I have a three level hierarchical model that has: hyper-hyper parameters (for populations) hyper parameters (for sub-populations) measurements The measurements are N(mu, sigma) where the mu and sigma are sampled from the hyper parameters. The hyper-parameters for each sub-population are mu ~ N(h…

Perhaps you could try sampling from the prior? pm.sample_prior_predictive will generate samples from your specified priors, and you can manually inspect them to see if they make sense. There could be something wrong with how your model specifies the priors.

Thanks. I have been doing that. I have seen that my variance parameters seemed like they could hit zero which could cause some of my divergences and chain failures. I have eliminated those, but my divergences remain. In traceplot() output, in the frequency charts, I see some of the variables, fo…

Thanks. I looked at the FAQ (might be a good idea to link from the PyMC3 docs to it), but am not sure I see how to proceed further, now that I have removed the chain failure (by ensuring that the variance parameters never go to zero), but my divergences remain. Is it correct to think that this cou…

+1 to making a FAQ to the docs! I’ll keep that in mind for a future PR :slightly_smiling_face: Here are some tips on divergences that seem applicable to your situation taken from my blog post . I might’ve already linked you to the post, but I can’t exactly recall, so better safe than sorry. Inspec…

Thanks for the advice. It seems like nudging my variance parameters (and hyper-parameters) up was necessary to fix the problem. Once I ruled out sigma of 0, the divergences went away. One remaining problem that I haven’t read about – I am getting complaints about there being not many effective sa…

By the way, sampling from the priors was very helpful. But it raised an issue with PyMC3 (at least I think it’s an issue). I don’t believe that there’s a plotting function for the output of sample_prior_predictive (if there is one, I didn’t find it in the docs). The sampler was neat and its outpu…

Again quoting straight from the blog post (we really should port all these tips into an FAQ, shouldn’t we?): The number of effective samples is smaller than XYZ for some parameters. Quoting Junpeng Lao on discourse.pymc3.io : “A low number of effective samples is usually an indication of strong au…

Thanks, again. Sorry to have taken so long to respond. Your advice about divergence was very helpful, and between your site and the two articles and the lecture video it pointed me to, I was able to solve my divergence problems. A quick follow-up – do you know of resources like that for dealing w…

Advice with hierarchical model

Questions

_eigenfoo September 16, 2018, 11:14pm 3

In fact, this is actually one of our FAQs, perhaps you could look here for more info: Frequently Asked Questions

Topic		Replies	Views
Multilevel model specification Questions	8	1112	January 2, 2020
There were 6 divergences after tuning. Increase `target_accept` or reparameterize Questions	5	4379	August 19, 2021
Sampling errors when running hierarchical model using hyperpriors Questions	0	320	October 26, 2021
Unused variable causing divergences? Questions	7	716	September 15, 2020
How to parametrize random effects in Nested ANOVA Questions	7	1377	April 6, 2018

Advice with hierarchical model

Related topics