Perhaps the answer depends on the context
of the research question or clinical question.
Is the test intended to be used as a
continuous variable or in a dichotomous approach?
But another consideration is what makes a
test “best”. In some clinical scenarios sensitivity is most
important; in others specificity is most important. A combined measure (ROC,
diagnostic odds ratio, diagnostic accuracy) may not be the best measure for a
specific use.
Diagnostic research may have a life cycle,
something like:
--hypothesis testing to find candidate
diagnostic approaches (specific tests, optimal cutoffs, etc.)
--study of clinical applications
--validation studies
Someone else on this list can probably
describe a more robust life cycle.
The outcomes of greatest interest and the
corresponding statistical approach may vary along the way.
Brian S. Alper, MD, MSPH
Editor-in-Chief, DynaMed
(www.ebscohost.com/dynamed)
From: Evidence based
health (EBH) [mailto:[log in to unmask]] On Behalf Of Roger Keller
Sent: Thursday, January 21, 2010
6:43 AM
To:
[log in to unmask]
Subject: ROC curves
Dear all,
I am comparing the performance of two diagnostic test scores. My
problem is that the curves cross each other.
If I compare the area under the curves using the whole scores (ten
points), then test A has a higher area under the curve than test B.
However, if I dichotomise them in their best cut-off points,
then test B has a better performance (SE+SP) than test A.
Does anyone has any advice on how to choose the best test? Consider
that they are similar in terms of reliability, ease of application, etc.
Thanks a lot.
________________________________
Roger Keller Celeste, CD, MSc, PhD
Departamento de Epidemiologia - IMS/UERJ
Rua São Francisco Xavier, 524 7o andar sala 7009.D
P Think before you print.
Veja quais são os assuntos do momento no Yahoo! + Buscados: Top
10 - Celebridades
- Música
- Esportes