Unbiased Prevalence Estimation with Multicalibrated LLMs

2026-04-23

AI Summary

This work addresses the systematic bias in traditional prevalence estimation under covariate shift, which arises from the assumption that measurement device error rates remain stable across distributions. To overcome this limitation, the study introduces multicalibration theory into prevalence estimation for the first time, achieving unbiased estimates under distributional shifts by calibrating predictions conditioned on key features. The proposed method is applicable to large language models and other classifiers and requires only calibration data covering important feature dimensions. Experimental results demonstrate that, in tasks such as state-level employment rate estimation and cross-national political text classification, the approach substantially reduces estimation bias—achieving near-zero error—and effectively overcomes the limitations of standard calibration and quantification methods when faced with distributional changes.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Bias, Fairness & Privacy
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Estimating the prevalence of a category in a population using imperfect measurement devices (diagnostic tests, classifiers, or large language models) is fundamental to science, public health, and online trust and safety. Standard approaches correct for known device error rates but assume these rates remain stable across populations. We show this assumption fails under covariate shift and that multicalibration, which enforces calibration conditional on the input features rather than just on average, is sufficient for unbiased prevalence estimation under such shift. Standard calibration and quantification methods fail to provide this guarantee. Our work connects recent theoretical work on fairness to a longstanding measurement problem spanning nearly all academic disciplines. A simulation confirms that standard methods exhibit bias growing with shift magnitude, while a multicalibrated estimator maintains near-zero bias. While we focus the discussion mostly on LLMs, our theoretical results apply to any classification model. Two empirical applications -- estimating employment prevalence across U.S. states using the American Community Survey, and classifying political texts across four countries using an LLM -- demonstrate that multicalibration substantially reduces bias in practice, while highlighting that calibration data should cover the key feature dimensions along which target populations may differ.

Problem

• prevalence estimation
• covariate shift
• measurement bias
• large language models
• calibration

Innovation

• multicalibration
• prevalence estimation
• covariate shift
• unbiased estimation
• large language models