[R] survSplit

[Thomas Lumley]

On Mon, 8 Nov 2004, Danardono wrote:

> I am just realized that  survival has the facility to do survival time 
> splitting survSplit
> after read some postings about  time dependency  in the list.
> Is it survSplit only for the survival data input (time,status)  and not for 
> the 'counting process' input (start,stop,status)?
>
> I take one example modified from the survSplit help:
>> data(aml)
>> aml3<-survSplit(aml,cut=c(5,10,50),end="time",start="start",event="status",episode="i",id="id")
>
>> coxph(Surv(time,status)~x,data=aml)
>
>               coef exp(coef) se(coef)    z     p
> xNonmaintained 0.916       2.5    0.512 1.79 0.074
>
> Likelihood ratio test=3.38  on 1 df, p=0.0658  n= 23
>
> #the same
>> coxph(Surv(time,status)~x,data=aml3)
>
>               coef exp(coef) se(coef)    z     p
> xNonmaintained 0.916       2.5    0.512 1.79 0.074
>
> Likelihood ratio test=3.38  on 1 df, p=0.0658  n= 63

This should NOT be the same, and on my computer is not.  You should have 
to use the counting-process syntax.  I get

> coxph(Surv(time,status)~x,data=aml3)
Call:
coxph(formula = Surv(time, status) ~ x, data = aml3)


                coef exp(coef) se(coef)    z     p
xNonmaintained 1.07      2.92    0.514 2.08 0.037

Likelihood ratio test=4.61  on 1 df, p=0.0317  n= 63
> coxph(Surv(start,time,status)~x,data=aml3)
Call:
coxph(formula = Surv(start, time, status) ~ x, data = aml3)


                 coef exp(coef) se(coef)    z     p
xNonmaintained 0.916       2.5    0.512 1.79 0.074

Likelihood ratio test=3.38  on 1 df, p=0.0658  n= 63



> BUT If I split aml3 further:
>
>> aml4<-survSplit(aml3,cut=c(9,12,40),end="time",start="start",event="status",episode="i",id="id2")
> #not the same!
>> coxph(Surv(start,time,status)~x,data=aml4)
>              coef exp(coef) se(coef)    z     p
> xNonmaintained 1.05      2.85    0.515 2.03 0.042
>
> Likelihood ratio test=4.38  on 1 df, p=0.0363  n= 105
>

Hmm.  I suspect this is a <= vs < bug of some sort in handling start 
times.

 	-thomas