasts of A vs B, x vs y, AB interaction xy would be highly significant?

These are all different hypothesis.  An F-test for A<B might not be significant if the p-value was 0.06, but the 1-sided t-test will have a p-value of 0.03 (I'm again assuming parametric tests for illustration).  I'm not sure what you mean about the interaction vs a test for Ax, Ay, Bx and By, if you're referring to simply testing whether those separate means are different from 0, that's a different hypothesis than the interaction test.  I think I already addressed your last t-test question as well.

Hope that helps,
Jeanette