TMB (version 1.7.4)

run_mcmc: MCMC sampling of TMB models

Description

[BETA VERSION] Draw samples from the posterior of a TMB model using a specified MCMC algorithm.

Usage

run_mcmc(obj, nsim, algorithm, params.init = NULL, covar = NULL, diagnostic = FALSE, ...)

Arguments

obj

A TMB model object.

nsim

The number of (dependent) samples to draw.

algorithm

A string specifiying an algorithm. Currently supported are:

• "RWM"the random walk Metropolis sampler
• "HMC"the Hamiltonian sampler (see Neal 2011)
• "NUTS"the No-U-Turn sampler (see Hoffman and Gelman 2014)

These algorithms require different arguments; see their help files for more information.

params.init

The initial parameter vector. The default of NULL signifies to use the starting values present in the model (i.e., obj$par).

covar

An optional covariance matrix which can be used to improve the efficiency of sampling. The lower Cholesky decomposition of this matrix is used to transform the parameter space. If the posterior is approximately multivariate normal and covar approximates the covariance, then the transformed parameter space will be close to multivariate standard normal. In this case the algorithm will be more efficient, but there will be overhead in the matrix calculations which need to be done at each step. The default of NULL specifies to not do this transformation.

diagnostic

Whether to return diagnostic information about chain. See individual algorithm for more information.

...

Further arguments to be passed to the algorithm. See help files for the samplers for further arguments.

Value

If diagnostic is FALSE, returns a data frame with posterior samples. Otherwise it returns a list containing the samples and properties of the sampler useful for diagnosing behavior and efficiency.

Details

This function is a top-level wrapper designed specifically to work with TMB models. There are several MCMC algorithms available for use. The user is responsible for specifying the model properly (priors, starting values, desired parameters fixed, etc.), as well as assessing the convergence and validity of the resulting samples (e.g., through the coda package) before making inference.

See Also

run_mcmc.hmc, run_mcmc.nuts, run_mcmc.rwm

Examples

## Not run: 
# ## Draw samples with the RWM, HMC and NUTS algorithms and compare.
# 
# ## Run the simple example, so that obj and opt are loaded into workspace
# runExample("simple")
# 
# ## Run RWM and two gradient based algorithms, using adative step size (eps)
# ## for each. Start from the MLE.
# rwm <- run_mcmc(obj=obj, nsim=500*8, algorithm='RWM', params.init=opt$par,
#             alpha=.08, diagnostic=TRUE)
# ## Thin it to better approximate the gradient methods
# rwm$par <- rwm$par[seq(1, nrow(rwm$par), by=8),]
# hmc <- run_mcmc(obj=obj, nsim=500, algorithm='HMC', L=8, params.init=opt$par,
#             diagnostic=TRUE, eps=0.1)
# nuts <- run_mcmc(obj=obj, nsim=500, algorithm='NUTS', params.init=opt$par,
#              diagnostic=TRUE, eps=0.1)
# 
# ## See how they compare via ACF
# par(mfrow=c(3,4))
# for(i in 1:4) acf(rwm$par[,i])
# for(i in 1:4) acf(hmc$par[,i])
# for(i in 1:4) acf(nuts$par[,i])
# ## End(Not run)