Jane,

The issue of degree of freedom aside, normalizing or not, the justification for concatenating all runs or different days of data is that you normally assume the same beta weight or same HRF corresponding to each stimulus/condition/task. With the option of -concat in 3dDeconvolve, different baselines among various runs are taken care of automatically. If you run 3dDeconvolve separately for each run or data from different day, you would most likely come up with different beta weights or different HRF's. If this is what you want to do, go for it.

If I apply the baseline correction to the raw data before I register the images, will the beta weights of the model still reflect % signal change?

If you mean normalization by 'baseline correction', yes, the beta weights would reflect the percent signal change of each regressor. If you don't normalize the data before running 3dDeconvolve, you would have to employ 3dcalc with transforming the beta weights to otbain their corresponding percent signal changes.

Do you recommend applying the baseline correction after I move the data (I'd rather not, as each day involves 9 runs, and it seems cumbersome to have to register each run separately, apply the normalization to each run, and then concatenate runs.)

Without any intrinsic difference involved, I would just simply normalize the original data separately for each run. Yes it is a little tedious to normalize the data for each run separately and then make concatenation, but the inhomogeneity of the scanning (different mean and drifting) among runs necessitates such a cubersome procedure.

Gang