Fit the base imputation model using a Bayesian approach

fit_mcmc() fits the base imputation model using a Bayesian approach. This is done through a MCMC method that is implemented in stan and is run by using the function rstan::sampling(). The function returns the draws from the posterior distribution of the model parameters and the stanfit object. Additionally it performs multiple diagnostics checks of the chain and returns warnings in case of any detected issues.

Usage

fit_mcmc(designmat, outcome, group, subjid, visit, method, quiet = FALSE)

Arguments

designmat: The design matrix of the fixed effects.
outcome: The response variable. Must be numeric.
group: Character vector containing the group variable.
subjid: Character vector containing the subjects IDs.
visit: Character vector containing the visit variable.
method: A method object as generated by method_bayes().
quiet: Specify whether the stan sampling log should be printed to the console.

Value

A named list composed by the following:

samples: a named list containing the draws for each parameter. It corresponds to the output of extract_draws().
fit: a stanfit object.

Details

The Bayesian model assumes a multivariate normal likelihood function and weakly-informative priors for the model parameters: in particular, uniform priors are assumed for the regression coefficients and inverse-Wishart priors for the covariance matrices. The chain is initialized using the REML parameter estimates from MMRM as starting values.

The function performs the following steps:

Fit MMRM using a REML approach.
Prepare the input data for the MCMC fit as described in the data{} block of the Stan file. See prepare_stan_data() for details.
Run the MCMC according the input arguments and using as starting values the REML parameter estimates estimated at point 1.
Performs diagnostics checks of the MCMC. See check_mcmc() for details.
Extract the draws from the model fit.

The chains perform method$n_samples draws by keeping one every method$burn_between iterations. Additionally the first method$burn_in iterations are discarded. The total number of iterations will then be method$burn_in + method$burn_between*method$n_samples. The purpose of method$burn_in is to ensure that the samples are drawn from the stationary distribution of the Markov Chain. The method$burn_between aims to keep the draws uncorrelated each from other.