 |
 |
Hi, This isn't all that unusual. In the simplest case where the F-test involves only a single contrast (numerator DF=1), the F-test will have p-values twice as large as the t-test (assuming the t-test is 1-sided). Earlier, Tom Nichols posted this great reference that describes the relationships between f-tests and t-tests well, when the numerator DF=2. The main issue relates to correlated contrasts, which I don't think is your problem, but it sheds light on other issues as well. http://www.amstat.org/publications/jse/v16n3/martin.html Cheers, Jeanette On Mon, Aug 26, 2013 at 7:30 AM, Diederick Stoffers <[log in to unmask]>wrote: > Hi, > > In a VBM analysis I am interested in the negative association between a > clinical score and GM volume in four different roughly equal-size groups. > > I created a full factorial model in which I can test for negative > associations between the clinical score and GM in the individual groups > using four separate t-contrasts and run an omnibus F-test over the four > individual contrasts. In one of the four groups a t-contrast shows a strong > relation between volume and the clinical score in a large brain cluster > (whole-brain, P<.001, FWE-corrected). In the omnibus F-test, a similar > cluster can be seen at P<.001, but this cluster doesn't even come close to > surviving FWE-correction (P=.4). > > I had a discussion with my PI about this, he says the effect from the > t-contrast probably doesn't show because the variances are unequal between > groups, which makes the F-test invalid. My feeling is that this is not an > issue, as the contrasts I am adding to the F-test are not comparing groups. > Can anyone comment on this and venture a guess as to why the cluster that > shows as very significant in the t-test doesn't even come close to > surviving FWE correction in the omnibus F-test? > > Many thanks, > > Diederick > >