On Wed, August 23, 2006 08:14, Arnaud Bervas wrote: I think your problem is related to the fact that, when you're in non-contiguous subsample mode, lags don't have an obvious temporal meaning anymore. Suppose you have these data time x 1 3 2 -2 3 4 if you restrict the sample by the rule x>0, and then do y=x-x(-1), you'll find that the value of y for period 3 is 4-3=1, because the "-2" simply isn't there anymore. Is this right? Is this wrong? Obviously, it's a convention; as such, it can be useful in some circumstances and deadly in others. But I'd say: if you want to compute differences, do "smpl full" first and ONLY AFTER do the subsampling. In your case, the changing sample depends on the fact that you re-generate the differences in subsample mode, which in turn leads to a different sub-sample, and so on. Riccardo "Jack" Lucchetti Dipartimento di Economia Facoltà di Economia "G. Fuà" Ancona