 |
 |
Hi Tim, what I was thinking of is a partial correlation I think. I am interested in the variance I get from the additional covariate (B) in areas that are active in my experimental task (A). The assumption here is that my additional covariate (B) modulates the extent to which (A) activates the brain and I'm interested in that component rather than wanting to get rid of it. Areas whose variance is explained by (B) but not by (A) may be interesting as well but I would very much like to distinguish the two because I am not quite sure what the latter means - an area whose variance is partially determined by (B, a task-unrelated behavioural covariate) but not by the task (A). I hope this makes sense now. Concerning the orthogonalisation, I am a bit confused as I don't seem to have the option of making the EVs orthogonal with respect to each other in the higher level analyses. Is there another way of doing it? Also, is the assumption that EVs are or are not orthogonal with the current higher level Feat options? Thanks a lot, Amande. Tim Behrens wrote: > Hi Amande, > > I'm afraid I still don't understand what you want to do, > > What do you mean by " a correlation with B _given_ A " ? > > Is it a partial correlation > > " a correlation of the data with B accounting for the effects A " ? > > or a multiple regression > > "How much of the variance in the data can a prescribe to both A and B > together?" > > I'm sure there will be a simple way of answering your question. The GLM is > amazingly flexible. I just can't see exactly what the question is yet.. > > > wrt non-orthogonal EVs > If your EVs are not orthogonal, then any shared variance will be dished > out between them in a way which is not easily predictable a priori. > However, if you are comparing conditions, this shared variance is > accounted for in the statisitics, so you can still accurately answer the > question "Where is the response to A bigger than that to B?" for example. > > Sorry I'm not more use > > Tim > > > > On Sun, 27 Jun 2004, Amande Pauls wrote: > > >>Hi, >> >>no, what I was wondering about is whether there is any way I can test >>whether there is a correlation with B (behavioural covariate) given A >>(task), e.g. by not making the EVs orthogonal. So really what I want to >>know is what it means if the two EVs are not orthogonal and whether this >>is in any way statistically meaningful. Such as activation in a certain >>area given RT on the task, or activation in a task given IQ or something. >> >>Amande >> >> >> >>Tim Behrens wrote: >> >> >>>... >>> >>>"At which voxels does my signal contain variance which can be explained by >>>my RTs _but not_ by my task EV ?" >>> >>>sorry - this might be confusing. >>> >>>more accurate is >>> >>>"At which voxels does my signal contain variance which can be explained by >>>my RTs after accounting for variance explained by my task EV ?" >>> >>> >>> >>>------------------------------------------------------------------------------- >>>Tim Behrens >>>Centre for Functional MRI of the Brain >>>The John Radcliffe Hospital >>>Headley Way Oxford OX3 9DU >>>Oxford University >>>Work 01865 222782 >>>Mobile 07980 884537 >>>------------------------------------------------------------------------------- >>> >>>---------- Forwarded message ---------- >>>Date: Wed, 23 Jun 2004 16:16:33 +0100 (BST) >>>From: Tim Behrens <[log in to unmask]> >>>To: FSL - FMRIB's Software Library <[log in to unmask]> >>>Subject: Re: [FSL] higher level analyses - using additional (behavioural) >>> covariates >>> >>>Hi Amande - I'm not sure whether I've understood this right, but it sounds >>>like what you want is exactly the opposite of the previous scenario. That >>>is >>> >>>"At which voxels does my signal contain variance which can be explained by >>>my RTs _but not_ by my task EV ?" >>> >>>If this is the case, you want to run the orthoganisation the other way >>>round. That is, you want to orthoganalise the RTs wrt the task covariates. >>> >>>This will remove from the RT EV, any variance which could be explained by >>>the task. >>> >>>Hope this is what you want >>> >>>T >>> >>> >>> >>> >>> >>> >>>On Wed, 23 Jun 2004, Amande Pauls wrote: >>> >>> >>> >>>>Hi there, >>>> >>>>thanks for the quick reply. >>>> >>>>I have another question about point (3). Say I don't want to factor out >>>>the additional covariate, but want to see whether there are areas whose >>>>variance reflects the additional covariate given the task. An example >>>>would be some motor task and I want to 'rank' people according to RTs >>>>I've got from them in some other task (or rather see whether activity in >>>>this task somehow reflects their prior motor performance). Do I use RTs >>> >>>>from the other task as additional covariate and then don't make the EVs >>> >>>>orthogonal? Or do I have to set that up in the contrasts somehow, after >>>>they have been orthogonalised? >>>> >>>>What I would like to know is whether my additional covariate, of >>>>interest or not, will correlate only with areas activated by the task, >>>>or whether what I see could also reflect baseline activity in some >>>>resting state network (like language areas etc). >>>> >>>>Thanks again. >>>>Amande >>>> >>>> >>>> >>>> >>>> >>>>Tim Behrens wrote: >>>> >>>> >>>>>Hi there >>>>> >>>>>On Tue, 22 Jun 2004, Amande Pauls wrote: >>>>> >>>>> >>>>> >>>>> >>>>>>(1) Is it possible to control for an additional covariate that I am not >>>>>>interested in but suspect to have an influence on the outcome of the >>>>>>experiment (like measures of intelligence)? By modelling all of them as >>>>>>an additional EV? Or by making it one per subject (like when allowing >>>>>>for individual differences in variance)? >>>>> >>>>> >>>>>Yep - you should use a single EV for each covariate of no interest (e.g. >>>>>one EV for IQ, ine for age etc. etc. ) >>>>> >>>>> >>>>> >>>>> >>>>>>(2) If I want to know whether there is a negative correlation between my >>>>>>additional covariate (modelled as a separate EV) and the data, do I need >>>>>>to set the contrast to -1? >>>>>> >>>>> >>>>> >>>>>Yes - absolutely right. >>>>> >>>>> >>>>> >>>>> >>>>>>(3) Having used an additional covariate (like the RT example in the >>>>>>webpages) and using that precise contrast, what does the result mean? >>>>>>I'm unclear on whether the brain area correlates both with RT and the >>>>>>task itself, or whether the level activity in that area somehow reflects >>>>>>RT, potentially independent of the task. Can I distinguish between those >>>>>>two cases, or make sure that my contrast reflects 'correlation with >>>>>>EV, given the task'? >>>>>> >>>>> >>>>> >>>>>It depends on the precise setup that you have chosen, but if you >>>>>orthogonalise such that RT explains the maximum possible variance (i.e. >>>>>orthogonalise all covariates of interest wrt RT) then the RT contrast >>>>>represents all the variance in the signal which _might possibly_ be >>>>>explained by RT. The copes of interest cannot then describe any variance >>>>>in the signal which could be ascribed to RT. >>>>> >>>>>JHupw this is clear >>>>> >>>>>Tim >>>> > ------------------------------------------------------------------------------- > Tim Behrens > Centre for Functional MRI of the Brain > The John Radcliffe Hospital > Headley Way Oxford OX3 9DU > Oxford University > Work 01865 222782 > Mobile 07980 884537 > ------------------------------------------------------------------------------- > -- -------------------------------------------------------------------- Amande Pauls University Laboratory of Physiology, Oxford, UK mailto:[log in to unmask]