PH models and probability estimates

[topepo]

There is currently some code in aaa_survival_prop.R to estimate survival probabilities for Cox PH models. It might be better to use survift(coxph_object) or basehaz(coxph_object) to get these since they take into account any stratification variables for that model. If that's the case, we might want to save the table produced by those functions in the model object and interpolate the step functions within each strata. Dealing with the strata is a pain, because of how survival encodes them, but some of that work is already done for the parametric model engine.

Also, I don't see any canonical ways of estimating the baseline hazard so that we can get type = "hazard". There as some methods that layer another model on top of the PH model results. It would be good to get community feedback on this.

We might also consider having a type = "cumulative hazard" method

@dincerti

[dincerti]

I would echo Max and suggest using survfit() to get survival predictions. It does a lot under the hood and can handle stratification, start/stop data, etc. There is also a summary.survfit() method can can interpolate for arbitrary time points.

type = "cumulative_hazard" would be a good idea and would work for most/all models. Perhaps type = "hazard" could be restricted to models where it is straightforward to compute (i.e., for parametric models)?

[topepo]

Do people ever want (instantaneous) hazard estimates for this model?

[hfrick]

update: survival probabilities for coxph models now use survfit()

leaving this open for the (instantaneous) hazard discussion

[dincerti]

I think hazards can be useful as a diagnostic (e.g., for understanding whether a parametric model seems reasonable, exploring non-proportional hazards). But I don't think they are nearly as useful as other summaries of the survival distribution, particularly in a prediction context.

[hfrick]

thanks for the feedback @dincerti !

[github-actions]

This issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.