Thanks for your thoughts.
1. The function is 'somewhat' linear. Small curvature but beginning and
end are similar. Off course I could fit more complex functions of time.
2. Do I really have to build (start, stop) datasets as I have time-varying
covariate effect but not time-varying covariate? Covariate seems to have an
effect to end point initially but starts to diminish after sometime.
3. Does this point hold in case of time-varying covariate effect also? I try to
get my hands on the book you mentioned
I have a competing risk setting where there is a competing risk of death. I
carry deaths forward to the end of followup time as we know the potential
followup time for every individual (subdistribution hazard regression).
A Proportional Hazards Model for the Subdistribution of a Competing Risk, Jason
P. Fine and Robert J. Gray Journal of the American Statistical Association Vol.
94, No. 446 (Jun., 1999), pp. 496-509
But I guess that does not have an effect on the matter at hand.
Mitja
On 8.10.2012, at 16.25, Terry Therneau wrote:
Dear All,
I have built a survival cox-model, which includes a covariate * time
interaction. (non-proportionality detected)
I am now wondering how could I most easily get survival predictions from my
model.
My model was specified:
coxph(formula = Surv(event_time_mod, event_indicator_mod) ~ Sex +
ageC + HHcat_alt + Main_Branch + Acute_seizure + TreatmentType_binary +
ICH + IVH_dummy + IVH_dummy:log(event_time_mod)
And now I was hoping to get a prediction using survfit and providing new.data
for the combination of variables
I am doing the predictions:
survfit(cox, new.data=new)
Some comments:
1. even though it is in the SAS manual and some literature, I have myself
never used
"X * log(time)" as a fix for lack of proportionality. Is it really
true that when you
use
fit <- coxph(Surv(event_time_mod, event_indicator_mod) ~ Sex +
ageC + HHcat_alt + Main_Branch + Acute_seizure +
TreatmentType_binary +
ICH + IVH_dummy)
zfit <- cox.zph(fit, transform="log")
plot(zfit[8])
that the estimated function is linear? I have not yet seen such a simple time
effect
and would find it interesting.
2. The code you wrote does not fit the time dependent model that you
suppose; it
treats event_time_mod as a fixed covariate. To fit the model see the relevant
vignette
for the survival package. Essentially the program has to build a large (start,
stop) data
set behind the scenes. (SAS does the same thing). Defining proper residuals
for said
data set is hard and the R code does not yet do this. (Last I checked, SAS did
the same
thing.)
3. The "survival curve" for a time dependent covariate is
something that is not
easily defined. Read chapter 10.2.4 of the Therneau and Grambch book for a
discussion of
this (largely informed by the many mistakes I've myself made.)
Terry Therneau
[[alternative HTML version deleted]]