Jim,

Although you tell 3dDeconvolve to ignore the first run of 244 TR's in the concatenated dataset, the program will not use the first run for analysis but still keep those 244 time points for time continuation purpose. Therefore it is not correct to change your -concat file anyway because 3dDeconvole would treat the original three-run dataset as from two runs. The invertibility problem of matrix X'X is most likely a chain reaction of this incorrect setup, and you can check this by adding an option of -xout in 3dDeconvolve to verify this (X will be spilt out on screen).

I'm not 100% sure, but I think that you should use the original concat file and also keep the glt matrices intact. I would have to dig into the codes to figure out whether this censorhsip would bring you the expected results (normally the -censor option is for individual abnormal time point to be excluded).

It might not be the best approach in your case, but I feel the following method would be easier and less troubling. If you've kept all your ANFI running scripts, simply create a new directory for each input dataset, and modify the one for concatenation, change the concat file and glt matrices in 3dDeconvovle. In doing so, you would preserve the original results from your previous running, and you would not have to change those output file names. It might waste some disk space, but it probably ends up saving your time.

Good luck,
Gang