Clean nuisance regressors by dropping rank-useless columns

Drops nuisance columns that are non-finite, zero-variance, or fail to increase QR rank after the block's baseline terms and earlier nuisance columns. Column order is respected, so when two columns are aliased the earlier column is kept.

Usage

clean_nuisance(
  nuisance_list,
  sframe,
  basis = c("constant", "poly", "bs", "ns"),
  degree = 1,
  intercept = c("runwise", "global", "none"),
  tol = sqrt(.Machine$double.eps),
  duplicate_threshold = 1 - sqrt(.Machine$double.eps),
  na_action = c("drop", "zero", "median")
)

Arguments

nuisance_list

A list of numeric matrices or data frames, one per block.

sframe

A sampling frame.

basis, degree, intercept

Baseline model settings used to construct the comparison baseline terms.

tol

Numeric tolerance passed to QR rank checks.

duplicate_threshold

Absolute correlation threshold used to flag duplicate or near-duplicate columns.

na_action

Character; how to handle NA values in nuisance_list columns before the diagnostics run. "drop" (default) leaves NAs in place so any column containing one is treated as non-finite. "zero" replaces NA with 0 (matching the fMRIPrep leading-row convention) and "median" replaces NA with the column median; both repair an isolated leading NA (e.g. in DVARS or framewise displacement) so the regressor is retained. NaN and Inf are never repaired and remain non-finite.

Value

A list with nuisance_list and report elements. Pass result$nuisance_list to baseline_model().