Dear DCM experts: a while back there was a discussion on the list about 2-step spatial smoothing (smoothing first the raw images then the con* images), the advantage being that for single-subject analyses one could use the first step to smooth just enough to implement random field theory, and the second step would bring smoothing up to a reasonable level for group comparisons. Given that DCM is typically carried out on the timeseries individually for each subject, presumably there might be an argument for using this 2-step procedure in DCM? My question is this. For a given peak in your (nice,smoothed) GLM analysis, one would presumably want to use individual subject peaks which fell closest to this RFX peak to extract data for entry into the DCM. Therefore: you don't actually need the individual peaks to pass any special threshold - just to be the highest peak in the neighbourhood. so....is there any reason why the DCM analysis should not be run on completely unsmoothed data? could one not simply define peaks of interest with a smoothed GLM analysis (as normal), go back, run the model on the unsmoothed data, and extract the timeseries from this unsmoothed analysis? I am interested in eliminating as much smoothing as possible, because my VOIs are reasonably close together (~20mm), and intrinsic connectivity (under normal smoothing: 8 x 8 x8) is extremely high (p< 1 x 10^-12 in some cases). I am wondering if it is possible for instrinsic connectivity to hit a sort of ceiling, such that there is no room for improvement (giving consequently unimpressive bilinear terms). Many thanks, Chris Christopher Summerfield [log in to unmask]