Audit Posterior Parity Across rstan and cmdstanr — audit_backend

Compares posterior summaries from two independently sampled fits. Mean differences are evaluated relative to the combined Monte Carlo standard error rather than requiring identical draws. Standard-deviation differences are reported separately. A parity pass is a computational consistency check, not evidence that either model is statistically or substantively adequate.

Usage

audit_backend_parity(
  rstan_fit,
  cmdstanr_fit,
  variables = NULL,
  mcse_multiplier = 3,
  absolute_tolerance = 0,
  relative_sd_tolerance = 0.1
)

Arguments

rstan_fit: A gp3bayes rstan fit, or a compatible posterior-summary data frame for testing/auditing.
cmdstanr_fit: A gp3bayes cmdstanr fit, or a compatible summary table.
variables: Optional parameter names to compare. When omitted, the package's approved population/group-scale parameter set is used.
mcse_multiplier: Multiplier applied to the combined MCSE of posterior means to define a sampling-noise comparison band.
absolute_tolerance: Minimum absolute tolerance for mean differences.
relative_sd_tolerance: Review threshold for relative posterior-SD differences.

Value

A gp3bayes_backend_parity_audit.

Examples

rstan_summary <- data.frame(
  variable = c("b_Intercept", "b_conditiontreatment"),
  mean = c(-0.6, 0.4), sd = c(0.20, 0.15),
  mcse_mean = c(0.01, 0.01)
)
cmdstanr_summary <- data.frame(
  variable = c("b_Intercept", "b_conditiontreatment"),
  mean = c(-0.59, 0.41), sd = c(0.21, 0.15),
  mcse_mean = c(0.01, 0.01)
)
audit_backend_parity(rstan_summary, cmdstanr_summary)
#> <gp3bayes_backend_parity_audit>
#>   Status: pass
#>   Parameters compared: 2
#>   Review parameters: 0
#>   Identical draws expected: FALSE