R: Convergence Diagnostics for Markov Chain Monte Carlo of Life...

e0.diagnose {bayesLife}

R Documentation

Convergence Diagnostics for Markov Chain Monte Carlo of Life Expectancy

Description

Function e0.diagnose runs convergence diagnostics of existing MCMCs, using the raftery.diag function from the coda package.

Usage

e0.diagnose(sim.dir, thin = 225, burnin = 10000, express = FALSE, 
    country.sampling.prop = NULL, keep.thin.mcmc = FALSE, verbose = TRUE)

Arguments

`sim.dir`	Directory with the MCMC simulation results.
`thin`	Thinning interval.
`burnin`	Number of iterations to be discarded from the beginning of the parameter traces.
`express`	Logical. If `TRUE`, the convergence diagnostics is run only on the country-independent parameters. If `FALSE`, the country-specific parameters are included in the diagnostics. The number of countries can be controlled by `country.sampling.prop`.
`country.sampling.prop`	Proportion of countries that are included in the diagnostics. If it is `NULL` and `express=FALSE`, all countries are included. Setting here a number between 0 and 1, one can limit the number of countries which are then randomly sampled. Note that for long MCMCs, this argument may significantly influence the run-time of this function.
`keep.thin.mcmc`	Logical. If `TRUE` the thinned traces used for computing the diagnostics are stored on disk (see `create.thinned.e0.mcmc`).
`verbose`	Logical switching log messages on and off.

Details

The function invokes the e0.raftery.diag function separately for country-independent parameters and for country-specific parameters. It results in two possible states: red, i.e. it did not converge, and green, i.e. it converged. The resulting object is stored in
‘{sim.dir}/diagnostics/bayesLife.convergence_{thin}_{burnin}.rda’ and can be accessed using the function get.e0.convergence.

Function has.mcmc.converged from the bayesTFR package can be used to check if the existing diagnostics converged.

Value

e0.diagnose returns an object of class bayesLife.convergence with components:

`result`	Table containing all not-converged parameters. Its columns include ‘Total iterations needed’ and ‘Remaining iterations’.
`lresult.country.independent`	Number of rows in `result` that correspond to country-independent paramters. These rows are groupped at the beginning of the table.
`country.independent`	Result of `e0.raftery.diag` processed on country-independent parameters.
`country.specific`	Result of `e0.raftery.diag` processed on country-specific parameters.
`iter.needed`	Number of additional iterations suggested in order to achieve convergence.
`iter.total`	Total number of iterations of the original unthinned set of chains.
`use.nr.traj`	Suggestion for number of trajectories in generating predictions.
`burnin`	Burnin used.
`thin`	Thinning interval used.
`status`	Vector of character strings containing the result status. Possible values: ‘green’, ‘red’.
`mcmc.set`	Object of class `bayesLife.mcmc.set` that corresponds to the original set of MCMCs on which the diagnostics was run.
`thin.mcmc`	If `keep.thin.mcmc` is `TRUE`, it is an object of class `bayesLife.mcmc.set` that corresponds to the thinned mcmc set on which the diagnostics was run, otherwise `NULL`.
`express`	Value of the input argument `express`.
`nr.countries`	Vector with elements `used` - number of countries used in this diagnostics, and `total` - number of countries that this `mcmc.set` object was estimated on.

Author(s)

Hana Sevcikova, Adrian Raftery