I’m developing a hierarchical mode for which I keep getting divergent transitions (even with adapt_delta set at 0.99). I see documentation on how to do a non-centered parameterization for rstan. Is there a way to implement this within rstanarm, and if so how? Thanks.
It is already hard-coded in the rstanarm package; you can’t even do a centered parameterization, which is occassionally useful.
Okay, thanks Ben. That being the case, is there an alternative to dealing with divergent transitions beyond increasing adapt_delta? Alternatively, is there a de minimis number of DTs that one can “get away with” and still reach reliable inferences?