Sorry...my message ended up as an attachment...this is the content:
Hi,
On the same 2-session, 2-run/session event-related fMRI data, I have run 2 similar first-level models -- first with the 4 runs modeled as they are in their separate blocks in one GLM, and then all runs and sessions concatenated, on which I used spm_fmri_concatenate() so that the high-pass filter and non-sphericity calculations are corrected and the there are regressors for run-/session-specific effects. I have the same regressors of interest in both analysis versions with the difference that for the concatenated model, I have one main regressor of interest for all the conditions in the task and the "real" regressors of interest have been included as parametric modulators of this "task" regressor.
For some subjects whose unconcatenated model estimated fine, the unconcatenated model can have many parameters (and sometimes all of the parameter modulators of interest) that could not be uniquely estimated. I do have unbalanced number of events across runs and subjects, as the regressors depend on participant responses, but I would have thought that the unconcatenated models would be what would run into estimation problems such as low number of events (which my data is plagued with), for example, seeing as the concatenated models would benefit from more events as all the runs are combined. I don't see any blatant errors in my model design, at least that jumps out at me. Is it possible to have such a scenario that the concatenated model variant has estimation problems whereas the unconcatenated variant does not, and if so, what are some causes?
Many thanks in advance,
g
|