[R] Subpopulations in Complex Surveys

[Thomas Lumley]

On Mon, 24 Feb 2003, Thompson, Trevor wrote:

> For svymean, can't you just pass the subpopulation into the design
> argument?
>

Yes, but you get the wrong standard errors (very slightly) in stratified
designs.

  If you compare the subsetting approach to the regression approach you
will find slightly different standard errors for the same point estimates.
Roughly speaking, you don't deserve as much benefit from the difference
between strata if some of the PSUs you originally sampled don't have any
of the subpopulation.

I've nearly finished an upgrade that I think handles this.

	-thomas