Proxy Variables and Redlining
What you'll learn: How seemingly innocent data fields can act as stand-ins for protected characteristics and enable discrimination.
The Hidden Discrimination Problem
A proxy variable is a feature that correlates strongly with a protected attribute (like race, gender, or religion) even though it appears neutral on the surface. When you build models or make decisions using proxy variables, you can inadvertently discriminate against protected groups—even if you never directly use the sensitive attribute itself.
Classic Example: Redlining
The term "redlining" comes from 1930s banking practices where lenders literally drew red lines around minority neighborhoods on maps and denied loans to anyone in those areas. Today's digital redlining works the same way: using zip code as a predictor might seem innocent, but if certain zip codes strongly correlate with racial composition due to historical segregation, you're effectively making race-based decisions without ever asking about race.
Common Proxy Variables
- Zip code or address → race, income, immigration status
- First or last names → ethnicity, religion, gender
- Shopping patterns → protected class membership
- Credit history location → historical discrimination effects
Why This Matters
Even with the best intentions, removing protected attributes from your dataset doesn't guarantee fairness. If your model learns that "applicants from zip codes 10001-10005 are higher risk," but those zip codes happen to be predominantly one ethnic group, you've just built discriminatory redlining into your algorithm.
Key Takeaway: Fairness requires examining not just direct use of protected attributes, but also identifying and addressing proxy variables that perpetuate historical discrimination patterns.