The new program 3dREMLfit is designed to solve linear systems,
like 3dDeconvolve does, but with allowance for serial (temporal)
correlation in the noise time series in each voxel.

3dREMLfit uses the REML (Restricted/Residual Maximum Likelihood)
method to estimate the two ARMA(1,1) parameters for each voxel's
serial correlation matrix. It then uses this estimated correlation
structure in a Generalized Least Square (GLSQ) solution of the
linear regression problem. 3dDeconvolve uses Ordinary Least Squares
(OLSQ) to fit the regression model to the FMRI time series dataset.

The principal difference from the results of 3dDeconvolve is that
the F- and t-statistics in individual subject activation maps will
be estimated more accurately, since the noise variance (which is
in the denominators of the formulas for these statistics) will be
estimated more accurately. The actual regression parameters (betas)
will not usually vary in any significant way, since OLSQ and GLSQ
are both unbiased.

At the group analysis level, where only the betas are carried forward
(e.g., in 3dANOVA or 3dLME), the analysis results will not be markedly
different from using 3dDeconvolve, since the betas are not strongly
affected by changing from OLSQ to GLSQ (i.e., from 3dDeconvolve
to 3dREMLfit).

At the individual subject level, the thresholded activation maps
from 3dREMLfit may differ from those generated by 3dDeconvolve,
particularly for slow block design studies. If you are using the
thresholded individual subject maps for any important purpose, you
should probably evaluate the use of 3dREMLfit.

3dDeconvolve is used as a front-end to run 3dREMLfit. 3dDeconvolve
generates a matrix file, containing all the regression columns,
stimulus labels, GLTs, etc., which is input to 3dREMLfit along with
the 3D+time dataset to be analyzed. It is not necessary to run
the analysis completely with 3dDeconvolve before starting 3dREMLfit:
3dDeconvolve can be instructed to stop after it outputs the matrix
file, using the '-x1D_stop' option.

The bucket datasets output by 3dREMLfit are designed to mirror the
structure of those from 3dDeconvolve, to make it easy to adapt your
scripts to use this program. In addition, when you run 3dDeconvolve,
it will print to the screen a command line that can be used to
run 3dREMLfit with the same input 3D+time dataset.

3dREMLfit is now available in source and binary formats at the AFNI
website. More details are available in the program's -help output,
as usual. A PowerPoint file about the program, including comparisons
with 3dDeconvolve activation maps and a brief discussion of how 3dREMLfit
relates to the SPM and FSL approaches to the same issue, is available at
  [afni.nimh.nih.gov]
  [afni.nimh.nih.gov]
Note that the PowerPoint .ppt file contains 2 .gif animations, which
you'll have to fetch as well to view the file properly. The presentation
is also available in PDF format as well, but that won't contain the
animations.

At some point in the near future, we will probably incorporate an
option to use 3dREMLfit into the afni_proc.py master script. We'll
see how the usage evolves before taking this step.

There will be an AFNI users' group meeting at the NIH in Bldg 10,
Room 4N222 (NIMH IRP conference room) at 10:30 on Tuesday,
21 Oct 2008, to discuss this and other new-ish features in the
AFNI imperium.