run_stage: Run a stage of the PMwG sampler

In pmwg: Particle Metropolis Within Gibbs

Run a stage of the PMwG sampler

Description

Run one of burnin, adaptation or sampling phase from the PMwG sampler. Each stage involves slightly different processes, so for the full PMwG sampling we need to run this three times.

Usage

run_stage(
  pmwgs,
  stage,
  iter = 1000,
  particles = 100,
  display_progress = TRUE,
  n_cores = 1,
  n_unique = ifelse(stage == "adapt", 100, NA),
  epsilon = NULL,
  p_accept = 0.8,
  mix = NULL,
  pdist_update_n = ifelse(stage == "sample", 50, NA)
)

Arguments

pmwgs	A Particle Metropolis within Gibbs sampler which has been set up and initialised
stage	The sampling stage to run. Must be one of 'burn', 'adapt' or 'sample'.
iter	The number of iterations to run for the sampler. For 'burn' and 'sample' all iterations will run. However for 'adapt' if all subjects have enough unique samples to create the conditional distribution then the stage will finish early.
particles	The default here is 1000 particles to be generated for each iteration, however during the sample phase this should be reduced.
display_progress	Display a progress bar during sampling.
n_cores	Set to more than 1 to use mclapply. Setting n_cores greater than 1 is only permitted on Linux and Mac OS X machines.
n_unique	A number representing the number of unique samples to check for on each iteration of the sampler (An initial test for the generation of the proposal distribution). Only used during the 'adapt' stage.
epsilon	A value between 0 and 1 that controls the extent to which the covariance matrix is scaled when generating particles from the previous random effect. The default will be chosen based on the number of random effects in the model.
p_accept	A value between 0 and 1 that will flexibly tune epsilon to achieve an acceptance ratio close to what you set p_accept to. The default is set at 0.8.
mix	A vector of floats that controls the mixture of different sources for particles. The function numbers_from_proportion is passed this value and includes extra details on what is accepted.
pdist_update_n	The number of iterations in the sample stage after which the proposal distribution will be recomputed.

Details

The burnin stage by default selects 50 parameter sample (selected through a Gibbs step) and 50 the previous random effect of each subject. It assesses each particle with the log-likelihood function and samples from all particles weighted by their log-likelihood.

The adaptation stage selects and assesses particle in the same was as burnin, however on each iteration it also checks whether each subject has enough unique random effect samples to attempt to create a conditional distribution for efficient sampling in the next stage. If the attempt at creating a conditional distribution fails, then the number of unique samples is increased and sampling continues. If the attempt succeeds then the stage is finished early.

The final stage (sampling) by default samples predominantly from the conditional distribution created at the end of adaptation. This is more efficient and allows the number of particles to be reduced whilst still getting a high enough acceptance rate of new samples.

Once complete each stage will return a sampler object with the new samples stored within it.

The progress bar (which is displayed by default) shows the number of iterations out of those requested which have been completed. It also contains additional information at the end about the number of newly generated particles that have been accepted. This is show as New(XXX average across subjects of newly sampled random effects accept rate. See accept_rate for more detail on getting individual accept rate values per subject.

Value

A pmwgs object with the newly generated samples in place.

Examples

library(rtdists)
sampled_forstmann$data <- forstmann
run_stage(sampled_forstmann, "sample", iter = 1, particles = 10)

pmwg documentation built on May 29, 2024, 12:09 p.m.