Offline segmentation of daily counts by penalised likelihood

Binary segmentation of the daily-count series: a segment is split at the candidate day maximising the summed ensemble log-likelihood of the two halves, and the split is kept only if it lowers the Bayesian information criterion, with the split charged for the parameters of one additional curve and the change location. The parameter count of an ensemble is the weight-averaged count of its members (three per curve with Poisson counts, four with negative binomial counts): the members are alternative descriptions of a single mean, not additive components. Each half is fitted by fit_ensemble_incidence().

Usage

bic_segment(
  x,
  d_min = 20L,
  family = c("poisson", "negbin"),
  weights = c("akaike", "inverse_mse"),
  beta = 1.5,
  step = 2L,
  max_cuts = 10L
)

Arguments

x

Daily counts.

d_min

Minimum segment length.

family, weights, beta

Passed to fit_ensemble_incidence().

step

Spacing of the coarse candidate grid (days); the optimum is then refined at unit spacing within step days.

max_cuts

Maximum number of change points.

Value

A list with cuts and segments.