arXiv · 2611.17578
Orthogonal Scores and Double/Debiased Machine Learning for Possibly Misspecified GMM
Abstract
We develop Neyman orthogonal scores for GMM models that remain valid when the moment conditions may fail to hold. Existing score constructions typically require the moment conditions to hold exactly, which may fail in over-identified models, where the number of moment functions exceeds the number of parameters. We construct \emph{misspecification-robust} orthogonal score that has a closed form; it is the conventional orthogonal score plus adjustment terms, for the estimation of the Jacobian and of the weighting matrix, which are proportional to the population moment and vanish under correct specification. The moment functions may depend on nonparametric or machine learning first steps, and we show that double/debiased machine learning (DML) combined with the proposed scores and cross-fitting provides uniformly valid inference for a low-dimensional GMM estimand, while the conventional DML standard errors are inconsistent when the moment conditions fail. We illustrate the results in the partially linear instrumental variables (PLIV) model and in the local average treatment effects (LATE) model with multiple instruments.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Byunghoon Kang. 2026-11-12. Orthogonal Scores and Double/Debiased Machine Learning for Possibly Misspecified GMM. https://searcharxiv.com/abs/2611.17578
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.