Tom, Thanks for the reply. Indeed it is a tricky thing when the data are unbalanced, but I'll give it a try using only year 1 and 2 data. On Thu, Apr 17, 2008 at 6:24 AM, Thomas Nichols <[log in to unmask]> wrote: > Dear Robert, > > Randomise only fits OLS GLM's, i.e. unweighted least squares fits. That > said, randomise can accommodate certain types of repeated measures data, the > simplest type of which is paired data. When you have just two time points, > follow the instructions for paired data analysis. When you have three or > more observations per subject, things get slightly trickier. > > If you have k>=3 repeated measures, you could simply treat the data like a > 2-way ANOVA, modeling one longitudinal factor, and one subject factor, but > since you're not doing a proper mixed effects model, the t-statistics won't > really be what you want (i.e. they won't be capturing the correct between > subject variation in the denominator). For the paired case, k=2, it happens > to works out exactly correct (i.e. OLS & MFX are the same), and the k=3 case > probably isn't too far off (if an assumption of compound symmetric > correlation--i.e. all equal correlations--within subject is true, and the > design is balanced, randomise's OLS results should match a full-blown mixed > effects model). But for imbalanced design (k varying between subject) and k > of 4 or more, you probably shouldn't trust OLS to be giving you sensible > answers. > > The safest approach for k>3 would be to create summary measures for each > subject, and then anlayze those with a simple second level model. For > example, if you're interested in slope, or change over time, fit a simple > linear regression model for each subject, and then model the images of slope > coefficients. > > Ideally there would be some scripts cobbled together to aid with such > longitudinal analyses, but does this give you an idea of what randomise can > and can't do? > > -Tom > > > > > > On Wed, Apr 16, 2008 at 3:50 AM, Robert Terwilliger <[log in to unmask]> > wrote: > > Dear FSL, > > > > We have been doing DTI analysis using TBSS successfully for some time > > now. Our cohort consists of normally developing adolescents. > > > > As a small example of our analysis consider a sample of FA images from > > six subjects (we have many more, but this is for simplicity), ages > > 12-17. I do a "within group" design in which the log of age is the > > regressor. The resulting design.mat and design.con files are as > > follows: > > > > ************** > > design.mat > > ************** > > /NumWaves 1 > > /NumPoints 6 > > /PPheights 1 > > /Matrix > > -0.145 > > -0.145 > > -0.065 > > 0.009 > > 0.143 > > 0.203 > > > > ************* > > design.con > > ************* > > /NumWaves 1 > > /NumContrasts 1 > > /PPheights 1 > > /Matrix > > 1 > > -1 > > > > The values in design.mat are the demeaned natural log of the subjects' > ages. > > > > So far, so good. > > > > Now fast forward a couple of years....This is actually a longitudinal > > study, with each subject scanned on an annual basis. As is common in > > longitudinal studies, not every subject is scanned every year, for a > > variety of reasons. > > > > If we consider only the first year scans as we are doing currently, > > each subject can be treated as an independent sample. However, now > > some subjects have had three scans, some have two, and a few didn't > > make it past the first year. > > > > Is there a way to set up a model in randomise where we can include > > multiple DTI scans from the same subject? This would violate the > > assumption of independence in the simple model, but we're looking for > > a way to account for the within-subject correlation in a mixed-model > > design. > > > > Many thanks, > > > > Robert Terwilliger > > Laboratory of Neurocognitive Development > > University of Pittsburgh > > > > > > > > -- > ____________________________________________ > Thomas Nichols, PhD > Director, Modelling & Genetics > GlaxoSmithKline Clinical Imaging Centre > > Senior Research Fellow > Oxford University FMRIB Centre