st: Survey analysis: Why does PSU affect proportions?

[Craig Anderson <clanders@uci.edu>]

I am using a published dataset (from CDC) and I want to verify that I  
can duplicate the published results before I begin my analysis.  I  
found that including clustering and stratification variables changed  
proportions, not only standard errors.  In the output that follows, I  
first used only the weights.  This gives results that agree with the  
published estimates of the proportions (produced using SUDAAN).  Next  
I used the PSU and STRATA, and get different proportions, but  
confidence intervals that agree with the published results, with  
minor variations in other runs.  I don't think the clustering by  
strata and psu should influence proportions, only their variance.

The dataset documentation says that a three-state cluster sample was  
used, but only the WEIGHT, PSU, and STRATA are included in the public  
dataset.  The FAQ (Survey-data analysis, first item) says that this  
sort of svyset statement will produce appropriate variance estimates  
for multistate designs.  Am I doing something wrong?

Craig L. Anderson, PhD (epidemiology)
Department of Emergency Medicine
University of California, Irvine



STATA output:

. clear

. set memory 16M
(16384k)

.
. use yrbs05

. svyset[pw=weight]

      pweight: weight
          VCE: linearized
     Strata 1: <one>
         SU 1: <observations>
        FPC 1: <zero>

. svy: tab q2 qn9, row se ci
(running tabulate on estimation sample)

Number of strata   =         1                  Number of obs       
=     13837
Number of PSUs     =     13837                  Population size    =  
13852.064
                                                Design df           
=     13836

-------------------------------------------------------
what is   |         never/rarely wore seat belt
your sex  |             1              2          Total
----------+--------------------------------------------
   female |         .0778          .9222              1
          |        (.004)         (.004)
          | [.0703,.0861]  [.9139,.9297]
          |
     male |          .125           .875              1
          |       (.0051)        (.0051)
          | [.1153,.1354]  [.8646,.8847]
          |
    Total |         .1017          .8983              1
          |       (.0033)        (.0033)
          | [.0954,.1083]  [.8917,.9046]
-------------------------------------------------------
  Key:  row proportions
        (linearized standard errors of row proportions)
        [95% confidence intervals for row proportions]

  Pearson:
    Uncorrected   chi2(1)         =   84.4318
    Design-based  F(1, 13836)     =   52.6547     P = 0.0000

. svyset psu [pw=weight], strata(stratum)

      pweight: weight
          VCE: linearized
     Strata 1: stratum
         SU 1: psu
        FPC 1: <zero>




. svy: tab q2 qn9, row se ci
(running tabulate on estimation sample)

Number of strata   =         4                  Number of obs       
=     13704
Number of PSUs     =        51                  Population size    =  
13730.762
                                                Design df           
=        47

-------------------------------------------------------
what is   |         never/rarely wore seat belt
your sex  |             1              2          Total
----------+--------------------------------------------
   female |         .0778          .9222              1
          |       (.0076)        (.0076)
          | [.0638,.0944]  [.9056,.9362]
          |
     male |         .1234          .8766              1
          |       (.0109)        (.0109)
          |  [.103,.1471]   [.8529,.897]
          |
    Total |         .1006          .8994              1
          |       (.0087)        (.0087)
          | [.0844,.1195]  [.8805,.9156]
-------------------------------------------------------
  Key:  row proportions
        (linearized standard errors of row proportions)
        [95% confidence intervals for row proportions]

  Pearson:
    Uncorrected   chi2(1)         =   78.7114
    Design-based  F(1, 47)        =   49.8319     P = 0.0000

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/