Bayesian model comparison - sjgershm/ccnl-fmri GitHub Wiki

To compare GLMs within a region of interest (ROI), first create a mask specifying the ROI. The mask can be saved as an image ('mask.nii'; this is easy to do using xjView), specified as a vector of voxel indices, or as a binary vector (whose length is the number of voxels in the volume).

We use the Bayesian information criterion as our model score, which can be computed by calling:

>> bic = ccnl_bic(EXPT,model,mask,subjects)

To convert BIC into log model evidence, use the transformation lme = -0.5*bic. The log model evidence can then be submitted to the Bayesian model selection tool in the mfit package:

>> [alpha,exp_r,xp,pxp,bor] = bms(lme)

The outputs of this function provide different ways of comparing models. One standard is to use the protected exceedance probability (pxp). See Rigoux et al. (2014) for more information.