I'm starting to write up some web pages about diagnostic tests, likelihood
ratios, Fagan nomogram, etc.
One of the examples that I want to discuss is sinusitis. There's a nice
write-up in
Williams JW Jr; Simel DL "Does this patient have sinusitis? Diagnosing acute
sinusitis by history and physical examination" JAMA, Sep 8 1993, 270(10)
p1242-6
But I don't have a good feel for what the pre-test odds would be (a) during
an epidemic (there is a lot this stuff going around); or (b) during other
times. Also, is sinusitis seasonal?
Can anyone give me a rough idea of what reasonable pre-test probabilities or
odds would be for a condition like sinusitis under epidemic or non-epidemic
conditions and for on-season/off-season (if that makes sense)?
I also wrote up a comment about pre-test odds
>To use this tool, you need to provide your best estimate of
>the probability of the disease prior to testing. This is
>usually related to the prevalance of the disease, though
>this may be modified up or down on the basis of certain
>risk factors that are present in your patient pool or
>possibly in this particular patient.
Does this seem reasonable? Can anyone help me elaborate a bit on this point?
Finally, I have a first draft of material in my Ask Professor Mean column
about the "Fagan nomogram" and "odds". I hope to add columns on "likelihood
ratio" and "diagnostic test" soon. If you have time to look at these and
give me some feedback, I would greatly appreciate it.
Thanks for everyone's help so far.
Steve Simon, [log in to unmask], Standard Disclaimer.
Ask Professor Mean: http://www.cmh.edu/stats/profmean.htm
-----Original Message-----
From: Simon, Steve, PhD
Sent: Friday, November 20, 1998 2:32 PM
To: 'Dr. Samuel Wiebe'; [log in to unmask]
Subject: RE: likelihood ratios
Sam Wiebe writes:
>I would appreciate comments on how to deal with empty cells
>in 2 by n tables when calculating Likelihood ratios. For
>example, adding 0.5 or 0.1 to each cell produces enormously
>different likelihood ratios. Thanks in advance.
I'm not an expert in likelihood ratios, but since nobody else answered.
An empty cell means that your likelihood ratio is either zero or infinity.
If that were really true, you would have a pretty darn good test. But you
and I both know that the true likelihood ratio is never going to be that
extreme. The empty cell is caused by sampling error, and ideally, you would
account for that sampling error by using confidence limits for your
likelihood ratio. In fact, you should use confidence intervals even when you
don't have an empty cell, but that's another story.
There are some formulas that you could use yourself for confidence limits
for a likelihood ratio, but they don't work when you have empty cells. The
methods that do work require sophisticated statistical software.
The only practical suggestion I would have is to pool adjacent categories so
that you have enough data to estimate a stable likelihood ratio.
Another possibility is to decide in advance what sort of likelihood ratio
you want to have and then add whatever constant gives you that desired
likelihood ratio. <grin>
Steve Simon, [log in to unmask], Standard Disclaimer.
Ask Professor Mean: http://www.cmh.edu/stats/profmean.htm
<http://www.cmh.edu/stats/profmean.htm>
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|