Skip to contents

Bayesian estimation diagnostics (leverage / influence / prior influence)

Source: R/analysis.R
estimation_diagnostics.Rd

Fit a univariate distribution to a sample with a Bayesian MCMC analysis and compute the three estimation diagnostics: observation leverage (from the posterior Hessian at the MAP), PSIS-LOO influence (per-observation Pareto-k), and prior influence (prior-vs-data log-likelihood share). Wraps the shared C++ BayesianAnalysis diagnostics.

Usage

estimation_diagnostics(
  data,
  distribution,
  sampler = "DEMCz",
  iterations = 3000L,
  output_length = 10000L,
  seed = 12345L,
  thinning_interval = -1L,
  thin_every = 10L
)

Arguments

data

numeric vector of observations.

distribution

distribution family name (any name the core distribution factory accepts).

sampler

MCMC sampler: "DEMCz" (default), "DEMCzs", "ARWMH", or "NUTS".

iterations

number of post-warmup MCMC iterations.

output_length

number of posterior samples retained.

seed

PRNG seed for the sampler.

thinning_interval

MCMC thinning interval; -1 (default) keeps the sampler's own default.

thin_every

prior-influence posterior thinning stride (default 10).

Value

A named list with three sub-lists: leverage (index, leverage, fit_influence, variance_influence, value arrays; prior_leverage; total_leverage, total_fit_influence, total_variance_influence), influence (pareto_k, elpd_loo arrays; count, mean_pareto_k, max_pareto_k, count_pareto_k_above_05/_07/_10, proportion_problematic, is_reliable), and prior_influence (count, prior_precision_share, total_prior_log_likelihood, total_data_log_likelihood, prior_to_data_ratio, is_prior_influential, mean_prior_precision_share).

Examples

peaks <- c(12500, 15300, 8900, 22100, 18700, 14200, 9800, 28500, 17400, 11600)
d <- estimation_diagnostics(peaks, "Normal", iterations = 100, output_length = 200,
                            seed = 12345, thinning_interval = 1)
d$influence$max_pareto_k
#> [1] 1.5