Paired DeLong Non-Inferiority Test

Tests whether a candidate model's AUC is non-inferior to a reference AUC, using a paired DeLong test on the same predictions.

**Common bugs this prevents**:

1. Using the wrong reference AUC (e.g., superseded Phase 2.1 instead of current Phase 2.1b)

2. Using an unpaired CI check instead of a formal paired DeLong test

3. Comparing AUCs from different sample denominators

Usage

test_noninferiority(
  response,
  predictor_candidate,
  predictor_reference,
  delta = 0.03,
  alpha = 0.05
)

Arguments

response

Binary outcome vector (0/1), same for both models

predictor_candidate

Candidate model predictions

predictor_reference

Reference model predictions. Must be on the SAME samples as predictor_candidate.

delta

Non-inferiority margin (positive value, e.g., 0.03). Candidate is non-inferior if its AUC is within delta of the reference.

alpha

Significance level (default 0.05 for one-sided test)

Value

A list with:

non_inferior

Logical: TRUE if candidate is non-inferior

auc_candidate

Candidate AUC

auc_reference

Reference AUC

delta_auc

Candidate - Reference AUC

delta

Non-inferiority margin used

se_diff

Standard error of AUC difference (paired DeLong)

z_stat

Z-statistic for non-inferiority

p_value

One-sided p-value

ci_lower_diff

Lower bound of 95% CI for AUC difference

Examples

if (FALSE) { # \dontrun{
result <- test_noninferiority(
  response = y_test,
  predictor_candidate = pred_reduced_panel,
  predictor_reference = pred_full_panel,
  delta = 0.03
)
cat("Non-inferior:", result$non_inferior, "\n")
} # }