Hi, folks:
Three questions about cox.zph()
a. The documentation says that the default time transform for this function
is 'km'. Indeed, when I look at the definition of cox.zph, it does set a
default of 'km' for transform. But whenever I invoke the function on a cox
object without specifying the transform, the result indicates that the
transform is 'identity'. Indeed, if I then plot this result, it is clear
that the 'identity' transform is being used. I only get 'km' if I specify
'km' explicitly. What's going on?
b. The documentation in S-Plus for cox.zph indicates that it returns as 'y'
the matrix of cumulated scaled Schoenfeld residuals for the specific
covariates under examination. But when one plots the cox.zph object, it is
clear that, with a significant covariate, the residuals don't sum to zero.
In Therneau's book, it appears that what is being returned by cox.zph and
and plotted by plot.cox.zph is actually the sum of the overall value of the
beta plus the cumulative scaled Schoenfeld residuals. This makes much more
sense to me. But even the book is a bit unclear on this point. Am I
correct that it is the cumulated residuals PLUS the overall beta that is
being returned and plotted?
c. Therneau's book suggests that to test for a slope to the cox.zph plot
one use the 'km' transform of the time scale so that the observations are
well spread out over the x axis. This seems sensible to me. Is it the
general consensus ot the community that one should (at least) test with the
'km' transforn, irrespective of what one chooses to plot?
Many thanks in advance for the wisdom you all can provide on the above
questions.
Larry Hunsicker
|