[Gelman-Rubin] - Invidual vs group

[Olivier Savrimoutoo]

Hello,

I am very new to using HDDM. I've fitted several models, and when extracting the R-hat values for one of them, I see values > 1.1 for two parameters for one subject.

Can the group-level parameters still be correctly interpreted in this case or is it imperative to re-run the model such that all parameters for all subjects converge?

Thanks a lot!

Olivier

[Liu]

Hi, being able to estimate subject parameters, not treating every subject as an average subject ( Wiecki et al. 2013), is one of the characteristics of HDDM, so all the R-hat values at group or subject levels need to be <1.1 for an acceptable convergence. 

You may try sample (20000, burn=200 (1/100 of the sample number), thin=2 (better than 5)) to get smaller R-hat values for parameters for all subjects, or fine-tune these three indices to get the numbers fit with your data. There are some recent publications using sample 20000 and thin =2 as well, such as this one https://www.sciencedirect.com/science/article/pii/S2666915323001324, smaller burn-in as 100 in https://academic.oup.com/cercor/article/30/10/5471/5851502

And if you are using Mac, then better to find a Win pc, because somehow the R-hat values from sample 5000 computed by Mac are larger than those from Win pc.  

Hope it helps and you will get a nice convergence with all values < 1.1  : )

One question for the number of burn-in to all the HDDM experts: the 1/100 possible ratio, burn-in vs. sample number, is gained after several attempts with different sample numbers, not sure if it is too small?...please let me know if it is a proper ratio, Thank you All.