Re: st: strata with single sampling units

[frone@ria.buffalo.edu]

Steve, 

Thanks.  Yes, it is inappropriate to ignore stratification, but I wasn't 
sure how to deal with all of the single certainty units.. 

You mentioned using fpc = 1 for the certainty units.  But what about the 
other units?--no information on the sampling rates.

I went back to the various documents for the data and found that they 
suggested the following code for SUDAAN:

PROC procname DESIGN=WR DATA=dsname; 
NEST stratum psu / MISSUNIT; 
WEIGHT weight; 

I haven't seen SUDAAN code in over 15 years, so I'm not sure what MISSUNIT 
is for.  But they aren't using fpc from what I can tell.

Then I went back to look at features in -svyset-.  An option is singleunit 
(missing/certainty/scaled/centered).  The keyword -certainty- causes 
strata  with single sampling units to be treated as certainty units, thus 
contributing nothing to the standard error.  It seems from the description 
of the sampling below and your response that  using -singleunit 
(certainty)- would be the way to deal with the single certainty units.  So 
I am now using:

svyset psu [pweight=weight], strata(stratum) singleunit(certainty);

This works in that I get variance estimates, and they are often much 
smaller than what one gets ignoring the stratum variable.  However, the 
odd thing is that when I try to replicate their published estimates 
(proportions) and SEs, I get basically the same SEs if I ignore the 
stratum variable.  The SEs I get using -singleunit (certainty)- tend to be 
much smaller.

Mike Frone 




sjsamuels@gmail.com 
Sent by: owner-statalist@hsphsun2.harvard.edu
03/04/2010 02:04 PM
Please respond to
statalist@hsphsun2.harvard.edu


To
statalist@hsphsun2.harvard.edu
cc

Subject
Re: st: strata with single sampling units






" Would I be safe in assuming that one would use weight and PSU, and 
ignore
strata in this case? "

No. Do not ignore the stratum specification.  By placing certainty
units in their unique strata you ensure that they contribute nothing
to the between-PSU component of variance.  If the purpose of your
study is to describe the population, then the first stage fpc for the
certainty units should be 1.

Steve Samuels
sjsamuels@gmail.com
18 Cantine's Island
Saugerties NY 12477
USA
845-246-0774



> When I went back to the documentation it is stated that the study
> used a multistage stratified design in which primary sampling units
> (PSUs) were stratified according to certain sociodemographic criteria. 
The
> sampling
> frame for housing units is the Census 2000/2001 Supplementary Survey
> (C2SS) and that
> for group quarters is the Census 2000 Group Quarters Inventory. The C2SS
> sample of
> 655 PSUs was selected at the first stage, including 401 
self-representing
> (SR) and 254
> non-self-representing (NSR) PSUs. All SR PSUs were selected with
> certainty. For the
> NSR sample, two PSUs were selected per stratum, with probability
> proportional to the
> size of the estimated 1996 population of the stratum.  NOTE: Sample was
> included from each of the 655 PSUs.
> However, to prevent potential respondent disclosure, some PSUs were
> collapsed so that the final
> data file shows 435 PSUs, 305 being SR and 130 being NSR.
>
> So there are 305 strata with one PSU.  Final sample is 43,093.
>
> Would I be safe in assuming that one would use weight and PSU, and 
ignore
> strata in this case?  No other design variables are provided.  Doing
> things this way seems to replicate SEs in the published tables  -- at
> least the few I tried to replicate.

*
*   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/


*
*   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/