Algorithmic Bias from Flawed Measurement Proxies

47 concepts38 papers10 proposals

When we train machine learning models to make high-stakes decisions—such as allocating healthcare resources, predicting criminal recidivism, or distributing social services—we rarely have access to direct measurements of the phenomenon we want to predict. Instead, we rely on observable stand-ins: proxy variables.

We want to measure a latent construct like health need, so we train a model on healthcare costs. We want to measure criminal behavior, so we train a model on arrest records.

This substitution introduces a fundamental vulnerability known as the construct gap. When a proxy variable systematically diverges from the true latent construct, the model does not learn to predict the real-world phenomenon. Instead, it optimizes for the institutional and systemic biases embedded within the proxy itself.

The Anatomy of a Construct Gap

If a proxy variable were simply the true construct plus random, un-correlated noise, standard statistical techniques could easily smooth over the errors. However, administrative and behavioral proxies are rarely neutral. They do not merely measure an individual's state; they measure an individual's interaction with a system.

Consider the domain of healthcare allocation. In a landmark study, Obermeyer (2019) analyzed a commercial algorithm used to identify patients for high-risk care management programs. The target construct was clinical health need, but the algorithm was trained to predict historical healthcare expenditure.

Because of systemic barriers to healthcare access, less money is historically spent on Black patients than on white patients, even when they suffer from the exact same level of illness. By using cost as a proxy for need, the algorithm codified this disparity, requiring Black patients to be significantly sicker than white patients to qualify for the same care. As Vokinger et al. (2021) and Yang et al. (2024) note, this pattern repeats across biomedicine whenever financial or utilization metrics substitute for actual clinical urgency.

[True Latent Construct]  ---> (Systemic Inequities / Barriers) ---> [Observable Proxy Label]
   (e.g., Health Need)             (e.g., Access to Care)              (e.g., Healthcare Cost)
                                                                                  |
                                                                                  v
                                                                        [Model Optimization]

A parallel failure occurs in the justice system. Risk assessment instruments designed to predict recidivism typically rely on arrest or conviction data as a proxy for actual re-offending behavior. However, Fogliato et al. (2021) demonstrate that arrest records fail as an unbiased ground truth due to severe geographic and situational sampling biases in policing practices.

When certain neighborhoods are policed more heavily, individuals in those neighborhoods are arrested at higher rates regardless of underlying behavior. As Butcher et al. (2022) and Tolan et al. (2021) argue, models trained on these administrative traces do not predict criminal risk; they automate and reinforce existing law enforcement patterns.

The Illusion of Mathematical Fairness

A major consequence of proxy contamination is that it invalidates standard algorithmic fairness metrics. In fairness literature, researchers frequently attempt to enforce constraints like predictive parity (ensuring the model is equally calibrated across groups) or error rate balance (equalizing false positive rates).

Chouldechova (2017) and Corbett-Davies et al. (2017) mathematically proved that it is impossible to satisfy all of these fairness criteria simultaneously if the base rates of the groups differ in the training data.

However, when the ground truth labels themselves are biased proxies, the entire evaluation framework becomes a mirage. Li (2024) argues that label bias "invisibilizes" societal discrimination: if a model's predictions perfectly match a biased proxy label, standard metrics will evaluate the model as perfectly accurate and fair, masking the underlying construct gap. As Fogliato et al. (2020) demonstrate, evaluating a model against a tainted baseline ensures that the resulting fairness guarantees are mathematically sound but socially hollow.

How Proxies Enter the Pipeline

Proxy bias is rarely an algorithmic glitch; it is a failure of upstream problem formulation. Through ethnographic observation of data science teams, Passi et al. (2019) showed that converting abstract organizational goals into measurable machine learning targets is a process of negotiation. Teams routinely shift their objectives to fit available administrative data, embedding ethical assumptions into the system's foundation before a single line of model code is written.

This translation process introduces subtle, secondary proxies that appear neutral but carry historical baggage:

  • Predictor Contamination: Feng (2025) notes that features like family medical history can act as proxies for socioeconomic or immigration status, penalizing individuals who lack access to documented longitudinal care.
  • Resource Allocation Mismatch: In social services, Coston et al. (2022) show that using service utilization costs to predict underlying human needs systematically underestimates the requirements of marginalized groups who face structural barriers to accessing those very services.

Statistical Remedies: Predict-Then-Debias

If high-quality, unbiased "gold-standard" labels are incredibly expensive or difficult to collect, and cheap administrative proxies are abundant but systematically biased, how can we perform valid scientific and statistical inference?

A growing body of research addresses this by designing frameworks that use a small, unbiased validation set to mathematically correct the errors of large-scale, biased proxies.

[Large Population]  ---> Collect Cheap, Biased Proxies (e.g., LLM Labels, Costs)
        |
        v (Sub-sample)
[Small Population]  ---> Collect Expensive, Gold-Standard Labels (e.g., Human Experts)
        |
        v
[Statistical Estimator] ---> Use Gold-Standard to "Debias" the Proxy-Based Predictions

For instance, when using Large Language Models (LLMs) or automated heuristics to label text data for downstream social science research, the automated labels often contain systematic, non-uniform biases. To resolve this, Egami et al. (2023) introduced Design-based Supervised Learning (DSL). DSL combines a massive set of biased automated predictions with a small, randomly sampled set of human-validated gold-standard labels. By using the gold-standard data to estimate and adjust for the systematic error of the proxy, DSL guarantees unbiased downstream statistical estimates and valid confidence intervals.

Similarly, in medical informatics, Ying et al. (2025) developed a semi-supervised framework to handle "Automated Computational Phenotypes" (ACPs)—algorithmic proxies derived from electronic health records that substitute for rigorous clinical diagnoses. They utilize a doubly robust estimator that leverages the abundant but imperfect ACPs alongside scarce gold-standard labels, maintaining statistical validity even under conditions of covariate shift.

Finally, Kluger (2026) formalized this paradigm within M-estimation theory through a multiwave "Predict-Then-Debias" framework. By adaptively sampling high-quality measurements across multiple waves based on the distribution of the cheap proxies, this approach dramatically boosts sample efficiency.

The core intuition across all these remedies is identical: we do not need to abandon cheap, biased proxies entirely. Instead, by treating them strictly as imperfect surrogates and anchoring them to a mathematically rigorous sub-sample of true gold-standard measurements, we can prevent institutional biases from permanently contaminating our models.

Key Foundational Work

Fair prediction with disparate impact: A study of bias in recidivism prediction instruments (2017) — contributed 2 concept elements to this direction.
A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions (2018) — contributed 2 concept elements to this direction.
Algorithmic decision making and the cost of fairness (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_proxy_variable_bias_proxy_label_bias_76898afa,
  title = {proxy variable bias (proxy label bias)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proxy-variable-bias-proxy-label-bias/rdr_76898afa11d64e8eaaf0cf69e4bc1d07},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_proxy_variable_bias_proxy_label_bias_76898afa_v1,
  title = {proxy variable bias (proxy label bias)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proxy-variable-bias-proxy-label-bias/rdr_76898afa11d64e8eaaf0cf69e4bc1d07/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake