control: Control the computational details of the imputation methods

In insightsengineering/rbmi: Reference Based Multiple Imputation

Control the computational details of the imputation methods

Description

These functions control lower level computational details of the imputation methods.

Usage

control_bayes(
  warmup = 200,
  thin = 50,
  chains = 1,
  init = ifelse(chains > 1, "random", "mmrm"),
  seed = sample.int(.Machine$integer.max, 1),
  ...
)

Arguments

warmup	a numeric, the number of warmup iterations for the MCMC sampler.
thin	a numeric, the thinning rate of the MCMC sampler.
chains	a numeric, the number of chains to run in parallel.
init	a character string, the method used to initialise the MCMC sampler, see the details.
seed	a numeric, the seed used to initialise the MCMC sampler.
...	additional arguments to be passed to rstan::sampling().

Details

Currently only the Bayesian imputation via method_bayes() uses a control function:

• The init argument can be set to "random" to randomly initialise the sampler with rstan default values or to "mmrm" to initialise the sampler with the maximum likelihood estimate values of the MMRM.

• The seed argument is used to set the seed for the MCMC sampler. By default, a random seed is generated, such that outside invocation of the set.seed() call can effectively set the seed.

• The samples are split across the chains, such that each chain produces n_samples / chains (rounded up) samples. The total number of samples that will be returned across all chains is n_samples as specified in method_bayes().

• Therefore, the additional parameters passed to rstan::sampling() must not contain n_samples or iter. Instead, the number of samples must only be provided directly via the n_samples argument of method_bayes(). Similarly, the refresh argument is also not allowed here, instead use the quiet argument directly in draws().

Note

For full reproducibility of the imputation results, it is required to use a set.seed() call before defining the control list, and calling the draws() function. It is not sufficient to merely set the seed argument in the control list.

insightsengineering/rbmi documentation built on Feb. 28, 2025, 3:34 a.m.