Large R-hat for boundary separation

[Nan Wang]

Hello everyone,

I was struggling with large R-hat values for boundary separation (other parameters are fine) no matter how I change the model (vary a with different factors or set it fixed across conditions). The data should be fine, I checked the response distribution and they look like Gaussian distribution. 

Did anyone run into similar problems before or have some possible solution for this?

Appreciation for any response! 

Nancy 

[Alexander Fengler]

Without knowing a bit more about your problem it is hard to give you a meaningful answer.

Could you let us know what the exact model is that you are fitting and some properties of your dataset?

I have seen all kinds of things, including problems with R-hat. Two categories of issues sort of empirically emerged for me.

1. Hierarchical models where the variance terms are much harder to fit --> creating problems with mixing for the respective group / individual level parameter

2. Identifiability issues, leading to strong parameter correlations, which the basic slice sampler (that's what HDDM uses), has trouble with and will also create serious issues with mixing

One thing to try is always to run much longer chains, but mileage varies depending on the exact problem.

Best,

Alex