------- Forwarded Message Follows -------
w.y.zhang Wrote:
>
>Allow me add a bit more salts to Prof Senn's comments. To my
>understanding relative risk (RR) and odd ratio (OR) are different.
>The latter is only an approximation for the former when the event
>rate is low (Altman DG BMJ1988;317:1318).
True, but you misunderstand the implication of this. I could just as
well say that the median is only an approximation to the mean if the
data are not too skewed. But this is not a reason for saying that you
should only ever use the median if the data are not skewed. You could
(and often would) argue the reverse: you should not use the mean if the
data are very skewed. The correct way to describe the relationship is to
say that the relative risk is only an acceptable approximation to the
odds ratio if the event rate is small.
>We use OR in case-control
>study because it can not produce event rate. In addition,
>case-control study is normally undertaken in the conditions with low
>event rate to save the budget and overcome the long term consuming
>(if the event is rare, say 1/10,000, it is difficult to prospectively
>observe this event during the short time period) . However, many
>meta-analyses based on prospective studies such as RCTs with high
>event outcomes, for example, efficacy of drug therapy, used OR
>instead of RR.
True. And with good reason. This is because NO random sampling is
involved in clinical trials. (Randomisation is involved but that is
about making sure that patients are comparable between groups not
representative.) As such the base rate in the population is NOT
estimable. From this point of view exactly the same problem arises as
with case control studies. It is only by falsely treating the trial as
representative that an "estimate" is produce. Clinical trials, as is the
case with all experiments, are about comparisons: what are needed are
reliable comparative measures. The odds ratio fits the bill.
>
>OR has some advantages but again they should not be overused.
>
>1.It can always take values between zero to infinity, which is not the
>case for RR. For example, if the baseline risk is bigger than 50%, it
>is impossible to double it with RR but with OR. This supply a
>mathematical superiority for OR in variety of conditions;
Exactly
>
>
>2.In addition, the existing multiple regression analysis such as
>logistic regression models to analyse association between event rate
>and risk factors actually work in terms of odds and report effects as
>odds ratio.
>
>Except for these, RR should be the better choice, particularly for
>RCTs with high event rate outcome.
This is just crazy. Your arguments should have led you to the opposite
conclusion. The only circumstance under which the RR is acceptable is
when the background risk is small and then as an approximation to the
OR.
Stephen
--
Stephen Senn
--------------------------------------------------
Professor Stephen Senn
Department of Statistical Science &
Department of Epidemiology and Public Health
University College London
Room 316, 1-19 Torrington Place
LONDON WC1E 6BT
Tel: +44 (0) 171 391 1698
Fax: +44 (0) 171 813 0280
Email: [log in to unmask]
webpage: http://www.ucl.ac.uk/~ucaksjs/
-------------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|