Empirical-Bayes shrinkage funnel / caterpillar

Visualizes empirical-Bayes shrinkage by drawing one row per facet level with the raw (pre-shrinkage) and shrunken estimates plus the shrinkage factor. Rows are ordered by absolute shrinkage so the levels that move most under the prior appear at the top.

Usage

plot_shrinkage_funnel(
  fit,
  facet = NULL,
  top_n = 30L,
  preset = c("standard", "publication", "compact", "monochrome"),
  show_ci = FALSE,
  ci_level = 0.95,
  draw = TRUE
)

Arguments

fit

An mfrm_fit augmented with empirical-Bayes shrinkage.

facet

Facet to draw (default: first non-person facet with shrinkage columns present).

top_n

Maximum number of rows to draw (default 30).

preset

Visual preset.

show_ci

Logical. When TRUE, draw approximate confidence-interval whiskers for raw and shrunken estimates when SE / ShrunkSE evidence is available.

ci_level

Confidence level used when show_ci = TRUE; default 0.95.

draw

If TRUE, draw with base graphics.

Value

An mfrm_plot_data whose data slot bundles the long Level, RawEstimate, ShrunkEstimate, ShrinkageFactor table. When show_ci = TRUE, the table also includes RawCI_Lower, RawCI_Upper, ShrunkCI_Lower, ShrunkCI_Upper, and CI_Level.

Details

Requires a fit produced via apply_empirical_bayes_shrinkage() or a fit_mfrm(..., facet_shrinkage = "empirical_bayes") run, so that fit$facets$others carries Estimate, ShrunkEstimate, and ShrinkageFactor columns.

See also

apply_empirical_bayes_shrinkage(), mfrmr_visual_diagnostics

Examples

if (FALSE) { # interactive()
toy <- load_mfrmr_data("example_core")
fit <- fit_mfrm(toy, "Person", c("Rater", "Criterion"), "Score",
                 method = "JML", maxit = 30)
fit_eb <- apply_empirical_bayes_shrinkage(fit)
p <- plot_shrinkage_funnel(fit_eb, draw = FALSE)
head(p$data$table)
# Look for: short segments (Raw and Shrunken close together) =
#   little pooling. Long segments fanning toward the centre = the
#   prior pulled the estimate strongly; this is most pronounced for
#   small-N levels. ShrinkageFactor near 1 means most of the
#   movement was driven by the prior rather than the data.
}