Hi Rowena,

The -D option is very useful as a shortcut, but you can still do the same analysis ignoring it altogether if you aren't completely sure of how it works. Without -D the design just needs to have an intercept -- it can be a column full of ones, or that column full of ones can be split into multiple columns that, when added together, are the same as a column full of ones. If this is done, the -D isn't needed.

Please, see more below:

On 5 June 2015 at 18:55, Rowena C. <[log in to unmask]> wrote:

Hi,

My apologies for riding on this thread and what seems to be a really hot topic. I'd like to get some relevant advice as well.

Anderson, you mentioned that the -D option can cause problems for the comparison of 3 or more groups. In that case, if I were to be comparing e.g. FA between disease subtype 1 vs. disease subtype 2 vs. control, I would probably not be selecting the -D option? Despite all the posts I have read, unfortunately there is quite a cloud of confusion surrounding the usage of the -D option, especially to someone who is not as experienced such as myself.

If there are 3 groups, the -D can still be used, but the design must use just 2 columns to code the 3 groups, otherwise the model becomes rank deficient. The contrasts then become a bit less obvious and perhaps the best is simply not use -D, and code 3 groups using 3 columns as usual.

Some other questions:

1. If I have matched all 3 groups on age and gender, I am assuming based on what I read, that it is also advisable to add age and gender as covariates?

Yes, even if the groups are matched, if these nuisance can explain some of the effects, they should still be included into the model.

In that case, would -D be needed, or should I forget about using -D and simply manually demean both covariates?

If you won't test the effect of these nuisance variables, but just compare groups, demeaning isn't needed. Only if you'd like to test the effect of age and sex, or test the mean for each group (which surely you won't for TBSS). You probably have seen already, but if not, please have a look at Jeanette's page on mean centering: http://mumford.fmripower.org/mean_centering/

That said, mean centering manually age and sex will not be harmful at all, and won't affect the results even for the cases that demeaning isn't necessary.

2. Assuming the above is done for my first TBSS run. However, after looking at the results, perhaps postdoc revealed significant differences between disease subtype 1 vs. disease subtype 2. Moving forward, I now decide I would like to control for other factors like illness duration and medication dosage, just between those two groups (as this is probably irrelevant to the control group). Do I now run a 2nd TBSS with just the two disease subtype groups and add the extra covariates (demeaned)?

Yes, that can be done (a second model, with just the two subgroups). However, if you'd like to correct for all these various hypotheses being tested in the same study (i.e., multiple contrasts), then having all in the same model would probably be a better approach, as correction over contrasts can then be done. Though it isn't yet available in randomise, it is available in another (experimental) tool called PALM.

Hope this helps.

All the best,

Anderson

Would greatly appreciate some clear advice on the above. Many thanks in advance for your time and effort :-)