Systemic Bias in Administrative Proxy Labels

This direction is merged into another direction: proxy variable bias (proxy label bias). Its last version stays readable and citable here.

Version 1 of 112 papers18 conceptsfrozen 2026-10-04

In high-stakes social domains like criminal justice and child welfare, machine learning models are often tasked with predicting latent qualities: "Will this person commit a crime?" or "Is this child at risk of maltreatment?" However, these latent qualities are rarely directly observable. Instead, researchers and practitioners use proxy labels—administrative records of human decisions, such as arrests, convictions, or foster care placements.

The core challenge in this research direction is that these proxies are not objective reflections of behavior. Instead, they are "noisy" labels where the noise is structurally biased by the very institutions the models are meant to assist.

The Gap Between Behavior and Record

The fundamental problem identified across these studies is the conflation of an event (e.g., a crime) with the institutional response to that event (e.g., an arrest). As Riccardo Fogliato and Alexandra Chouldechova argue in Fairness Evaluation in Presence of Biased Noisy Labels, we cannot observe "crime" directly. When we use "arrest" as a proxy for recidivism, we are not just measuring the individual's behavior; we are measuring the behavior of the police.

If policing practices are concentrated in specific neighborhoods or directed at specific demographics, the "ground truth" labels in the dataset will be denser for those groups, regardless of the underlying rate of actual criminal activity. A model trained on this data does not learn to predict crime; it learns to predict which individuals are most likely to be caught in the existing net of law enforcement.

Administrative Actions as Self-Fulfilling Prophecies

In child welfare, the stakes are similarly high. The Allegheny Family Screening Tool (AFST), analyzed by Alexandra Chouldechova and Rhema Vaithianathan in A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions, uses foster care placement as a proxy for child maltreatment.

The intuition here is subtle but critical: foster care placement is a human decision made by workers within a system. If the system has historically been more likely to remove children from certain types of households (e.g., low-income or minority families) for the same level of risk, the model will treat those removals as "correct" outcomes to be replicated. By using administrative actions as ground truth, the model "bakes in" the subjective judgments and systemic biases of past practitioners, effectively automating institutional history rather than objective risk assessment.

Scaling the Analysis of Systemic Bias

To move beyond small-scale studies, researchers are developing massive datasets to capture the nuances of these institutional processes. Elliott Ash and Naman Goel introduced the WCLD dataset in WCLD: Curated Large Dataset of Criminal Cases from Wisconsin Circuit Courts, which tracks 1.5 million cases over five decades.

This scale allows researchers to see how "ground truth" shifts not just by individual behavior, but by geography and judge behavior. When recidivism is defined by legal convictions, the label becomes a composite of:

  1. The individual's actions.
  2. The prosecutor's decision to charge.
  3. The judge's sentencing patterns.
  4. The availability of legal counsel.

By treating these labels as the output of a complex social system rather than a simple binary fact, this research direction highlights why "accuracy" can be a misleading metric. A model can be highly accurate at predicting a biased proxy while being fundamentally unfair—and factually incorrect—about the underlying human reality.

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_systemic_bias_in_administrative_proxy_la_22606b37,
  title = {Systemic Bias in Administrative Proxy Labels},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/systemic-bias-in-administrative-proxy-labels/rdr_22606b37013d43108a9d601fcf7b4046},
  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_systemic_bias_in_administrative_proxy_la_22606b37_v1,
  title = {Systemic Bias in Administrative Proxy Labels},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/systemic-bias-in-administrative-proxy-labels/rdr_22606b37013d43108a9d601fcf7b4046/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}