When a Stand-In Carries the Bias

A common attempt to prevent bias is to simply leave a sensitive attribute out of the data — don't give a hiring tool a candidate's age, or don't give a lending tool an applicant's neighborhood. That sounds like it should work. In practice, it often doesn't, because other, seemingly neutral fields can stand in for the excluded one.

These are called proxy variables. A candidate's zip code can correlate strongly with race or income even with no explicit demographic field present. The name of a school someone attended can correlate with class background. A model doesn't need to be told the sensitive category directly — it can reconstruct a close approximation of it from several ordinary-looking fields combined.

This is why "we didn't include that field" isn't, by itself, a guarantee of fairness. Removing one obvious variable doesn't remove the pattern if the pattern is smeared across several other variables that quietly correlate with it. Actually checking for this takes deliberate testing — looking at outcomes across groups, not just checking which fields were in the input.

Why doesn't simply removing a sensitive field (like age or neighborhood) from training data guarantee an unbiased model?

A field like zip code that correlates strongly with a removed sensitive attribute and can stand in for it is called a ____ variable.

Removing an obviously sensitive field from a dataset guarantees that a model trained on it can't reconstruct that pattern from other fields.