A1891
Title: Universal inference for incomplete models
Authors: Yi Zhang - Jinan University (China) [presenting]
Hiroaki Kaido - Boston University (United States)
Abstract: A growing number of empirical models exhibit set-valued predictions. Such models are attractive because they avoid ad-hoc assumptions on model components that are not understood by the analyst. A tractable inference method with finite-sample validity is developed for such settings. The proposed approach provides robust tests and confidence intervals for counterfactual objects and other functionals of the underlying parameter, and accommodates discrete and continuous covariates, and nuisance parameters. The procedure builds on a robust version of the universal inference framework of Wasserman et al. (2020). Relative to existing inference approaches based on moment restrictions, it does not rely on resampling or simulation and can be implemented without additional functional form assumptions such as linearity.
