group_fairness {SPARRAfairness}

group_fairness

Description

Estimates group fairness metric according to a specification vector of the form

Usage

group_fairness(
  specs,
  scores,
  target,
  group1,
  group2,
  cutoffs = seq(min(scores, na.rm = TRUE), max(scores, na.rm = TRUE), length = 100)
)

Arguments

specs

specification vector; see description

scores

vector of risk scores

target

vector of values of target (which risk score aims to predict)

group1

indices of group 1

group2

indices of group 2

cutoffs

score cutoffs at which to estimate metric (default 100 evenly-spaced)

Details

c(A1,B1,C1,A2,B2,C2)

encoding a probability

P(A1,B1,C1|A2,B2,C2)

where

A1/A2 are events coded by 1:'score>= cutoff'; 0: 'score<cutoff' and NA: 1/TRUE B1/B2 are events coded by 1:'target=TRUE'; 0: 'target=FALSE' and NA: 1/TRUE C1/C2 are events coded by 1:'group=g'; and NA: 1/TRUE

For example, specs=c(NA,1,NA,0,NA,1) would encode false omission rate:

P(target=TRUE|score<cutoff,group=g)

Value

matrix of dimension length(cutoffs)x4, with (i,2g-1)th entry the relevant fairness metric for group g at the ith cutoff value and (i,2g)th entry the approximate standard error of the (i,2g-1)th entry

Examples

# See vignette