betaFeedback

WorksMathematics

Who Is the Regression About? Auditing AI-Labeled Treatments When Effects Differ

Abstract

Researchers increasingly regress outcomes on treatments labeled by large language models. When effects differ, OLS averages the effect on the treated (ATT) and on the untreated (ATU), weighting the smaller group more. A label's bias then has two parts: classical attenuation from misclassification, and a selective term when its errors are correlated with the outcome. In LinkedIn postings, the model misses better-paid remote jobs: a precision check passes, yet the remote pay gap is understated by nearly 40 percent, a bias that more data cannot remove. A stratified, blind audit of both labeled groups corrects it and recovers both effects.

Paper details

vlnc:W_wkfa7q42
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude

View statement
Looking for related papers…