R help - Aug 2008 - Help on competing risk package cmprsk with time dependent covariate

Dear R users,


I d like to assess the effect of "treatment" covariate on a disease
relapse risk with the package cmprsk.
However, the effect of this covariate on survival is time-dependent
(assessed with cox.zph): no significant effect during the first year of
follow-up,
then after 1 year a favorable effect is observed on survival (step
function might be the correct way to say that ?). 
For overall survival analysis
I have used a time dependent Cox model which has confirmed this positive effect
after
1 year.
Now I m moving to disease relapse incidence and a similar time dependency seems
to be present.

what I d like to have is that: for
patients without "treatment" the code for "treatment"
covariate is
always 0, and for patients who received "treatment" covariate I d like
to have it = 0 during time interval 0 to 1 year, and equal to 1 after 1
year. Correct me if I m wrong in trying to do so.


First, I have run the following script (R2.7.1 under XPpro) according to
previous advices:

library(cmprsk)
attach(LAMrelapse)
fit1<- crr(rel.t, rel.s, treatment, treatment, function(uft)
cbind(ifelse(uft<=1,1,0),ifelse(uft>1,1,0)), failcode=1,
cencode=0, na.action=na.omit, gtol-06, maxiter)
fit1

where:
rel.t = time to event (in years)
rel.s = status , =1 if disease relapse, =2 if death from non disease
related cause (toxicity of previous chemotherapy), =0 if alive &
not in relapse
treatment binary covariate (value: 0 or 1) representing the treatment to test
(different from chemotherapy above, with no known toxicity)
I have not yet added other covariates in the model.


this script gave me the following result:
> fit1 <- crr(relcmp.t, relcmp.s, treatment, treatment, function(uft)
cbind(ifelse(uft <= 1, 1, 0), ifelse(uft > 1, 1, 0)), failcode = 1,
cencode = 0,
    na.action = na.omit, gtol = 1e-006, maxiter = 10)
> fit1
convergence:  TRUE 
coefficients:
[1] -0.6808  0.7508
standard errors:
[1] 0.2881 0.3644
two-sided p-values:
[1] 0.018 0.039

...That I dont understand at all since it looks like if "treatment"
covariate had also a significant effect of the first period of time !? 
This is absolutely not the case. 
So I m surely wrong with a part of this script... cov2 and tf are
pretty obscure for me in the help file of the package. I would really
appreciate advices regarding these 2 "terms". 

I was thinking that I might changed : cbind(ifelse(uft <= 1, 1, 0),
ifelse(uft > 1, 1, 0)                   into:        cbind(ifelse(uft <=
1, 0, 1), ifelse(uft > 1, 1, 0)

But since I only have one covariate (treatment) to test, shouldnt I only write
the following:
fit1<- crr(rel.t,
rel.s, treatment, treatment, function(uft) ifelse(uft<=1,0,1)), failcode=1,
cencode=0, na.action=na.omit, gtol-06, maxiter)

which gives me :
> fit1
convergence:  TRUE 
coefficients:
[1]  0.06995 -0.75080
standard errors:
[1] 0.2236 0.3644
two-sided p-values:
[1] 0.750 0.039

which, if I understand things
correctly (I m not sure at all !) confirms that before 1 year, the effect of
"treatment" covariate is not
significant, but is significant after 1 year of follow up. But there I m again
not sure of the result I obtain...

any help would be greatly appreciated with cov2 and tf

thanks for  if you have some time for this,


Philippe Guardiola


     
_____________________________________________________________________________

o.fr
	[[alternative HTML version deleted]]

Hello,

Something i don't understand
in your question.
Is treatment a time-dependent covariate?
That is, do patients receive the treatment
at the beginning of the study or later?

cmprsk cannot handle time-dependent covariates.

But if treatment is a baseline covariate,
but has a time-varying effect (i.e. does the subdistribution hazard 
ratio varies with time?), your solution
to assess that is weird, because you will transform
your baseline covariate into a time-dependent one,
thus considering all the patients to receive no treatment
the first year. For sure, the treatment wont have any
effect for the first year.
To assess a time-varying effect on competing risks,
i would either follow the cmprsk documentation, including
an interaction with functions of time, or use the comp.risk
function in the timereg package, which fits more flexible
models for the cumulative incidence functions.

