Matched AUC difference across repeated CV strata

Computes fixed-direction AUC for two methods on exactly aligned held-out rows and applies singlesample_corrected_repeated_cv() to the stratum-level paired differences. The default preserves the historical group-collapsed estimand; analysis_level = "profile" gives every held-out profile equal weight while retaining provenance-group-safe folds. Larger scores must already indicate the positive class according to a training-frozen orientation.

Usage

singlesample_matched_pair_auc(
  y,
  score_a,
  score_b,
  stratum_id,
  sample_id,
  group_id,
  expected_strata,
  analysis_level = c("group", "profile"),
  margin = 0.05,
  conf_level = 0.95
)

Arguments

y

Binary held-out outcome.

score_a, score_b

Finite held-out scores for methods A and B on the same rows.

stratum_id

Repeated-CV stratum per row, such as `seed::fold`.

sample_id

Held-out profile identifier. Each profile may occur only once in a stratum.

group_id

Biological/provenance group identifier. Group labels must be homogeneous.

expected_strata

Complete prespecified repeated-CV strata.

analysis_level

Either "group" (collapse repeated profiles within biological/provenance group) or "profile" (equal profile weight).

margin

Positive relevance margin in AUC units.

conf_level

Confidence level for the nominal corrected t interval.

Value

A list with one `strata` data frame and one corrected `summary` list.