Posted by B. Douglas Ward on April 26, 2001 at 12:07:49:
In Reply to: choice of the length of estimated IRF in 3dDeconvolve posted by Giuseppe Pagnoni on April 25, 2001 at 10:39:31:
Giuseppe Pagnoni:
1. With TR = 2 sec, and the US -> next CS interval varying from 4 to 7 sec.,
the input stimuli change at sub-TR times. That is, the input stimuli are
not constant within each TR.
(Restricting the US -> next CS interval to be a multiple of TR, so that the
input stimuli change only at multiples of TR, would greatly simplify the
analysis. However, in the following I will not assume this.)
Therefore, before running 3dDeconvolve, you should first use 3dTshift to align
all slices of the 3d+time dataset to time offset 0.
2. Since changes in the input stimulus functions occur at sub-multiples of TR,
it will be necessary to use the -stim_nptr option of 3dDeconvolve. For example,
if the stim functions change at multiples of 1 sec, then the 'Regular' and
'Delayed' stim functions will have to be specified at intervals of 1 sec (i.e.,
nptr=2). Note that the minlag and maxlag for these stim functions will also
have to be specified in units of (TR/nptr) = 1 sec. (See option -stim_nptr
on p.23 of 3dDeconvolve.ps)
To model a hemodynamic response that lasts about 10 sec., you could use the
following commands:
-stim_file 1 Regular.1D -stim_maxlag 1 10 -stim_nptr 1 2 -stim_label 1 Regular \
-stim_file 2 Delayed.1D -stim_maxlag 2 10 -stim_nptr 2 2 -stim_label 2 Delayed \
In this case, each IRF is estimated at 11 points (0 to 10 seconds).
For a longer response (you mentioned 30 sec), the maxlag would have to be
increased accordingly. However, please keep in mind that there is only a
limited amount of data. To estimate a 31 pt. IRF for each stim function, in
addition to 6 motion parameters, 1 global average, and 2 baseline parameters,
that's a total of 71 regression parameters (and only 150 time points).
3. It is important to select the proper value for maxlag. This parameter
is for modeling the time delay due to the hemodynamic response. For the
finger-tapping paradigm, a maxlag of 10 secs. seemed sufficient. Of course,
this would vary from experiment-to-experiment, not to mention voxel-to-voxel
within the same experiment. Constructing the correct model, something that
would be generally applicable across all voxels, is part art and part science.
You should try to use the minimum value for maxlag that permits a good fit to
the data.
The Deconvolution plugin should be used, both before and after running program
3dDeconvolve, to visually verify the adequacy of the fit for voxels that are
deemed to be active. The Deconvolution plugin lets you interactively change
the model parameters (minlag, maxlag, etc.), and displays the full model fit
on top of the actual data. Make sure that you use the same input parameters
as you would use for 3dDeconvolve. In particular, if you are using sub-TR
stimuli, the NPTR option (at far right of window) must be set accordingly.
4. Using program 3dDeconvolve, you can compare the IRF's for Regular and
Delayed directly with the -glt option.
In order to use the -glt option, use must create a file which contains
the GLT matrix which specifies the statistical test.
For example, if you wish to compare the "area under the curve" for the
Regular and Delayed IRF's, this requires a GLT matrix such as:
glt1.mat =
0 0 1 1 1 1 1 1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 0 0 0 0 0 0 0
It is very important to understand that each column of the matrix corresponds
to a parameter in the full model. In the present case, the first 2 columns
correspond to the baseline parameters (constant and slope). The next 11 columns
correspond to the 11 pt. IRF for Regular. The following 11 columns correspond
to the 11 pt. IRF for Delayed. The final 7 columns correspond to the motion
parameters and global average.
With this in mind, the above GLT matrix represents the following test
of hypotheses:
Ho: Area(Regular IRF) - Area(Delayed IRF) = 0, vs.
Ha: Area(Regular IRF) - Area(Delayed IRF) <> 0
The commands -glt 1 glt1.mat -glt_label 1 "DiffArea"
will produce 2 sub-bricks. The first, "DiffArea LC[0]" will contain the
difference in the areas of the two IRF's. The second, "DiffArea F-stat"
will contain the F-stat for statistical significance of this difference.
You can also test for the point-wise difference in the IRF's . This is
accomplished by the following 11 row GLT matrix:
glt2.mat =
0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0
This GLT matrix represents the following test of hypotheses:
Ho: Regular IRF(t) = Delayed IRF(t), vs.
Ha: Regular IRF(t) <> Delayed IRF(t)
The commands -glt 11 glt1.mat -glt_label 2 "DiffIRF"
will produce 12 sub-bricks. The first 11 sub-bricks will contain the
point-by-point difference between the Regular and Delayed IRF's. The 12th
sub-brick will contain the F-stat for the statistical significance of this
set of 11 differences.
5. You can create two 3d+time datasets containing the estimated IRF's
at each voxel location with the following commands:
-iresp 1 Regular.IRF \
-iresp 2 Delayed.IRF \
As mentioned above, you are estimating quite a few parameters relative to the
amount of data. Do not be surprised if the estimated IRF's appear quite
noisy. One way that you could smooth the IRF's is to average them over
ROI's (say, regions of activation). Alternatively, you could use program
3dNLfim to fit a smooth curve to each voxel's IRF. But that's another story.
6. For more details, and numerous examples, see the documentation contained
in file 3dDeconvolve.ps.
Doug Ward