Fw: Weights -> estimate with relative standard error

[David Spaugh]

For the mailing list:

 

Yesterday morning I forgot to use reply-all when responding to Ben's email.  Therefore the reply did not reach the list. 

 

Although the issue is now closed, I am forwarding yesterday's email to the list in case other readers were interested in the correspondence. 

 

Ben has already replied that CSPLAN ANALYSIS & CSTABULATE commands do not exist. 

 

In closing - PSPP is great software with ample features; imminently useful.  Thanks to Ben and to the community in general for supporting endeavors such as these. 

 

DS

 

 

----- Original Message -----

From: David Spaugh

To: Ben Pfaff

Sent: Thursday, May 19, 2011 8:07 AM

Subject: Re: Weights -> estimate with relative standard error

Thanks to Marshall for recognition of the specific issue and John for the evaluation script (which I will study today). 

 

In reply to Ben's query, here is some relevant info from the NAMCS documentation:

 

2008 NAMCS is 28,741 visit-records obtained by multistage probability design. 1st stage is selection of 112 primary sampling units (PSU's, ie counties, or cities, etc).  Typically, the second stage consists of a probability sample of practicing physicians from the PSU's.  Consenting physicians complete detailed surveys for patient's seen during their "survey week".  CDC compiles the data. 

 

Estimation procedure included: Inflation of reciprocals by sampling probabilities, Adj for non-response, ratio adj, and weight smoothing. 

 

Each record is assigned an inflation factor called the "patient visit weight." By aggregating the PVW's on the 28,741 records, the user can obtain the estimated total of 955,968,818 office visits made in the United States in 2008.  Aggregation of weights from a segregated class of records yields estimated national incidence rates for visits accompanied by the class characteristic. 

 

*Up to this point, use of the data is easy for a novice:  Open the file with PSPP, export as CSV, open with Excel, segregate the class, then aggregate the class weights. 

 

*The "problem" that arises: the reported figure needs to be accompanied by an estimated "variance" or RSE, otherwise the audience has no means of assessing the quality of the reported figure.  Continuing from the doc file:

 

The National Center for Health Statistics considers an estimate to be reliable if it has a relative standard error of 30 percent or less.  It should be noted that estimates based on fewer than 30 records are also considered unreliable.

 

To estimate variance...two sample design variables (CSTRATM and CPSUM) are provided for use with software utilizing a "between PSU" variance estimator or "ultimate cluster" sampling design. 

 

Examples using CSTRATM and CPSUM in Stata, SUDAAN’s 1-stage WR (with replacement) design option, and SAS’s PROC surveymeans applications are presented below:

Stata

The pweight (PATWT), strata (CSTRATM), and PSU (CPSUM) are set with the svyset command as follows:

Stata 8: svyset [pweight=patwt], psu(cpsum) strata(cstratm)

Stata 9 and Stata 10: svyset cpsum [pweight=patwt], strata(cstratm)

SUDAAN 1-stage WR Option:

The program below provides a with replacement ultimate cluster (1-stage) estimate of standard errors for a cross-tabulation.

PROC CROSSTAB DATA="" DESIGN=WR FILETYPE=SAS; NEST CSTRATM CPSUM/MISSUNIT;

SAS - PROC SURVEYMEANS

PROC SURVEYMEANS DATA="" CLUSTER CPSUM; STRATA CSTRATM;

SPSS

This code pertains to SPSS Inc.’s Complex Samples 12.0 module. It would be used with the "Analysis Preparation Wizard" component of that module. The PLAN FILE statement would be invoked in statistical runs, as in the example for CSTABULATE shown below.

CSPLAN ANALYSIS

/PLAN FILE=’DIRECTORY\PLANNAME.CSAPLAN’

/PLAN VARS ANALYSISWEIGHT=PATWT

/PRINT PLAN

/DESIGN STAGELABEL= ‘ANY LABEL’ STRATA=CSTRATM CLUSTER=CPSUM

/ESTIMATOR TYPE=WR.

CSTABULATE

/PLAN FILE=’DIRECTORY\PLANNAME.CSAPLAN’

/TABLES VARIABLES = var1 var2

/CELLS POPSIZE

/STATISTICS SE

/MISSING SCOPE = TABLE CLASSMISSING = EXCLUDE.

***

Thanks!

DS

----- Original Message -----

From: "Ben Pfaff" <address@hidden>

To: "David Spaugh" <address@hidden>

Cc: <address@hidden>

Sent: Wednesday, May 18, 2011 9:01 PM

Subject: Re: Weights -> estimate with relative standard error

> "David Spaugh" <address@hidden> writes:
>
>> Apparently SPSS will do this with a script partly provided on page 89 of the
>> NAMCS file documentation.  They mention "SPSS Complex Samples 12.0 Module".
>
> Can you pass along this script?  Then we can talk about whether
> PSPP implements all of the necessary features.
> --
> Ben Pfaff
> http://benpfaff.org