Statalist


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

Re: st: multiple weights per person in GEE?


From   "Ariel Linden" <ariel.linden@gmail.com>
To   <statalist@hsphsun2.harvard.edu>
Subject   Re: st: multiple weights per person in GEE?
Date   Fri, 24 Jul 2009 07:20:34 -0700

Good catch Steve! Thank you and Joseph, I think the descrepancy between
Stata's GLM and SAS GEE using weights has been solved.
 
The results of the SAS GEE seem to match that of Stata's GLM, when you look
at the empirical SE's. Here are the output tables for Stata GLM and SAS
GENMOD (both using the IPTW weights and independence correlation matrix and
cluster(ID):
 
*** Stata Output***

. glm rx sccenrolled [pweight = iptw], family(gaussian) link(identity)
vce(cluster personnumber)
          
                        (Std. Err. adjusted for 7868 clusters in
personnumber)	
-------------------------------------------------------------

           		Robust	 		
          rx |      Coef.   Std. Err.      z    P>|z|     [95% Conf.
Interval]	
-------------+--------------------------------------------------------------
--	 	
 TX |   2.476136   .4841721     5.11   0.000     1.527176    3.425096	
       _cons |   1.145776   .0553768    20.69   0.000      1.03724
1.254313	
---------------------------------------------------------------



****SAS Output****

GEE Model Information	
Correlation Structure	 Independent	
Subject Effect	 PERSONNU (7868 levels)	
Number of Clusters	 7868	
Clusters With Missing Values	 1	
Correlation Matrix Dimension	 24	
Maximum Cluster Size	 24	
Minimum Cluster Size	 23	

 

Analysis Of GEE Parameter Estimates	
Empirical Standard Error Estimates	
Parameter	 Estimate	 Standard Error	 95% Confidence Limits	 Z
Pr > |Z|	
Intercept	 1.1458	 0.0554	 1.0372	 1.2543	 20.69	 <.0001	
TX	 2.4761	 0.4841	 1.5272	 3.4250	 5.11	 <.0001	

 
 
 
Date: Thu, 23 Jul 2009 11:55:54 -0400

From: sjsamuels@gmail.com

Subject: Re: st: multiple weights per person in GEE?

Because the SAS output shows the SCALE parameter, I think that Ariel

reported the model-based, not the empirical, standard errors. If I

recall, model-based standard errors ignore the clustering (REPEATED

statement) and so, are just OLS. The ratio of the total sample size

to the number of clusters is about 25:1, with a square root of 5:1.

This would account for some, but not all, of the standard-error

discrepancies. Ariel, please show the second part of the GENMOD

results.

- -Steve


*
*   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/



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