I agree, but would go further. One should work with whatever scale
is most nearly additive. For binary outcomes, for example, this is
likely to be the log-odds ratio scale. The trouble with many of these
other scales is that they only work if 1) the background risk is
homogenous in the population being studied and 2) is the same in
the population to which the study results are to be applied.
However, clinical trials never take a random sample of individuals
and it is often very diffcult to try and work out what population the
study might be a representative sample of. What are needed,
therefore, are "portable" measures of effect: measures that
translate reasonably well from studies on (of necessity) atypical
populations. In drug development this is done quite commonly for
example in bioequivalence studies, carried out on healthy
volunteers, whereby formulations are compared using the difference
in log-AUC. For some treatment it is possibly that the difference in
absolute AUC will be of greater clinical relevance but amongst
arguments for sticking with log AUC are that it is a measure that
will plausibly give similar results in patients and volunteers.
If necessary, additive measures can be translated into a clinically
relevant measure at the point of application given knowledge of the
background risk. There was a nice paper illustrating how this could
be done by Irwig and Glaziou in the BMJ a few years back.
Stephen Senn
Date sent: Thu, 20 May 1999 12:02:49 +0200
Organization: Psychiatric Hospital, University of Munich, Germany
Subject: Re: NNTs with asymmetric confidence intervals
From: "Rolf R. Engel" <[log in to unmask]>
To: [log in to unmask]
Copies to: [log in to unmask]
Send reply to: "Rolf R. Engel" <[log in to unmask]>
> This is a nice comment. Would it not be easier to work with ARRs from
> the beginning? Is NNT really simpler to transmit to clinicians as ARR? I
> doubt it.
>
> Rolf
>
> W Y Zhang schrieb:
> >
> > Yes, there is a way to make NNT and its CIs symmetric, that is to
> > convert the values back to their originals, ie., absolute risk
> > difference (ARR). Douglas Altman (BMJ 1998;317:1309-1312) suggested
> > to present NNT and CIs by relabelling ARR scale. From there you can
> > see a symmetric graph but asymmetric label just like log value
> > conversion.
> >
> > Weiya
> >
> > > When calculating the NNT for a treatment that is barely statistically
> > > significant and has a large NNT, the confidence interval is wide and very
> > > asymmetrical. For example in women with prior vertebral fractures and
> > > femoral neck BMD at least 2.1 standard deviations below young normal, the
> > > NNT with alendronate for 3 years to prevent one hip fracture =91 (95% CI
> > > 59-4545)
> > >
> > > I assume that if I were to able to calculate a mean NNT, it would be higher
> > > than 91, perhaps 200-300. Is their a formula for recalulating the NNT to
> > > adjust for this asymmetry? and is this a reasonable thing to do?
> > >
> > > Thanks
> > > Mick Braddick
> > > Clinical Epidemiologist
> > > Group Health Cooperative
> > > Seattle
> > > USA
> > >
>
>
> --
> ----------------------------------------------------------
> Prof. Dr. Rolf R. Engel ([log in to unmask])
> Psychiatric Hospital, University of Munich
> Nussbaumstr. 7, D-80336 München
> Tel ..49 89 51605561, Fax ..49 89 51605562
> ----------------------------------------------------------
--------------------------------------------------
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/
-------------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|