Dear colleagues, I have been recently looking at some articles by authors who estimate the overall survival of their patients based on time-dependent variables (eg, surgery of metastasis, thrombosis, etc.). Normally the effect of these variables is communicated by hazard ratios, which are usually obtained by dividing the observation periods into intervals (eg, by the tmerge function in R, or by using other procedures in STATA). However, the use of time-dependent variables is a violation of the precepts of the Kaplan-Meier method for estimating the percentage of events in specific time periods, insofar as the variable that serves to stratify the population is not always present at baseline, but appears capricious at any time of follow-up. The most logical a priori way of communicating the effect would be computing time from the time-dependent variable to the event of interest. But if we want to communicate a result from the beginning of the follow-up, is there any kind of plausible and common sense solution, or should it be considered as an absurd question and one should limit to report the hazard ratio?
It is not an absurd question and a situation that is not unusual. It can be solved by putting the time-dependent covariate in a Cox proportional hazards regression model. I am not familiar with R or STATA but in SAS, it is a direct application of PROC PHREG. See sas.support and search for PHREG. A discussion, with an example, are given there. Conceptually it should provide guidance for other software.
In Stata you need to set up your data for survival analysis using the
STSET command
Then use STCOX to fit a Cox regression model
If you type in the Stata command line window
help stset
help stcox
you will see details with examples on how to proceed.
I hope this helps,
Nikos
Have you considered using multicriteria analysis for your problem? Maybe it's an easier way to get at the intended result.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics