Skip to contents

Combines out-of-fold prediction error with rescaled BIC, CAIC, or CICOMP penalties. The tuning constant lambda controls the contribution of the information-criterion component.

Usage

select_dimension_cv_icomp(
  X,
  y,
  sdr_method = c("dr", "sir", "save", "phd"),
  cov_method = c("sample", "ridge", "oas", "lw", "mec"),
  d_max = 10,
  v = 5,
  nslices = 6,
  standardize = TRUE,
  stabilize = TRUE,
  stabilization = c("eigenfloor", "ridge", "nearest_pd"),
  complexity = c("C1", "C1F"),
  lambda = 1,
  seed = NULL,
  cov_args = list(),
  stabilization_args = list(),
  sdr_args = list(),
  ...
)

Arguments

X

Numeric predictor matrix or data frame.

y

Numeric continuous response vector.

sdr_method

SDR method.

cov_method

Covariance estimator.

d_max

Maximum candidate structural dimension.

v

Number of cross-validation folds.

nslices

Number of response slices.

standardize

Logical. Standardise within each training fold.

stabilize

Logical. Stabilise the estimated covariance matrix.

stabilization

Covariance stabilisation method.

complexity

Covariance complexity measure.

lambda

Non-negative information-criterion weight.

seed

Optional random seed.

cov_args

Named list of covariance-estimator arguments.

stabilization_args

Named list of stabilisation arguments.

sdr_args

Named list of SDR-kernel arguments.

...

Backward-compatible component arguments.

Value

A list containing selected dimensions, the aggregated cross-validation table, and fold-level results.