s-news
[Top] [All Lists]

kappa stuff - NON-SPLUS

To: "S-PLUS Newsgroup (E-mail) (E-post)" <s-news@lists.biostat.wustl.edu>
Subject: kappa stuff - NON-SPLUS
From: Jan Brogger <Jan.Brogger@med.uib.no>
Date: Mon, 27 Nov 2000 16:13:24 +0100
Reply-to: "Jan.Brogger@med.uib.no" <Jan.Brogger@med.uib.no>
Apologies for non-SPLUS related post.
>From:  SoSoSanta@aol.com [SMTP:SoSoSanta@aol.com]
>I'm using Kappa values to look at interrater reliability for a 
psychological
>tool and I am having a few problems.

>The tool uses yes/no questions and thus I have problems with Kappa scores 
not
>reflecting very good levels of agreement when many of the rating scores 
are
>Yes and there are few no's. Any suggestions on playing around with the
>values? I thought of keeping the yes/no ratio the same and just getting 
the
>total yes and no scores to be about 50% each of all the scores.This seems 
to
>get a more realistic Kappa value, but I'm not sure if it is allowed?

The dependence of Cohen's kappa on prevalence has been noted. There is no 
easy solution. There is too little data in a 2x2 table to estimate 
chance-correct agreement (kappa). So you have to make assumptions, which 
may be plausible depending on circumstances. If the assumptions are 
plausible, just use Cohen's kappa.

Two articles discuss kappa-type measures in general (1,2). IMHO, they're 
required reading for anybody using kappa. Article (3) summarizes some of 
the discussion about kappa and prevalence. There is a large body of work 
(4-6) about prevalence and kappa which are, IMHO, made significantly less 
interesting by articles 1 and 2:

Regarding "equalizing" prevalences:

We tried this (7) for purposes of comparing subgroups (smokers vs exsmokers 
vs neversmokers) but I'm not sure how valid it would be to equalize 
prevalences between different questions. Also, you run into sample size 
problems, especially for rare prevalences when the problem is most acute.

If it were possible, you could construct a scale out of your yes-no 
questions - then you'd have more degrees of freedom and could estimate more 
parameters.

Yours,
Jan Brogger
PhD student, Respiratory Epidemiology Group, University of Bergen, Norway

(1) Guggenmoos-Holzmann, I. (1996). "The meaning of kappa: probabilistic
concepts of reliability and validity revisited." J Clin Epidemiol 49(7):
775-782.
(2) Guggenmoos-Holzmann, I. and R. Vonk (1998). "Kappa-like indices of 
observer
agreement viewed from a latent class perspective." Stat Med 17(8 Apr 30):
797-812.
(3) Cook RJ.  Kappa. In:  Armitage P, Colton P, editors. Encyclopedia of 
biostatistics. John Wiley & Sons Ltd, Chichester; 1998. p. 2160-2169.
(4) Byrt T, Bishop J, Carlin JB.  Bias, prevalence and kappa. J Clin 
Epidemiol 1993; 46(5):423-429.
Feinstein AR, Cicchetti DV.  High agreement but low kappa: I. The problems 
of two paradoxes. J Clin Epidemiol 1990; 43(6):543-549.
(5) Cicchetti DV, Feinstein AR.  High agreement but low kappa: II. 
Resolving the paradoxes. J Clin Epidemiol 1990; 43(6):551-558.
(6) Lantz CA, Nebenzahl E.  Behavior and interpretation of the kappa 
statistic: resolution of the two paradoxes. J Clin Epidemiol 1996; 
49(4):431-434.
(7) Brogger J, Bakke PS, Gulsvik A.  Comparison of two respiratory symptoms 
questionnaries. Int J Tuberc Lung Dis 2000; 4(1):83-90.

<Prev in Thread] Current Thread [Next in Thread>
  • kappa stuff - NON-SPLUS, Jan Brogger <=