The 1st row of your table doesn't look correct:
> Coef Tstat SD
> -0.1947 -0.1641 0.1186
EITHER t-statistic or SD should be 10 times bigger, OR Coef should be 10 times smaller.
The output from -sresp is the standard deviation of the coef estimate (the square root of the variance of the coef estimate), which sometimes is also called standard error of the estimate. Again the focus here is about the estimate of each coefficient, not the mean of a sample. Basically the SD column on your table can obtained by dividing Coef by Tstat.
> Other papers that do deconvolution appear to show "(between-subject SD at
> each time point)/sqrt(number of subjects)" as error bars. Is this the right
> approach to plot deconvolution time-series?
Now you're talking about group level, which is totally different from the standard deviation (error) of the coef estimate above. Since each subject has a beta (coefficient), you can easily calculate the standard deviation at each time point across subjects for such a plot, for example using 3dMean -sd ...
Gang