I have a question in my mind, but I have never be such clear.

For the 3dDeconvolve step (of course, based on each subject), if I really do not collapse the regressors, I have about 70~80 regressors. Is that fine for a 3dDeconvolve? In detail,in my experiment, there are 5 runs, each run has 72 trials. The design is (6 SOAs x 6 Event conditions) x 10 repetitions/observations. As I have 6 SOA x 6 Event condition, so for the stimuli onset, I have 36 regressors. Then for the response stage, also 36 regressors. For them, totally 36+36=72 regressors. If I also include the error regressor, miss regressor, and then 6 motion regressors, then totally 80 regressors.

For each regressor, it has about 10 observations, i.e., each run has 2 observations for that event. For instance, the timeonsets for one regressor go below.

Again, to decrease the regressor number, how about I collapse one first, then collapse the other? Suppose I have 120 regressors - they are from stimuli stage, and response stage. Thus I can first collapse the response stage regressors, so in the GLM (3dDeconolve), finally I have 60 onset regressors + 1 response regressors = 61 regressors. By this way, I can analyze the onset effect.

Then I collapse the onset regressors, but use the individual response stage regressors, so finally another 61 regressors. 1 onset regressor + 60 response regressors = 61 regressors. By this way, I can analyze the different response effect.

Is this way fine? I think it would be confusing if in a GLM (3dDeconolve), if there are too many regressors. But again, as I said, I really do not know how many is “too many”, hence not good.

How long are the stimulus events? If they are short,
having 10 events per class might not be sufficient to
obtain robust beta weight estimates.

How many regressors is “too much” is hard to say in a
precise way, but 120 certainly seems like a lot,
particularly if they are not from TENT functions.

Modeling it 60+1 at a time may be a possible way to
evaluate the robustness of the model, to some degree.
But given that it is possibly desirable suggests that
the model might not be robust. If the 120 term model
actually fails, testing as 60+1 is surely not a good
way around that.

Is 3dDeconvolve complaining about the 120 parameter
version? If so, what are the messages?

Thanks so much for your reply. My stimulus events lasts about 1.5s.

I was silly to think of using too many regressors. In my current model, there are 42 regressors (6 motion stimuli files are included). But now I meet a trouble, and can not find the way to correct it. It is very strange. I do not find any problem (script, also the time stimuli files), but the script can not go!

** ERROR: ‘-stim_base 13’ illegal with ‘-stim_times’
** ERROR: ‘-stim_base 14’ illegal with ‘-stim_times’
** ERROR: ‘-stim_base 15’ illegal with ‘-stim_times’
** ERROR: ‘-stim_base 16’ illegal with ‘-stim_times’
** ERROR: ‘-stim_base 17’ illegal with ‘-stim_times’
** ERROR: ‘-stim_base 18’ illegal with ‘-stim_times’
** FATAL ERROR: Can’t continue after above problems!
** Program compile date = Aug 13 2014

The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.