The problem posed was the following:
> fit <- coxph(....
> zp <- cox.zph(fit)
> plot(zp[1])
Error in plot.window(xlim, ylim, log, asp, ...) :
need finite 'ylim' values
There were 21 covariates, but over 7000 events, so the Cox model fits ok.
After sending me the data, the problem, which I haven't seen before, is thus:
In order to put a smooth curve onto the plot, cox.zph calls lm essentially
thus
lm( resid ~ ns(x, df=4)
where x is the set of death times, and resid is a scaled Schoenfeld residual.
For the data set in question x is of length 7109, and the lm call fails due to
a
singular fit! This leads to a missing value and then eventually to the error
message reported above. Further exploration reveals that 4022 of the x values
are log(.01), with the other 3087 spread out from 0 to log(365) in an
approximately Gaussian shape. The residuals are in two clumps from -4 to 0 and
2 to 5, each clump dense, the histogram has two unimodal humps.
I don't see why lm is failing.
I have fixed plot.cox.zph so that it now prints an error message
"Spline fit is singular, try a smaller degrees of freedom"
when the fit is singular. This will at least be less confusing. For the
data in hand
plot(zp[1], df=2)
works fine.
The repair will migrate to the R code in due course. (I am currently working
on
some other updates to the survival package).
Terry Therneau
______________________________________________
[email protected] mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
