1
$\begingroup$

When doing cox proportional hazards regression one often has competing risks. The typical approach for this is to fit separate cox proportional hazards models for each risk, censoring the competing risk.

How do you do model selection in this scenario? Should the two models be the same? Do I build a model based on one of the risks and use the same model for the other risk? Do I build a model based off a completely different response (e.g. overall survival)?

$\endgroup$
  • 1
    $\begingroup$ You need to define what "build a model" means. If it means estimating regression coefficients then your approach to fit a series of models with different target outcomes seems sound. If you mean "variable selection" then you have to question why you would spend time doing variable selection and not using subject matter expertise to specify the model. $\endgroup$ – Frank Harrell Nov 27 '14 at 13:36
  • $\begingroup$ @FrankHarrell I was referring to variable selection. This is for a project in school on a real dataset and we don't have access to the "subject matter experts". However, I think "you should only consult the experts" is kind of a cop out answer because that belittles the entire field of model selection. $\endgroup$ – bdeonovic Nov 27 '14 at 14:15
  • 2
    $\begingroup$ The entire field of model selection is full of false promises. Please read elsewhere on this site about the many pitfalls and high probability of getting a misleading result. Or just look at stata.com/support/faqs/statistics/stepwise-regression-problems $\endgroup$ – Frank Harrell Nov 27 '14 at 17:23
  • $\begingroup$ I didn't suggest stepwise procedures either, but rather AIC/BIC type of criteria, but you have a valid argument and I can appreciate your experience. Happy Thanksgiving :) $\endgroup$ – bdeonovic Nov 27 '14 at 23:32
  • 1
    $\begingroup$ BIC will underfit. AIC works well if you have say 3 competing models to choose from. AIC is just a (smarter) restatement of $P$-values so doesn't solve the fundamental problem if the number of competing models exceeds a few. $\endgroup$ – Frank Harrell Nov 28 '14 at 12:24
3
$\begingroup$

As you say, one approach is to use cause-specific hazard functions, treating the observations experiencing a competing events as censored. The problem with such approach is that it doesn't have a direct interpretation in terms of survival probabilities, that are obviously reduced by the the other risk: you are basically diseregarding the information about the events due to it. However, such approach replies to an important etiological question, i.e. how each risk contributes to non-survival: it is epidemiologically interesting because it's based on real probabilities, i.e., in each moment, the one of dying among people that may actually die.

However, you wondered whether you should consider overall survival in modeling competing events. This is another approach, based on Fine and Gray (1999) cause-specific ${subdistribution}$ (i.e. the cumulative incidence function). Here, the drawback is that they consider in the risk set for a given event also individuals that may not have the event itself (because the competing one has already occurred previously), thus, their estimates are not based on probability in a proper sense. However, their approach has a direct interpretation in terms of survival probabilities: it allows to calculate the probability of each event. For example, a Fine and Gray approach may lead us to conclude that, in a given sample, with a given covariate pattern, $25\%$ of patients will die for cardiological reasons, $20\%$ of patients for other reasons, and $55\%$ will survive until the end of the follow-up.

Fine, J. P., & Gray, R.J. 1999. A proportional hazards model for the subdistribution of a competing risk. Journal of the American Statistical Association,94(446), 496–509.

| cite | improve this answer | |
$\endgroup$

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.