When: Wednesday, January 28, 2026 from 10 – 11 am

Where: Raitt 223 and on Zoom

Title: Drawing Rhinoceroses with Algorithms: Inference on Predicted Data and its implications for Demography (Kentaro Hoffman)

Machine learning is increasingly used in demography to predict quantities that were once directly observed. Yet predictions are often treated as data, a practice that can lead to biased estimates and misleading uncertainty. This talk introduces Inference on Predicted Data (IPD), a framework for conducting valid statistical inference when outcomes are generated by black-box prediction models rather than measured directly.

I illustrate IPD through an application to verbal autopsies, where causes of death are inferred from free-text narratives using modern NLP methods, including large language models. While these models can achieve high predictive accuracy, naïvely using predicted causes of death in downstream analyses produces distorted demographic patterns. IPD-based corrections leverage a small amount of labeled data to recover valid estimates and uncertainty, even under prediction error and distribution shift.

The results highlight a key lesson for computational demography: accurate predictions alone are not enough for reliable population inference.

Kentaro Hoffman (University of Washington) is a statistician whose research focuses on inference with AI-generated and predicted data, uncertainty quantification, and responsible machine learning. He was previously a postdoctoral scholar at the University of Washington and Johns Hopkins University, working with Tyler McCormick, Peter Searson, and Scott Zeger. His work lies at the intersection of statistics, machine learning, and computational demography, with applications including verbal autopsies, global mortality estimation, electronic medical records, and active learning.