Describe method keyword arguments

[joshualeond]

I had asked in the Turing slack channel if there were a simple way to adjust the precision of the describe methods output. I was told there isn't and to open a feature request. Currently we see something like this with the describe method:

julia> describe(chn)
2-element Array{ChainDataFrame,1}

Summary Statistics

│ Row │ parameters │ mean    │ std      │ naive_se  │ mcse      │ ess     │ r_hat    │
│     │ Symbol     │ Float64 │ Float64  │ Float64   │ Float64   │ Any     │ Any      │
├─────┼────────────┼─────────┼──────────┼───────────┼───────────┼─────────┼──────────┤
│ 1   │ m          │ 1.04149 │ 0.794192 │ 0.0251146 │ 0.0927405 │ 57.9072 │ 1.00897  │
│ 2   │ s          │ 1.95544 │ 1.46153  │ 0.0462175 │ 0.115658  │ 115.065 │ 0.999489 │

The keyword argument would adjust the level of precision on this output. Something like this, with an additional digits argument:

julia> describe(chn, digits = 2)
2-element Array{ChainDataFrame,1}

Summary Statistics

│ Row │ parameters │ mean    │ std      │ naive_se  │ mcse      │ ess     │ r_hat    │
│     │ Symbol     │ Float64 │ Float64  │ Float64   │ Float64   │ Any     │ Any      │
├─────┼────────────┼─────────┼──────────┼───────────┼───────────┼─────────┼──────────┤
│ 1   │ m          │ 1.04    │ 0.79     │ 0.03      │ 0.09      │ 57.91   │ 1.01     │
│ 2   │ s          │ 1.96    │ 1.46     │ 0.05      │ 0.12      │ 115.06  │ 1.0      │

While we're discussing the output of the describe method. Maybe it would be advantageous to include a depth argument as well. An example of this is in the rethinking package in R. Here's a description of the argument in the rethinking::precis function:

depth	If 1, suppresses vectors and matrices of parameters. If 2, displays all parameters

A quick example of the differences with depth:

precis(m10.9, depth = 1)
#> 6 vector or matrix parameters omitted in display. Use depth=2 to show them.
#>     Mean StdDev lower 0.89 upper 0.89 n_eff Rhat
#> bm -0.09   0.09      -0.23       0.05   350    1

precis(m10.9, depth = 2)
#>       Mean StdDev lower 0.89 upper 0.89 n_eff Rhat
#> a[1]  0.68   0.11       0.52       0.86   482    1
#> a[2]  0.63   0.12       0.44       0.82   480    1
#> a[3] -0.58   0.08      -0.71      -0.46  1066    1
#> a[4] -0.62   0.09      -0.76      -0.47   680    1
#> a[5] -1.06   0.10      -1.22      -0.90   934    1
#> a[6] -2.65   0.16      -2.89      -2.40  1104    1
#> bm   -0.09   0.09      -0.23       0.05   350    1

[goedman]

Hi Josh,

Thanks for the feedback!

Having a digits=... argument is a good suggestion.

On your 2nd question, the approach taken in MCMCChains is slightly different in that you can adjust the chain's name_map, e.g. in your example:

    chns = set_section(chn, Dict(
      :pooled => ["a[$i]" for i in 1:6]
    )

It means an extra step, but given a slightly more elaborate example:

    chns = set_section(chn, Dict(
      :parameters => ["mu", "tau"],
      :thetas => ["theta.$i" for i in 1:8],
      :etas => ["eta.$i" for i in 1:8],
      :internals => ["lp__", "accept_stat__", "stepsize__", "treedepth__", "n_leapfrog__",
        "divergent__", "energy__"]
      )
    )

for models with lots and lots of parameters it allows finer control.