Is it possible to perform a multiclass Cox proportional hazards model? I'm interested in finding the probability of self cure on consumer loans, or just study how self-curing clients behave so as to identify future customers that behave the same way. My dependent variable of interest has 3 classes: -not cured (0) -cured after being contacted (1) -self cured (2)

I was thinking maybe two regressions have to be run in this case (similarly as in the multinomial logit), one where it's 0 vs 1, and another where it is 0 vs 2, maybe another 1 vs 2?

I tried looking for this but I didn't find anything, so I'm wondering if it's possible at all?


This seems to be a fairly straightforward "competing risks" analysis. If "cured" is your event, then you simply code the event differently depending on whether it was a "self-cure" or "cure-after-being-contacted". In general, instead of a coding {0,1} for {censored, event} you code {0,1,2,...} for {censored, eventType1, eventType2, ...}. That's handled pretty simply in the R survival package, for example. See Section 2.3 of the main survival vignette of that package.

You don't want to do this analysis separately on different types of events. As the vignette says (Section 2.3.2):

A common mistake with competing risks is to use the Kaplan-Meier separately on each event type while treating other event types as censored.

See the vignette for further details. If you want more background, do a search on "competing risks" in survival analysis.

| cite | improve this answer | |
  • $\begingroup$ Thank you! This seems to be what I was looking for, I didn't know such a thing existed. I wonder though, isn't Multi-state data in section 2.4 more ideal to my model? $\endgroup$ – amestrian Aug 28 at 17:17
  • $\begingroup$ Also, do you know if these competing risks models account for recurrent events as well? Because that's the shape of my data (many stages for the same unit) $\endgroup$ – amestrian Aug 28 at 17:20
  • $\begingroup$ @amestrian multi-state models with arbitrary transition patterns and competing risks are possible to model this way. See the multi-state vignette for the survival package. Of course it gets more complicated the more transitions that you allow, and data formatting might get dicey. I don't have any substantive experience with such models, but now you know the buzz-words for which to search. $\endgroup$ – EdM Aug 28 at 18:01

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.