DCMstructContent.txt

DCM struct

DCM.Y.
        y       time series 166x4
        name    region names (1x4 cell)
        dt      2 (seconds)
        X0      states 166x6 (sin and cos transforms -> as we had HPF before, to account for drifts etc.)
        Q       covariance matrices (covariance modelled as param which estimated * this matrix) (cell, large sparse)
        
DCM.a           fixed connections 4x4 double
DCM.b           modulatory inputs 4x4x2 -> why 2D? (because intentional and accidential)
DCM.c           4x2 driving inputs (intentional and accidential)
DCM.d           empty! -> for non-linear DCMs

DCM.U           input
        name    accidential, intentional
        dt      0.0625 microtime resolution (subsampling of 2s in 32 steps)
        u       ...x2 sparse double input (stores only nonzero entries (which are 1))
        idx     2x2 ??
        
DCM.xY          1x4 struct
        xyz     coordinates - different from specified -> only wants to look at coordinates where something significant -> shift until there)
        XYZmm   why different sized 3x... doubles for all regions? (probably because cut for some images if at border or if our centre at midpoint of voxel or not -> different # voxels included)
        others  ??
        
DCM.v           # slices (166)
DCM.n           # regions (4)

DCM.TE          echo time (why 0.04? -> how to set?) - probably specified in study data but should not matter much
DCM.delays      if slice timing correction done -> can set all to 0, else takes into account when each region has been recorded (as our regions around in same height and about in the middel (where 0 is), doesn't matter, 0 ok)
DCM.options     how to set?


.y              resulting series
.R              what remains as noise
look at var(y)/var(y+R)
free energy F
.EP posterior params



To estimate:
> spm_dcm_estimate(’DCM_mod_bwd’);

to concatenate X0:
combine blocks (put them "on diagonal")

Q are just diagonal matrices of '1' -> assumes Gaussian noise for each region
e.g. again 4 Qs, prolong each by factor 6 (assumes that same noise everywhere) or 24 such matrices, each size 24*166 -> estimate 24 noise params, concatenate them