Skip to contents

Estimate the Log-Evidence of a Model

calculate.ernest_sampler.Rd

Calculate the marginal likelihood of a given model and return the estimates in a tidy tibble().

Usage

# S3 method for class 'ernest_sampler'
calculate(
  x,
  add_points = c("none", "unit", "original", "both"),
  add_progress = FALSE,
  ...
)

Arguments

x

An object of class ernest_sampler.

add_points

A string, either "none", "unit", "original", or "both". If "none", no additional columns are added. If "unit"or "original", the parameter values associated with each point are added, in their respective units.

add_progress

Adds columns for the number of calls to the likelihood function between each iteration.

...

Must be empty.

Value

NULL, if x does not contain any iterations. Otherwise, a tibble with at least these columns:

  • .iter: The iteration number.

  • log_lik, log_vol: The log likelihood and estimated prior volume associated with the point extracted from the live set.

If log_lik and log_vol are well-specified (see details), then these additional columns are added:

  • log_z: The log-evidence estimate.

  • log_z_var: The variance of the log-evidence estimate.

  • log_weight: The posterior log-weight of the point.

  • h: Information, reported through the Kullback–Leibler divergence.

If add_points is not "none", columns are added with names given by the 'ptypes' argument for the sampler. Points in unit scale have names prefixed with unit_.

Finally, if add_efficiency is TRUE, columns .calls, .id, and .sampler are added, reporting the number of likelihood calls between iterations, the index of the point within the live set upon removal, and the number of updates to the ernest_lrps object, respectively.

Details

Nested sampling must generate a series of increasing likelihood values and a series of strictly decreasing contour volume estimates. If neither of these hold, ernest will warn the user and return a stripped down tibble with no evidence estimates or log weights. In this case, the user should check the behaviour of the sampler used to generate points, or report the issue to the package authors.