Maybe there are 3 discussions going on here, so it might be good
to try and clarify them.

1. 3dDeconvolve allows 'zero' regressors for the convenience of
scripting. The options -allzero_OK and -GOFORIT are to make life
easier (so that the same command can be run, producing the same
output sub-bricks). If an input regressor has no stimulus times,
the output for it will be zero.

Separate from that is the question of what to do with the results
in group analysis, as Gang has addressed.

2. Gang's point about having an equal number of time points seems
to be slightly different from Michael's question, but is quite
relevant. I belive his point is that for a given regressor, the
number of stimuli might vary quite a bit across subjects (in this
case). 3dMEMA would be able to deal with that since t-stats are
carried forward. The ANOVA programs would not take it into account,
and the assumption of small within-subject variability might be
violated to some degree.

3. Michael may be referring to the case of subjects not having data
at all for some regressors, so an ANOVA might not be appropriate.

Gang, I guess you still suggest 3dLME.R for that case, yes?

- rick