Proxy variables are features that are not themselves protected characteristics but correlate strongly enough with protected characteristics to serve as surrogates. Postcode correlates with ethnicity and socioeconomic status. University name correlates with social class. Name correlates with gender and ethnicity. A model that excludes protected characteristics from its inputs but includes strong proxies can still discriminate.
The detection method computes correlation between each feature and each protected characteristic using the appropriate measure: Pearson for continuous-continuous pairs, point-biserial for continuous-binary pairs, Spearman for ordinal pairs, and mutual information as a non-linear alternative for any pair type. Features with correlation coefficients above a defined threshold are flagged for review. A threshold of 0.3 is a common starting point, though the Technical SME calibrates this to the domain; in domains where even modest proxy effects carry serious consequences (employment, credit), a lower threshold may be appropriate.
The output is a correlation matrix: each feature against each protected characteristic, with the correlation statistic and confidence interval. Features exceeding the threshold are flagged, but the flag does not automatically mean removal. The Technical SME conducts a justification review for each flagged feature. Some features may have strong predictive value for the legitimate intended purpose and may be retainable if the proxy risk is mitigated through fairness constraints during training.
Column-level lineage supports proxy detection by revealing indirect relationships. A derived feature such as “risk_score” may not correlate directly with a protected characteristic, yet if its constituent source features (for example, postcode and annual income) are themselves proxies, the derived feature inherits the proxy risk. Lineage enables the Technical SME to trace these indirect pathways.
Key outputs
- Proxy variable correlation matrix
- Flagged features register (exceeding threshold)
- Correlation methodology and threshold documentation