Date:         Fri, 8 Oct 2010 09:43:45 -0400
Reply-To:     "Simon, Lorna" <Lorna.Simon@UMASSMED.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "Simon, Lorna" <Lorna.Simon@UMASSMED.EDU>
Subject:      FW: kappa vs. interclass correlation coefficient
Content-Type: text/plain; charset=us-ascii

Received this answer to my question and it wasn't posted to the list, so I thought I'd forward it along.

-----Original Message-----
From: Zack, Matthew M. (CDC/ONDIEH/NCCDPHP) [mailto:mmz1@cdc.gov]
Sent: Friday, October 08, 2010 7:57 AM
To: Simon, Lorna
Subject: RE: kappa vs. interclass correlation coefficient

Kappa can be used for estimating agreement corrected for chance for
three or more categories.
The problem with kappa when used for multiple categories or continuously
valued variables
split at arbitrary cutpoints is that kappa decreases with an increase in
the number of
categories and that kappa will differ depending on the arbitrary
cutpoints.

One informative article on this topic is the following:

  Maclure M, Willett WC.  Misinterpretation and misuse of the kappa
statistic.
   American Journal of Epidemiology 1987 Aug;126(2):161-169.

Matthew Zack

P.S.:  The intraclass correlation coefficient, not the interclass
correlation coefficient,
is a more relevant statistic to compare the kappa statistic to.

-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of
Simon, Lorna
Sent: Thursday, October 07, 2010 10:24 AM
To: SAS-L@LISTSERV.UGA.EDU
Subject: kappa vs. interclass correlation coefficient

A colleague of mine told me that kappa could not be used for
non-dichotomous variables and that the interclass correlation
coefficient should be used instead.  A quick web search failed to
support this. I'd like the opinion of some of the statisticians in the
group on this. thanks.