Hi, That should be enough to have different group variances. Sorry missed the interaction f-test on the first pass. b is the F-test you'd use for the interaction. For (a) you're only testing whether one or the other slope (covarA or covarB) differs from 0. In (b) significance implies that one slope (between your data and covar) is larger than the other, but you do not know which. If the f-test is significant, you get the directional info from the individual t-tests [ (0 0 1 -1) and (0 0 -1 1)]. Cheers, Jeanette On Mon, Jan 28, 2013 at 3:36 PM, Esther <[log in to unmask]> wrote: > Hi, > > thanks for your advice. What sample size would you consider large enough > (we got 80/100)? And could you just tell me, which F-test in model 2 is > correctly testing for the group*covar interaction effect? a) or b)? > > a) > groupA groupB covarA covarB F-test > contrast1 0 0 -1 0 x > contrast2 0 0 0 1 x > > However, we are not too sure about the F-test in 2). Maybe it should be > just: > > b) > groupA groupB covarA covarB F-test > contrast1 0 0 -1 1 x > > Thanks a lot, > > Esther > > =========================================== > > Hi, > > if the "group" column is set to all 1's, both of these models are > identical. Unless you have large sample sizes in your 2 groups, I wouldn't > use the "group" column to indicated different groups. This allows > different between-subject variance estimates in your two groups, but with > lower sample sizes I've found the single estimate is better than two and my > guess is that the individual group variance estimates are more variable due > to the smaller sample size. Either way your inferences are valid. > > Cheers, > Jeanette > > On Mon, Jan 28, 2013 at 3:05 AM, Esther <[log in to unmask]> > wrote: > > Hi, > > we have two groups with a covariate of interest and want to test the > group*cov interaction effect in FEAT. > As far as I can tell, there are two ways to do that in FSL: > > 1) > > intercept GroupA-B covar group*covar > 1 -1 -3 3 > > 1 -1 -2 2 > > 1 1 1 1 > > 1 1 4 4 > > > with the following interaction-testing contrast: > > intercept GroupA-B covar group*covar F-test > contrast 0 0 0 1 x > > which has the disadvantage that I can't model separate variances. So I > could also do: > > 2) > > groupA groupB covarA covarB > 1 0 -3 0 > > 1 0 -2 0 > > 0 1 0 1 > > 0 1 0 4 > > > with the following contrast: > > groupA groupB covarA covarB F-test > contrast1 0 0 -1 0 x > contrast2 0 0 0 1 x > > However, we are not too sure about the F-test in 2). Maybe it should > be just: > > groupA groupB covarA covarB F-test > contrast1 0 0 -1 1 x > > Which one is correctly testing for the group*covariate interaction > effect? > And would you expect the 1) and 2) model results to differ and if so, > which model would you prefer (and why)? > > Thanks for your help, > > Esther >