 |
 |
Hi Steve - Thanks for your quick reply. Yes, you're right: EVs 3&4 and 5&6 should be swapped in the description (example design matrix shows test version first, followed by scan order; sorry for mixing this up). For every session, I used a similar but slightly different version of a memory task. This may cause some variation in signal intensity which I wanted to correct for using EVs coding for test version. The design works well (i.e. does not become rank deficient) if I only specify EVs 1-6 (plus 18 EVs coding for within-subject variation / pairedness: 24 EVs in total). So I think that the rank-deficiency is probably not caused by information shared between EVs coding for test version and the remaining EVs. The problems start as soon as I enter EV 25 , i.e. an additional EV coding for demeaned behavioral test scores. For this EV, I have to enter three identical values for each subject. Maybe that this EV and EVs coding for pairedness have mutual information? Eventually, I'm abIe to circumvent problems of rank deficiency by collapsing paired data (calculating differential images for all subjects separately and entering these in a 3rd level analysis; see below). By now however, none of my original (uncorrelated) effects survive, perhaps because data collapsing resulted in a loss of DOF. I hope there is a way around this, since correlating cognitive status to response to treatment seems quite interesting. Thanks, Rutger. ----- Original Message ----- From: "Stephen Smith" <[log in to unmask]> To: <[log in to unmask]> Sent: Saturday, July 24, 2004 8:28 PM Subject: Re: [FSL] Repeated Measures Design: Correlating with Behavioral Measures > Hi - I think I'm following the description, though I'm guessing that > EVs 3 and 4 should be swapped in your description. I don't understand > what "testversion" means, but it seems likely that the information > contained in testversion is rank deficient when compared with EVs 1-4 > - is that possible? > > Cheers. > > > On Thu, 22 Jul 2004, Goekoop, R. wrote: > > > Dear all, > > > > I'm attempting to correlate functional effects of medication treatment > to > > behavioral data. The study involves 18 patients, 3 sessions for each > patient > > (corresponding to 3 medication regimes (A, B, C); so 3 measures for > each > > patient). > > > > For this, I specified a higher level design involving 54 inputs (18 x > 3), 1 > > group, and 24 EVs. The first two EVs code for (randomized) medication > > regimes (B > A and C > A respectively; contrasts between these EVs > yields C > > <> B). The second two EVs code for scanorder (session 2 > 1; session 3 > > 1 > > respectively). Next two EVs code for testversion (2 > 1 and 3 > 1 > > respectively; randomized); The remaining 18 EVs put weights to data > derived > > from the same subject (to account for pairedness of the data), i.e.: > > > > Inputs: Group EV1 EV2 EV3 EV4 EV5 EV6 EV7 > EV8 > > EV9.... > > > > 1 1 -1 -1 -1 -1 0 1 1 > 0 > > 0 > > 2 1 0 1 1 0 -1 -1 1 > 0 > > 0 > > 3 1 1 0 0 1 1 0 1 > 0 > > 0 > > > > 4 1 0 1 -1 -1 1 0 0 > 1 > > 0 > > 5 1 -1 -1 1 0 0 1 0 > 1 > > 0 > > 6 1 1 0 0 1 -1 -1 0 > 1 > > 0 > > > > 7 ... 54: Etc. > > > > Now, I want to add an additional EV coding for (demeaned) MMSE scores > (a > > neuropsychological scale), e.g.: > > > > EV 55: > > > > 3.5 > > 3.5 > > 3.5 > > > > -2.5 > > -2.5 > > -2.5 > > > > ...etc > > > > The design however becomes rank-deficient (something to the power of > -17) > > as soon as I add this EV. > > > > I tried various approaches, the first one being to collapse my data by > > calculating differential images of regime-types (i.e. B > A and C > A) > in a > > second level analysis, and entering these images as inputs in a > third-level > > analysis for correlation with behavioral data. However, as soon as I > try to > > regress out effects of either scanorder or testversion at second level > (at > > which level the differential images are calculated), the design > becomes rank > > deficient (again something ^ -17). I therefore stuck to calculating > > differential images at 2nd level and then tried to regress out effects > of > > scanorder and testversion at 3rd level, while at the same time > correlating > > data with behavioral scores. > > > > For this, I separated differential inputs (B > A) from (C > A) inputs > and > > ran two separate third level analyses on these separate datasets. I > thought > > this to be necessary, since putting these inputs together in one > design > > would require additional EVs coding for pairedness of inputs derived > from > > the same subject. This would bring the total number of EVs to 27, with > > number of inputs = 36, which I thought would not work well (please > correct > > me if I'm wrong). > > > > Both separate 3rd level analyses worked well, however by now, > collapsing of > > data had reduced the total number of inputs for each design to 18, > while the > > number of EVs was 8. By now, nothing of my original effects of > treatment > > survived (effects are rather small). Might this be due to > data-collapsing > > and loss of DOF? Do I just demand too much of my data, or would there > > perhaps be an alternative solution to this problem? Any help would be > > greatly appreciated, > > > > Thanks, > > > > Rutger. > > > > Drs. R. Goekoop, MD. > > Department of Neurology > > Vrije Universiteit Medical Centre > > De Boelelaan 1117, P.O. Box 7057 > > 1007 MB Amsterdam, the Netherlands > > Phone: 0031-20-4440316 > > E-mail: <mailto:[log in to unmask]> > > > > Stephen M. Smith DPhil > Associate Director, FMRIB and Analysis Research Coordinator > > Oxford University Centre for Functional MRI of the Brain > John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK > +44 (0) 1865 222726 (fax 222717) > > [log in to unmask] http://www.fmrib.ox.ac.uk/~steve