Burn-in settings and general model issue of LibBi

[MollyCYM]

Thank you for helping us to have a check on the following two issues we currently encounter in our project about stochastic epidemic compartment models:

1. H1N1-Brownian Motion Model: this model is an H1N1-SEIR compartment model with transmission rate beta modelled as a Brownian motion. The issue we have here is although we set in the synthetic data experiment that parameter sigma=0.4 (true value and red line in the trace plot), and prior of this parameter is set to be a wider one-Uniform(0,1) (green lines), the proposal is a random walk, the resulted posterior samples gained from LibBi (blue line-posterior median) goes to a very tiny value: 0.04. Is this happened due to we set something wrong in our R code?

2. In order to exclude some parameter fitting errors in our model, we’re conducting some experiments on another Covid-Brownian motion model. We want to check if the posterior samples correctly start from the initial values we gave to each parameter. The current way we did is to remove the adapt_proposal line in our code and followed with sample(nsamples=10000, thin=1). But the resulting trace plot tells us this seems to be incorrect (e.g. tau’s initial value is set to be 0.8, but from the trace plot, it starts from 0.3 ). So is there a way to correctly discard burn-in in LibBi?

Thank you in advance for your assistance of using LibBi in our research model, meanwhile, we have uploaded a small pdf file with this issue ticket, from where you can check our model description and corresponding R code (if you need R files to run them, please feel free to contact me by [email protected]).

All the best,
Molly
Model Description & R code.pdf

[sbfnk]

Could you put a reproducible example here that to illustrate the behaviour you're describing? It's really hard to investigate this without being able to follow exactly what sequence of commands you're trying.