On Fri, 12 Jan 2001 12:09:38 +0100, Tor A Strand
<[EMAIL PROTECTED]> wrote:
> Logistic regression and Relative risk (RR)
>
> In some statistical applications (eg. SAS) RR is an option in logistic
> regression. This is rarely used in Scientific literature where some authors
> even report odds ratios (OR) as RR. This is seen even in recent papers in
> the New England Journal of Medicine.
>
> Is there any interpretation difficulties when logistic regression models
> report RR in stead of OR and why is this option seldom used.
Yes, difficulties. For example.
Risk A=90%, Risk B= 99%; implies RR= 90% (not impressive, if no one
shows you the raw numbers).
However, the OR is 0.1 or 10.0 (impressive).
When the Risk is small, the RR and OR are about the same,
and the same math might be applied for tests.
There is not any easy math to apply with a large Risk;
it is horrible to try to figure with, for the same essential reason
that you can't tell whether 90% is impressive or not:
you have to deal with the rate, which is built-in for *some* useful
effect for the OR. (Unlike using the Pearson r for an effect-size,
even the OR is still a little awkward for translating to a test.)
An expedient for the RR is to compute tests or CIs with the OR and
adapt them (see: Mantel-Haenszel test across levels of a factor).
Here's a guess. A paper that computes the OR and mis-calls it a RR
is better paper than one (since about 1980) that uses an actual RR.
[ ... snip, Q about RR in STATA]
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
