 |
 |
Dear Jones, Here is how I would do: First, use as the time marker a number that closer to what your hypothesis is, e.g., if the visits at 0, 6 and 15 months are to assess the efficacy of a treatment applied before or at t=0, these time indicators are probably fine. If the visits are to assess, e.g., the effect of age, then perhaps you could replace these values for the actual age of the subjects. You can use years with decimal places, or express the age in months, or maybe even in weeks if it's for preterm or newborns. The design matrix can be arranged as: (Data) (Grp) (Time,controls) (Time,patients) (Intercept,subj1) (Intercept,subj2) (Intercept,subj3) (Intercept,subj4) (Subj1,Visit1) 1 a11 0 1 0 0 0 (Subj1,Visit2) 1 a12 0 1 0 0 0 (Subj1,Visit3) 1 a13 0 1 0 0 0 (Subj2,Visit1) 2 a21 0 0 1 0 0 (Subj2,Visit2) 2 a22 0 0 1 0 0 (Subj2,Visit3) 2 a23 0 0 1 0 0 (Subj3,Visit1) 3 0 a31 0 0 1 0 (Subj3,Visit2) 3 0 a32 0 0 1 0 (Subj3,Visit3) 3 0 a33 0 0 1 0 (Subj4,Visit1) 4 0 a41 0 0 0 1 (Subj4,Visit2) 4 0 a42 0 0 0 1 (Subj4,Visit3) 4 0 a43 0 0 0 1 The timings for each subject and visit is indicated as "a", which could be the 0-6-15 months as you have or the age at scan. The contrasts to compare the two groups are [1 -1 0 0 0 0] and [-1 1 0 0 0 0]. The contrasts to test if the slopes for each group are positive (increase in FA over time) or negative (decrease in FA) are [1 0 0 0 0 0], [-1 0 0 0 0 0], [0 1 0 0 0 0] and [0 -1 0 0 0 0]. Define a file with the exchangeability groups as in the column "Grp" above, i.e., one group per subject. When running randomise, include the option --permuteBlocks, so that the subjects are permuted while keeping the order of the visits. There are probably other ways of accomplishing a similar result, e.g., computing a slope per subject, then using the slopes in randomise to compare groups. I think the above is a more direct solution. Hope this helps! All the best, Anderson 2013/2/25 James Cole <[log in to unmask]> > Dear FSLers, > I'm hoping to run TBSS on diffusion data collected over three visits (0, 6 > and 15 months) on a sample of 85 subjects (45 patients, 40 controls). I'm a > little stumped as how to set up the appropriate design matrix and contrasts > for Randomise and would greatly appreciate any guidance. > From what I understand about Randomise and having looked through the FSL > archive, there isn't a formal way to model change within subjects about 3+ > timepoints, in order to compare this change between groups. My attempt at a > design matrix (simplified to 3 controls, 3 patients) got as far as this: > > con-time1 con-time2 con-time3 pat-time1 pat-time2 > pat-time3 con1 con2 con3 pat1 pat2 pat3 > 1 0 0 0 0 0 1 0 0 0 > 0 0 > 1 0 0 0 0 0 0 1 0 0 > 0 0 > 1 0 0 0 0 0 0 0 1 0 > 0 0 > 0 1 0 0 0 0 1 0 0 0 > 0 0 > 0 1 0 0 0 0 0 1 0 0 > 0 0 > 0 1 0 0 0 0 0 0 1 0 > 0 0 > 0 0 1 0 0 0 1 0 0 0 > 0 0 > 0 0 1 0 0 0 0 1 0 0 > 0 0 > 0 0 1 0 0 0 0 0 1 0 > 0 0 > 0 0 0 1 0 0 0 0 0 1 > 0 0 > 0 0 0 1 0 0 0 0 0 0 > 1 0 > 0 0 0 1 0 0 0 0 0 0 > 0 1 > 0 0 0 0 1 0 0 0 0 1 > 0 0 > 0 0 0 0 1 0 0 0 0 0 > 1 0 > 0 0 0 0 1 0 0 0 0 0 > 0 1 > 0 0 0 0 0 1 0 0 0 1 > 0 0 > 0 0 0 0 0 1 0 0 0 0 > 1 0 > 0 0 0 0 0 1 0 0 0 0 > 0 1 > > Does this look correct? As for the contrasts, pairwise comparisons are > feasible (i.e. visit 1 > visit 2, visit 1 > visit 3 etc), but want I really > would like it to model change across all three timepoints and look for > group differences in patterns of change. How would this be possible using > Randomise? > Many thanks, > James > > > James Cole PhD > Huntington's Disease Research Group > UCL Institute of Neurology >