Re: Are graded clinical signs more reliable than dichotomized?

John Uebersax wrote:
Roland wrote:


Instead of dichotomising these clinical signs I think that a grading
of the intensity of the sign must retain more of the diagnostic
information


Correct


One problem is of course that there is no definition for the grades of
the variable. However the dichotomised variable has the same problem


Correct


and I assume that the agreement between two examiners will be larger
if graded variables are used instead of dichotomised.


No. There will be *more* agreement with *fewer* categories. The fewer
the
categories, the less opportunity for disagreement.

That is technically correct but when you multiple the probability of error times the magnitude of the error, the problem is worse for a binary variable.


I tend to agree with those who suggest that having more categories is
better
from a *research* or *statistical* standpoint. But clinical practice
evolves
to optimize several factors. Some are optimized by having more
categories,
some by fewer. There is a tradeoff, and the solution to the tradeoff
is
different for each application.

Generally speaking, having more categories improves every aspect of prediction and decision making, and using more categories will also take better account of residual confounding, making some predictors less necessary. Sometimes a predictor is important because it is a substitute for the information lost by cutting another predictor.


I believe there is such a tradeoff between
reliability/agreement/reproducibility (better with fewer levels) and
accuracy/precision (better with more levels).

Don't see this when the cost or magnitude of an error is taken into account.


The solution to this tradeoff is different for each application. That
is why we see staging in some areas, but yes/no distinctions in others.

One example where dichotomous ratings might be better is a screening
test. It is logistically better to screen patients into two groups.
Then a more refined test can be given to those who screen positive.

Even then there are advantages to more continuous scales, because what you do depends on the degree of positivity, and areas with more resources will be able to treat more subjects (thereby using in effect a different cutpoint).

Frank Harrell


Hope this helps.
--
John Uebersax PhD

.