If -dmbase is on (the default), then all the baseline model columns have their mean removed, except for the constant ones. (To be exact, for polort > 0, the portions of the vectors corresponding to each Pn(x), for n=1..polort, for each imaging run -- from -concat -- have their mean calculated and subtracted for that run's part of the vector; for -stim_base vectors, the mean over the full length of the vector is calculated and subtracted from each element.)

The purpose is to improve the condition number (numerical stability) of the regression matrix X. The baseline subspace (the vector span of the baseline model columns) is not altered by this, since the constant column(s) remain. Since the activation model F statistics are really comparing how well subspaces fit the data, using -dmbase will not affect the statistics of the activation model fits, since the baseline subspace is not affected (only its set of basis vectors). However, it will affect the baseline model coefficients.

This feature was added because some people here use as a regressor the residuals left from the 3dvolreg program (in addition to the movement estimates). I don't think this is a good idea, but they do; however, this column of data is usually nearly constant, so with the P0(x)=1 column, the X matrix might be nearly singular.

You can disable this with -nodmbase, if you like. Circumstances when you might want to use -nodmbase are:
- when you want the coefficients of the baseline model to reflect the input baseline stimuli precisely, so that you can use these coefficients for some other purpose
- that's the only circumstance I can think of actually

If this is too much mathematical jargon for you, then don't use the word 'exactly' in your future questions.

bob cox