Le mer. 20 janv. 2021 à 15:05, Sven Schreiber <svetosch@gmx.net> a écrit :

Am 20.01.2021 um 14:26 schrieb Artur Bala:
I don't know what Stata does there (although I suspect it is documented),

Yes, sure. To be frank I haven't researched this yet as I was surprised by the difference. I mean, the user does not necessarily need to be aware of the technical details and for such a "classical" model as WLS, he/she expects the same result from different software I think.

but in principle it seems easy enough to compute a fitted value yhat even for any obs (y-X-constellation) which wasn't part of the estimation sample. I mean that's what we do for prediction, no? Does a (static) "fcast" work there in gretl?

That's right too. My point is that I'm trying to replicate an iterative WLS algorithm for outliers correction first implemented in Stata. And if Stata and gretl mak different treatments for weight 0 I wonder if I should adjust the algorithm in order to come up with the same results as Stata - and then need to scrutinize how Stata deals with it - or prefer the way gretl handles the problem - as I'm doing currenlty, even if the results will be different.

Best,

Artur