 |
 |
Hi Philipp, I find it easiest to think about a t- or F-contrast in terms of the null hypothesis that it specifies. > I have a basic questions on the interpretation of two different F contrasts > in a model with 8 regressors (no.1-4: of no interest; no.5-8: "of interest"): I'll refer to these as [A B C D E F G H] below > Contrast A: > 0 0 0 0 1 1 1 1 This specifies the null hypothesis E+F+G+H = 0, tested against the alternative hypothesis that this sum is non-zero (for an F-contrast; or >0 for a t-contrast). > Contrast B: > 0 0 0 0 1 > 0 0 0 0 0 1 > 0 0 0 0 0 0 1 > 0 0 0 0 0 0 0 1 This has the null hypothesis E = F = G = H = 0, i.e. you can interpret each row as a separate contrast, here each specifying that one regressor (with a 1 in its column) is 0. If any of E-H were significant for a particular region then this null would be rejected (i.e. that region would be declared significant) > The questions on the regressors is/should be: are there brain areas > significantly associated with any of the regressors? To ask if any of them are interesting, your null hypothesis would be that none of them are. Hence your contrast B answers this question. > In earlier postings explanations can be found stating that the hypothesis > in Contrast A is much more restrictive than in B. I'd say B has a more restricted null hypothesis or reduced model -- all E-H are zero; whereas A could include both this case and for example the case that E and F are large and positive but cancelled out by equally large negative G and H, along with many other situations. Though see below for whether some data might be significant for A and not B. > b. If one of the regressors does not carry any weight, how much of a > "dilution" effect is present in the two contrasts? If, say, E explains very little variance, then both E+F+G+H=0 and E=F=G=H=0 would be dictated by the F,G,H, which is probably what you'd want. You'd only need to worry about E messing things up if you were looking for *all* E-H to be significant, rather than any of them. I'm not aware of a contrast that would test that; I think you'd have to test each regressor with a separate contrast and then look at the conjunction null of these. Anyway, that's not the question you said you were interested in. > c. What is the meaning of a region appearing in B but not A? (I think it > indicates that rather a single regressors brings in the effect, but I am > not sure) Hard to pin down, since various combinations of regressors could reject B's null but not A's. For example, the above mentioned case of positive/negative ones cancelling out. I think it would also be possible to find regions that were significant for A but not B, if for example all of E-H were positive but small, possibly B would conclude that none is significantly different from 0, while A would conclude that their sum was. I hope this helps. Some of the finer points are rather difficult, and I may well have got them wrong, but I'm fairly confident in the simpler point that B is the right contrast for your question. Best, Ged.