0
$\begingroup$

How does one determine the probability that an event will eventually occur given competing risks?

Given the following state machine, I am trying to determine the probability that an observation in state A at time $t_i$ eventually ends up in state C (as opposed to state D).

enter image description here

I am considering using Weibull-AFT to estimate the hazard functions $h_{AB}(t)$, $h_{BC}(t)$, $h_{AD}(t)$, $h_{BD}(t)$ for each of the transitions. This would give me the shape $k$ and scale $\lambda$ parameters for each of the underlying Weibull distributions.

What I would like to do is get probabilities for the AB vs. AD transitions and probabilities for the BC vs. BD transitions, then treat the state model like a Markov chain.

From there I get stuck. I have looked into Fine and Gray methods, but these seem to be more focused on the likelihood of an event happening during the next time period. I have also considered taking the ratio of the integral of the competing hazard functions, but it seems like I am making things up at this point.

$\endgroup$
2
  • $\begingroup$ Are all the state transitions irreversible, as implied by the single-headed arrows? $\endgroup$
    – EdM
    Commented Mar 14 at 20:11
  • $\begingroup$ @EdM that is correct $\endgroup$
    – JoeBass
    Commented Mar 14 at 21:06

1 Answer 1

0
$\begingroup$

I'd recommend reading the multi-state vignette in the R survival package for guidance. What follows is mostly gleaned from that.

Fine-Gray won't be helpful. First, you don't have a simple competing-risks situation. Second, even if you did, the transitions in Fine-Gray are modeled separately in a way that can lead to the sum of final probabilities over all states being different from 1.

Direct multi-state modeling, as explained in Section 3 of the vignette in the context of a Cox proportional hazards model, can handle all of your data together. That can provide plots of estimated probability in state as a function of time. As you are willing to use a parametric Weibull model with its underlying proportional hazards assumption, that can give you an even more flexible way to proceed.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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