Synthetic predictive-policing dataset with a known disparity — morie_fairness_simulate_biased_crime

Generates per-record data with area, group, true_outcome (group-independent Bernoulli at base_rate), detected (group-dependent), and risk_score (0–500, shifted up by bias * 100 points for non-reference groups). The bias input is the ground truth the audits should recover.

Usage

morie_fairness_simulate_biased_crime_data(
  n = 2000L,
  groups = c("A", "B"),
  group_props = NULL,
  n_areas = 20L,
  base_rate = 0.3,
  bias = 0.5,
  seed = 0L
)

Arguments

n: Number of records.
groups: Character vector of group labels (the first entry is treated as the reference group).
group_props: Optional sampling proportions.
n_areas: Number of areas (>= number of groups).
base_rate: Reference-group favourable-outcome rate in 0–1.
bias: Injected disparity in -1–1.
seed: Reproducibility seed.

Value

A data.frame with columns area, group, true_outcome, detected, risk_score.

Examples

d <- morie_fairness_simulate_biased_crime_data(n = 100L, seed = 1L)
head(d)
#>      area group true_outcome detected risk_score
#> 1 area_01     B            0        1   260.5669
#> 2 area_09     B            0        0   184.4432
#> 3 area_10     A            1        0   224.3807
#> 4 area_00     A            0        1   272.8203
#> 5 area_17     B            0        0   297.6111
#> 6 area_12     A            1        0   246.0729