metrosamp: Perform Metropolis sampling with batch averaging
In JGCRI/metrosamp: Basic Metropolis Sampler for Markov Chain Monte Carlo

Description Usage Arguments Details Value To Do

View source: R/metrosamp.R

This function performs a basic Metropolis sampling of a user-supplied log-posterior function. The sampling is done in batches, with the batch means returned as the output. Setting the batch length to 1 will produce unbatched samples.

1 2	metrosamp(lpost, p0, nsamp, batchlen, scale = NULL, debug = FALSE, lp0 = NA)

`lpost`	Log-posterior function
`p0`	Starting parameters for sampling, OR a `metrosamp` structure from a previous run.
`nsamp`	Number of batches to run
`batchlen`	Number of samples per batch
`scale`	MC step scaler; this will be multiplied by a vector of standard normal deviates to get the proposal step. Optional if a `metrosamp` structure was supplied for `p0`; required otherwise.
`debug`	Flag to turn on additional debugging information.
`lp0`	Log-posterior for the starting parameters (p0). If not supplied it will be calculated automatically.

The output metrosamp structure will be a list with the following elements:

samples: Matrix (nsamp x nparam) of parameter samples
samplp: Vector (nsamp) of log-posterior values for the samples
accept: Probability of accepting a proposal, averaged across all samples
plast: Last parameter set. Can be used to continue sampling where the last run left off.
scale: Scale factor used in the calculation. Also useful for continuing a run.

If the debug flag is set, the output will have some additional elements that can be used to diagnose the sampling procedure. If batch sampling is in use, then most of these will pertain to the last proposal evaluated in each batch. Intermediate proposals within a batch are not returned. Therefore, when debugging proposals, it is best to use a batch length of 1.

proposals: The proposal parameters evaluated by the sampler.
proplp: Log-posterior for the proposals. The same notes apply as to the proposals entry.
prop_accepted: Flag indicating whether each proposal was accepted.

A run can be continued by passing the metrosamp structure from the previous run as the p0 argument. If this is done, then the scale parameter may be omitted, and the new run will use the same scale as the old. If a scale parameter is supplied, then it will override the scale parameter stored in the old structure.

A metrosamp structure of Monte Carlo outputs (described in Details).

Store the last log-posterior value, so continuation runs don't have to recompute it.
Add code to compute MCSE.
Add code to compute Neff.
Add option to run functions on MC samples.
Allow covariance matrix for scale parameter.