No, you're argument is not circular. You want to solve Bayes' theorem for
the prevalence (or pretest probability) using your known values for
sensitivity, specificity and PPV (or post test probability). You will
simply have to rearrange the equation, which is usually set up to solve for
the post test probability. The proportion of patients who recall is the
post test probability, which is known to you. That will be different in
the intervention group and the control group, and that difference will
result in different prevalence solutions for the two groups. The prevalence
or pretest probability is the proportion of doctors who asked the patients
about smoking. If this is not obvious to you now, contact me and I will
help you set up the equations.
I'm sure you are aware that the accuracy of this method depends on the
accuracy of your assessment of sensitivity and specificity. Your actual
primary outcome is the proportion of patients who recall. Knowing that
recall is not a perfect measurement of what the doctors asked, you are
attempting to adjust the recall to be more accurate using previously
ascertained sensitivity and specificity and Bayes' theorem. This then gives
you a secondary, adjusted outcome. There is some question as to the
importance of this secondary outcome. The important patient outcome would
seem to be more related to the recall outcome. If the doctor asked, but it
made so little impression on the patient that there is no recall, what good
is that. There is asking that is effective, and asking that is not. It
would seem to be the former that you are most interested in, and that is
given directly by the recall proportion.
David L. Doggett, Ph.D.
Senior Medical Research Analyst
Technology Assessment Group
ECRI, a non-profit health services research organization
5200 Butler Pike
Plymouth Meeting, PA 19462-1298, USA
Phone: +1 (610) 825-6000 ext.5509
Fax: +1(610) 834-1275
E-mail: [log in to unmask]
> -----Original Message-----
> From: owen dempsey [SMTP:[log in to unmask]]
> Sent: Monday, March 27, 2000 6:23 AM
> To: Evidence-Based-Health
> Cc: Tim Coleman; Richard Neal; Michael. [log in to unmask] wakeha. northy.
> nhs. uk; Dawood Dassu
> Subject: accuracy of test and confidence in trial outcomes
>
> dear all
> a problem is puzzling us
> we have a RCT design; to simplify a bit: the primary outcome of interest
> is in effect a test (patient recall of being asked about their smoking
> by their GP); data suggests a sensitivity of approx 90% and specificity
> 80%. The PPV (which is the necesaary link between the test and the
> outcome of interest) will therefore vary depending upon the prevalence
> of the target condition (whether a patient was asked about their smoking
> at the last consultation with a GP). The trial intervention is designed
> to affect this target condition (to increase the prevalence); therefore
> the PPV of the test will vary in a way that we cannot predict. (and
> whose only measure is the PPV; which becomes a circular argument))
> This seems to introduce another source of error affecting our ability to
> be confident about the effect of the intervention.
> Is this a significant problem or am I missing something?
> Have others had to deal with this; if so how?
> presumably this is a common trial problem when the outcome is measured
> via a test (less than 100% accurate)of any sort; has anything been
> written about this?
> TIA; you're my last hope
> O
> Owen Dempsey
> General Practitioner
> Senior Research Fellow
> Centre for Research in Primary Care
> Hallas Wing
> Nuffield Institute for Health
> 71-75 Clarendon Road
> University of Leeds
> Leeds LS2 9PL
> work: 01484-460298 (practice)
> work 0113-233-4835 (Leeds office)
> home: 01484-654794
> e-mail: [log in to unmask] (home)
> [log in to unmask] (Leeds) << File: Owen Dempsey.vcf >>
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|