Hi JEsper, thanks a lot!
so what you are saying here:
"If you can arrange the scanning so that each subject has all his/her 
scans on the same scanner (which would be good) you will not need to 
model scanner since that will be implicit in your subject/block regressors."
-- which will be the case, because each patient will stay at his center 
and his scanner--: then we would not have to model scanner at all??

best agnes

 Jesper Andersson schrieb:
> Dear Agnes,
>
>> regarding longitudinal , multi-center trials, using DTI and FSL:
>> We are planning a longitudinal study on stroke patients , and want to 
>> acquire DTI at different time points during recovery/training.
>> Different centers are involved, so obviously different scanners.
>> How would one best correct for the different scanners, in the 
>> analysis? (FA, tracts);  and is it possible to combine PHilips and 
>> Siemens data?
>> (1.5 versus 3 T is not possible at all, I guess;  so just within the 
>> 3 T and the 1.5 T data)
>
> it would obviously be best if everything could be kept constant, but 
> that is clearly not possible for a study like this.
>
> Therefore I would say that you can vary almost any parameter you want 
> (i.e. field-strength, manufacturer etc) as long as you model it in 
> your design and is aware of the consequences of doing so. By modeling 
> it I mean for each scanner including a regressor with 1 if the scan in 
> question was performed on that scanner and a 0 otherwise. The 
> consequences of doing this is that you may potentially lose 
> sensitivity (depending on what scans are performed on what scanners).
>
> To make this concrete, lets say you have acquired all your "initial 
> time points" on one scanner, then you get a grant to buy a new scanner 
> and acquire all the subsequent scans (time point two) on the new 
> scanner. You will then end up with a design where you have identical 
> "time" and "scanner" regressors which means that you have no 
> possibility to disambiguate between the 2. The highest sensitivity 
> (best design) would be the one where "scanner" and "time" are 
> orthogonal, i.e. when there is a equal number of time-point 1 and 2 on 
> all scanners.
>
> If you can arrange the scanning so that each subject has all his/her 
> scans on the same scanner (which would be good) you will not need to 
> model scanner since that will be implicit in your subject/block 
> regressors.
>
> I wouldn't worry about modeling different variances for different 
> scanners (I would assume the variance was dominated by the 
> inter-subject variance).
>
>> as for scanning parameter: 60 directions, plus 6 volumes without 
>> diffusion weighting as the standard; do we have to choose the exact 
>> same TR and TE on all scanners? are there parameters that are optimal 
>> for longitudinal measurements?
>
> 60+6 is a very rich dataset, possibly slight overkill for TBSS (though 
> I'll let someone with more experience comment on that one). There are 
> as far as I know no special considerations for longitudinal data.
>
> Good luck Jesper
>