TODO

[snotskie]

• NA imputation - if an NA should be a zero, make it a zero
• define and justify what we mean by fair, valid
• means and SDs for sri for each subgroup, sbst

[snotskie]

Dalex fairness metrics are defined by barocas in https://fairmlbook.org/

We are failing the independence metric, formally:
$$\hat y \perp G$$
or
$$P(\hat y \mid G) = P(\hat y)$$.
That is, the distribution of the predicted values is substantively (even if not significantly) different across genders, resulting in disparate outcomes.

However, we do pass the separation metric, formally:
$$\hat y \perp G \mid y$$
or
$$P(\hat y \mid G, y) = P(\hat y \mid y)$$.
That is, even though the distribution of the predicted values is substantively different across genders, so too were the true values, avoiding (to the extent measured here) disparate treatment.