Dear Martin,
I would exclude that patient from your analysis.
For your other questions:
1 - Do the subtraction just before randomise. That is, follow all the steps up until randomise, then do the subtraction and merge the subtracted images, then run randomise.
It is possible that you could get some extra accuracy from doing within-subject registrations, but that would require a significant modification to how the pipeline works, so to start with I would just try it as I described here and check to see if you think the registrations are good enough with the standard method. As it will project onto the skeleton, it will already be pretty robust to small deviations.
2 - Do demean your covariates. However, I would not use -D, as you are probably modelling the mean within your design matrix anyway, and using -D is an unnecessary complication in this case. For the groups I would normally use a two group model and either have a single covariate or split covariate, depending on if you are interested in looking at the different degrees of correlation in the groups (or slopes in the groups; or group by slope interaction; as these all mean the same thing).
I hope this helps.
All the best,
Mark
On 27 Jun 2013, at 19:19, Martin Reiss-Zimmermann <[log in to unmask]> wrote:
> Dear experts,
>
> I'm trying to perform a tbss-analysis of a group of patients with suspected idiopathic normal pressure hydrocephalus and I ran into a problem (once again). We acquired for each patient a total of two decent DTI scans (MGH_sequence, 2 mm iso, 60 dir) - one before and one after lumbar drainage.
>
> I followed the recommendations of tbss, starting with bet, eddy_correct (the age before blip up/down) and dtifit. Then I did tbss_1 and tbss_2_reg -n. The second step took about 30 to 36 hours (too bad, we don't have SGE).
> After running tbss_3 I looked at the skeleton and had to find out, that in one patient the estimated skeleton runs right through the parietal part of both lateral ventricles.
>
> I knew that I have to use a study specific template (2nd step with the -n option) but even that wasn't enough for one patient, which I truly understand looking at the raw data - this one really does have HUGE ventricles.
>
> So my question for the interested audience is: Should I exclude the patient or can I proceed tbss?
>
> Proceeding tbss-analysis leads me to more questions I would like to ask (and that I didn't find an answer reviewing the forum history):
> I want to analyze to whole group and therefore could use the paired t-Test (ALL patients had exactly two scans). Since I also want to use co-variates (respondes positive to lumbar drainage, age, drained CSF volume) I would follow the approach suggested by Prof. Mumford: https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=fsl;e8a2690f.1305
>
> Question no1: Following this approach, I have to subtract examination 2 from examination 1 (i.e. FA_after.nii.gz minus FA_before.nii.gz). At what stage do I do that? Do I run tbss all the way up to tbss_4 and then subtract? If I do so - what about the transforms that are probably slightly different for the single examinations (before, after) performed by tbss before? Or do I perform fslmaths-subtraction right after dtifit and start tbss from the beginning?
>
> Question no2: Having a very limited experience of statistics reading the forum and recommended sites I would demean the covariates "age" and "CSF volume" and also use the -D option in randomize. But what about the covariate "response to lumbar drainage" - actually this covariate devides the group into two subgroups - can I perfom randomize using a covariate or do I rather have to perform randomize with each subgroup for itself?
>
> Once again, thank you very much for your kind help.
>
> Best,
> Martin
|