OK, it means that you have multicollinearity issue with some combinations of minlag and maxlag, which makes the linear system insolvable with the corresponding design matrix.
Thhis is some intrinsic problem about the experiment design in which there is not enough randomization of the stimuli. With the data already collected, there is not much you could do about it unless you experiment with various minlag and maxlag as you have already tried.
This unfortunate situation reiterates the importance of analyzing experimental design before real scanning, e.g., running RSFgen and/or 3dDeconvolve with '-nodata' option.
Gang