Chunming,

I found that by the plug-in deconvolution,and DC fit,the modeled curve fit the original time series to some extent, and below the baseline.So this meant the deactivation should not be doubted?

Well, if you are sure the signal is below the baseline across the whole session for that specific stimulus, then you might get deactivation. If this is the case across all the subjects, it would be more convincing.

In addition, due to the task was naming, in some time points,the movement was big,such as head jerk.So should I add motion parameters into the GLM?

There are a couple of approaches to deal with head movement. If you see big spikes, try to use 3dDespike. If they are a couple of isolated time points, you can use option -censor in 3dDeconvolve. Adding motion parameters as extra regressors is another solution. You might have to try all of them and see which works best.

And if I added time lag, such as maxlag=3, but only the first lag(1 0 0 0)was examined, by which deactivation decreased significantly,was it reasonable?

Did you set up 3dDeconvolve with minlags = 0 and maxlags = 3? I thought that you had a block design, no? I am not sure what your question is here.

normalizing the data before deconvolve might underestimate the percent change of signal in some cases

Yes, it would underestimate the percent signal change, but the amount most of the time would be negligible. I would not be worried about it unless you see the baseline constant is way off from 100 (e.g. bigger than 110).

Gang