Best regards,
Arthur Allignol


Philippe Guardiola wrote:
> Dear R users,
> 
> 
> I d like to assess the effect of "treatment" covariate on a
disease relapse risk with the package cmprsk.
> However, the effect of this covariate on survival is time-dependent
> (assessed with cox.zph): no significant effect during the first year of
follow-up,
> then after 1 year a favorable effect is observed on survival (step
> function might be the correct way to say that ?). 
> For overall survival analysis
> I have used a time dependent Cox model which has confirmed this positive
effect after
> 1 year.
> Now I m moving to disease relapse incidence and a similar time dependency
seems to be present.
> 
> what I d like to have is that: for
> patients without "treatment" the code for "treatment"
covariate is
> always 0, and for patients who received "treatment" covariate I d
like
> to have it = 0 during time interval 0 to 1 year, and equal to 1 after 1
> year. Correct me if I m wrong in trying to do so.
> 
> 
> First, I have run the following script (R2.7.1 under XPpro) according to
previous advices:
> 
> library(cmprsk)
> attach(LAMrelapse)
> fit1<- crr(rel.t, rel.s, treatment, treatment, function(uft)
> cbind(ifelse(uft<=1,1,0),ifelse(uft>1,1,0)), failcode=1,
> cencode=0, na.action=na.omit, gtol-06, maxiter)
> fit1
> 
> where:
> rel.t = time to event (in years)
> rel.s = status , =1 if disease relapse, =2 if death from non disease
> related cause (toxicity of previous chemotherapy), =0 if alive &
> not in relapse
> treatment > binary covariate (value: 0 or 1) representing the treatment
to test
> (different from chemotherapy above, with no known toxicity)
> I have not yet added other covariates in the model.
> 
> 
> this script gave me the following result:
>> fit1 <- crr(relcmp.t, relcmp.s, treatment, treatment, function(uft)
cbind(ifelse(uft <= 1, 1, 0), ifelse(uft > 1, 1, 0)), failcode = 1,
cencode = 0,
>     na.action = na.omit, gtol = 1e-006, maxiter = 10)
>> fit1
> convergence:  TRUE 
> coefficients:
> [1] -0.6808  0.7508
> standard errors:
> [1] 0.2881 0.3644
> two-sided p-values:
> [1] 0.018 0.039
> 
> ...That I dont understand at all since it looks like if
"treatment"
> covariate had also a significant effect of the first period of time !? 
> This is absolutely not the case. 
> So I m surely wrong with a part of this script... cov2 and tf are
> pretty obscure for me in the help file of the package. I would really
> appreciate advices regarding these 2 "terms". 
> 
> I was thinking that I might changed : cbind(ifelse(uft <= 1, 1, 0),
ifelse(uft > 1, 1, 0)                   into:        cbind(ifelse(uft <=
1, 0, 1), ifelse(uft > 1, 1, 0)
> 
> But since I only have one covariate (treatment) to test, shouldnt I only
write the following:
> fit1<- crr(rel.t,
> rel.s, treatment, treatment, function(uft) ifelse(uft<=1,0,1)),
failcode=1,
> cencode=0, na.action=na.omit, gtol-06, maxiter)
> 
> which gives me :
>> fit1
> convergence:  TRUE 
> coefficients:
> [1]  0.06995 -0.75080
> standard errors:
> [1] 0.2236 0.3644
> two-sided p-values:
> [1] 0.750 0.039
> 
> which, if I understand things
> correctly (I m not sure at all !) confirms that before 1 year, the effect
of "treatment" covariate is not
> significant, but is significant after 1 year of follow up. But there I m
again not sure of the result I obtain...
> 
> any help would be greatly appreciated with cov2 and tf
> 
> thanks for  if you have some time for this,
> 
> 
> Philippe Guardiola
> 
> 
>      
_____________________________________________________________________________
> 
> o.fr
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at r-project.org 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.
>