Assess Numerical Stability of Correlated Profile Completion

Repeats the existing conditional truncated-MVN Gibbs profile completion for fixed posterior draws and a fixed, reproducibly selected subset of incomplete events. It does not refit Stan or alter the fitted model. The comparison therefore isolates finite Gibbs-chain length. The conditional completion algorithm is the correct profile-completion algorithm; this diagnostic checks whether its finite Monte Carlo run length is adequate for a particular fit, especially when posterior residual correlations are high.

Usage

assess_gibbs_profile_stability(
  fitted_model,
  schedules = list(baseline = list(burnin = 30L, kept = 50L), medium = list(burnin = 60L,
    kept = 100L), long = list(burnin = 100L, kept = 250L)),
  n_posterior_draws = 200L,
  max_events = 250L,
  seed = 123L,
  tolerance = 0.01,
  event_selection = c("stratified", "all")
)

Arguments

fitted_model

Object returned by [fit_bayesian_multivariate_probit()].

schedules

Named list of `list(burnin=, kept=)` schedules. It must contain `baseline`, `medium`, and `long`.

n_posterior_draws

Number of fixed posterior draws to use.

max_events

Maximum number of incomplete events to assess.

seed

Seed used once to select draws/events and then reset before each schedule. Thus schedules share posterior draws, events, and RNG streams.

tolerance

Maximum aggregate difference for status `"pass"`; twice this value is the `"warning"` upper bound.

event_selection

Currently `"stratified"` or `"all"`.

Value

A structured list with summary, event/aggregate comparisons, selected events, schedules, posterior draw indices, and a separate stability status.

Details

The default 0.01/0.02 thresholds are numerical-stability conventions for this diagnostic, not universal statistical thresholds.