cpt_robustness: Does a changepoint survive a change of noise model?

In ggchangepoint: Combines Changepoint Analysis with 'ggplot2'

Does a changepoint survive a change of noise model?

Description

Re-runs a detector under each setting of its noise-model argument (family for stepR's SMUCE, variant for NSP, the cost for the changepoint engines, ...) and reports which changepoints appear under every setting and which only under some. A changepoint that survives a change of noise model is evidence; one that survives a bootstrap (cpt_stability()) or a change of algorithm (cpt_consensus()) may only be a feature of the noise, which resampling and other algorithms reproduce.

Usage

cpt_robustness(
  x,
  method = "pelt",
  over = "noise_model",
  settings = NULL,
  margin = 5,
  seed = NULL,
  ...
)

## S3 method for class 'ggcpt_robustness'
print(x, ...)

## S3 method for class 'ggcpt_robustness'
tidy(x, ...)

## S3 method for class 'ggcpt_robustness'
autoplot(object, ...)

Arguments

x	A numeric series, or a ggcpt fit (its series, method and penalty are used).
method	The detector, when x is a series.
over	What to vary. Only "noise_model" for now.
settings	Optional named list of settings, each a list of arguments for cpt_detect(), replacing the method's own sweep. The first is the reference.
margin	Positions within which two changepoints count as the same. Defaults to 5.
seed	Optional seed, scoped to this call.
...	Further arguments for every cpt_detect() call.
object	A ggcpt_robustness object (for autoplot()).

Value

A ggcpt_robustness object: a list with settings (one row per setting: setting, n_cp, powered and error), changepoints (one row per changepoint of the reference setting: cp, found_by (how many other powered settings find it within margin), of (how many powered alternatives there are) and survives), union and intersection (distinct locations found by any and by every powered setting) and fits. With print() and autoplot().

Powerless settings are reported, not counted

A setting that finds nothing where the reference finds several is not a vote against them: measured, NSP's variant = "ar" removes the false positives of autocorrelated noise by removing detection altogether (0 of 2 real changes found). Such a setting is marked powered = FALSE and left out of found_by.

See Also

cpt_assumptions(), cpt_sensitivity() for sweeping any argument.

Other inference: cpt_abruptness(), cpt_assumptions(), cpt_attribute(), cpt_attribute_event(), cpt_confint_coverage(), cpt_effect(), cpt_gof(), cpt_k_weights(), cpt_null_power(), cpt_screen(), cpt_test_at(), cpt_test_null()

Examples

set.seed(1)
x <- c(rnorm(100), rnorm(100, 2)) + as.numeric(arima.sim(list(ar = 0.6), 200))
cpt_robustness(x, method = "pelt")

ggchangepoint documentation built on Nov. 5, 2026, 9:06 a.m.