 |
 |
Thanks Donald, really helpful. I have a follow up question: Is there any way to compare 3 groups (say GroupA, GroupB, GroupC) with only the 2 first groups having sub-groups (ie GroupA1, GroupA2, GroupB1, GroupB2, GroupC)? I understand I cannot use the fullfactorial design in that case, because my GroupC does not have sub-group. Any idea/suggestions? Thanks. Yann On 30 Jan 2014, at 1:12 pm, MCLAREN, Donald <[log in to unmask]> wrote: > See below. > > > > On Wed, Jan 29, 2014 at 7:54 PM, Yann Quidé <[log in to unmask]> wrote: > Dear all, > > I apologize if the question has been previously answered. > > We have 2 groups A and B, that we sub-divised in 2 sub-groups 1 and 2. We therefore tried to compare them using a full factorial design entered as A1, A2, B1, B2. > First question: is it the best way to proceed? > > It depends on your question... > > > We specifically wanted to look at the interaction between group and sub-group, and therefore looked at the F-test. However, using SPM8, we cannot correct our results at the cluster level. > > I believe SPM12 has the correction for the F-test. You could use 2 t-tests and correct them separately. > > Second question: how can we report these results, if not corrected? Just by mentioning them but not really discuss them? What is the (statistical) validity of uncorrected results on a F-test? > > You'd need to use post-hoc tests to properly determine the significance. I think the t-test option would be better though. > > > When looking at the corresponding t-tests, we could therefore correct our maps for multiple comparisons. > Third question: Even if we correct these t-tests at the cluster-level, using the same initial threshold that the F-test (say p<0.005 uncorr), is there any other (additional) correction to add (ie by dividing the initial threshold by the number of test AND multiple comparison at the cluster level)? > > In your case, there is only 1 t-test that you can run. You have the positives and negatives that have to be corrected separately. Correcting them separately is the common approach in neuroimaging and is the same as a one-tailed test reported in statistical packages. You could also change the threshold to half, then you get the two-tailed test reported in statistical packages. > > Or is this too conservative? What would be the best method to present the results with sufficient correction? > > The first t-test, included in SPM contrast manager by default [1 -1 -1 1]), did not give us any surviving cluster (just with cluster level correction). However, we created the "reverse" contrast ([-1 1 1-1]; still cluster-corrected), that gave us a lot of information! > Fourth question: we tentatively interpreted this as "there was greater difference between B1 and B2, than in A1 and A2. Is that right? If not, what is the meaning of such interaction, and what means the "direction" of the contrast (ie [1 -1 -1 1] or [-1 1 1 -1])? > > The interpretation is correct. > > -1*A1+1*A2+1*B1+-1*B2 > becomes > (B1-B2)-A1+A2 > becomes > (B1-B2)-(A1-A2) > becomes > (B1-B2)>(A1-A2) > > > Thanks for your help! > > Best, > > Yann >