SBC shows failures when rejecting individual data points

sbc_rejection_per_datapoint.R

library(SBC)

library(cmdstanr)

library(bayesplot)
library(posterior)

library(future)
plan(multisession) 

options(SBC.min_chunk_size = 10)

m_code <- " data {
  int<lower=0> N;
  array[N] real y;
}

parameters {
  real mu;
}

model {
  mu ~ normal(0, 2);
  y ~ normal(mu, 1);
}
"

backend <- SBC_backend_cmdstan_sample(cmdstan_model(write_stan_file(m_code)), iter_warmup = 800, iter_sampling = 800)


N <- 10
generator <- SBC_generator_function(function() {
  mu <- rnorm(1, 0, 2)
  
  # Reject observations one at a time
  y <- rep(-Inf, N)
  repeat {
    bad_indices <- y < -1
    if(!any(bad_indices)) {
      break
    }
    y[bad_indices] <- rnorm(sum(bad_indices), mu, 1)
  }
  list(
    variables = list(mu = mu),
    generated = list(N = N, y = y)
  )
})

set.seed(2323455)
datasets <- generate_datasets(generator, 100)

results <- compute_SBC(datasets, backend, keep_fits = FALSE)

plot_ecdf_diff(results)