Home / 13. Fairness / Protected attributes & proxies

Protected attributes & proxies

Deleting gender or race from the feature list does not remove their influence if proxies remain.

Protected attribute $G$ (gender, race, disability, …) may be correlated with other features $X$ that still enter the model. The decision $\hat{Y}$ can still depend on $G$ indirectly.

Zip code proxy

You drop race from the loan model but keep zip code. If zip code correlates with race, decisions can still track race. “We did not use $G$” is not a fairness proof.

Key takeaways

  • Proxies leak sensitive information.
  • Need criteria beyond “don’t use $G$”.
  • Evaluate outcomes by group.