Hi Paddy I will try to answer your queries - but Jean or Karl or one of the people involved in coding this may have something to add. I'm performing MVB on various regions of interest and can't help feeling > that the size of the search volume one selects will impact quite > drastically on the log evidence for a particular spatial prior. Is there > any rule of thumb for determining the ideal search volume given the size of > the ROI? > As far as I know there is no reason that ROI size per se will increase the evidence - I would expect this to depend on the nature of the psychological variable to be predicted and the relationship between it and the regions within the ROI. If you use small ROIs within a specialised region, you may observe size-dependency, but if you extend the ROI outside the regions 'interested' in your process then evidence should stop increasing. The Bayesian model evidence naturally takes account of model complexity, so there will be a penalty to increasing ROI size. However, because the greedy search MVB uses finds a sparse solution on the pattern weights within, you may not see an obvious relationship. I'd be very interested to hear any comments from those who really understand the estimation nuts and bolts here :) And on top of this, is there anywhere in which the four different spatial > priors are expained in Layman's terms? > Here is a stab at this: 1. Sparse - patterns are voxels, so the model is a combination of individual voxels, ie distributed activity 2. Smooth - patterns are voxels with local gaussian weighted support (default FWHM 4mm), ie the model is a set of locally smooth clusters 3. Compact - patterns are defined around voxels which have the maximum variance in activity (after variance attributable to design confounds has been removed), in order. Each pattern is the principal eigenvariate of the maximum variance voxel (then next to max variance voxel, and so on) and its closest neighbours within the ROI. For details see spm_mvb_U.m. 4. Support - patterns are simply voxel data adjusted for design confounds (5. Singular - patterns are singular variates obtained from an SVD of the ROI voxel data adjusted for design confounds - this may not be available in the lastest release?) My understanding of 3-5 is based on a reading of the pattern definition code in spm_mvb_U.m. I recommend you have a look there yourself. See our recent paper for some discussion of interpretation of the 'sparse' (distributed) and 'smooth' (clustered) models ( http://www.sciencedirect.com/science/article/pii/S1053811911009888). I hope this helps Alexa -- Dr. Alexa Morcom Centre for Cognitive & Neural Systems Centre for Cognitive Ageing & Cognitive Epidemiology Psychology, University of Edinburgh http://www.ccns.sbms.mvm.ed.ac.uk/people/academic/morcom.html The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336