Date: Mon, 13 Oct 1997 08:32:56 -0400
Reply-To: Christianna Williams <christianna.williams@YALE.EDU>
Sender: "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From: Christianna Williams <christianna.williams@YALE.EDU>
Subject: Re: Effects of weighting...
Content-Type: text/plain; charset="us-ascii"
I feel like I should put in my $0.02 on this. You need to keep straight the
effect of weighting vs. the effect of a stratified sampling design. Both
have impacts on variance estimates. While the fix suggested for the
"we would run significance tests on unweighted data and report weighted
percentages, now someone is suggesting that we go ahead and run the sig
tests on weighted data, but each weight gets divided by the weighted
total. In this way our weighted sample is no bigger than the unweighted
sample, and sigs are not inflated"
will work if it is not a stratified/survey sample, it is probably not
adequate if there was a complex sampling design. Adjusting for the design
variables by just adding them to the model is not sufficient -- that doesn't
account for the fact of clustering within strata. That is why something
like SUDAAN is needed so that the standard errors take into account both the
sample weights and the sampling design.
Christianna S. Williams
Yale University Program on Aging
129 York Street, Suite 1N
New Haven, CT 06511
voice: (203) 764-9827
fax: (203) 764-9831