0
$\begingroup$

I am attempting to extrapolate a Extended Cox Model (cox model with a time-varying covariate) from 6 years to 30 years in R. Is there a way to do this?

I have thought of a mediocre way of doing this: Plot two plots, one with the slopes from the KM curve for each year vs. time and one with the downward shifts of the KM curve for each year and fit a line for each plot. Use those lines to estimate the slopes and shifts for the years past 6. I am also not quite sure how to nicely code this in R.

Another idea is to repeat the hazard at year 6 over and over again. This would be an extremely conservative approach, but may get the job done.

I know this may not be the "fanciest" way of doing things. But, aside from fitting a distribution, I have run out of ideas.

$\endgroup$
  • $\begingroup$ You will need to fit a parametric survival model. You can use an ad hoc approach but properties have not been studied. $\endgroup$ – Frank Harrell Aug 24 '15 at 18:24
  • $\begingroup$ That is what I have seen. I am just wondering if either of these ideas that I have would make sense to do. $\endgroup$ – Emily R. Aug 24 '15 at 19:20
  • $\begingroup$ If these are human data, you will probably have to deal with the competing risk of death from other causes. $\endgroup$ – EdM Aug 24 '15 at 19:52
2
$\begingroup$

Some assumptions need to be made to extrapolate (far) beyond the range of the data you have. These assumptions could either be parametric models or something like repeating the hazard of the last year (or last 2 years) over and over. The latter approach is quite a strong assumption (more or less extrapolation with an exponential distribution based on the average hazard rate of the last year, if the hazard rate did not vary too much throughout the year). One possibility to avoid assuming one particular model and to reflect the uncertainty about the appropriate model, is to entertain a range of plausible models and to do (e.g. Bayesian) model averaging. Another possibility is to use information external to your main dataset, because extrapolating to 30 years from just 6 years may be very challenging unless there are reasons to believe the underlying process would not change too much (or if we know how we expect it to change). It is hard to say anything about that without knowing what the unity under risk are.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.