On Wed, 20 Jan 2021, Riccardo (Jack) Lucchetti wrote:
On Wed, 20 Jan 2021, Allin Cottrell wrote:
> I've committed a "fix" which brings us in line with stata (and
> R). Prior to that, we were netting out observations with zero
> weights from n in calculating standard errors if and only if the
> weights series was a 0/1 dummy. Yet in the model printout we were
> always reporting a number of observations net of the
> zero-weighted ones. Now we always net out points with zero weight
> from the start.
>
> I think this is defensible. Data points with zero weight have no
> effect on the parameter estimates by construction. Points with
> very small weights may or may not have a non-negligible effect on
> the estimates; that'll depend on their leverage as well as the
> size distribution of the weights (perhaps they're all tiny?).
Hm. To be honest, I'm not 100% comfortable with this. My gut
feeling is that using weights to surreptitiously perform
subsampling feels a bit like cheating. I don't like very much the
idea that you can dramatically alter an essential statistic such
as a hypothesis test simply by replacing 0 with 1.0e-9 or vice
versa.
In my view, if you want some observations to be out of your
sample, you should do it the "proper" way, that is, by using the
smpl command. [...]
I take your point. The principled way to effect this would be to
disallow non-positive weights in WLS (as small as you like, but not
exact zero). That would put us at odds with other popular
statistical software but would be defensible. Anyone else care to
weigh in on this?
Allin