by Kenneth Enevoldsen | 2021-05-15
Propublica and COMPAS
Well-calibrated, statistical parity
Balance for the positive class
Balance for the negative class
A person in group $a$ and a person in group $b$ should have equal probability to be assigned to either class positive or negative class.
$$ P(R = + | A=a) = P(R = + | A = b) \quad \forall a, b\in A $$
Where $R$ is the predicted response variable.
The average score received by a positive case (unknown to the model) should be the same in each group.
$$ E(S |Y=+, A=a) = E(S |Y=+, A=b) \quad \forall a, b\in A $$
Where $S$ is the probability score, $Y$ is the actual response variable.
Similar to before:
$$ E(S |Y=-, A=a) = E(S |Y=-, A=b) \quad \forall a, b\in A $$
Where $S$ is the probability score, $Y$ is the actual response variable.
“Consider an instance of the problem in which there is a risk assignment satisfying the [three fairness conditions]. Then the instance must either allow for perfect prediction […] or have equal base rates."
The approximate version of 1.1.
approximately fair $\Rightarrow$ approximately perfect prediction or approximately equal base rate
One of the following must hold:
the test’s probability estimates are systematically skewed upward or downward for at least one gender
the test assigns a higher average risk estimate to healthy people, in one gender than the other
the test assigns a higher average risk estimate to carriers of the disease in one gender than the other