The Statistical Cost of Second-Order Orthogonality
Juan Carlos Escanciano, Joël R. Terschuur
Abstract
Debiased machine learning removes the first-order effect of nuisance estimation, but the remaining quadratic bias can invalidate inference when nuisance functions are hard to learn. Second-order orthogonal moments remove this bias at the cost of additional variance. We characterize this cost as the minimum variance of a correction that offsets the quadratic sensitivity of the moment to nuisance errors, and show that exact correction is possible only when this cost is finite. Because the cost varies across moments, the first-order efficient moment need not minimize total variance. We therefore choose the moment and its correction jointly, and derive the resulting minimum variance, a second-order analogue of the semiparametric efficiency bound. The optimal moment may trade first-order efficiency for a lower correction cost. In a binary choice logit model, it uses information beyond the conditional likelihood only in well-populated covariate cells; in difference-in-differences with continuous covariates, exact correction forgoes the efficiency gains from exclusion restrictions. In the partially linear model with continuous controls, the cost is infinite for every moment, so that only part of the bias can be removed. For such cases, we propose a data-driven adaptive rule that chooses how much bias to remove within a bound on variance inflation, and we show that inference remains valid after selection. In simulations for the partially linear model, the coverage of nominal 95\% intervals rises from 32--80\% to 94--97\% in all designs but a stress test.
Create a lesson
Related papers
Measuring climate credibility? Talk, emissions, and corporate debt composition
Georgios Chortareas, Fangyuan Kou, Alexia Ventouri
Structural Fine-Tuning under Economic Restrictions
Yi Niu, Yuecheng Su, Ethan Yang
Informative Local Average Treatment Effects Using Propensity Scores
Martin Huber, Ying-Ying Lee
Network-Adaptive Instrumental Variables
Yi Niu, Zhengyan Xu
Design-Based Inference for Fuzzy Difference-in-Differences
Pedro Picchetti
Exploiting Common Volatility Dynamics in High-Dimensional Portfolio Selection
Dominik Ammon, Julian Heinz