Date:         Mon, 22 Oct 2007 13:11:44 -0700
Reply-To:     "Nordlund, Dan (DSHS/RDA)" <NordlDJ@DSHS.WA.GOV>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "Nordlund, Dan (DSHS/RDA)" <NordlDJ@DSHS.WA.GOV>
Subject:      Re: intraclass correlation with complex survey data
In-Reply-To:  <200710221927.l9MGNIw5005629@mailgw.cc.uga.edu>
Content-Type: text/plain; charset=iso-8859-1

> -----Original Message-----
> From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On
> Behalf Of Jennifer Rose
> Sent: Monday, October 22, 2007 12:27 PM
> To: SAS-L@LISTSERV.UGA.EDU
> Subject: intraclass correlation with complex survey data
>
> Hi All,
>
> I was wondering if anyone can tell me how I might obtain an intraclass
> correlation coefficient for clustered data that originate
> from a complex
> survey design. Specifically, I have a binary outcome, clustered data
> (children nested within families), 3 stratification variables
> and 1 sample
> weight variable. For clarification, I have used the following
> code to run a
> logistic regression that ignores the fact that the data are clustered.
>
> proc surveylogistic;
> stratum sw1 sw2 sw3 ;
> weight w1;
> by sample;
> class  new1(ref='0') site(ref='XX') /param=ref;
> model  new2(event ='1')= site new1 site*new1;
> run;
>
> The question now is whether or not I should be taking into account the
> clustering of the data, so I want to calculate an intraclass
> correlation
> coefficient (ICC) on my outcome variable (new2). Is there a
> SAS routine
> that can give me an ICC that takes into account the survey
> weights? If so,
> and my ICC warrants taking into account the clustering of the
> data, can I
> run an analysis in SAS (i.e., GEE) that takes into account the survey
> weights?
>
> Thanks in advance for your help,
> Jen Rose

Jen,

I suspect that David Cassell will post some very useful advice (if he is not out galavanting around :-).  But let me begin by suggesting that since you have survey data from a clustered design, you should be using the SAS survey procs to analyze the data (accounting for the clustering and any other design effects).  It seems to me that there really isn't any reason to look at intra-class correlations to decide if you should take clustering into account. The answer is that you should.

Is there some reason you don't want to use the survey procs to analyze your data?

HTCT (since I am pretending to be David),

Dan

Daniel J. Nordlund
Research and Data Analysis
Washington State Department of Social and Health Services
Olympia, WA  98504-5204