Robust Fairness Auditing with Unobserved Protected Classes

This direction is merged into another direction: sensitive attribute dependency. Its last version stays readable and citable here.

Version 1 of 110 papers24 conceptsfrozen 2026-10-04

In an ideal world, auditing an algorithm for bias is straightforward: you compare the outcomes for different demographic groups—such as race or gender—and measure the disparity. In practice, however, these "protected attributes" are often missing from datasets due to privacy regulations (like GDPR) or institutional policies.

To fill this gap, practitioners often turn to proxies: they use names to guess ethnicity or ZIP codes to guess socioeconomic status. But as a growing body of research shows, treating these guesses as ground truth doesn't just add a little noise—it systematically hides the very bias we are trying to find.

The Proxy Trap: Why Naive Estimation Fails

The most common mistake in fairness auditing is "plug-in" estimation: replacing a missing label with a proxy's prediction and proceeding as if the data were perfect. Zhaowei Zhu, Yuanshun Yao, and their colleagues demonstrate that this approach leads to a dangerous "false sense of security" [1].

The intuition is that proxy error acts as a "smoothing" filter. If a model discriminates against Group A, but your proxy mislabels 20% of Group A as Group B, the observed outcomes for the two groups will look more similar than they actually are. This error is not random; it is proportional to the inaccuracy of the proxy and the magnitude of the true underlying disparity. Consequently, a system can appear fair under audit while remaining discriminatory in reality [2].

This problem is particularly acute in ad delivery. Basileal Yoseph Imana and Aleksandra Korolova point out that if an auditor uses the same flawed inference tools as the platform they are auditing, the errors cancel out in a way that makes the platform's bias invisible to the auditor [3].

From Point Estimates to Robust Bounds

If we cannot know the exact demographic labels, how can we trust a fairness score? Nathan Kallus, Angela Zhou, and Xiaojie Mao propose shifting the goal from finding a single number to finding a range [4].

Using a framework called partial identification, they calculate the tightest possible upper and lower bounds on fairness metrics that are consistent with the observed data. Instead of claiming "the disparity is 2%," a robust audit might conclude "given our proxy's known error rates, the true disparity must be between 5% and 12%." This provides a more honest assessment of risk, ensuring that practitioners don't ignore potential harms just because they lack perfect data.

Calibrating with Weak Proxies

While one proxy might be misleading, multiple "weak" proxies can be surprisingly powerful. Zhaowei Zhu and Yang Liu developed a calibration algorithm that uses three different inaccurate proxies to provably recover true fairness metrics [5].

The core idea is that by analyzing the patterns of disagreement between different proxies, we can mathematically "subtract" the noise. This allows for high-precision auditing without ever needing to collect sensitive personal data, effectively using mathematical rigor to protect user privacy while enforcing social accountability.

The Risk of Paradoxical Unfairness

The danger of missing data extends beyond auditing to mitigation. If you try to "fix" a model's bias using a noisy proxy, you might actually make the real-world bias worse.

Jeremias Sulam and Beepul Bharti highlight that if a proxy's error rates are imbalanced—for example, if it's better at identifying men than women—standard fairness interventions will produce suboptimal results [6]. Optimizing for fairness against a flawed proxy can paradoxically increase the worst-case unfairness toward the actual demographic groups [7].

To solve this, they introduce post-processing methods that incorporate the specific error imbalances of the proxy predictors. By accounting for how the proxy fails, these methods can minimize true unfairness even when the sensitive attribute is never directly observed.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_fairness_auditing_with_unobserved_13b58584,
  title = {Robust Fairness Auditing with Unobserved Protected Classes},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-fairness-auditing-with-unobserved-protected-classes/rdr_13b5858418b04a4d98125f741acb5a25},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_fairness_auditing_with_unobserved_13b58584_v1,
  title = {Robust Fairness Auditing with Unobserved Protected Classes},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-fairness-auditing-with-unobserved-protected-classes/rdr_13b5858418b04a4d98125f741acb5a25/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}