Measuring Fairness Without the Attribute

Every fairness diagnosis requires knowing who belongs to which group, and organisations frequently do not have that data, because collecting it was avoided for privacy reasons, because it was never relevant to the product, or because collecting it is restricted. The result is a widespread situation where an obligation to demonstrate non-discrimination coexists with an inability to measure it.

The available approaches

Ask, with consent and a stated purpose. Voluntary self-identification collected specifically for fairness assessment, stored separately from operational systems, with access restricted to the assessment. It is the highest-quality option and it produces non-response that is itself patterned, so the responding sample is not representative of the population.

Impute from proxies. The best-known method, Bayesian Improved Surname Geocoding, combines surname and geographic distributions from census data to produce a probability distribution over race, and is used in fair lending analysis by regulators and firms. Analogues exist for other attributes and for other jurisdictions with their own reference data.

Use aggregate constraints. Where individual membership is unknown but group proportions are known for a region or cohort, disparity can sometimes be bounded without individual assignment.

Test with audits. Correspondence studies, sending matched applications differing only in a signal of group membership, measure discrimination in outcomes directly without needing the attribute for the whole population. It is the strongest evidence available and it is expensive and covers only the scenarios tested.

What proxy error does

Imputation is noisy, and the noise does not average out in the way intuition suggests. Measurement error in the group variable generally attenuates estimated disparity, so a proxy-based analysis tends to understate the difference between groups. That is the direction that matters: the method is conservative about finding a problem, which means a null result is weak evidence and a positive result is strong.

The error is also not uniform. Surname-based inference is more accurate for some populations than others, so the attenuation differs by group, which can reorder which group appears worst off.

Reporting honestly

A fairness result built on imputed attributes should state the method, its known accuracy characteristics, the direction of the expected bias, and the uncertainty in the estimate. A point estimate of a disparity ratio, presented without that context, is more precise than the evidence supports, and the precision is what makes it persuasive to people who will not read the methodology.

When it breaks

Proxies encode the geography of segregation. Geographic inference works because populations are spatially segregated, which means the proxy is most accurate exactly where segregation is strongest, and the method inherits the structure it is measuring.

Imputation creates a sensitive dataset. A file of inferred race attached to customer records is personal data of a sensitive category, inferred without consent, and it is a liability that has to be governed. Some organisations decline to create it for that reason alone, which is a defensible position with a real cost.

A null result is not evidence of no disparity. Given attenuation and noise, failing to detect a difference is a weak conclusion. Reporting it as "no disparity found" rather than "no disparity detectable with this method at this power" is the misstatement that most often follows this kind of analysis.

Not measuring is a choice with consequences. Declining to estimate group membership does not make disparity absent; it makes it invisible, and invisible disparity persists. The tradeoff between the privacy cost of measurement and the harm of unmeasured discrimination is a genuine one, and it should be decided deliberately rather than by default.