Bias Amplification Score (parity gap x Gini of group rates)

BAS = delta_parity * G, where delta_parity is the demographic-parity gap of the worst-affected group and G is the Gini coefficient of the per-group favourable-outcome rates. Large only when a directional disparity coincides with high cross-group inequality.

BAS = Delta_parity * G, where Delta_parity is the demographic parity gap of the worst-affected group and G is the Gini coefficient of the per-group favourable-outcome rates. Large only when a directional disparity coincides with high overall inequality.

Usage

morie_fairness_bias_amplification(
  y_pred,
  group,
  privileged = NULL,
  favorable = 1
)

morie_fairness_bias_amplification(
  y_pred,
  group,
  privileged = NULL,
  favorable = 1
)

Arguments

y_pred

Vector of decisions/assignments, one per individual.

group

Vector of protected-attribute values (e.g. race).

privileged

The reference group. If NULL (default) the highest-rate group is used and a warning is attached.

favorable

The value of y_pred counted as the favourable outcome (default 1).

Value

A morie_fairness_result; headline value is BAS.

A named list: value (BAS), bias_amplification_score, demographic_parity_gap, gini, rates, privileged, warnings, interpretation.

Details

Reimplemented from Barman & Barman, arXiv:2603.18987.

Reimplemented from Barman & Barman, "Unmasking Algorithmic Bias in Predictive Policing" (arXiv:2603.18987).

Examples

pred <- c(1, 1, 1, 1, 0, 0, 0, 0)
race <- c(rep("A", 4), rep("B", 4))
res <- morie_fairness_bias_amplification(pred, race, privileged = "A")
res$value # -0.5  (parity gap -1.0 times Gini 0.5)
#>    B 
#> -0.5