Survival models in rstanarm


Ok, great. Yeah, the last point from @lcomm would explain the behaviour I was seeing. It was increasing the dimension of the simplex that seemed to enforce the smoothing, rather than changing the concentration parameter. I had \alpha_i = 1 for all i=1,...k in my example.

I’m not sure whether it’s desirable. Maybe it is? But it definitely wasn’t intuitively expected, at least not my me. I expected that increasing the degrees of freedom for the spline basis would increase flexibility, but that prior distribution had other ideas.



I am not used with survival models, but I found that estimating the \alpha vector as hierarchical parameters might solve this dimensionality problem while providing an interesting amount of shrinkage. But I’m affraid it would increase the sagpling time and model complexity!


Do you have example code?


Sorry, I was not clear. The idea is simply to estimate the \alpha vector instead of fixing it.