Audit whether an algorithm's area risk ranking matches realised outcomes

Ranks areas by predicted risk and by realised outcome rate (rank 1 = highest), forms rank_gap = outcome_rank - risk_rank per area (positive = over-predicted), and averages the gap within each group. A Spearman correlation summarises overall calibration.

Usage

morie_predpol_calibration_audit(areas, mean_risk, outcome_rate, group)

Arguments

areas: Area identifiers (one per area).
mean_risk: Mean predicted risk per area.
outcome_rate: Realised outcome rate per area.
group: Majority protected-attribute label per area.

Value

A named list: value (worst per-group mean gap), spearman, spearman_pvalue, group_rank_gap, worst_group, rank_gap, warnings, interpretation.

Examples

res <- morie_predpol_calibration_audit(
  areas = c("d1", "d2", "d3", "d4", "d5", "d6"),
  mean_risk = c(90, 80, 70, 30, 20, 10),
  outcome_rate = c(10, 20, 30, 70, 80, 90),
  group = c("X", "X", "X", "Y", "Y", "Y")
)
res$group_rank_gap$X # 3  (group X over-predicted)
#> [1] 3
res$spearman # -1 (perfectly miscalibrated)
#> [1] -1