Dear Chen, > I have 2 questions concerning randomise inference and DTI analysis. In > Jones et al paper (NeuroImage 2005;26:546-554) mention about the problem > in different smoothing parameter and parametric statistic inference. The problem that Derek describes isn't really related to the issue of ariance smoothing. In his paper he simply shows (I think, I've only read his poster on the same topic) that depending on what filter width (for smoothing data, not just the variance) you choose you will be differentially sensitive to different size (extent) activations/differences. So, for example if you choose a 20mm FWHM filter you will be more sensitive to large spatially extended differences in FA compared to if you use 5mm. > My > question is > 1. If I use non-parametric approach and use variance smoothing under > randomise, is it still valid to overcome the non-normally disturbed > residuals problem ? what's the effect of variance smoothing ? It is valid. The effect of variance smoothing is to render the data distributed a little more like a normal distribution (rather than a low df t distribution) thereby increasing sensitivity for low df data. There is then also a potential risk of loosing some of the fine spatial detail of the changes/differences, but in practice I hardly think that is an issue. > 2. In randomise, if I calculate cluster-based threshold at one time, > should I redo the analysis again if I change the c_thresh ? Yes. Good luck Jesper