search

audrey-b / sims

An R package to simulate data from JAGS or R Code
https://poissonconsulting.github.io/sims/
Other
1 stars 2 forks source link

Monitors in sims_simulate #29

Closed audrey-b closed 4 years ago

audrey-b commented 4 years ago

So far, each time I've used sims_simulate in an application, I always end up having to tweak to latent=NA, stochastic = NA and set monitors manually to get the nodes I want. This is because in a lot of models some of the latent variables are actually observed; and also we may be observing a transformation of a random node (i.e. deterministic nodes). I think that from a practical point of view, it would make more sense to set the default to latent=NA and stochastic = NA. This way one can simply set monitor= to whatever they need. Otherwise, an error message is thrown out when trying to set monitors to a different subset of nodes than the one reported by default. I find that it is not intuitive to have to change latent=NA and stochastic = NA to get the desired result and that could deter users.

On the other hand, with very simple models that do not have latent variables (e.g. a regression model), having latent=NA and stochastic = NA by default would produce the same results as the current setup.

Thoughts?

joethorley commented 4 years ago

The distinction is kind or arbitrary from a true Bayesian perspective (data are just observed parameters) which fits with your practical experience. Do you think we should get rid of latent and stochastic entirely?

joethorley commented 4 years ago

@audrey-b - I don't have time to address for the foreseeable future - feel free to explore in a branch.