Dataset-Identity AUC ("Bias Floor")

Fits a logistic regression using ONLY cohort/dataset dummy variables as predictors and reports its AUC against the case/control label. This is the empirical lower bound on how much of an apparent classifier AUC can be explained by cohort identity alone, before any biomarker is used.

If the bias floor is high (e.g., > 0.70), any pooled classifier AUC on the same data inherits that share of discrimination from cohort provenance rather than biology. A responsible biomarker claim must show AUC materially above the bias floor, on data where the floor itself has been mitigated.

Usage

os_bias_floor_auc(y, cohort, cv_folds = 5L, seed = 42L)

Arguments

y

Binary outcome vector (0/1 or factor with two levels). Length n.

cohort

Vector or factor of cohort/dataset identity. Length n.

cv_folds

Integer. If non-NULL, also computes K-fold cross-validated AUC. Default 5. Set to NULL to skip CV.

seed

Random seed for CV folds. Default 42L.

Value

A list with class "os_bias_floor":

apparent_auc

Logistic regression AUC on the training data (upper bound — optimistic by design)

cv_auc

Cross-validated AUC, if requested

per_cohort

Data frame: cohort, n, case rate, mean predicted probability

interpretation

Character string summarizing the floor

Examples

if (FALSE) { # \dontrun{
y <- c(rep(1, 317), rep(0, 200), rep(1, 40), rep(0, 100))
cohort <- c(rep("A", 317), rep("A", 200), rep("B", 40), rep("B", 100))
bf <- os_bias_floor_auc(y, cohort)
print(bf)
} # }