Via: "Stephen M. Perle, D.C." <[log in to unmask]> 06/30/99 02:08PM
>>>
Regarding what:
Allan Meyers wrote:
>
> A colleague has asked me to ask:
>
> I am not aware of a correction/adjustment for Kappa for low prevalence
> items, and
> would be interested in this method and reference. However, intuitively, it
> would
> seem that a high prevalence item would use the same methodology, since Kappa is
> based on a k by k table, and the calculations do not depend on which column/row
> (marginals) were high/low or how they are labeled. If there are different
> methods
> posed to adjust/correct for Kappa statistics in high prevalence and low
> prevalence
> scenarios, I would like to review these for use in our current data.
> Currently, I
> resort to showing "percent agreement" in addition to Kappa when these situations
> arise so that readers can interpret the biased Kappa.
>
> Many thanks.
>
> *************************************
>
> Allan R. Meyers, Ph.D., Professor
> Department of Health Services
> Boston University Schools of Medicine and Public Health
> Research Director,
> New England Regional Spinal Cord Injury Center
> 715 Albany Street, TW-349
> Boston, MA 02118, USA
> Telephone: 617 638 4510
> FAX: 617 638 5374
>
Dr Meyers;
I appreciated your dilemma regarding low/high prevalence. I address
that in an article
published in J Clin Epidemiol 1996; 24(4):431-434. Since, for a 2x2
matrix, (and probably for
kxk matrices as well) the equations are symmetrical, there is,
basically, no difference between
low and high prevalence. A low prevalence of a condition can be seen as
a high prevalence of
normals. There is a correction factor I just learned about from a
communication from a Dr.
Sestini; it is an article by:
Agresti, A. Ghosh, G.Bini, M.
> Raking Kappa: Describing Potential Impact of Marginal
> Distributions on Measures of Agreement.
> Biometrical journal. 1995 v 37 n 7 p.811
I have ordered the article and await its arrival.
Using Po is important...in fact, as I argue in my article cited above,
one cannot appropriately
interpret K without a knowledge of Po, and researchers often fail to
publish Po with K results.
In fact, we recommend that every publication on K provide three critical
values: Po, P++ and
S(D), then the entire contingency table can be reproduced and all
parameters computed, should
the reader so desire. Po, in fact, completely determines the three
characteristic K values that
accompany every 2x2 K calculation, Kmax, Kmin & Knor, which provide a
range and "ideal"
K against which to compare Ko (observed K). I am excited to learn that
someone has finally
worked out the algebra for providing a correction for prevalence in
ordinary experimental
situations.
This, of course, raises other concerns regarding the determination of
prevalence, versus knowing
what the underlying prevalence is by independent estimates, and how
those two issues get
resolved for any given experimental setting.
Hope all this helps.
Please send me your e-mail address, as I am not a subscriber to the chat
system you use for K
discussions.
sincerely
Charles A. Lantz, DC, PhD
Director of Research
Life Chirorpactic College West
San Lorenzo CA 94580
[log in to unmask]
510-889-1700x439
--
_____________________________________________________________________
Stephen M. Perle, D.C. "A man who knows that
Assistant Professor of Clinical Sciences he is a fool is not
University of Bridgeport College of Chiropractic a great fool."
Bridgeport, CT 06601 Chuang Tzu
E-mail: [log in to unmask]
http://www.bridgeport.edu/chiro/
_____________________________________________________________________
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|