I think the error in spm_spm.m occurs at the stage where the hyperparameters are being estimated. SPM does a very liberal (generous) significance test to be able to focus on voxels with a chance of signal, so there's a "hidden" thresholding going on. So if the error occurs, it either means there's a mistake in the data processing pipeline, or the signal is extremely weak or nonexistent.
Here's a post of mine in a related thread:
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=SPM;a018ed15.1101
As to why it might depend on the amount of smoothing: smoothing will change the statistics a bit, so one amount of smoothing might be enough to push some voxels above the threshold.
|