admixturegraph (version 1.0.2)

run_metropolis_hasting: Run a Metropolis-Hasting MCMC to sample graph parameters.

Description

The MCMC performs a random walk in transformed parameter space (edge lengths are log transformed and admixture proportions inverse Normal distribution transformed) and from this explores the posterior distribution of graph parameters.

Usage

run_metropolis_hasting(model, initial_state, iterations, no_temperatures = 1, cores = 1, no_flips = 1, max_tmp = 100, verbose = TRUE)

Arguments

model

Object constructed with make_mcmc_model.

initial_state

The initial set of graph parameters.

iterations

Number of iterations to sample.

no_temperatures

Number of chains in the MC3 procedure.

cores

Number of cores to spread the chains across. Best performance is when cores = no_temperatures.

no_flips

Mean number of times a flip between two chains should be proposed after each step.

max_tmp

The highest temperature.

verbose

Logical value determining if a progress bar should be shown during the run.

Value

A matrix containing the trace of the chain with temperature 1.

Details

Using the posterior distribution of parameters is one approach to getting parameter estimates and a sense of their variability. Credibility intervals can directly be obtained from sampled parameters; to get confidence intervals from the likelihood maximisation approach requires either estimating the Hessian matrix for the likelihood or a boot-strapping approach to the data.

From sampling the likelihood values for each sample from the posterior we can also compute the model likelihood (the probability of the data when integrating over all the model parameters). This gives us a direct way of comparing graphs since the ratio of likelihoods is the Bayes factor between the models. Comparing models using maximum likelihood estimtates is more problematic since usually graphs are not nested models.