A1813
Title: Policy learning with unstructured data
Authors: Yue Fang - The Chinese University of Hong Kong (China) [presenting]
Abstract: Unstructured data, such as text and images, are increasingly central to empirical policy problems, yet most policy learning theory assumes low-dimensional, structured features. Policy learning when the available information takes the form of high-dimensional embeddings, from which researchers construct low-dimensional features to design welfare-maximizing policies, raises a key econometric challenge: the estimated features act as generated regressors and can introduce first-order bias into welfare evaluation. The bias is characterized by deriving the pathwise derivative of the welfare functional, revealing how the estimation error propagates through nuisance functions and the induced policy. A debiased policy learning method is proposed that removes the first-order bias and restores regret guarantees with the same rate as if the low-dimensional features were directly observed. The framework is illustrated using popular supervised and unsupervised feature-extraction methods, including average-derivative-type constructions, principal component analysis, latent semantic analysis, and clustering. An empirical application studies the use of satellite-based embeddings to guide mosquito net allocation and improve health outcomes in Nigeria.
