Ecological modeling in Stan

[Matthijs Hollanders]

Hi everyone,

I know Bayesian ecological modeling is largely dominated by JAGS and NIMBLE because gradient-based methods like Hamiltonian Monte Carlo (HMC) don't allow discrete latent parameters. For me this was a major obstacle but I decided to learn how to do it and discovered that constructing ecological models in a state-space formulation like hidden Markov models actually lends itself to an implementation of the forward algorithm that scales with increasing complexity, including hierarchical observation processes with multiple hidden states. I wrote a blog post about how to do this in Stan and I hope some people here wanting to transition to Stan find it useful. Stan is a great language with an incredible team of developers and has some really nice features for some our big and slow models, such as within-chain parallelisation that spreads the work of HMC chains across all the cores on your computer (typical parallel workflows just use one core per chain, whereas this methods can perform the work of individual chains across many cores). I'd be happy to answer questions and comments about the post here.

Thanks!

Matt