You don't want the tvc option. Using time-varying covariates in Stata's survival routines is less about the command and more about data set-up. You need to know how to use stset with multiple lines of data per subject.

In the simplest survival data set, every line has a time to event, and an indicator for whether the event was the one of interest (coded 1) or censorship (coded 0). Stset will assume an entry time of zero, and an event time of t, and create the st variables used by the routines. For time varying covariates, changes in covariate values create a censorship event. They enter the study alive, awaiting a procedure, are censored when they have the procedure, and start a new line of data where the entry time is the censoring time from the first line, with a new covariate value indicating post-procedure. If you have dates for the entry, procedure time, event time, you can use stset first to create one line per subject with event time, and secondly use stsplit to create the multi-line data at procedure time.

It is a good idea to use the cluster option, treating this as repeated measures on each subject. It may not make any difference, since you have at most one event per subject, but it will not do any harm either.

P

If I remember well tvc() stands for time varying coefficents (I know it sound strange) and is supposed to allow for changes in proportionality of hazard over time (and function texp() defines this interaction). The thing that you need to do is to set the data properly i.e. you need to have multiple observations per individual every time value of your time varying covariate changes (this you can do easily with help of stsplit() command). Then during the stset() you need to use option id() to tell stata that separate observations having the same id are just the changes of its characteristics over time.

I've tried with this, but not sure if it is completely correct. I'd rather prefer to avoid stsplit but just if it is posible...

stset OS, failure(STATUS) id(id) // set survival time

stsplit , at(failures) // create the new dataset

generate TPHtvc=2 //create the time-varying covariate

replace TPHtvc=TPH if OS>OS1

stcox varX TPHtvc

stjoin

Same thing is already discussed earlier on Statalist, the helping horsemen behind using Stata including me.

http://www.stata.com/statalist/archive/2011-07/msg00687.html

HTH

You don't want the tvc option. Using time-varying covariates in Stata's survival routines is less about the command and more about data set-up. You need to know how to use stset with multiple lines of data per subject.

In the simplest survival data set, every line has a time to event, and an indicator for whether the event was the one of interest (coded 1) or censorship (coded 0). Stset will assume an entry time of zero, and an event time of t, and create the st variables used by the routines. For time varying covariates, changes in covariate values create a censorship event. They enter the study alive, awaiting a procedure, are censored when they have the procedure, and start a new line of data where the entry time is the censoring time from the first line, with a new covariate value indicating post-procedure. If you have dates for the entry, procedure time, event time, you can use stset first to create one line per subject with event time, and secondly use stsplit to create the multi-line data at procedure time.

It is a good idea to use the cluster option, treating this as repeated measures on each subject. It may not make any difference, since you have at most one event per subject, but it will not do any harm either.

P

There is video demonstration on youtube with a stata do file which you could hugely benefit from for your enquiry.Just enter cox regression on YouTube search box. Good luck!

Check the following link. This might help you.

www.stata.com/stata11/st_stcox.pdf‎

Hi everyone! Thank you for all your comments. I guess you all have a bit of reason. According to what Mr Ogórek and Mr Wileyto pointed, I have worked on the stset-stsplit syntax. I wanted to avoid stsplit because it seems risky not to be able to do a stjoin at the end...may be I'm too cautious but I don't like my dataset appearance change. In any case, now that it seems no other choice..I'm starting to understand the idea but still have some trouble.

For instance, how many rows should I have if my cohort is 274 patients and the number of transplants are16? I guess 290...

I tried this syntax inspired in the "webuse stanford" example at "Examples of splitting data at designated times" (help stsplit)

*create unique person identifier.

generate id = _n

*Change missings in OS1 to 0, in a new variable called OS2.

generate OS2=0

replace OS2=OS1 if OS1

It's survival analysis in stata by Ani Katchova, posted on the 18th of February 2013.Hope that helps!

Sorry, in my last comment I forgot to tell that the syntax that I pasted simplifies the original problem because I don't take into account the type of transplantation, just if it has been done or not . To address the original question my proposal is as follows (in the last post I tell where it comes from)

*create unique person identifier

generate id = _n

*Recode missings from OS1 into 0, in a new variable called OS2.

generate OS2=0

replace OS2=OS1 if OS1