 |
 |
Hi Andrew, Yep, same logic applies to VBM too. All the best, Anderson Am 25.02.14 23:48, schrieb Andrew Martin: > Hi Anderson > Thanks for your help. I assume the covariate would also need to be > transformed for VBM analysis too? > Regards > Andrew > ------------------------------------------------------------------------ > *From:* FSL - FMRIB's Software Library [[log in to unmask]] on behalf > of Anderson M. Winkler [[log in to unmask]] > *Sent:* Tuesday, 25 February 2014 11:10 PM > *To:* [log in to unmask] > *Subject:* Re: [FSL] FA contrast design > > Hi Andrew, > > There are more than one way of doing, and I think I'd do as this: > > design.mat: > 1 0 a1 > 1 0 a2 > 1 0 a3 > 1 0 a4 > 0 1 0 > 0 1 0 > 0 1 0 > 0 1 0 > > design.con: > 0 0 1 > 0 0 -1 > > The 1st and 2nd EVs coding for group, and the 3rd coding the variable > that has data for only one of the groups (a1..a4). There is no need to > mean center EV3 for this contrast, but doing so won't harm and and > will allow to additionally test a contrast as [1 0 0] without having > to make other modifications. I would center EV3 within group, so that > for the 2nd group, EV3 would continue to be all zero. > > All the best, > > Anderson > > > Am 25.02.14 11:34, schrieb Andrew Martin: >> Thanks >> I planned on doing a transformation but still uncertain how to set >> out the contrast as I only have data for my clinical group. >> Regards >> Andrew >> ------------------------------------------------------------------------ >> *From:* FSL - FMRIB's Software Library [[log in to unmask]] on behalf >> of Anderson M. Winkler [[log in to unmask]] >> *Sent:* Tuesday, 25 February 2014 8:03 PM >> *To:* [log in to unmask] >> *Subject:* Re: [FSL] FA contrast design >> >> Hi Andrew, >> >> This range is too wide. The model should contain an intercept (or the >> data & design should be mean-centered). Imagine then that the GLM >> will have to find a single scalar that, multiplied by this variable >> of interest, will put it into about the same range as FA (0-1). The >> regression coefficient would have to be very small, about >> 2.5*10^(-7). The variance of FA data in practice may be small, but >> not all that small across most of the brain, say, between 10^(-4) and >> 10^(-2). Then to have a modest effect with a t-statistic of, say, 1.9 >> in the voxels with the smallest variance (so, being optimistic), the >> number of subjects that would have to participate of the study would >> be about (1.9*sqrt(10^(-4))/(2.5*10^(-7)))^2, that is, something >> about 5.8*10^9, or 5.8 billion people (!). And we aren't even talking >> about correction for multiple testing in the image. >> >> If the wide range is due to outliers, consider where they arose from >> and either try solve the problem or remove them from the analysis. >> Consider also if this measurement that varies so widely is a valid >> one. If they are not outliers, and the measurement is correct and >> presumably of an useful quantity, consider applying a transformation, >> e.g., take the logarithm and use it instead. Of course, the >> hypothesis changes (no longer a linear effect from the original >> measurement, but an effect in a log-scale, which indeed is often more >> plausible than a linear one). >> >> All the best, >> >> Anderson >> >> >> Am 25.02.14 01:33, schrieb Andrew Martin: >>> Dear experts >>> I have 2 groups and a variable with a large range (0-4,000,000) for >>> one of the groups. How could I set up the design to look at >>> correlations in my FA data with this variable in the 1 group? >>> Regards >>> Andrew >> >