Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Jackknife standard errors using replicate weights in Stata 12.1


From   Stas Kolenikov <skolenik@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Jackknife standard errors using replicate weights in Stata 12.1
Date   Tue, 19 Feb 2013 18:17:23 -0600

If you have a complex sample with stratification and clustering, then
running estimates as if they were -svyset [pw=weight0]- without PSUs
and strata gives you the results that are simply wrong. I think you
need to try to figure out what's going within each replicate by
specifying -noisily- option of -svy jackknife-.

One of the conceptual problems with the jackknife (and circumventing
it depends on the particular implementation, and may differ between
WesVar and Stata) is that some of the jackknife replicates will
produce estimates that are identical to the original estimates. Stata
highlights them in output:

. webuse nhanes2jknife, clear

. svy , subpop( if region==1)  : proportion race
(running proportion on estimation sample)

Jackknife replications (62)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
ssssssssssssssssssssssssssssssssssss..............    50
ssssssssssss

Survey: Proportion estimation

Number of strata =      31       Number of obs    =      10351
                                 Population size  =  117157513
                                 Subpop. no. obs  =       2096
                                 Subpop. size     =   24237893
                                 Replications     =         14
                                 Design df        =        -17

--------------------------------------------------------------
             |              Jackknife
             | Proportion   Std. Err.     [95% Conf. Interval]
-------------+------------------------------------------------
race         |
       White |   .9477102    .011452             .           .
       Black |   .0459008   .0107839             .           .
       Other |    .006389   .0023199             .           .
--------------------------------------------------------------

"s" in place of the dot indicates that Stata found the estimate to be
identical to the original one, so Stata did not really trust the
results, and omitted that particular replicate. There were only 14
replicates here in which the estimate was different from the original
one, and the standard errors were based on these 14 replicates. This
is the right approach, as comparison with the linearized standard
errors shows:

. webuse nhanes2, clear

. svy , subpop( if region==1)  : proportion race
(running proportion on estimation sample)

Survey: Proportion estimation

Number of strata =       7        Number of obs    =      2096
Number of PSUs   =      14        Population size  =  24237893
                                  Subpop. no. obs  =      2096
                                  Subpop. size     =  24237893
                                  Design df        =         7

--------------------------------------------------------------
             |             Linearized
             | Proportion   Std. Err.     [95% Conf. Interval]
-------------+------------------------------------------------
race         |
       White |   .9477102   .0114516      .9206315    .9747888
       Black |   .0459008   .0107834      .0204021    .0713995
       Other |    .006389   .0023198      .0009035    .0118745
--------------------------------------------------------------
Note: 24 strata omitted because they contain no subpopulation
      members.

Stata produced a more informative message this time.

This may be one of the problems you could have with the jackknife. If
you posted more of the syntax and output, identifying what you
perceive as problems, we could give you more detailed feedback.

-- 
-- Stas Kolenikov, PhD, PStat (SSC)  ::  http://stas.kolenikov.name
-- Senior Survey Statistician, Abt SRBI  ::  work email kolenikovs at
srbi dot com
-- Opinions stated in this email are mine only, and do not reflect the
position of my employer


On Tue, Feb 19, 2013 at 9:36 AM, Annelies Blom <blom@survex.de> wrote:
> Dear colleagues,
>
> for some analyses that I am currently conducting I have a dataset with a complex
> sample design with replicate weights. The replicate weights were produced with a
> jackknife(1) procedure. Various analyses are to be conducted with this setup.
> Other project partners conduct the analyses in WesVar and our results have to be
> comparable to theirs. I  prefer using Stata, but am encoutering problems.
>
> An example of my svyset and analysis command:
> svyset [pw = weight0], jkrw(weight1 - weight80, multiplier(1))
> svy jackknife: proportion age
>
> The following setup and command yields the same results:
> svyset [pw = weight0], jkrw(weight1 - weight80, multiplier(1)) vce(jackknife)
> svy: proportion age
>
> The problem:
> The standard errors calculated by Stata with these commands are unplausibly
> small. In addition, for some variables Stata does not give any SEs at all (or no
> confidence intervalls). The SEs when not specifying jackknife SEs (i.e. when
> calculating Taylor linearized SEs, which is the default) are much more
> plausible. And they are complete.
>
> Via the Stata search function I found
> <http://www.ats.ucla.edu/stat/stata/library/replicate_weights.htm>
> These researchers state that "although the command will run (and run faster)
> without the jackknife option after the svy, you will get linearized standard
> errors instead of the jackknife standard error.  This jackknife standard error
> matches the standard errors produced by both SUDAAN and WesVar."
> They refer to Stata 9, while I am working in Stata 12.1. According to them, if I
> do not specify jackknife SEs, my results are not be comparable to the WesVar
> results.
>
> My questions:
> - Are the jackknife SEs calculated by Stata 12.1 correct? If so, why do I get
> missing SEs?
> - How can I solve these problems while at the same time remaining comparable
> with analyses conducted in WesVar?
>
> With best wishes,
> Annelies
>
> ---
> Annelies Blom, Ph.D.
> Survex - Survey Methods Consulting
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index