0
$\begingroup$

I'm interested in evaluating the effect of various factors on the lifespan of borrower-lender relationships using a Cox Time-Varying Proportional Hazards regression.

However, as the relationships we observe are defined by the existence of a loan filing which covers a 5-year span (the loan can be terminated early, extended, or a follow-up loan filed), the relationships are most likely to end at the 5 year mark (or at least in 5-year intervals). Does this present a potential complication for a Cox Proportional Hazards model, or is the framework able to properly account for situations like this where there's a pattern in when an event might occur?

Basically, is it a problem for a CPH regression that the base time component possibly follows a relatively fixed interval?

$\endgroup$
1

1 Answer 1

1
$\begingroup$

A Cox model can have problems with multiple tied event times, as outlined here. The exact correction for ties can be computationally expensive. The Efron and Breslow approximations don't work well when most of the events happen at the same time.

If you had a limited number of fixed intervals then a discrete-time survival model would be appropriate. It seems, however, that other events can happen at any time during the standard 5-year window.

For either a Cox model or a discrete-time model, you need to decide whether you think that the associations of covariates with the risk of ending the borrower-lender relationship are the same for each type of termination. This might be handled better with a competing-risks model, maybe a Markov multi-state model, that allows for differences in baseline hazards and in covariate associations and that can accommodate both continuous and discrete event times. I don't have experience with those, but the R rstpm2 package seems to provide the necessary tools.

$\endgroup$

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.