Re: st: brr and replicate weights specification

[Stas Kolenikov <skolenik@gmail.com>]

In your -svyset-, you should use the fully processed weight rather
than the 0/1 variables (Hadamard matrix, essentially) that was
provided to you. (I am surprised to hear that anybody could do that;
at the data provider side, they have the capacity to produce the fully
appropriate BRR weights which would be

(original probability weight) times (0 or 2 for the current BRR
replicate) times (post-stratification and non-response adjustment
specific to the particular BRR replicate)

That is, if you do any adjustments of the weights, the similar
procedure should be followed with BRR. That the data provider did not
do that is odd, to say the least.

Anyway, back to your question: you will have something like

foreach x of varlist <0/1 weight modification variables> {
   gen brrw`x' = pweight*2*`x'
}
svyset [pw=pweight], vce(brr) brrw( brrw* )

You are right that you don't need the stratification and clustering
information on top of this. However, if you do have that information,
you can

svyset PSU [pw=pweight] , str( strata )

and check that you indeed have 2 PSUs/stratum with

svydes

If you don't, then you would rather want to use the second statement
than the first one.

Generally, all variance estimation methods are asymptotically (in the
number of PSUs) equivalent to one another, so the considerations to
choose BRR or jackknife or linearization or the bootstrap or
generalized variance functions or whatever have you generally come
from non-statistical considerations. Usually replication weights are
provided in an attempt to mask the PSU identifiers with the hope of
better identity protection of the survey respondents. However
providing both strata/PSU and BRR weights in a single data set does
not serve that purpose. So it is unclear to me as to what the idea
behind these weights is.

On Wed, Apr 21, 2010 at 2:06 PM, Ian Breunig <ianbreunig@gmail.com> wrote:
> Dear Statalist,
>
> I'm attempting to use svy: brr estimation.  My data set is the MEPS
> data set supplied by AHRQ.  MEPS supplies a sampling weight and a
> standard set of 128 BRR replicates (= 1 or 0).  They instruct to
> create replicate weights as BRRwt = BRR * 2 * pweight  (where BRR is
> an element of [1,128]) then use these as balanced replicate wieghts.
>
> My question is,  when I declare my - svyset - statement should the
> brrweight() option contain the BRR1wt calculated above or the 128
> replicates that are equal to 1 or 0?  If the former one is correct do
> I still need to declare my sampling weight in the  - svyset -?
>
> The [SVY] manual does not explicitly explain this.  Does the brr
> command create the BRRwt calculated above using the pweight and 128
> replicates defined, or does it use the BRRwt calculated and supplied
> by me and then ignore the sampling weight.
>
> Lastly, I understand from the [SVY] documentation that I should ignore
> the stratum and cluster variables supplied by MEPS when declaring -
> svyset - when using - brr -.  Can you confirm this, or does it hurt to
> put those in as well?
>
> Thanks ahead of time,
>
> Ian Breunig
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>



-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/