Plot SBC rank ECDF with simultaneous confidence band

Plots the empirical CDF of SBC ranks against the uniform CDF (the diagonal), with a simultaneous confidence band following Säilynoja, Bürkner, and Vehtari (2022). The band is calibrated so that, under correct calibration, the entire ECDF stays within it with probability alpha; deviations anywhere along the band therefore indicate miscalibration at level 1 - alpha.

Ranks are normalized per task: each task's ranks are scaled by that task's own support, (rank + 0.5) / n_ranks with n_ranks = support + 1 (kept post-thinning draws + 1). When tasks in a panel have different supports (e.g. rank_metric() with thin = "auto" under autocorrelation), pooling on the panel maximum would squash small-support tasks' ranks toward zero and manufacture apparent miscalibration; per-task normalization avoids that artifact, and a warning notes that the simultaneous band, which assumes iid ranks on a common support, is then approximate. Legacy results without a recorded n_ranks fall back per task to n_draws; with a historical thinning stride > 1 the true support is unknown, so that fallback only bounds it.

Usage

plot_rank_ecdf(ranks, alpha = 0.95, by = NULL)

Arguments

ranks: A tibble from sbc_ranks(), or a bayesim_simulation_result.
alpha: Coverage level of the simultaneous confidence band (default 0.95).
by: Optional character vector of condition columns to facet by. These columns are preserved by sbc_ranks() for simulation results. Using by computes a separate ECDF and simultaneous band per condition cell instead of pooling ranks across cells.

Value

A ggplot object.

References

Säilynoja T, Bürkner PC, Vehtari A (2022). Graphical test for discrete uniformity and its applications in goodness-of-fit evaluation. Statistics and Computing, 32(2).

Examples

if (FALSE) { # \dontrun{
plot_rank_ecdf(sbc_ranks(result))
plot_rank_ecdf(sbc_ranks(result), alpha = 0.99)
plot_rank_ecdf(sbc_ranks(result), by = "data_n")
} # }