Abstract

We construct simultaneous confidence intervals for solutions to conditionalmoment equations. The intervals are built around a class of nonparametricregression algorithms based on subsampled kernels. This class encompassesvarious forms of subsampled random forest regression, including GeneralizedRandom Forests (Athey et al., 2019). Although simultaneous validity is oftendesirable in practice -- for example, for fine-grained characterization oftreatment effect heterogeneity -- only confidence intervals that conferpointwise guarantees were previously available. Our work closes this gap. As aby-product, we obtain several new order-explicit results on the concentrationand normal approximation of high-dimensional U-statistics.

 

Quick Read (beta)