0
$\begingroup$

I've built a parametric survival model (survreg in R) to predict injury recovery duration using features such as age and treatments. I'm looking at trying model-based boosting using mboost with Weibull, loglog, or lognormal distribution (as described here). Can anyone recommend other alternatives? Has anyone tried neural network methods to predict time to event / duration data?

$\endgroup$

1 Answer 1

0
$\begingroup$

Yes, neural-network modeling of survival data dates back about 3 decades at least. You might want to consult this PubMed entry and its links.

Such survival analysis suffers from the difficulty, inherent in neural-network modeling, of understanding how particular predictors contribute to the differences in outcomes.

Boosting can be useful, as you can let the data help take into account interactions among predictors that you might not have considered. You at least get some estimate of which predictors are important. You don't, however, know the specific forms of their associations with outcome.

Semi- or fully parametric survival regression models typically provide much better interpretability, and with them you can use your understanding of the subject matter to address specific hypotheses of interest.

$\endgroup$
1
  • $\begingroup$ Thanks EdM. I’ve just tried mboost with gaussian and weibull distributions as well as a regression tree (rpart with anova). Surprisingly (or perhaps not) the baseline parametric weibull survival model outperformed them all on both RMSE and median error. Some hyper-parameter tuning may be necessary. $\endgroup$
    – Dan
    Mar 14 at 23:32

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.