Dear List,
I have been thinking about diagnostic tests, and seem to have run into
a problem. My apologies for what might seem like an overly simple
approach to what is mostly common knowledge, but I am worried by
what I have found, and wanted to make sure it wasn’t because of
something stupid.
My understanding of the current approach, advocated by Sackett et al.,
is to generate a pre-test (PreT) probability, convert it to a PreT odds
ratio, multiply by the Liklihod ratio to generate the PostT odds, and
convert these back to PostT probabilities. This is something I am fairly
familiar with, and have done many times.
My problem came when I started to look at the 95% CI’s around the
PreT prob and the LR. Both are (or should) be quoted as a fiqure, ±
limits. If you run the test once, then these limits are statistical
expressions of precision. But if we look at a populations basis (say,
all those with chest pain who then have an ECG) then the distribution
describes what range of values we might find each time we do the
test, and how often each occurs.
Using a very basic Monte Carlo approach, I have simulated a pretest
population, and applied an LR (with a confidence interval) to it, to
generate a post-test prob. Population. Does this sound reasonable?
Has anyone else done this? What am I doing wrong?
All help much appreciated,
Matt Williams (MRCP(UK))
Liverpool, England
|