Dear Doug!
Thanks for you clear answer. When I thought about you suggestions some other questions came up. Please, let me know, if there is no theoretical answer to it.
1) We are using a mixed design, which can be seen somehow as combination of both a block and event-related design. There is a baseline condition (coded a zero in my stimfiles) in between "blocks" of a pure event-related presentation of the stimuli (three stimuli). The stimuli occur at 1/2 TR of 3sec and are totally randomized. The interval within "blocks" between stimuli varies from 0 to 3 TR (also subsampled as sub-TRs or 1/2TR). The mean interval is about 2 sec. Null events ("Fixation Crosses") are presented between the occurence of the 3 stimuli. During the occurence of these event-related "blocks" subjects have to perform 2 tasks (One is a control task). Only one task has to be performed for each "block". The blocks are balanced for both runs ( 183 TR each). The baseline condition is present at at about 50% of the time. During the pretesting of this design with the -nodata option of 3dDeconvole the parameters for estimation and detection eficasy looked quite good.
So formally spoken, lets say I call baseline B, the task T, and the control task C the runs would look something like this:
B T B T B C B C first run
B C B C B T B T second run
To get started I defined 6 regressors representing 6 columns of in my stimfile looking something like this (binaries representing 1/2 TR):
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0 for B followed
1 0 0 0 0 0 by T
0 1 0 0 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 1 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 0
0 1 0 0 0 0 for T followed by
0 0 0 0 0 0 B again
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0 etc.
The first 3 columns represent stimuli during the TASK and the second 3 columns during the control task.
Sorry for that long introduction, but I think it is important that you exactly know, what I have to deal with...
Well, my first question is, if you think that for this kind of design a deconvolulion or a multiple linear regression approach would be more suitable.
I actually tried the deconvolution approach at the very beginning and was not too content with it analyzing two subjects. I shifted therefore to the use of waver, which seemed to work better. But I also improved meanwhile preprocessing steps, which could motivate me to go back again, if you think this would be more approriate.
2) The second question is that, if I would use the deconvolution approach, I have to use the -ntpr option and that seemed to cost some statistical power. This is why I used waver, because I could get around of this option, because I used a time vector as input.
Well, considering these two operation modes of 3dDeconvolve, I have in my case the "trade off" of using the -ntpr option in the deconvolution approach and I do not have this with waver.
Do you think, this will change your suggestions?
Thanks in advance for your patience with my questions...
Lukas
B. Douglas Ward wrote:
>
>
> Hello Lukas:
>
> As I see it, there are basically two operating modes for
> 3dDeconvolve:
>
> 1) Multiple linear regression -- Use program 3dDeconvolve to
> model the fMRI
> signal as a linear sum of the input stimulus functions. In
> this case, you
> might want to "preconvolve" the experimental input binary
> sequence(s) with
> the assumed shape for the hemodynamic response (say, a gamma
> variate function),
> using program waver. This is then used as the input stimulus
> function for
> program 3dDeconvolve. (You might want to use the "-peak"
> option of waver, to
> keep the amplitudes reasonble.) Also, in this case, you
> would normally set
> the 3dDeconvolve maxlag = 0.
>
> 2) Deconvolution -- Use program 3dDeconvolve to estimate the
> hemodynamic
> response (aka, IRF) for each of the input stimuli. In this
> case, the input
> stimulus functions should usually be the "raw" binary
> sequences. That is,
> the binary sequence(s) should NOT be preconvolved with the
> assumed hemodynamic
> response, since you are attempting to estimate the
> hemodynamic response (IRF).
> So, no waver. Also, you should set maxlag > 0.
>
>
> The multiple linear regression approach is usually more
> appropriate for
> block type designs. The deconvolution approach is usually
> more appropriate
> for event related designs.
>
> There are probably exceptions to all of the above statements.
>
> Doug Ward