Jo had this problem with her third-world data, which is a panel with
unequally spaced waves. Monica suggested
replace time2 = 1 if d73 == 1 ;
replace time2 = 2 if d78 == 1 ;
replace time2 = 3 if d83 == 1 ;
replace time2 = 4 if d88 == 1 ;
replace time2 = 5 if d93 == 1 ;
replace time2 = 6 if d98 == 1 ;
tsset panchayat time2 ;
While this will technically work, one should be VERY careful. Given that
Jo's observations differ in their timing by two years (86, 88, ... 96), [as
opposed to Monica's, which appear to be quinquennia] keep in mind that the
xtabond estimator is based on differencing the time series (which is why
there were 'no observations' -- a biannual series on an annual calendar has
no defined first differences). Creating a time variable in which t+1 = 2
years down the road is fine, but you want to be very careful not to
interpret any findings in 'per annum' terms, since for this dataset one unit
of time = 2 calendar years. For instance, if anything in the model ends up
being an estimate of a growth rate, it should be decompounded to annual for
comparison.