Hi,
> 1. What is the appropriate multiple comparisons correction for vertex analysis results in randomize?:
> - 1a Should TFCE be 2D or 3D?
The results are on a surface so the 2D option is probably better tuned, but either option is valid (in fact all statistical thresholding options are valid, but they have different intrinsic sensitivities, and hence why there are different options).
> - 1b In FDR, by "adjusted" do you mean the Yekutieli D, Benjamini Y (1999) adjustment?
Yes, that's right.
> 2. How to interpret results?
> - 2a With dichotomous variables, is the contrast result one-directional?
All the t-tests are uni-directional but F-tests are bi-directional, so it depends.
> So, if I code Group1 as +1 and Group2 as -1, and my contrast is: +1 (as in the example on the FSL website), I see both regions that have significant expansion AND regions that have significant atrophy in Group1 as compared to Group2?
For an F-test, yes. But for a t-test, no.
> And, since the new version of vertex analysis doesn't produce the surfaces with arrows anymore, how do I tell, for a region that is significant, which is the direction of the difference?
You can use the t-test results together with information on their signs (using the information about inside/outside sign as given on the wiki page, together with the relative signs in the contrast).
> - 2b With covariates: are results interpreted similarly to the dichotomous variables -- so I only need 1 contrast (e.g., +1) and not that AND its reverse (e.g., "-1")?
Again, it depends if you use an F-test or a t-test.
> I guess the intuition would be that there are 4 possibilities (1, regions that expand in positive correlation with the covariate; 2, regions that atrophise in positive correlation with the covariate; 3, regions that expand in negative correlation with the covariate; 4, regions that atrophise in negative correlation with the covariate) but the latter 2 are entailed by the former 2?
For a t-test you can only see changes that are positively related or negatively related. The mean value of the vertex analysis values should be removed in the analysis (either explicitly in the design matrix or using the -D option in randomise) and so you do not have separate positive-only or negative-only scenarios. The values at each vertex will be a mixture of positive and negative values, and it is simply whether they increase or decrease when the covariate increases that determines the sign.
> Thank you very much. Any help is greatly appreciated.
I hope this has helped.
All the best,
Mark
|