This is a second post in my series on taming divergences in Stan models, see the first post in the series for a general introduction. Also see guide to Stan warnings Standard caveat: I am not an expert on Stan, I consider myself just an advanced user who likes to explain things.

Sometimes you are developing a model that has multiple variants: maybe you want to consider several different link functions somewhere deep in your model, or you want to switch between estimating a quantity and getting it as data or something completely different. In these cases, you might have wanted to use optional parameters and/or data that apply only to some variants of your model.

I have just finished my submission to StanCon Helsinki 2018 which deals with my ongoing work on Bayesian modelling for gene regulation. I submitted late, because I procrastinated a lot on it and I recently have terrible time management in general. So we will se if it will even get considered. View the compiled submission or check the Genexpi-Stan GitHub repository, to view the source code.

Although my time with the Stan language has been enjoyable, there is one thing that is not fun when modelling with Stan. And it is the dreaded warning message: There were X divergent transitions after warmup. Increasing adapt_delta above 0.8 may help.

This is a neat trick I found on Tyler Morgan-Wall’s Twitter and is originally attributed to Joe Cheng. You can run any Shiny app without blocking the session. My helper function to run ShinyStan without blocking is below: launch_shinystan_nonblocking <- function(fit) { library(future) plan(multisession) future( launch_shinystan(fit) #You can replace this with any other Shiny app ) } Hope that helps!

Not long ago, I came across a nice blogpost by Kahtryn Morrison called A gentle INLA tutorial. The blog was nice and helped me better appreciate INLA. But as a fan of the Stan probabilistic language, I felt that comparing INLA to JAGS is not really that relevant, as Stan should - at least in theory - be way faster and better than JAGS.

So I recently started programming in Elm and also did some stuff in Rust. Honestly, it was mostly a hype-driven decision, but the journey was definitely worth it. I also noticed that although those two languages differ in their target audience and use cases, they made many similar design decisions. I think this is no coincidence.

I love thinking about other people’s statistical problems! So get in touch if you need help with: Developing and debugging statistical models (particularly Stan and INLA) Exploratory data analysis Statistical graphics Small Inquiries If I have time, I will answer to small inquiries on Stan Discourse, tag me in as @martinmodrak to get my attention. You can check out list of my previous answers.

This is a personal webpage/blog of Martin Modrák, an academic researcher in bioinformatics, currently at Institute of Microbiology of the Czech Academy of Sciences, laboratory of bioinformatics. I also do a bit of statistical consulting on other people’s projects. Get in touch! I also used to do research in AI for computer games. And I have a defunct software engineering blog.