[R] Schoenfeld residuals

[Laura Bonnett]

Hi,

Thank you for your comments and apologies for the delay in replying.
rem.Rcens =1 for the censored variables.  The problem arises because I
am not strictly looking at time to death.  Instead I am looking at
time to 12-month remission in epilepsy.  Therefore a lot of people
have the same event i.e. they successfully achieve 12-month remission
from day 1 of the treatment.

I think I shall avoid the problem by 'excluding' patients with
immediate 12-month remission i.e. I will look at patients with
immediate success in a separate analysis to patients with delayed
success.

Thanks for your help,

Laura


2009/4/6 Terry Therneau <therneau at mayo.edu>:
> Laura Bonnett was kind enough to send me a copy of the data that caused the
> plotting error, since it was an error I had not seen before.
>
> 1. The latest version of survival gives a nicer error message:
>
>> fit <- coxph(Surv(rem.Remtime, rem.Rcens) ~ all.sex, nearma)
>> cfit <- cox.zph(fit)
>> plot(cfit)
> Error in plot.cox.zph(cfit) :
>   Spline fit is singular, try a smaller degrees of freedom
>
> 2. What's the problem?
>  There are 1085 events in the data set (rem.Rcens==1), and of these 502 are
> tied events on exactly day 365.  The plot.cox.zph function tries to fit a
> smoothing spline to the data to help the eye; the fit gives weight 1 to each
> death and having this high a proportion of ties creates problems for the
> underlying regression.
>
> 3.
>> plot(cfit, df=2)
>  Warning messages:
> 1: In approx(xx, xtime, seq(min(xx), max(xx), length = 17)[2 * (1:8)]) :
>  collapsing to unique 'x' values
> 2: In approx(xtime, xx, temp) : collapsing to unique 'x' values
>
>  These warning messages are ignorable.  I'll work on making them go away.
>
>
> 4. A shot in the dark -- is perchance the variable rem.Rcens=1 a marker of a
> censored observation, and the events are 0?  (A whole lot of events at 1 year is
> suspicious, but half censored at one year is believable.) Then the proper coxph
> code is
>
>> fit2 <-  coxph(Surv(rem.Remtime, rem.Rcens==0) ~ all.sex, nearma)
>
>        Terry Therneau
>
>
>
>