A new econometric toolkit builds debiased moments from machine-learned components so researchers can obtain root‑n inference even with complex conditional moment restrictions; under an approximate sparsity assumption the Lasso-computed 'Orthogonal IVs' deliver valid GMM estimates, validated in simulations and a firm-level production-function application.

Main Finding

The paper provides a practical, theory-backed algorithm to automatically construct Neyman-orthogonal (locally robust) moments in very general models defined by a finite number of conditional moment restrictions (CMRs). The core innovation is an estimable construction of Orthogonal Instrumental Variables (OR-IVs): residualized functions of conditioning variables obtained by solving a Lasso-type program with generated regressors (estimated conditional expectations). Combining these OR-IVs with the model residuals yields debiased moments and a Debiased GMM (DGMM) estimator that attains √n-consistency and asymptotic normality even when nuisance components are estimated with modern Machine Learning (ML) tools (provided mild rate and sparsity conditions hold). Monte Carlo experiments and an application to semiparametric production-function estimation (Chilean firm panel) illustrate improved bias and coverage relative to naive (non-orthogonal) procedures.

Key Points

• Problem addressed: using ML for high-dimensional or nonparametric nuisance estimation typically introduces regularization bias that breaks √n inference for finite-dimensional structural parameters in CMR models.
• OR-IV concept: instruments that are “residualized” versions of known functions of conditioning variables, orthogonal to nuisance-direction derivatives (the range of a linear operator built from CMRs). OR-IVs are necessary components for Neyman-orthogonal moments in this class of models.
• Main technical device: represent computing the orthogonal projection (residualization) as a mean-squared projection of starting IVs onto a (possibly high-dimensional) set of regressors (reduced-form conditional expectations). Estimate those regressors with ML, then recover OR-IVs by solving a Lasso problem with generated regressors (ℓ1-penalty allows p > n under approximate sparsity).
• Two-step estimator: (1) estimate nuisance objects (reduced forms, other nonparametric components) via ML with cross-fitting; (2) form OR-IVs via Lasso residualization and compute DGMM for the finite-dimensional parameters. Standard sandwich standard errors apply.
• Asymptotics & requirements: derives rates for OR-IV estimators and shows DGMM is √n-consistent and asymptotically normal when (i) the parameter has a finite efficiency bound, (ii) approximate sparsity holds for the Lasso representation, and (iii) nuisance estimators converge faster than n−1/4 (achievable by many ML methods). Cross-fitting is used to control overfitting.
• Empirical evidence: Monte Carlo in production-function setting shows DGMM reduces bias and achieves closer-to-nominal coverage versus naive approaches; application to Chilean firm panel indicates substantive sensitivity of production-parameter estimates to first-stage modeling choices, underscoring the value of debiased inference.
• Limitations/assumptions: relies on approximate sparsity, correctness of the Lasso linear approximation for the orthogonal projection, the ability to estimate conditional expectations accurately (rate requirement), and the finite-efficiency-bound condition for √n-estimability.

Data & Methods

• Model class: semiparametric models defined by a finite number of CMRs, possibly with different conditioning sets and endogenous regressors; structural parameter is finite dimensional and may enter residuals nonlinearly.
• Algorithm (practical summary):
• Choose a set of starting instruments/functions of conditioning variables (e.g., Z, Z^2, other transforms).
• Estimate the relevant reduced-form conditional expectations (regressors needed for the projection) using ML learners (Random Forests, Boosting, Lasso, Neural Nets, etc.), with cross-fitting/sample-splitting.
• Form a Lasso (ℓ1-penalized least squares) problem projecting the starting instruments onto the dictionary of estimated regressors (these are generated regressors). The residuals from that projection are estimated OR-IVs.
• Build Neyman-orthogonal moments as products of model residuals and the OR-IVs; combine and solve a (debised) GMM in the second step.
• Use the usual sandwich variance formula for inference (no bootstrap required for first-stage regularization bias).
• Theoretical contributions:
  • Characterizes OR-IVs via orthogonality to the range of a Fréchet-derivative-derived linear operator coming from the CMRs.
  • Shows the orthogonal projection reduces to a mean-squared minimization over a (possibly large) basis, justifying the Lasso approach under approximate sparsity.
  • Provides convergence rates for the estimated OR-IVs and uses them to prove √n-consistency and asymptotic normality of DGMM under stated conditions.
• Empirical implementation: production-function (proxy-variable) example—nuisance policy functions estimated by off-the-shelf ML; DGMM applied to jointly estimate input elasticities; Monte Carlo and a Chilean firms panel used to demonstrate performance.

Implications for AI Economics

• Enables credible use of flexible ML methods in structural/economic models: researchers can estimate complex nonparametric nuisance components with ML while still performing valid √n inference on focal structural parameters.
• Automates orthogonalization step: the paper gives a practical recipe (ML reduced-form + Lasso residualization) to produce OR-IVs, which can be plugged into GMM pipelines—this is well suited to incorporate into automated/productionized causal-ML toolchains.
• Broad applicability: the method covers many canonical economic settings (production functions, demand, auctions, dynamic discrete choice, simultaneous equations, missing-data problems) where CMRs arise, allowing richer functional forms and higher-dimensional controls without sacrificing inferential validity.
• Practical guidance: default to constructing OR-IVs (via the paper’s method) rather than using ad hoc instruments when ML is used in first stages; standard sandwich SEs suffice under the conditions outlined, simplifying applied work.
• Caveats and directions:
  • Structural inference remains contingent on approximate sparsity and obtaining the required ML rates (n−1/4). For settings where those fail (e.g., dense high-frequency features or where the linear projection is poorly approximated), performance may degrade.
  • Computational and implementation considerations: estimating many conditional expectations and solving Lasso with generated regressors adds complexity; careful cross-fitting and tuning are needed.
  • Opportunities: extend to alternative penalties/regularizers beyond ℓ1, study robustness when sparsity is weak, and combine with modern deep-learning based reduced-form estimators (with theory for their rates) to expand applicability in very high-dimensional or unstructured-data contexts.
• Overall: this work materially lowers a major barrier to integrating ML into structural economic estimation—providing an automated, theoretically justified route to debiased inference that will likely be useful for empirical work and for AI-augmented econometric pipelines.