Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: brrweight(), jkrweight(), and vce() options of svyset
From
Stas Kolenikov <[email protected]>
To
[email protected]
Subject
Re: st: brrweight(), jkrweight(), and vce() options of svyset
Date
Wed, 12 May 2010 13:27:30 -0500
On Wed, May 12, 2010 at 10:57 AM, Sergiy Radyakin
<[email protected]> wrote:
> 2Stas: thank you for your message. Is there any command that will use
> "something from jkrweights()" and "something from brrweights()" in one go?
No, that is meaningless and impossible.
> Also, you write "for each
> particular estimation command you run, you will have a choice of
> whether you want to
> specify the -vce(linearized)-, -vce(brr)- or -vce(jknife)-". Could you
> please demonstrate this
> with the -svy- commands?
webuse nhanes2brr, clear
merge 1:1 sampl using http://www.stata-press.com/data/r11/nhanes2jknife.dta
drop _merge
svyset
* amazingly, Stata picked up the jackknife weights into the design!
merge 1:1 sampl using http://www.stata-press.com/data/r11/nhanes2.dta
drop _merge
svyset
* but it failed to pick up the PSUs into the design. Ah well, let's force it:
svyset psu [pw=final] , strat( strata ) jkr( jkw* ) brrw( brr* )
svy : mean bp*
svy , vce( linearized ) : mean bp*
svy , vce( brr ) : mean bp*
svy , vce( jknife ) : mean bp*
The differences in the standard errors are insubstantial, although if
we were talking about estimation of the standard error for Gini
coefficient, I may have some hope that the BRR standard errors work
OK, but I certainly won't trust linearized standard errors.
> 2everyone: Let me rephrase the question: given only ONE svyset command
> in the program,
> (and possibly multiple svy estimation commands following it) does it
> ever make sence to
> allow for both brrweight() and jkrweight() options simultaneously in
> this command. If so, can you demonstrate such a case?
As I said in the previous post, if you study the properties of
different variance estimators (performance in different designs, small
sample properties, etc.), it might make sense. I personally prefer to
-svyset- my data differently each time I want to try a different
variance estimation method, just to be sure Stata does not have any
extra room for (mis)interpreting/defaulting to the options in the way
different from intended.