0
$\begingroup$

I am trying to figure out the following. Could you please comment on this?

Goal is to compare survival between two groups. Let's assume we have data with one-year follow up and have a simple model for getting age-adjusted estimates:

time | status ~ group + age

Is it reasonable to run models with different follow-up lengths? E.g. 1, 2, 6 and 12 months? By reasonable, I mean does it give some extra or valuable information?

My understanding is as follows:

If this is a Cox model

If the assumption of proportional hazards is met with all follow-ups, there is no point run models with different follow-ups due to the PH assumption. For example, one group always has a constantly higher/lower hazard ratio.

If this is an Accelerated Failure Time (AFT) model

If the proportional hazards assumption is not met, I would use AFT. Is it now reasonable to run models with different follow-ups? For example, group A and B may have similar 1- and 3-month survival but patients in group B die more likely after the 3rd month.

$\endgroup$
1
$\begingroup$

You almost certainly don't want to be building different models on time-restricted subsets of data. That's just throwing away information.

What can make a lot of sense is to build a model based on the entire data set and then evaluate its predictions at various times. In general, it's not a bad idea to check the calibration of a survival model that way, comparing predicted and observed survival probabilities at a range of times.

Furthermore, your understanding of the subject matter and the reasons for building your model might suggest particular times at which to compare the groups. If so, choose those particular times.

Whether your model is AFT or PH isn't really an issue. You think about the relationships between covariate values and outcomes differently between those classes of models, but a survival curve is still a survival curve regardless of how you modeled it.

$\endgroup$
1
  • $\begingroup$ Thank you for the really good and clear answer! I tried to do predictions from a model; however, the result seemed a bit weird. Could you have a comment on that? stackoverflow.com/questions/66459653/… $\endgroup$ – st4co4 Mar 4 at 7:46

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.