Hi there,
I have a curious problem that I tried to get to the bottom of for a few days now, and I'm ready to admit lack of any further ideas.
I have run an experiment with 4 different stimulus types. I produced model responses by convolving the 4 stimulus lists with a model HRF. When I now decorrelate the measured response based on these 4 regressors, I get reasonable results for 3 of them, and an unreasonable result for the fourth. The partial F-values for this stimulus are all exactly 1000. When I remove one of the other three from the model, or declare one of the other three a baseline stimulus, this problem seems to disappear.
None of the stimulus timecourses looks suspicious, specifically I can see no difference between the troublesome one and the other three. But it has to be a problem with the stimulus timecourse, because when I swap the models around, the problematic F-scores go with this timecourse.
The X-matrix seems OK, the individual timecourses of the 4 stimuli add up correctly, so everything should work. In addition, I don't understand how a constant F-score is possible in the first place, based on the math.
What am I missing?
Thanks for any help,
Kai