Dear All,
In prognostic studies with censored endpoint, before using the Cox
regression model, we generally need to encode continuous characteristics.
Some people use a procedure which consists in modelling the effect of the
characteristic on the hazard of failure using nonparametric spline functions
(Hastie & Tibshirani, 1995). Then, they use this plot to choose a convenient
encoding of the characteristic (separation in class, left continuous...).
I have a trouble with this procedure because I think that using procedures
that rely on the prognostic to encode characteristics (like optimized
cut-off point method) are not reproducible and can thus induce a bias.
What is your opinion about that ? Is this method a reliable way to encode
characteristics? Have you any other suggestion?
Thank you in advance.
Cedric
______________________________________________
CEDRIC MAHE
Service de Biostatistique et d'Epidemiologie
Institut Gustave-Roussy
39 rue Camilles Desmoulins
94805 Villejuif
France
tel : 33 1 42 11 50 94
fax : 33 1 42 11 52 58
e-mail : [log in to unmask]
______________________________________________
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|