number of variables in an mcmc object?

[njtierney]

This is my current thinking for n_var for stanfit and jags classes:

#' @export
n_var.stanfit <- function(x){
    # note - is "lp__" a parameter we care about?
    # what about y[1] and y[2]?
    length(x@model_pars)
}

#' @export
n_var.jags <- function(x){
    x$state() %>%
        purrr::pluck(1) %>%
        length()
}

You can test on: example_stan_model and example_jags_model - the models for which are given in data-raw

I'm not sure if this is a good idea the way it is currently written...also, the number of variables is kind of a pretty big question! if things are nicely laid out in an mcmc.list then it is clear, but things like jags and stanfit models are perhaps less clear.

What do you think, @samclifford ?

[samclifford]

So this is operating on a model object rather than on a set of samples? If that's the case, surely we'd be pulling out the number of partially observed nodes from the graph.

[njtierney]

• jags = operating on the jags model
• .stanfit = the fitted model

we'd be pulling out the number of partially observed nodes from the graph.

That sounds about right, but I'm not sure how to retrieve the graph structure from the jags/stan models? Maybe I can wrap up this feature and then we can add methods for n_var for stanfit and jags later.

[samclifford]

See my recent push to the glance branch. My thinking is that if a node has an value upon compilation then it is likely to be a parameter in the model. For nodes which have size k we can count it as k parameters or as 1 using the unroll_vectors logical.