[R] Actual vs. predicted survival times

Terry Therneau therneau at mayo.edu
Wed Feb 8 14:57:26 CET 2012


The Cox model predicts two things: the relative hazard (death rate)
associated with each variable, and a predicted survival curve for any
particular variable combination.

   The predicted survival curve will look like a Kaplan-Meier curve:
multiple small steps, and will only rarely go all the way to zero.
There is not a natural way to give "a predicted survival time" for each
subject.  There isn't any program to do what you ask because there isn't
a good answer to "what" it should compute.  This troublesome fact has
led to a lot of work to define some residuals for a Cox model in other
ways: see chapters 4-7 of the book by Therneau and Grambsh for a partial
summary of ideas.  (A little too much to summarize in an email).

Terry Therneau

---- begin included message ----
I would like to produce a plot like the attached - although simplified
to actual vs. Predicted survival time with distinguishing marks for
censored and observed points.  I have a dataset and have fitted a Cox
model to it.  In an attempt to visualise how accurate the model is it
would be ideal if I could plot the actual survival times against the
predicted survival times.

I have been looking on the internet to see if there are ways to do this
in R.  The only post I found
(https://stat.ethz.ch/pipermail/r-help/2009-February/189888.html) that
seemed directly relevant suggested that I shouldn't be generating
survival times at all.  Given that, I was concerned about proceeding but
I would like to have access to a plot to make a decision on its
usefulness.
 ....



More information about the R-help mailing list