search

ImperialCollegeLondon / covid19model

Code for modelling estimated deaths and cases for COVID19.
MIT License
944 stars 271 forks source link

Divergent transitions in stan fits #115

Open ounsley opened 4 years ago

ounsley commented 4 years ago

Hi,

Thanks for the all the great work, and any thoughts on this issue that you can provide are greatly appreciated.

Describe the bug I am experiencing various warnings whilst trying to run the base.r model with the latest code (v5). Of most concern is the presence of divergent transitions in the fits. I note that in report 13 you mention that "In fitting we also ensured that the MCMC sampler experienced no divergent transitions".

Is the presence of the dts a result of changes made ot the most recent version of base.stan and has consideration been made to rectify this?

Would increasing adapt_delta help with this, or does this require reparameterisation?

I have tried to identify parameters of which idicate the presence of pathologies associated with these divergences, but have not been successful.

To Reproduce run: Rscript base.r -F

Expected behavior No divergent transistions

Desktop (please complete the following information):

ounsley commented 4 years ago

I have seen significant drops in the number of divergent transitions when inceasing adapt delta from .95 to .99, from around 200 - 300 down to less than 10 from a run of 5000 iterations (2000 warmup) and 5 chains.