¥ Implements Solution #2: estimate
correlation parameters and use GLSQ
H REML is
a (partially nonlinear) method
for simultaneously estimating variance + correlation parameters and
estimating regression fit parameters (b s)
H Each
voxel gets a separate estimate of its own correlation
parameters (a,b)
oEstimates
of a
and b
can be spatially smoothed before they are used to compute the b s
oCan also
input a
and b
directly and skip their estimation (the slow part), if desired, and use those
values to compute the b
s
oVariance estimate uses
pre-whitened residuals to keep DOF=N – m
H Even if correlation decay parameter a
was the same for all voxels, relative amount of white noise (measured by b)
mixed in would vary spatially
o Sample analyses using 1-parameter AR(1) and MA(1) models
shown later
¥ Inputs to 3dREMLfit
H Run 3dDeconvolve
first to setup .xmat.1D
matrix file, GLTs, etc.
o DonÕt have to let 3dDeconvolve
finish analysis: -x1D_stop
o 3dDeconvolve
also outputs a command line to run 3dREMLfit
with the same 3D+time dataset and the matrix file just created
H Then, input matrix file and 3D+time
dataset to 3dREMLfit
¥ Output datasets are structured to be
similar to those in 3dDeconvolve
H It should be easy to adapt scripts that
use 3dDeconvolve output
files (e.g., for group analysis)
to use the new software