I think one of the main reasons might be that it is common in survival analysis for people to fit a regression model on the hazard scale (due to the popularity of the Cox model), but the outcome is observed as a time-to-event. So the approach probably doesn’t feel very familiar to people who are used to more common regression modelling frameworks like, say, a GLM(er) where applying the inverse link function to your linear predictor gets you a predicted value for the outcome.
That’d be great! I’m still pushing changes regularly, so might break things occasionally, but hopefully it will run without too many bugs from now on. I noticed the estimates from the M-splines model, when fit to the Breast Cancer data, were a bit different to the R&P or B-spline estimates, and increasing the df for the M-splines didn’t really improve things (I think this was true when running the model with your .stan file or with
stan_surv) – but this only seemed to happen for that one dataset. A simulated dataset seemed to return similar estimates for the M-splines, B-splines, or R&P models, I think. Would be interested to hear what you observe.
(If you want to test using simulated data, then you could download the simsurv package).
Yeah, happy to collaborate on the code. Will just be figuring out the logistics! I’ve made a new class
stansurv in rstanarm that inherits the class
stanreg and have added a
prior_summary method for those objects today. Still some other methods to add (including
summary and perhaps a
basehaz_summary method). For predictions, hopefully we can just use most of the structure from the
posterior_survfit function from
stan_jm (although the stan_jm stuff is a lot more messy and dense, but still should be a bit useful). Also things like time-dependent effects, AFT models, trying out the constrained likelihood thing… I think we could each work on aspects of those, without confusing each other too much with conflicting commits? The method you mentioned here:
could be quite cool to implement too, if possible – perhaps you could try add it to the surv.stan file I’ve started in rstanarm? In any case, this thread is probably the wrong place for this discussion! – if your keen to collaborate on the rstanarm code then we could move this discussion to GitHub. I’ll open an issue on the rstanarm repo so that we can discuss ideas there…