Why is divergent trasition bad?
- divergent transitions result in a biased sample – at the end of a divergent transition, the parameter estimates are unchanged, effectively, the same draw reported twice.
- divergent transitions indicate that the model is ill-specified w.r.t the data.
How to resolve?
a. find out more information via reported diagnostics – RStan – https://mc-stan.org/rstan/reference/check_hmc_diagnostics.html, CmdStan and CmdStanR, CmdStanPy have “diagnose” functions.
b. run warmup longer and see if the problem goes away, or increase adapt_delta to .95
c. reparameterize – these sections of the User’s Guide should help:
https://mc-stan.org/docs/2_26/stan-users-guide/hierarchical-priors-section.html
https://mc-stan.org/docs/2_26/stan-users-guide/reparameterization-section.html
For c, draw pairs plots to see which parameters are highly correlated accordingly.
Comment is the energy for a writer, thanks!