Yes, it's possible to have dramatic change in those coefficient estimates for unchanged regressors because sub-dividing some of your original regressor may change the correlation landscape among some regressors. You can use ExamineXmat.R, an R program to identify where exactly the issue lies. If you have R installed on your computer, type the following inside R:
> source ('/Users/ziad/abin/ExamineXmat.R')
or from terminal prompt:
ExamineXmat.R
and then select the design matrix generated by 3dDeconvolve.
Gang